def main(IMAGE_PATH,DESTINATION_PATH):
# DESTINATION_PATH_ = DESTINATION_PATH.split(".")
# DESTINATION_PATH_[1] = "npy"
# DESTINATION_PATH_ = ".".join(DESTINATION_PATH_)
im = cv2.imread(IMAGE_PATH)
# rotate_and_save(im,0,DESTINATION_PATH)
increment_by = 4
n = 5
im2 = []
# im2 = np.load(DESTINATION_PATH_).tolist()
# print(len(im2))
for i in range(1,n+1,1):
pos_rot_im = ndimage.rotate(im,i*increment_by)
pos_rot_im = scan_and_crop.main(pos_rot_im)
im2.append(pos_rot_im)
neg_rot_im = ndimage.rotate(im,-i*increment_by)
neg_rot_im = scan_and_crop.main(neg_rot_im)
im2.append(neg_rot_im)
# rotate_and_save(im,i*increment_by,DESTINATION_PATH)
# rotate_and_save(im,-i*increment_by,DESTINATION_PATH)
# print(len(im2))
# print(DESTINATION_PATH_)
# print(len(im2))
im = scan_and_crop.main(im)
return [im,im2]
# cv2.imwrite(DESTINATION_PATH,im2)
def find_field_in_stack(stack, x, y, z, yaw, pitch, roll, height, width):
""" Get a cutout of the given height, width dimensions in the rotated stack at x, y, z.
:param np.array stack: 3-d stack (depth, height, width)
:param float x, y, z: Center of field measured as distance from the center in the
original stack (before rotation).
:param yaw, pitch, roll: Rotation angles to apply to the field.
:param height, width: Height and width of the cutout from the stack.
:returns: A height x width np.array at the desired location.
"""
# Rotate stack (inverse of intrinsic yaw-> pitch -> roll)
rotated = ndimage.rotate(stack, yaw, axes=(1, 2), order=1) # yaw(-w)
rotated = ndimage.rotate(rotated, -pitch, axes=(0, 2), order=1) # pitch(-v)
rotated = ndimage.rotate(rotated, roll, axes=(0, 1), order=1) # roll(-u)
# Compute center of field in the rotated stack
R_inv = np.linalg.inv(create_rotation_matrix(yaw, pitch, roll))
rot_center = np.dot(R_inv, [x, y, z])
center_ind = np.array(rotated.shape) / 2 + rot_center[::-1] # z, y, x
# Crop field (interpolate at the desired positions)
z_coords = [center_ind[0] - 0.5] # -0.5 is because our (0.5, 0.5) is np.map_coordinates' (0, 0)
y_coords = np.arange(height) - height / 2 + center_ind[1] # - 0.5 is added by '- height / 2' term
x_coords = np.arange(width) - width / 2 + center_ind[2]
coords = np.meshgrid(z_coords, y_coords, x_coords)
out = ndimage.map_coordinates(rotated, [coords[0].reshape(-1), coords[1].reshape(-1),
coords[2].reshape(-1)], order=1)
field = out.reshape([height, width])
return field
def scale_image(image,angle,ratio_compression_y_over_x,interpolation_value=0):
"""Corrects for scaling distortions in image plane, e.g. from astigmatism
in the optical system. Returns the scaled image with the same rotation."""
image_rot = rotate(image,angle,order=interpolation_value)
image_norm = zoom(image_rot,[ratio_compression_y_over_x,1],order=interpolation_value)
image_final = rotate(image_norm,-angle,order=interpolation_value)
return image_final
def loci_medfilter(self):
log = self.log
if self._file_exists('red','cube_loci_medfilter'): return
ft = lambda i: '%d:%.2d' % ((time.time()-i)/60,(time.time()-i)%60)
pa = self.pa
for chan in ['red','blue']:
cube = self._read_fits(chan,'cube_medfilter')
bp,cube = nt.reset_nans(cube)
log.info('Applying loci method ...(Can take a while......)')
ti = time.time()
cube = loci_subtract(cube, pa)
log.info(' '+ft(ti))
sz = np.shape(cube)
for i in range(sz[0]):
log.info('rotating cube_medfilter['+str(i+1)+'/'+str(sz[0])+'] %.2f'%-pa[i])
cube[i,:,:] = nd.rotate (cube[i,:,:],pa[i], reshape=False)
bp[i,:,:] = nd.rotate (bp[i,:,:],pa[i], reshape=False)
#cube[i,:,:] = nd.rotate (cube[i,:,:],-pa[i], reshape=False)
#bp[i,:,:] = nd.rotate (bp[i,:,:],-pa[i], reshape=False)
cube = nt.restore_nans(cube,bp)
self._write_fits(cube,chan,'cube_loci_medfilter')
self._write_fits(np.median(cube,axis=0),chan,'loci_medfilter_medcrunch')
fs = lambda cube: np.sum(cube,axis=0) / np.sum(np.isfinite(cube),axis=0)
self._write_fits(fs(cube),chan,'loci_medfilter_sumcrunch')
del cube,bp
def loci_sdi(self):
log=self.log
ft = lambda i: '%d:%.2d' % ((time.time()-i)/60,(time.time()-i)%60)
if self._file_exists('','cube_loci_sdi'): return
pa = self.pa
cube = self._read_fits('','cube_sdi')
#cube = self.cube_diff
sz = np.shape(cube)
bp,cube = nt.reset_nans(cube)
log.info('Applying loci_subtract...(Can take a while......)')
ti = time.time()
cube = loci_subtract(cube, self.pa)
#bp = loci_subtract(bp, self.pa)
log.info(' '+ft(ti))
for i in range(sz[0]):
log.info('rotating cube_diff['+str(i+1)+'/'+str(sz[0])+'] %.2f'%-pa[i])
cube[i,:,:] = nd.rotate (cube[i,:,:],pa[i], reshape=False)
bp[i,:,:] = nd.rotate (bp[i,:,:],pa[i], reshape=False)
#cube[i,:,:] = nd.rotate (cube[i,:,:],-pa[i], reshape=False)
#bp[i,:,:] = nd.rotate (bp[i,:,:],-pa[i], reshape=False)
cube = nt.restore_nans(cube,bp)
self._write_fits(cube,'','cube_loci_sdi')
self._write_fits(np.median(cube,axis=0),'','loci_sdi_medcrunch')
fs = lambda cube: np.sum(cube,axis=0) / np.sum(np.isfinite(cube),axis=0)
self._write_fits(fs(cube),'','loci_sdi_sumcrunch')
def _rotate_cube(self,cube):
log = self.log
pa = self.pa
sz = np.shape(cube)
# Now rotate this cube
to=time.time()
for k in range(sz[0]):
ti=time.time()
bp,cube[k,:,:] = nt.reset_nans(cube[k,:,:])
cube[k,:,:] = nd.rotate (cube[k,:,:], pa[k], reshape=False)
#cube[k,:,:] = nd.rotate (cube[k,:,:], -pa[k], reshape=False)
#Now rotate the bp matrix
#bp = nd.rotate (bp, -pa[k], reshape=False)
bp = nd.rotate (bp, pa[k], reshape=False)
# And restore the Nans to cube
cube[k,:,:] = nt.restore_nans(cube[k,:,:],bp)
#_form_line('rotating ,k,nk,to,t1,t2,pa[k])
t = time.time()
line = 'rotating ['+str(k+1)+'/'+str(sz[0])+']'
log.info(line+'%.2f %.2f %.2f' % (pa[k],t-ti,t-to))
return cube
def rotate_images(self, data, angle=-1):
if angle == -1:
angle = self.rng.random_integers(360)
if data.ndim <= 2:
data = ndimage.rotate(data, angle=angle)
else:
for i in xrange(data.shape[0]):
data[i] = ndimage.rotate(data[i], angle=angle)
return data
def testRigidTransformationMultiscale(dTheta, displacement, level):
inImg=misc.imread('T2sample.png')[...,0]
left=ndimage.rotate(inImg, -0.5*dTheta)#Rotate symmetricaly to ensure both are still the same size
right=ndimage.rotate(inImg, 0.5*dTheta)#Rotate symmetricaly to ensure both are still the same size
right=ndimage.affine_transform(right, np.eye(2), offset=-1*displacement)
rightPyramid=[i for i in transform.pyramid_gaussian(right, level)]
leftPyramid=[i for i in transform.pyramid_gaussian(left, level)]
rcommon.plotPyramids(leftPyramid, rightPyramid)
beta=estimateRigidTransformationMultiscale(leftPyramid, rightPyramid)
print 180.0*beta[0]/np.pi, beta[1:3]
return beta
def testRigidTransformEstimation(inImg, level, dTheta, displacement, thr):
left=ndimage.rotate(inImg, dTheta)
right=ndimage.rotate(inImg, -dTheta)
left=ndimage.affine_transform(left , np.eye(2), offset=-1*displacement)
right=ndimage.affine_transform(right, np.eye(2), offset=displacement)
rightPyramid=[i for i in transform.pyramid_gaussian(right, level)]
leftPyramid=[i for i in transform.pyramid_gaussian(left, level)]
sel=level
beta=estimateRigidTransformation(leftPyramid[sel], rightPyramid[sel], 2.0*dTheta, thr)
return beta
def oversample(I):
samples = []
samples.append(I)
samples.append(ndimage.rotate(I, 90))
samples.append(ndimage.rotate(I, 180))
samples.append(ndimage.rotate(I, 270))
samples.append(np.fliplr(I))
samples.append(np.flipud(I))
samples.append(np.fliplr(samples[1]))
samples.append(np.flipud(samples[1]))
return samples
def align_image_to_ellipse(coeffs, image):
"""
Given the coefficients of an ellipse in 2D and a binary
image, return the angle required to align the image to the
principal axes of the ellipse (with the longest axis
as the first major 'hump' on the left).
"""
coeff_a, coeff_b, coeff_c = coeffs[:3]
# Calculate tan(angle) for the angle of rotation of the major axis
preangle = coeff_b / (coeff_a - coeff_c)
if not np.isinf(preangle):
# Take the arctan and convert to degrees, which is what
# ndimage.rotate uses.
angle = radians_to_degrees(-0.5 * np.arctan(preangle))
# Order = 0 prevents interpolation from being done and screwing
# with our object boundaries.
rotated = ndimage.rotate(image, angle, order=0)
# Pull out the height/width of just the object.
try:
height, width = rotated[ndimage.find_objects(rotated)[0]].shape
except IndexError:
raise EllipseAlignmentError("Can't find object after " \
+ "initial rotation.")
else:
angle = 0.
height, width = image.shape
# we want the height (first axis) to be the major axis.
if width > height:
angle -= 90.0
rotated = ndimage.rotate(image, angle, order=0)
# Correct so that in budding cells, the "major" hump is always
# on the first.
if np.argmax(rotated.sum(axis=1)) > rotated.shape[0] // 2:
angle -= 180.0
rotated = ndimage.rotate(image, angle, order=0)
# Do a find_objects on the resultant array after rotation in
# order to _just_ get the object and not any of the extra
# space that's been added.
try:
bounds = ndimage.find_objects(rotated)[0]
except IndexError:
raise EllipseAlignmentError("Can't find object after final rotation.")
return rotated[bounds], angle
def generate_overlay_image(ct_slice, labelmap_slice, window_width=None, window_level=None, opacity=None):
""" This function that takes as input a 2D ct slice and a labelmap and
returns an overlay
Parameters
----------
ct_image : array, shape ( N, N )
A description
labelmap_slice : array, shape ( N, N )
A description
window_width : int
window_level : int
opacity : float. Not yet implemented
Returns
-------
overlay : array, shape ( N, N, 3)
Array with the r,g,b, pixel values of the overlayed image
"""
assert ct_slice.shape[0] == labelmap_slice.shape[0], "CT slice and label disagree in dimension"
assert ct_slice.shape[1] == labelmap_slice.shape[1], "CT slice and label disagree in dimension"
length = labelmap_slice.shape[0]
width = labelmap_slice.shape[1]
overlayed_image = np.squeeze(ndimage.rotate(labelmap_slice, -90))
ct_image = np.squeeze(ndimage.rotate(ct_slice, -90))
# shift all values to positive
min_ct = np.float(np.min(ct_image))
ct_image = ct_image - min_ct
if window_width == None:
max_ct = np.float(np.max(ct_image))
window_width = max_ct
window_level = max_ct / 2
else:
window_level = window_level - min_ct
rgbArray = np.zeros((length, width, 3), "uint8")
grey_val = ((ct_image.astype(np.float)) * 256.0 / (window_width + window_level)).astype(np.uint8)
rgbArray[..., 0] = rgbArray[..., 1] = rgbArray[..., 2] = grey_val
color_val = overlayed_image > 0
rgbArray[color_val, 0] = 255
return rgbArray
def rotate_images(pixel_array, mask):
param = (np.random.randint(-15,15))
pixel_array = rotate(pixel_array, axes=(2,1), angle=param, reshape=False)
mask = rotate(mask, axes=(0,1), angle=param, reshape=False)
rotated_data = {
"pixel_array" : pixel_array,
"mask" : mask,
"params" : param
}
return rotated_data
def transformImage(self, image, aligndata, refimage): alignedimage = image shift = (aligndata['xshift'], aligndata['yshift']) alignedimage = ndimage.shift(alignedimage, shift=shift, mode='wrap', order=1) ### due to the nature of Radon shifts, it cannot tell the difference between 0 and 180 degrees alignedimage1 = ndimage.rotate(alignedimage, -aligndata['rotangle'], reshape=False, order=1) alignedimage2 = ndimage.rotate(alignedimage, -aligndata['rotangle']+180, reshape=False, order=1) cc1 = self.getCCValue(alignedimage1, refimage) cc2 = self.getCCValue(alignedimage2, refimage) if cc1 > cc2: return alignedimage1 else: return alignedimage2
def loci_asdi(self):
"""
Combine sdi and adi method on the shift_medfiltered cubes
"""
log = self.log
if self._file_exists('','asdi_medcrunch'): return
ft = lambda i: '%d:%.2d' % ((time.time()-i)/60,(time.time()-i)%60)
pa = self.pa
# in case l1,l2 are not defined.
#self._get_filters_scaling(self.idir+self.inputs[0])
l1 = self.l1
l2 = self.l2
cube_red = self._read_fits(('blue','red')[l1 < l2],'cube_shift_medfilter')
cube_blue = self._read_fits(('red','blue')[l1 < l2],'cube_shift_medfilter')
# Lets not use bpr because we run out of memory
#bpr,cube_red = nt.reset_nans(cube_red)
cube_red = np.nan_to_num(cube_red)
log.info('Applying loci method ...(Can take a while......)')
ti = time.time()
#
cube_red = loci_subtract(cube_red, pa, np.nan_to_num(cube_blue))
log.info(' '+ft(ti))
del cube_blue
#cube_red = nt.restore_nans(cube_red,bpr)
self._write_fits(cube_red,'','cube_asdi')
#bpr,cube_red = nt.reset_nans(cube_red)
line = 'rotating combined sdi,adi cube'
cube_red2=cube_red.copy()
sz = np.shape(cube_red2)
for i in range(sz[0]):
log.info(line+'['+str(i+1)+'/'+str(sz[0])+']')
cube_red[i,:,:] = nd.rotate (cube_red[i,:,:], pa[i], reshape=False)
#cube_red[i,:,:] = nd.rotate (cube_red[i,:,:], -pa[i], reshape=False)
#cube_red = nt.restore_nans(cube_red,bpr)
self._write_fits(np.median(cube_red,axis=0),'','asdi_medcrunch')
del cube_red
for i in range(sz[0]):
#Do counter rotate (only noise)
cube_red2[i,:,:] = nd.rotate (cube_red2[i,:,:], -pa[i], reshape=False)
#cube_red2[i,:,:] = nd.rotate (cube_red2[i,:,:], pa[i], reshape=False)
self._write_fits(np.median(cube_red2,axis=0),'','asdi_counter_medcrunch')
del cube_red2
def plot_overlays(atlas, b0, cmaps):
plt.rcParams.update({'axes.labelsize': 'x-large',
'axes.titlesize': 'x-large'})
if b0.shape == (182, 218, 182):
x = [78, 90, 100]
y = [82, 107, 142]
z = [88, 103, 107]
else:
shap = b0.shape
x = [int(shap[0]*0.35), int(shap[0]*0.51), int(shap[0]*0.65)]
y = [int(shap[1]*0.35), int(shap[1]*0.51), int(shap[1]*0.65)]
z = [int(shap[2]*0.35), int(shap[2]*0.51), int(shap[2]*0.65)]
coords = (x, y, z)
labs = ['Sagittal Slice (YZ fixed)',
'Coronal Slice (XZ fixed)',
'Axial Slice (XY fixed)']
var = ['X', 'Y', 'Z']
# create subplot for first slice
# and customize all labels
idx = 0
for i, coord in enumerate(coords):
for pos in coord:
idx += 1
ax = plt.subplot(3, 3, idx)
ax.set_title(var[i] + " = " + str(pos))
if i == 0:
image = ndimage.rotate(b0[pos, :, :], 90)
atl = ndimage.rotate(atlas[pos, :, :], 90)
elif i == 1:
image = ndimage.rotate(b0[:, pos, :], 90)
atl = ndimage.rotate(atlas[:, pos, :], 90)
else:
image = b0[:, :, pos]
atl = atlas[:, :, pos]
if idx % 3 == 1:
ax.set_ylabel(labs[i])
ax.yaxis.set_ticks([0, image.shape[0]/2, image.shape[0] - 1])
ax.xaxis.set_ticks([0, image.shape[1]/2, image.shape[1] - 1])
min_val, max_val = get_min_max(image)
plt.imshow(atl, interpolation='none', cmap=cmaps[0], alpha=0.5)
plt.imshow(image, interpolation='none', cmap=cmaps[1], alpha=0.5,
vmin=min_val, vmax=max_val)
fig = plt.gcf()
fig.set_size_inches(12.5, 10.5, forward=True)
return fig
def warp_ellipse_to_circle(cube, a, b, pa, stop_if_huge=True):
'''
Warp a SpectralCube such that the given ellipse is a circle int the
warped frame.
Since you should **NOT** be doing this with a large cube, we're going
to assume that the given cube is a subcube centered in the middle of the
cube.
This requires a rotation, then scaling. The equivalent matrix is:
[b cos PA b sin PA]
[-a sin PA a cos PA ].
'''
if cube._is_huge:
if stop_if_huge:
raise Warning("The cube has the huge flag enabled. Disable "
"'stop_if_huge' if you would like to continue "
"anyways with the warp.")
else:
warn("The cube has the huge flag enabled. This may use a lot "
"of memory!")
# Let NaNs be 0
data = cube.with_fill_value(0.0).filled_data[:].value
warped_array = []
for i in range(cube.shape[0]):
warped_array.append(nd.zoom(nd.rotate(data[i], np.rad2deg(-pa)),
(1, a / b)))
warped_array = np.array(warped_array)
# We want to mask outside of the original bounds
mask = np.ones(data.shape[1:])
warp_mask = \
np.isclose(nd.zoom(nd.rotate(mask, np.rad2deg(-pa)),
(1, a / b)), 1)
# There's probably a clever way to transform the WCS, but all the
# solutions appear to need pyast/starlink. The output of the wrap should
# give a radius of b and the spectral dimension is unaffected.
# Also this is hidden and users won't be able to use this weird cube
# directly
warped_cube = SpectralCube(warped_array * cube.unit, cube.wcs)
warped_cube = warped_cube.with_mask(warp_mask)
return warped_cube
def simple_adi(img_cube, img_params):
median = np.median(img_cube, axis=0)
rotation = [param[1][-1] for param in img_params]
I = []
I_without_sub = []
img_cube2 = copy.deepcopy(img_cube)
for (img, angle) in zip(img_cube, rotation):
derotated_without_sub = ndimage.rotate(img, math.degrees(angle), reshape=False)
derotated = ndimage.rotate(img-median, math.degrees(angle), reshape=False)
I.append(derotated)
I_without_sub.append(derotated_without_sub)
#plotfast.image(np.array(img_cube))
I = np.median(np.array(I), axis=0)
return I
def rand_aug(im,z):
ang = np.float32(np.random.random(1)*10.0-5.0); ang = ang[0]
rgb = np.float32(np.random.random(3)*0.4+0.8);
ctrst = np.float32(np.random.random(1)+0.5); ctrst = ctrst[0]
im = rotate(im,ang,reshape=False,mode='nearest',order=1).copy()
z = rotate(z,ang,reshape=False,mode='nearest',order=1).copy()
for j in range(3):
im[:,:,j] = im[:,:,j] * rgb[j]
im[...] = np.minimum(255.0,(im[...] ** ctrst) * (255.0**(1.0-ctrst)))
return im,z
def boxMaskStack(bmstackf, partdatas, box, xmask, ymask, falloff, imask=None, norotate=False):
from appionlib.apSpider import operations
from appionlib import apEMAN
import os
# create blank image for mask using SPIDER
maskfile = "boxmask.spi"
operations.createBoxMask(maskfile,box,xmask,ymask,falloff,imask)
# convert mask to MRC
apEMAN.executeEmanCmd("proc2d boxmask.spi boxmask.mrc",verbose=False,showcmd=False)
os.remove("boxmask.spi")
maskarray = mrc.read("boxmask.mrc")
# box particles
maskedparts = []
for i in range(len(partdatas)):
if norotate is True:
rotatemask = maskarray
else:
angle = (-partdatas[i]['angle'])-90
rotatemask = ndimage.rotate(maskarray, angle=angle, reshape=False, order=1)
maskedparts.append(rotatemask)
# write to stack
apImagicFile.writeImagic(maskedparts, bmstackf)
os.remove("boxmask.mrc")
return bmstackf
def create_transformed_array(array_original, reflectz, reflecty, reflectx, swapxy, angle, rotation_order, scaling_factor=None):
array = array_original
if array.ndim > 3:
try:
array = array.reshape(array.shape[-3:])
except:
raise ValueError("Can't transform ndarray with more than 3 dimensions with length > 1. "
"This array's shape is {}".format(array.shape))
if scaling_factor is not None:
array = array * scaling_factor
if reflectz:
array = array[::-1, :, :]
if reflecty:
array = array[:, ::-1, :]
if reflectx:
array = array[:, :, ::-1]
if swapxy:
array = array.transpose((0, 2, 1))
if angle > 0:
new_array = ndimage.rotate(array.astype(np.float32),
angle,
axes=(1, 2),
order=rotation_order,
cval=0)
else:
new_array = array
if new_array.dtype != array_original.dtype:
print('dtype mismatch: new_array.dtype = {0}, array_original.dtype = {1}'
.format(new_array.dtype, array_original.dtype
))
return new_array
def largest_region(imData):
belowMeanFilter = np.where(imData > np.mean(imData), 0., 1.0)
dialated = morphology.dilation(belowMeanFilter, np.ones((3, 3)))
regionLabels = (belowMeanFilter * measure.label(dialated)).astype(int)
# calculate common region properties for each region within the segmentation
regions = measure.regionprops(regionLabels)
areas = [(None
if sum(belowMeanFilter[regionLabels == region.label]) * 1.0 / region.area < 0.50
else region.filled_area)
for region in regions]
if len(areas) > 0:
regionMax = regions[np.argmax(areas)]
# trim image to the max region
regionMaxImg = trim_image(
np.minimum(
imData*np.where(regionLabels == regionMax.label, 1, 255),
255))
# rotate
angle = intertial_axis(regionMaxImg)[2]
rotatedRegionMaxImg = ndimage.rotate(regionMaxImg, np.degrees(angle))
rotatedRegionMaxImg = trim_image(trim_image(rotatedRegionMaxImg, 0), 255)
else:
regionMax = None
rotatedRegionMaxImg = None
angle = 0
return regionMax, rotatedRegionMaxImg, angle, regionLabels, regions, areas, belowMeanFilter, dialated
def plot_rotation(imData):
def plot_bars(x_bar, y_bar, angle):
def plot_bar(r, x_center, y_center, angle, pattern):
dx = r * np.cos(angle)
dy = r * np.sin(angle)
plt.plot([x_center - dx, x_center, x_center + dx],
[y_center - dy, y_center, y_center + dy], pattern)
plot_bar(10, x_bar, y_bar, angle + np.radians(90), 'bo-')
plot_bar(30, x_bar, y_bar, angle, 'ro-')
def plot_subplot(pawprint):
x_bar, y_bar, angle = intertial_axis(pawprint)
plt.imshow(pawprint, cmap=cm.gray)
plot_bars(x_bar, y_bar, angle)
return angle
plt.figure()
angle = plot_subplot(imData)
plt.title('Original')
plt.figure()
plot_subplot(ndimage.rotate(imData, np.degrees(angle)))
plt.title('Rotated')
def shared_dataset(data_xy, augument=False, borrow=True):
data_x, data_y = data_xy
data_x_npy = numpy.asarray(data_x, dtype=theano.config.floatX)
data_y_npy = numpy.asarray(data_y, dtype=theano.config.floatX)
if augument:
data_x_npy = numpy.concatenate((data_x_npy, data_x_npy))
data_y_npy = numpy.concatenate((data_y_npy, data_y_npy))
for x in range(data_x_npy.shape[0] / 2):
sqrt = numpy.sqrt(data_x_npy.shape[1])
rot = random.uniform(5, -5)
rotated = numpy.reshape(data_x_npy[x, :], (sqrt, sqrt))
rotated = srot.rotate(rotated, rot)
rotated = rotated[(rotated.shape[0] - sqrt) / 2: (rotated.shape[0] - sqrt) / 2 + sqrt,
(rotated.shape[1] - sqrt) / 2: (rotated.shape[1] - sqrt) / 2 + sqrt]
data_x_npy[x, :] = numpy.reshape(rotated, (sqrt * sqrt,))
print data_x_npy.shape
# add required 1 vector for bias
# shared_x = theano.shared(numpy.hstack((data_x_npy,
# numpy.ones((data_x_npy.shape[0], 1), dtype=tensor.dscalar()))),
# borrow=borrow)
shared_x = theano.shared(data_x_npy, borrow=borrow)
shared_y = theano.shared(data_y_npy, borrow=borrow)
return shared_x, tensor.cast(shared_y, 'int32')
def clip_image(image,roi=[],rotangle=0.0,cp=False):
"""
Clip an image given the roi
Parameters:
-----------
* roi is a list [c1,r1,c2,r2] = [x1,y1,x2,y2]
Note take x as the horizontal image axis index (col index),
and y as vertical axis index (row index).
Therefore, in image indexing:
image[y1:y2,x1:x2] ==> rows y1 to y2 and cols x1 to x2
* rotangle is the angle to rotate the image by
* cp is flag to indicate if a new copy of the image is generated
"""
if len(roi) != 4:
roi = [0,0,image.shape[1], image.shape[0]]
else:
[c1,r1,c2,r2] = _sort_roi(roi)
# rotate
if rotangle != 0:
image = ndimage.rotate(image,rotangle)
if cp == True:
return copy.copy(image[r1:r2, c1:c2])
else:
return image[r1:r2, c1:c2]
def _transform(self, images_org):
if self.image_options["crop"]:
resize_size = int(self.image_options["resize_size"])
y = np.random.permutation(range(resize_size/2, images_org.shape[0]-resize_size/2))
y = int(y[0])
y1 = int(y - resize_size/2.0)
y2 = int(y + resize_size/2.0)
x = np.random.permutation(range(resize_size/2, images_org.shape[1]-resize_size/2))
x = int(x[0])
x1 = int(x - resize_size/2.0)
x2 = int(x + resize_size/2.0)
image = images_org[y1:y2, x1:x2,...]
if self.image_options["resize"]:
resize_size = int(self.image_options["resize_size"])
image = misc.imresize(image, [resize_size, resize_size], interp='nearest')
if self.image_options["flip"]:
if(np.random.rand()<.5):
image = image[::-1,...]
if(np.random.rand()<.5):
image = image[:,::-1,...]
if self.image_options["rotate_stepwise"]:
if(np.random.rand()>.25): # skip "0" angle rotation
angle = int(np.random.permutation([1,2,3])[0] * 90)
image = rotate(image, angle, reshape=False)
return np.array(image)
def morph_mean(src, angle):
dst = ndimage.rotate(src, angle)
strel = cv2.getStructuringElement(cv2.MORPH_RECT, (1, 16))
dst = cv2.morphologyEx(dst, cv2.MORPH_OPEN, strel)
mean = cv2.mean(dst)[0]
print mean
return mean
def boxerRotate(imgfile, parttree, outstack, boxsize):
"""
boxes the particles with expanded size,
applies a rotation to particle,
reduces boxsize to requested size,
and saves them to a imagic file
"""
# size needed is sqrt(2)*boxsize, using 1.5 to be extra safe
bigboxsize = int(math.ceil(1.5*boxsize))
imgarray = mrc.read(imgfile)
bigboxedparticles = boxerMemory(imgarray, parttree, bigboxsize)
boxedparticles = []
boxshape = (boxsize,boxsize)
apDisplay.printMsg("Rotating particles...")
for i in range(len(bigboxedparticles)):
if i % 10 == 0:
sys.stderr.write(".")
bigboxpart = bigboxedparticles[i]
partdict = parttree[i]
### add 90 degrees because database angle is from x-axis not y-axis
angle = partdict['angle']+90.0
rotatepart = ndimage.rotate(bigboxpart, angle=angle, reshape=False, order=1)
boxpart = imagefilter.frame_cut(rotatepart, boxshape)
boxedparticles.append(boxpart)
sys.stderr.write("done\n")
apImagicFile.writeImagic(boxedparticles, outstack)
return True
def HOGpicture(w, bs, positive=True): # w=feature, bs=size, # construct a "glyph" for each orientaion bim1 = np.zeros([bs,bs]) bim1[:,round(bs/2):round(bs/2)+1] = 1 bim = np.zeros([bim1.shape[0],bim1.shape[1], 9]) bim[:,:,1] = bim1 for i in range(2,10): bim[:,:,i-1] = nd.rotate(bim1, -(i-1)*20, reshape=False) #crop? # make pictures of positive weights bs adding up weighted glyphs shape_ = w.shape if positive: w[w<0] = 0 else: w[w>0] = 0 # im = np.zeros([bs*shape_[0], bs*shape_[1]]) im = np.zeros([bs*shape_[1], bs*shape_[0]]) for i in range(1,shape_[0]): for j in range(1,shape_[1]): for k in range(9): # im[(i-1)*bs:i*bs, (j-1)*bs:j*bs] += bim[:,:,k]*w[i,j,k] im[(j-1)*bs:j*bs,(i-1)*bs:i*bs] += bim[:,:,k]*w[i,j,k] return im
def lena_example():
im = rotated_lena()
# find corners on grayscale image
# use negative so that boundary is 0
bw = im[:, :, 0]
bw = 255 - bw
corners, outline = guess_corners(bw)
angle = mean_rotation(corners) * 180./np.pi
im2 = clear_border(im, outline)
im2 = ndimage.rotate(im, angle, reshape=False, mode='nearest')
cropped = keyboard_crop(im2)
fig, (ax1, ax2, ax3) = plt.subplots(3)
ax1.imshow(im)
ax1.plot(corners[:, 1], corners[:, 0], 'o')
ax2.imshow(im2)
ax3.imshow(cropped)
fig, ((ax1, ax2), (ax3, ax4)) = plt.subplots(2, 2)
ax1.imshow(cropped[:20, :20])
ax2.imshow(cropped[:20, -20:])
ax3.imshow(cropped[-20:, :20])
ax4.imshow(cropped[-20:, -20:])
return cropped
def rotation(Pfad, image_path, lower, upper):
image = cv2.imread(Pfad + image_path)
# create border
row, col = image.shape[:2]
bottom = image[0:row, 0:col]
mean = cv2.mean(bottom)[0]
bordersize = 5
border = cv2.copyMakeBorder(image,
top=bordersize,
bottom=bordersize,
left=bordersize,
right=bordersize,
borderType=cv2.BORDER_CONSTANT,
value=[mean, mean, mean])
## edge detection on image with border
gray = cv2.cvtColor(border, cv2.COLOR_BGR2GRAY)
gray = cv2.GaussianBlur(gray, (5, 5), 0)
edged = cv2.Canny(gray, 75, 200)
# find angle
cnts, hierarchy = cv2.findContours(edged.copy(), cv2.RETR_EXTERNAL,
cv2.CHAIN_APPROX_SIMPLE)
# contour with the largest area is possibly the plate
max_area = 0
max_cnt = None
for cnt in cnts:
area = cv2.contourArea(cnt)
if (area > max_area):
max_area = area
max_cnt = cnt
if max_cnt is not None:
min_rect = cv2.minAreaRect(max_cnt)
(midpoint, widthheight, angle) = min_rect
# Get the image size
# NumPy stores image matricies backwards
image_size = (border.shape[1], border.shape[0])
image_center = tuple(np.array(image_size) / 2)
# Convert the OpenCV 3x2 rotation matrix to 3x3
rot_mat = np.vstack(
[cv2.getRotationMatrix2D(image_center, angle, 1.0), [0, 0, 1]])
rot_mat_notranslate = np.matrix(rot_mat[0:2, 0:2])
# Shorthand for below calcs
image_w2 = image_size[0] * 0.5
image_h2 = image_size[1] * 0.5
# Obtain the rotated coordinates of the image corners
rotated_coords = [
(np.array([-image_w2, image_h2]) * rot_mat_notranslate).A[0],
(np.array([image_w2, image_h2]) * rot_mat_notranslate).A[0],
(np.array([-image_w2, -image_h2]) * rot_mat_notranslate).A[0],
(np.array([image_w2, -image_h2]) * rot_mat_notranslate).A[0]
]
# Find the size of the new image
x_coords = [pt[0] for pt in rotated_coords]
x_pos = [x for x in x_coords if x > 0]
x_neg = [x for x in x_coords if x < 0]
y_coords = [pt[1] for pt in rotated_coords]
y_pos = [y for y in y_coords if y > 0]
y_neg = [y for y in y_coords if y < 0]
right_bound = max(x_pos)
left_bound = min(x_neg)
top_bound = max(y_pos)
bot_bound = min(y_neg)
new_w = int(abs(right_bound - left_bound))
new_h = int(abs(top_bound - bot_bound))
# translation matrix to keep the image centred
trans_mat = np.matrix([[1, 0, int(new_w * 0.5 - image_w2)],
[0, 1, int(new_h * 0.5 - image_h2)], [0, 0, 1]])
# Compute the tranform for the combined rotation and translation
affine_mat = (np.matrix(trans_mat) * np.matrix(rot_mat))[0:2, :]
# Apply the transform --> (-0° to -45° License Plates, which are tilted to the left and -46° to -90° License Plates, which are tilted to the right)
rotation1 = cv2.warpAffine(image,
affine_mat, (new_w, new_h),
flags=cv2.INTER_LINEAR)
# correction of rotation for -45 > angle >= -90
rotation2 = ndimage.rotate(rotation1, 90)
# if angle is not between -20° and -70°, keep original image in pipeline
if (lower >= angle >= -45.0):
rotated_image = rotation1
rotated_image = Image.fromarray(rotated_image)
rotated_image = rotated_image.convert('L')
rotated_image.save(Pfad + image_path)
elif (-45 > angle >= upper):
rotated_image = rotation2
rotated_image = Image.fromarray(rotated_image)
rotated_image = rotated_image.convert('L')
rotated_image.save(Pfad + image_path)
else:
image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
cv2.imwrite(Pfad + image_path, image)
else:
image = cv2.cvtColor(image, cv2.COLOR_BGR2GRAY)
cv2.imwrite(Pfad + image_path, image)
def plot_vehicle_nice_mv(ax, predictions, dt, max_hl=10, ph=6, map=None, x_min=0, y_min=0):
prediction_dict, histories_dict, futures_dict = prediction_output_to_trajectories(predictions,
dt,
max_hl,
ph,
map=map)
assert (len(prediction_dict.keys()) <= 1)
if len(prediction_dict.keys()) == 0:
return
ts_key = list(prediction_dict.keys())[0]
prediction_dict = prediction_dict[ts_key]
histories_dict = histories_dict[ts_key]
futures_dict = futures_dict[ts_key]
if map is not None:
ax.imshow(map.fdata, origin='lower', alpha=0.5)
cmap = ['k', 'b', 'y', 'g', 'r']
line_alpha = 0.7
line_width = 0.2
edge_width = 2
circle_edge_width = 0.5
node_circle_size = 0.3
a = []
i = 0
node_list = sorted(histories_dict.keys(), key=lambda x: x.id)
for node in node_list:
h = node.get(np.array([ts_key]), {'heading': ['°']})[0, 0]
history_org = histories_dict[node] + np.array([x_min, y_min])
history = histories_dict[node] + np.array([x_min, y_min]) + 5 * np.array([np.cos(h), np.sin(h)])
future = futures_dict[node] + np.array([x_min, y_min]) + 5 * np.array([np.cos(h), np.sin(h)])
predictions = prediction_dict[node] + np.array([x_min, y_min]) + 5 * np.array([np.cos(h), np.sin(h)])
if node.type.name == 'VEHICLE':
for t in range(predictions.shape[2]):
sns.kdeplot(predictions[0, :, t, 0], predictions[0, :, t, 1],
ax=ax, shade=True, shade_lowest=False,
color=line_colors[i % len(line_colors)], zorder=600, alpha=1.0)
r_img = rotate(cars[i % len(cars)], node.get(np.array([ts_key]), {'heading': ['°']})[0, 0] * 180 / np.pi,
reshape=True)
oi = OffsetImage(r_img, zoom=0.08, zorder=700)
veh_box = AnnotationBbox(oi, (history_org[-1, 0], history_org[-1, 1]), frameon=False)
veh_box.zorder = 700
ax.add_artist(veh_box)
i += 1
else:
for t in range(predictions.shape[2]):
sns.kdeplot(predictions[:, t, 0], predictions[:, t, 1],
ax=ax, shade=True, shade_lowest=False,
color='b', zorder=600, alpha=0.8)
# Current Node Position
circle = plt.Circle((history[-1, 0],
history[-1, 1]),
node_circle_size,
facecolor='g',
edgecolor='k',
lw=circle_edge_width,
zorder=3)
ax.add_artist(circle)
"""
import scipy.ndimage as nd
import scipy.misc as misc
from imreg.models import model
from imreg.metrics import metric
from imreg.samplers import sampler
from imreg.visualize import plot
from imreg import register
# Form some test data (lena, lena rotated 20 degrees)
image = register.RegisterData(misc.lena())
template = register.RegisterData(nd.rotate(image.data, 20, reshape=False))
# Form the registrator.
affine = register.Register(model.Affine, metric.Residual,
sampler.CubicConvolution)
fullSearch = []
# Image pyramid registration can be executed like so:
pHat = None
for factor in [30., 20., 10., 5., 2., 1.]:
if pHat is not None:
scale = downImage.coords.spacing / factor
# FIXME: Find a nicer way to do this.
pHat = model.Affine.scale(pHat, scale)
def step(self, action):
X, Y = np.meshgrid(self.moveCenters, self.moveCenters)
coords = np.stack([np.reshape(Y, [-1]), np.reshape(X, [-1])], axis=0)
halfSide = self.blockSize / 2
position = action % self.num_moves
pickplace = action / (self.num_moves * self.num_orientations)
orientation = (action - pickplace * self.num_moves *
self.num_orientations) / self.num_moves
# orientation = action / self.num_moves # DEBUG
# if PICK
if pickplace == 0:
# if not holding anything
if self.state[1] == 0:
ii = coords[0, position]
jj = coords[1, position]
jjRangeInner, iiRangeInner = np.meshgrid(
range(jj - halfSide + self.gap, jj + halfSide - self.gap),
range(ii - halfSide + 1, ii + halfSide - 1))
jjRangeOuter, iiRangeOuter = np.meshgrid(
range(jj - halfSide, jj + halfSide),
range(ii - halfSide, ii + halfSide))
# if there's something in this spot
if np.sum(self.state[0][iiRangeOuter, jjRangeOuter, 0]) > 0:
im = self.state[0][iiRangeOuter, jjRangeOuter, 0]
# if it's a separate object
shape = np.shape(im)
jjGap1, iiGap1 = np.meshgrid(
range(0, shape[1]),
np.r_[range(0, 1),
range(shape[1] - 1, shape[1])])
jjGap2, iiGap2 = np.meshgrid(
np.r_[range(0, 1),
range(shape[1] - 1, shape[1])],
range(0, shape[1]))
if (np.sum(im[iiGap1, jjGap1]) < 1) and (np.sum(
im[iiGap2, jjGap2]) < 1):
if orientation == 0:
imRot = np.int32(im > 0.5)
# imRot = np.int32(ndimage.rotate(im, 45, reshape=False)>0.5)
elif orientation == 1:
# imRot = np.int32(im>0.5)
imRot = np.int32(
ndimage.rotate(im, 45, reshape=False) > 0.5)
elif orientation == 2:
# imRot = np.int32(im>0.5)
imRot = np.int32(
ndimage.rotate(im, 90, reshape=False) > 0.5)
# imRot = np.int32(ndimage.rotate(im, 135, reshape=False)>0.5)
elif orientation == 3:
# imRot = np.int32(im>0.5)
# imRot = np.int32(ndimage.rotate(im, 90, reshape=False)>0.5)
imRot = np.int32(
ndimage.rotate(im, 135, reshape=False) > 0.5)
else:
print("error in orientation 1")
# if there's a gap for the fingers
shape = np.shape(imRot)
jjGap, iiGap = np.meshgrid(
np.r_[range(0, self.gap),
range(shape[1] - self.gap, shape[1])],
range(0, shape[0]))
if np.sum(imRot[iiGap, jjGap]) < 1:
self.holdingImage = imRot
self.state[
1] = 1 # set holding to contents of action target
self.state[0][iiRangeOuter, jjRangeOuter,
0] = np.zeros([
np.shape(iiRangeOuter)[0],
np.shape(jjRangeOuter)[1]
])
# if PLACE
elif pickplace == 1:
if self.state[1] != 0:
ii = coords[0, position]
jj = coords[1, position]
jjRangeOuter, iiRangeOuter = np.meshgrid(
range(jj - halfSide, jj + halfSide),
range(ii - halfSide, ii + halfSide))
if True not in (self.state[0][iiRangeOuter, jjRangeOuter, 0] >
0): # if this square is empty
im = self.holdingImage
if orientation == 0:
imRot = im
# imRot = ndimage.rotate(im, -45, reshape=False)
elif orientation == 1:
# imRot = im
imRot = ndimage.rotate(im, -45, reshape=False)
elif orientation == 2:
# imRot = im
imRot = ndimage.rotate(im, -90, reshape=False)
# imRot = ndimage.rotate(im, -135, reshape=False)
elif orientation == 3:
# imRot = im
# imRot = ndimage.rotate(im, -90, reshape=False)
imRot = ndimage.rotate(im, -135, reshape=False)
else:
print("error in orientation 2")
# self.state[0][iiRangeOuter,jjRangeOuter,0] = np.copy(self.holdingImage)
self.state[0][iiRangeOuter, jjRangeOuter, 0] = imRot
self.state[1] = 0 # set holding to zero
else:
print("error: action out of bounds")
# check for termination condition
reward = 0
done = 0
# check for two horizontal, vertical, or diagonally adjacent blocks
if self.state[1] == 0: # if both objs on the grid
bounds = self.getBoundingBox(self.state[0][:, :, 0])
if (bounds[1] - bounds[0]) < (self.blockSize * 0.7):
if (bounds[3] - bounds[2]) < (self.blockSize * 3.0):
done = 1
reward = 10
if (bounds[1] - bounds[0]) < (self.blockSize * 3.0):
if (bounds[3] - bounds[2]) < (self.blockSize * 0.7):
done = 1
reward = 10
im = np.int32(
ndimage.rotate(self.state[0][:, :, 0], 45, reshape=True) > 0.5)
bounds = self.getBoundingBox(im)
if (bounds[1] - bounds[0]) < (self.blockSize * 0.9):
if (bounds[3] - bounds[2]) < (self.blockSize * 3.0):
done = 1
reward = 10
if (bounds[1] - bounds[0]) < (self.blockSize * 3.0):
if (bounds[3] - bounds[2]) < (self.blockSize * 0.9):
done = 1
reward = 10
if self.episode_timer > self.max_episode:
self.episode_timer = 0
done = 1
self.episode_timer += 1
return self.state, reward, done, {}
def plot_vehicle_nice(ax, predictions, dt, max_hl=10, ph=6, map=None, x_min=0, y_min=0):
prediction_dict, histories_dict, futures_dict = prediction_output_to_trajectories(predictions,
dt,
max_hl,
ph,
map=map)
assert (len(prediction_dict.keys()) <= 1)
if len(prediction_dict.keys()) == 0:
return
ts_key = list(prediction_dict.keys())[0]
prediction_dict = prediction_dict[ts_key]
histories_dict = histories_dict[ts_key]
futures_dict = futures_dict[ts_key]
if map is not None:
ax.imshow(map.fdata, origin='lower', alpha=0.5)
cmap = ['k', 'b', 'y', 'g', 'r']
line_alpha = 0.7
line_width = 0.2
edge_width = 2
circle_edge_width = 0.5
node_circle_size = 0.3
a = []
i = 0
node_list = sorted(histories_dict.keys(), key=lambda x: x.id)
for node in node_list:
history = histories_dict[node] + np.array([x_min, y_min])
future = futures_dict[node] + np.array([x_min, y_min])
predictions = prediction_dict[node] + np.array([x_min, y_min])
if node.type.name == 'VEHICLE':
# ax.plot(history[:, 0], history[:, 1], 'ko-', linewidth=1)
ax.plot(future[:, 0],
future[:, 1],
'w--o',
linewidth=4,
markersize=3,
zorder=650,
path_effects=[pe.Stroke(linewidth=5, foreground='k'), pe.Normal()])
#print(node.id, future.shape)
for t in range(predictions.shape[2]):
sns.kdeplot(predictions[0, :, t, 0], predictions[0, :, t, 1],
ax=ax, shade=True, shade_lowest=False,
color=line_colors[i % len(line_colors)], zorder=600, alpha=0.8)
vel = node.get(np.array([ts_key]), {'velocity': ['x', 'y']})
h = np.arctan2(vel[0, 1], vel[0, 0])
r_img = rotate(cars[i % len(cars)], node.get(np.array([ts_key]), {'heading': ['°']})[0, 0] * 180 / np.pi, reshape=True)
oi = OffsetImage(r_img, zoom=0.025, zorder=700)
veh_box = AnnotationBbox(oi, (history[-1, 0], history[-1, 1]), frameon=False)
veh_box.zorder = 700
ax.add_artist(veh_box)
i += 1
else:
# ax.plot(history[:, 0], history[:, 1], 'k--')
for t in range(predictions.shape[2]):
sns.kdeplot(predictions[0, :, t, 0], predictions[0, :, t, 1],
ax=ax, shade=True, shade_lowest=False,
color='b', zorder=600, alpha=0.8)
ax.plot(future[:, 0],
future[:, 1],
'w--',
zorder=650,
path_effects=[pe.Stroke(linewidth=edge_width, foreground='k'), pe.Normal()])
# Current Node Position
circle = plt.Circle((history[-1, 0],
history[-1, 1]),
node_circle_size,
facecolor='g',
edgecolor='k',
lw=circle_edge_width,
zorder=3)
ax.add_artist(circle)
key = cv2.waitKey(1)
# exit on 'q' 'esc' 'Q' //salida a la 'q' 'esc' 'Q'
if key in [27, ord('Q'), ord('q')]:
break
# resize the captured frame for face detection to increase processing speed //
# cambiar el tamano del fotograma capturado para la deteccion de cara a aumentar la velocidad de procesamiento
resized_frame = cv2.resize(frame, frame_scale)
processed_frame = resized_frame
# Skip a frame if the no face was found last frame // Saltar de un marco de la cara no se encontro ultimo fotograma
if frame_skip_rate == 0:
faceFound = False
for rotation in current_rotation_map:
rotated_frame = ndimage.rotate(resized_frame, rotation)
gray = cv2.cvtColor(rotated_frame, cv2.COLOR_BGR2GRAY)
# return tuple is empty, ndarray if detected face // retorno tupla esta vacia, array si se detecta la cara
faces = face_cascade.detectMultiScale(
gray,
scaleFactor=1.3,
minNeighbors=5,
minSize=(30, 30),
flags=cv2.cv.CV_HAAR_SCALE_IMAGE)
# If frontal face detector failed, use profileface detector
# // Si la deteccion de caras frontal fallo, perfil de usuario cara de deteccion otro xml
faces = faces if len(faces) else sideFace_cascade.detectMultiScale(
gray,
slika1 = cv2.addWeighted(slika1, 0.6, prazna_slika, 1, 0.0)
return slika1
##PRETVORBA VIDEO V SLIKE
vidcap = cv2.VideoCapture('detection.mp4')
success, image = vidcap.read()
count = 0
alpha = 2 # Contrast control (1.0-3.0)
beta = -140 # Brightness control (0-100)
img_array = []
while success:
if count % 2 == 0:
try:
image = rotate(image, -90)
image = image[500:1000, 0:720]
slika_ceste = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
slika_ceste = cv2.convertScaleAbs(slika_ceste,
alpha=alpha,
beta=beta)
for x in range(500):
for y in range(720):
if (slika_ceste[x, y][0] <
130) or (slika_ceste[x, y][1] <
130) or (slika_ceste[x, y][2] < 130):
slika_ceste[x, y][0] = 0
slika_ceste[x, y][1] = 0
slika_ceste[x, y][2] = 0
upper = int(min(255, (1.0 + sigma) * v))
edges = cv2.Canny(new_img, lower, upper, apertureSize=3)
cv2.imwrite(sys.argv[1] + "_edges.jpg", edges)
#cv2.imshow('canny_edge', img_edges)
key = cv2.waitKey(0)
angles = []
lines_image = img_before.copy()
lines = cv2.HoughLinesP(edges, 1, np.pi / 180, 80, 30, 10)
for k in range(0, len(lines)):
for x1, y1, x2, y2 in lines[k]:
angle = math.degrees(math.atan2(y2 - y1, x2 - x1))
if angle != 0 and angle != 90 and angle != -90:
print("Angle is %f" % angle)
angles.append(angle)
cv2.line(lines_image, (x1, y1), (x2, y2), (0, 255, 0), 10)
cv2.imwrite(sys.argv[1] + "_houghlines.jpg", lines_image)
#cv2.imshow('After', img_before)
key = cv2.waitKey(0)
median_angle = np.median(angles)
print("Median angle is: %f" % median_angle)
img_rotated = ndimage.rotate(img_before, median_angle)
cv2.imwrite(sys.argv[2], img_rotated)
yStartPos = (shape[0] - ysize)
finalY = True
return slices
def rename_files():
for file in glob.glob('{}{}/*_.bmp'.format(inputDirectory, 'images')):
os.rename(file, file.replace("_.bmp", ".bmp"))
if __name__ == '__main__':
files = glob.glob('{}{}/*.bmp'.format(inputDirectory, 'images'))
print("Found {} files".format(len(files)))
file_count = 1
for file in files:
fileName = os.path.basename(file)
image = plt.imread(file)
csv_values = []
dataArrays = []
read_csv(fileName)
if not csv_values:
continue
imageSlices = slice_image(image)
imageSliceRotated = slice_image(ndimage.rotate(image, 90))
plot_image_rotate(imageSlices, imageSliceRotated, image)
if file_count % 10 == 0:
print("Processed {} files".format(file_count))
file_count += 1
def fix_frame(frame):
frame = ndimage.rotate(frame, -120, reshape=False)
frame = images.crop_and_mask(frame, crop_result.bbox, crop_result.mask)
return frame
def get_crop_result(vid):
frame = vid.read_next_frame()
frame = ndimage.rotate(frame, -120, reshape=False)
crop_result = images.crop_rectangle(frame)
vid.set_frame(0)
return crop_result
import cv2
import numpy as np
import math
import pytesseract
from scipy import ndimage
## (1) read
img = cv2.imread(r"C:\Users\Internship004\Desktop\rot3.jpg")
gray = cv2.cvtColor(img, cv2.COLOR_BGR2GRAY)
img_edges = cv2.Canny(gray, 100, 100, apertureSize=3)
# Using Houghlines to detect lines
lines = cv2.HoughLinesP(img_edges, 1, math.pi / 180.0, 100, minLineLength=100, maxLineGap=5)
angles = []
for x1, y1, x2, y2 in lines[0]:
angle = math.degrees(math.atan2(y2 - y1, x2 - x1))
angles.append(angle)
# Getting the median angle
median_angle = np.median(angles)
# Rotating the image with this median angle
img_rotated = ndimage.rotate(img, median_angle)
cv2.imwrite(r"C:\Users\Internship004\Desktop\orientation_corrected.jpg", img_rotated)
print(pytesseract.image_to_string(img_rotated))
def video(
*,
scene,
resolution_ratio,
frame_rate,
exp_time,
drift_angle,
drift_velocity,
angle_velocity,
ccd_size,
start,
):
"""
Get video frames from input scene, applying appropriate motion blur
Args:
scene (ndarray): high resolution input scene
resolution_ratio (float): downsample factor to low resolution images
frame_rate (float): video frame rate
exp_time (float): experiment duration
drift_angle (float): linear drift direction (deg)
drift_velocity (float): linear drift velocity (pix / s)
angle_velocity (float): camera rotation rate (deg / s)
ccd_size (int): size of square detector ccd (pixels)
start (tuple): start location of detector in scene
"""
num_frames = exp_time * frame_rate
def coord(k):
return np.array(
(start[0] - k * drift_velocity * np.sin(np.deg2rad(drift_angle)) *
resolution_ratio / (frame_rate),
start[1] + k * drift_velocity * np.cos(np.deg2rad(drift_angle)) *
resolution_ratio / (frame_rate))).astype(int).T
# FIXME check box bounds correctly, need to account for rotation
assert (0 <= coord(0)[0] < scene.shape[0]
and 0 <= coord(0)[1] < scene.shape[1]
and 0 <= coord(num_frames)[0] < scene.shape[0]
and 0 <= coord(num_frames)[1] < scene.shape[1]
), f"Frames drift outside of scene bounds \
({coord(0)[0]}, {coord(0)[1]}) -> ({coord(num_frames)[0]}, {coord(num_frames)[1]})"
# calculate the middle points for all frames
mid = coord(np.arange(num_frames + 1))
# initialize frame images
frames = np.zeros((num_frames, ccd_size[0], ccd_size[1]))
# calculate each frame by integrating high resolution image along the drift
# direction
for frame in tqdm(range(num_frames), desc='Frames', leave=None,
position=1):
hr_size = np.array(ccd_size) * resolution_ratio
hr_frame = np.zeros(hr_size)
# calculate middle coordinates for the shortest line connecting the
# middle coordinates of the consecutive frames
path_rows, path_cols = line(mid[frame][0], mid[frame][1],
mid[frame + 1][0], mid[frame + 1][1])
total_rotation = exp_time * angle_velocity
angles = total_rotation * np.sqrt(
(path_rows - mid[0][0])**2 +
(path_cols - mid[0][1])**2) / np.linalg.norm(mid[-1] - mid[0])
if len(path_rows) > 1:
path_rows, path_cols = path_rows[:-1], path_cols[:-1]
for row, col, angle in zip(path_rows, path_cols, angles):
# accelerate algorithm by not rotating if angle_velocity is 0
if angle_velocity == 0:
slice_x = slice(row - hr_size[0] // 2,
row + (hr_size[0] + 1) // 2)
slice_y = slice(col - hr_size[1] // 2,
col + (hr_size[1] + 1) // 2)
hr_frame += scene[slice_x, slice_y]
else:
# diameter of circumscribing circle
circum_diam = int(np.ceil(np.linalg.norm(hr_size)))
slice_x = slice(row - circum_diam // 2,
row + (circum_diam + 1) // 2)
slice_y = slice(row - circum_diam // 2,
row + (circum_diam + 1) // 2)
unrotated = scene[slice_x, slice_y]
hr_frame += crop(rotate(unrotated, angle, reshape='largest'),
width=hr_size)
# scale collected energy of subframes
hr_frame /= frame_rate * len(path_rows)
frames[frame] = rescale(hr_frame,
1 / resolution_ratio,
anti_aliasing=False)
return frames, mid
return outputs.index(max(outputs))
'*****************************************************************'
imgFull = cv2.imread(
"C:/New Folder/PCB inspection/Inspecting SMD/123clahe.jpg", 0)
template = cv2.imread(
"C:/New Folder/PCB inspection/Inspecting SMD/template12.jpg", 0)
GoldenImg = cv2.imread("C:/New Folder/PCB inspection/Inspecting SMD/123.jpg")
template1 = cv2.imread(
"C:/New Folder/PCB inspection/Inspecting SMD/template2NewCol.jpg", 0)
data = TemplateMatching(imgFull, template, 100)
data1 = TemplateMatching(imgFull, template1, 45)
rotated = ndimage.rotate(template, 90)
data2 = TemplateMatching(imgFull, rotated, 10)
fourier_desc = FourierDescriptor(data, GoldenImg, template, 0)
fourier_desc1 = FourierDescriptor(data1, GoldenImg, template1, 1)
fourier_desc2 = FourierDescriptor(data2, GoldenImg, rotated, 2)
database = np.vstack((fourier_desc, fourier_desc1, fourier_desc2))
normalizedFourier = np.zeros(database.shape, np.float)
for i in range(0, MIN_DESCRIPTOR):
normalizedFourier[:, i] = NormalizeFourier(database[:, i])
normalizedFourier[:, MIN_DESCRIPTOR] = database[:, MIN_DESCRIPTOR]
dataset = normalizedFourier
n_inputs = len(dataset[0]) - 1
n_outputs = len(set([row[-1] for row in dataset]))
network = initialize_network(n_inputs, 30, n_outputs)
train_network(network, dataset, 0.4, 5000, n_outputs)
HRpQCT_Files = [
File for File in os.listdir(HRpQCT_Path) if File.endswith('.mhd')
]
HRpQCT_Files.sort()
## Load uCT mask and rotate and segment it
uCT_Mask = sitk.ReadImage(uCT_Path + uCT_Mask_File)
Spacing = uCT_Mask.GetSpacing()
Origin = uCT_Mask.GetOrigin()
Direction = uCT_Mask.GetDirection()
uCT_Mask_Array = sitk.GetArrayFromImage(uCT_Mask)
DSCs = pd.read_csv(Results_Path + 'DSCs.csv')
BestAngle = int(DSCs.loc[DSCs['DSC'].idxmax(), 'Angle'])
uCT_Mask_Array = ndimage.rotate(uCT_Mask_Array,
-BestAngle, (1, 2),
reshape=True)
uCT_Mask_Array = np.rot90(uCT_Mask_Array, 2, (0, 1))
Otsu_Filter = sitk.OtsuThresholdImageFilter()
Otsu_Filter.SetInsideValue(0)
Otsu_Filter.SetOutsideValue(1)
Segmentation = Otsu_Filter.Execute(uCT_Mask)
R_Threshold = Otsu_Filter.GetThreshold()
uCT_Mask_Array[uCT_Mask_Array <= R_Threshold] = 0
uCT_Mask_Array[uCT_Mask_Array > 0] = 1
## Shift uCT mask according to initial translation
ParameterMapFile = open(Results_Path + 'InitialTranslation.txt', 'r')
Text = ParameterMapFile.read()
Start = Text.find('TransformParameters')
def get_text_lines_from_image(src):
'''
将图片按文本单行切割
:param src:
:return:图片数组
'''
#调整大小
src = adjust_size(src)
temp = src.copy()
#调整水平
src = cv2.Canny(src, 100, 200)
src, slope = adjust_slope(src)
#src = cv2.erode(src,cv2.getStructuringElement(cv2.MORPH_CROSS,(1, 3)) )
#src = cv2.dilate(src,cv2.getStructuringElement(cv2.MORPH_CROSS,(1, 3)) )
src = cv2.dilate(src, cv2.getStructuringElement(cv2.MORPH_RECT, (40, 3)))
src = cv2.erode(src, cv2.getStructuringElement(cv2.MORPH_RECT, (40, 3)))
src = cv2.erode(src, cv2.getStructuringElement(cv2.MORPH_RECT, (5, 5)))
src = cv2.dilate(src, cv2.getStructuringElement(cv2.MORPH_RECT, (6, 5)))
src = 1 * (src > 128)
labels_open, nbr_objects_open = measurements.label(src)
#调整水平
block_size = 40
temp = threshold_adaptive(temp, block_size, offset=10)
temp = numpy.array(temp, 'uint8') * 255
temp = cv2.bitwise_not(temp)
if slope != 0:
temp = ndimage.rotate(temp, slope)
#旋转后像素会平滑，重新二值化
temp = cv2.bitwise_not(temp)
temp = threshold_adaptive(temp, block_size, offset=20)
temp = numpy.array(temp, 'uint8') * 255
temp = cv2.bitwise_not(temp)
lines = []
image = numpy.zeros(numpy.array(src).shape)
count = 0
for i in range(1, nbr_objects_open + 1):
test = temp.copy()
test[labels_open != i] = 0
box = bounding_box(test)
x, y, w, h = box
if h < 10 or w < 3:
continue
#忽略靠近上下边的区域
'''if y<2:
continue
if y+h> len(temp)-2:
continue'''
data = test[y:y + h, x:x + w]
lines.append(data)
copy = src.copy() * 255.
copy[labels_open != i] = 0
box = bounding_box(copy)
x, y, w, h = box
toerode = w / 3
if toerode <= 1:
continue
copy = cv2.erode(
copy, cv2.getStructuringElement(cv2.MORPH_RECT, (toerode, 1)))
copy = cv2.dilate(
copy, cv2.getStructuringElement(cv2.MORPH_RECT, (toerode, 1)))
copy = 1 * (copy > 128)
sub_labels_open, sub_nbr_objects_open = measurements.label(copy)
if (sub_nbr_objects_open > 1):
for i in range(1, sub_nbr_objects_open + 1):
test = temp.copy()
test[sub_labels_open != i] = 0
box = bounding_box(test)
#count+=1
#image[sub_labels_open == i] = count
x, y, w, h = box
if h < 10 or w < 3:
continue
#忽略靠近上下边的区域
if y < 2:
continue
if y + h > len(temp) - 2:
continue
data = test[y:y + h, x:x + w]
lines.append(data)
'''figure()
subplot(221)
imshow(temp)
subplot(222)
imshow(image)
subplot(223)
imshow(labels_open)
show()'''
return lines
def main():
# Ground Truth Paths
json_path = r'/nfs/masi/nathv/miccai_2020/micro_methods_hcp_mini/data_list_compart_dti.json'
json_path = os.path.normpath(json_path)
all_data = json.load(open(json_path))
#tr_data = all_data["train"]
#val_data = all_data["validation"]
test_data = all_data["test"]
gt_paths_dict = test_data['output']
method_list = ['BS_2003', 'IVIM', 'MC_SMT', 'NODDI_WATSON', 'DTI']
# Predicted Data Paths
base_pred_path = r'/nfs/masi/nathv/miccai_2020/bottleneck_compart_dti_test/bn_15/predicted_volumes'
base_pred_path = os.path.normpath(base_pred_path)
# Load the Mask Data
mask_data = load_nifty(test_data['mask'], data_type='float32')
mask_bool = np.array(mask_data, dtype='bool')
# TODO Slice number is stored here
slice_num = 72
# DTI Analysis
pred_dti_path = os.path.join(base_pred_path, 'dti.nii.gz')
pred_dti = load_nifty(pred_dti_path, 'float32')
plt.figure(1, figsize=(8, 12))
metric_nums = len(gt_paths_dict['DTI'])
plot_counter = 1
for idx, each_vol_path in enumerate(gt_paths_dict['DTI']):
gt_vol = load_nifty(each_vol_path, 'float32')
plt.subplot(1, metric_nums, plot_counter)
plt.imshow(rotate(np.squeeze(gt_vol[:, :, slice_num]), 90))
plt.axis('equal')
plt.colorbar()
plt.clim(0, 1)
plt.title('GT')
plt.subplot(3, metric_nums, plot_counter + metric_nums)
plt.imshow(rotate(np.squeeze(pred_dti[:, :, slice_num, idx]), 90))
plt.axis('equal')
plt.colorbar()
plt.clim(0, 1)
plt.title('Predicted')
diff_img = np.abs(
np.squeeze(gt_vol[:, :, slice_num]) -
np.squeeze(pred_dti[:, :, slice_num, idx]))
plt.subplot(3, metric_nums, plot_counter + metric_nums * 2)
plt.imshow(rotate(diff_img, 90))
plt.axis('equal')
plt.colorbar()
plt.clim(0, 0.1)
plt.title('Abs Difference')
plot_counter = plot_counter + 1
plt.clf()
plt.close(1)
# IVIM Analysis
pred_dti_path = os.path.join(base_pred_path, 'ivim.nii.gz')
pred_dti = load_nifty(pred_dti_path, 'float32')
plt.figure(1, figsize=(12, 12))
metric_nums = len(gt_paths_dict['IVIM'])
plot_counter = 1
for idx, each_vol_path in enumerate(gt_paths_dict['IVIM']):
gt_vol = load_nifty(each_vol_path, 'float32')
plt.subplot(3, metric_nums, plot_counter)
plt.imshow(rotate(np.squeeze(gt_vol[:, :, slice_num]), 90))
plt.axis('equal')
plt.colorbar()
plt.clim(0, 1)
plt.title('GT')
plt.subplot(3, metric_nums, plot_counter + metric_nums)
plt.imshow(rotate(np.squeeze(pred_dti[:, :, slice_num, idx]), 90))
plt.axis('equal')
plt.colorbar()
plt.clim(0, 1)
plt.title('Predicted')
diff_img = np.abs(
np.squeeze(gt_vol[:, :, slice_num]) -
np.squeeze(pred_dti[:, :, slice_num, idx]))
plt.subplot(3, metric_nums, plot_counter + metric_nums * 2)
plt.imshow(rotate(diff_img, 90))
plt.axis('equal')
plt.colorbar()
plt.clim(0, 0.1)
plt.title('Abs Difference')
plot_counter = plot_counter + 1
plt.clf()
plt.close(1)
# MC SMT Analysis
pred_dti_path = os.path.join(base_pred_path, 'mc_smt.nii.gz')
pred_dti = load_nifty(pred_dti_path, 'float32')
plt.figure(1, figsize=(8, 12))
metric_nums = len(gt_paths_dict['MC_SMT'])
plot_counter = 1
for idx, each_vol_path in enumerate(gt_paths_dict['MC_SMT']):
gt_vol = load_nifty(each_vol_path, 'float32')
plt.subplot(3, metric_nums, plot_counter)
plt.imshow(rotate(np.squeeze(gt_vol[:, :, slice_num]), 90))
plt.axis('equal')
plt.colorbar()
plt.clim(0, 1)
plt.title('GT')
plt.subplot(3, metric_nums, plot_counter + metric_nums)
plt.imshow(rotate(np.squeeze(pred_dti[:, :, slice_num, idx]), 90))
plt.axis('equal')
plt.colorbar()
plt.clim(0, 1)
plt.title('Predicted')
diff_img = np.abs(
np.squeeze(gt_vol[:, :, slice_num]) -
np.squeeze(pred_dti[:, :, slice_num, idx]))
plt.subplot(3, metric_nums, plot_counter + metric_nums * 2)
plt.imshow(rotate(diff_img, 90))
plt.axis('equal')
plt.colorbar()
plt.clim(0, 0.2)
plt.title('Abs Difference')
plot_counter = plot_counter + 1
plt.clf()
plt.close(1)
# Ball Stick Analysis
pred_dti_path = os.path.join(base_pred_path, 'ball_stick.nii.gz')
pred_dti = load_nifty(pred_dti_path, 'float32')
plt.figure(1, figsize=(20, 12))
metric_nums = len(gt_paths_dict['BS_2003'])
plot_counter = 1
for idx, each_vol_path in enumerate(gt_paths_dict['BS_2003']):
gt_vol = load_nifty(each_vol_path, 'float32')
plt.subplot(3, metric_nums, plot_counter)
plt.imshow(rotate(np.squeeze(gt_vol[:, :, slice_num]), 90))
plt.axis('equal')
plt.colorbar()
plt.clim(0, 1)
plt.title('GT')
plt.subplot(3, metric_nums, plot_counter + metric_nums)
plt.imshow(rotate(np.squeeze(pred_dti[:, :, slice_num, idx]), 90))
plt.axis('equal')
plt.colorbar()
plt.clim(0, 1)
plt.title('Predicted')
diff_img = np.abs(
np.squeeze(gt_vol[:, :, slice_num]) -
np.squeeze(pred_dti[:, :, slice_num, idx]))
plt.subplot(3, metric_nums, plot_counter + metric_nums * 2)
plt.imshow(rotate(diff_img, 90))
plt.axis('equal')
plt.colorbar()
plt.clim(0, 0.2)
plt.title('Abs Difference')
plot_counter = plot_counter + 1
plt.clf()
plt.close(1)
# NODDI Analysis
pred_dti_path = os.path.join(base_pred_path, 'noddi.nii.gz')
pred_dti = load_nifty(pred_dti_path, 'float32')
plt.figure(1, figsize=(18, 12))
metric_nums = len(gt_paths_dict['NODDI_WATSON'])
plot_counter = 1
for idx, each_vol_path in enumerate(gt_paths_dict['NODDI_WATSON']):
gt_vol = load_nifty(each_vol_path, 'float32')
plt.subplot(3, metric_nums, plot_counter)
plt.imshow(rotate(np.squeeze(gt_vol[:, :, slice_num]), 90))
plt.axis('equal')
plt.colorbar()
plt.clim(0, 1)
plt.title('GT')
plt.subplot(3, metric_nums, plot_counter + metric_nums)
plt.imshow(rotate(np.squeeze(pred_dti[:, :, slice_num, idx]), 90))
plt.axis('equal')
plt.colorbar()
plt.clim(0, 1)
plt.title('Predicted')
diff_img = np.abs(
np.squeeze(gt_vol[:, :, slice_num]) -
np.squeeze(pred_dti[:, :, slice_num, idx]))
plt.subplot(3, metric_nums, plot_counter + metric_nums * 2)
plt.imshow(rotate(diff_img, 90))
plt.axis('equal')
plt.colorbar()
plt.clim(0, 0.2)
plt.title('Abs Difference')
plot_counter = plot_counter + 1
plt.show()
print('Debug here')
return None
from PIL import Image
import numpy as np
from matplotlib import pyplot as plt
from matplotlib import cm
from scipy import ndimage
im = Image.open('../data/grenouille.jpg')
rouge, vert, bleu = im.split()
z = np.array(rouge)
zrr = ndimage.rotate(z, 30.)
zrn = ndimage.rotate(z, 30., reshape=False)
ny, nx = z.shape
nyrr, nxrr = zrr.shape
nyrn, nxrn = zrn.shape
fig = plt.figure(0) # On cree une figure
plt.clf()
ax1 = fig.add_subplot(3, 1, 1)
plt.imshow(z, origin="upper")
plt.xticks([0, nx])
plt.yticks([0, ny])
ax2 = fig.add_subplot(3, 1, 2)
plt.imshow(zrr, origin="upper")
plt.xticks([0, nxrr])
plt.yticks([0, nyrr])
ax2 = fig.add_subplot(3, 1, 3)
plt.imshow(zrn, origin="upper")
plt.xticks([0, nxrn])
plt.yticks([0, nyrn])
plt.show()
# diff = np.abs(pp-ll)
# sort_idxs = np.argsort(-diff)
# sd_test = sd_test[sort_idxs]
# stest = stest[sort_idxs]
ppp = st[:, 0]
lll = st[:, 1]
mistakes = ppp != lll
sd_test = sd_test[mistakes]
stest = stest[mistakes]
# for i, dat, pred, label in zip(range(100), sd_test[100:200], stest[100:200, 0], stest[100:200, 1]):
for i, dat, pred, label in zip(range(100), sd_test[:100], stest[:100, 0],
stest[:100, 1]):
f_guess = f'image_{i}.jpg'
f_label = f'image_{i}_pred_{pred}_label_{label}.jpg'
dat = ndimage.rotate(dat, 90)
scipy.misc.imsave(os.path.join(path_guess, f_guess), dat)
scipy.misc.imsave(os.path.join(path_labeled, f_label), dat)
a_pred = int(pred > 1.11)
a_label = int(label > 1.11)
squared_error = (pred - label)**2
writer.write_row(i=i,
prediction=pred,
label=label,
prediction_amyloid=a_pred,
label_amyloid=a_label,
correct=int(a_pred == a_label),
squared_error=squared_error)
print(i, pred, label)
plt.show()
if __name__ == "__main__":
#Img = Image.readFrom("person_toy/00000001.jpg")
#Img45 = Image.readFrom("person_toy/00000001.jpg")
#Img90 = Image.readFrom("person_toy/00000001.jpg")
#Img45.im = rotate(Img45.im, 45)
#Img90.im = rotate(Img90.im, 90)
Img = Image.readFrom("pingpong/0000.jpeg")
Img45 = Image.readFrom("pingpong/0000.jpeg")
Img90 = Image.readFrom("pingpong/0000.jpeg")
Img45.im = rotate(Img45.im, 45)
Img90.im = rotate(Img90.im, 90)
CornerAlgo1 = HarrisCornerDetector(3, 1e6)
CornerAlgo2 = HarrisCornerDetector(3, 1e5)
CornerAlgo3 = HarrisCornerDetector(3, 1e4)
CornerAlgo4 = HarrisCornerDetector(3, 1e7)
H1 = CornerAlgo1.get_Hvalues(Img, sigma=1)
H2 = CornerAlgo2.get_Hvalues(Img, sigma=1)
H3 = CornerAlgo3.get_Hvalues(Img, sigma=1)
H4 = CornerAlgo4.get_Hvalues(Img, sigma=1)
Hr45 = CornerAlgo1.get_Hvalues(Img45, sigma=1)
Hr90 = CornerAlgo1.get_Hvalues(Img90, sigma=1)
mask1 = CornerAlgo1.evaluate(H1)
mask2 = CornerAlgo2.evaluate(H2)
def image(imagename, filtertype):
flag = 0
log = db.log
ts = datetime.datetime.now()
filename='assets/'+imagename+'.jpg'
img = misc.imread(filename)
img[:] = np.max(img,axis=-1,keepdims=1)/2+np.min(img,axis=-1,keepdims=1)/2
if filtertype == "greyscale":
plt.imsave('test.jpg',img)
flag=1
elif filtertype == "lowpass" and request.args.get('value'):
blurred = gaussian_filter(img, sigma = float(request.args.get('value')))
plt.imsave('test.jpg',blurred)
flag=1
if filtertype == "crop":
def crop_center(img,cropx,cropy):
y,x,c = img.shape
startx = x//2 - cropx//2
starty = y//2 - cropy//2
return img[starty:starty+cropy, startx:startx+cropx]
y,x,z = img.shape
if y>x:
cropped = crop_center(img,x,x)
else:
cropped = crop_center(img,y,y)
plt.imsave('test.jpg',cropped)
flag=1
elif filtertype == "dx":
sobelx = cv2.Sobel(img,cv2.cv2.CV_8U,1,0,ksize=5)
plt.imsave('test.jpg',sobelx)
flag=1
if filtertype == "dy":
sobely = cv2.Sobel(img,cv2.cv2.CV_8U,0,1,ksize=5)
plt.imsave('test.jpg',sobely)
flag=1
elif filtertype == "downsample" and request.args.get('value'):
downsampled = rescale(img, 1.0 / float(request.args.get('value')), anti_aliasing=False)
plt.imsave('test.jpg',downsampled)
flag=1
if filtertype == "rotate" and request.args.get('value'):
rotated = ndimage.rotate(img, int(request.args.get('value')))
plt.imsave('test.jpg',rotated)
flag=1
te = datetime.datetime.now()
if request.args.get('value') and flag == 1:
log.insert({'imagename':imagename, 'cmd':filtertype, 'value':float(request.args.get('value')),'request_timestamp':str(ts),'processing_time':str(te-ts)})
elif flag == 1:
log.insert({'imagename':imagename, 'cmd':filtertype, 'value':'null','request_timestamp':str(ts),'processing_time':str(te-ts)})
if flag == 1:
return send_file('test.jpg',mimetype='image/gif')
else:
return 'Error Processing File'
def main():
global X_train, y_train, X_valid, y_valid, X_test, y_test, EPOCHS, BATCH_SIZE, learning_rate, feed_forward,\
fc_keep_prob_value, conv_keep_prob_value, saver, CONTINUE, conv1_depth_val, conv2_depth_val, conv_filter_size,\
available_training_data_set, image_depth, acc_list, train_acc_list, loss_list, TRAIN, TEST, max_accuracy, max_accuracy_epoch
global accuracy_operation, x, y, conv_keep_prob, fc_keep_prob
# TODO: Fill this in based on where you saved the training and testing data
training_file = "./traffic_signs_data/train.p"
validation_file = "./traffic_signs_data/valid.p"
testing_file = "./traffic_signs_data/test.p"
with open(training_file, mode='rb') as f:
train = pickle.load(f)
with open(validation_file, mode='rb') as f:
valid = pickle.load(f)
with open(testing_file, mode='rb') as f:
test = pickle.load(f)
X_train, y_train = train['features'], train['labels']
X_valid, y_valid = valid['features'], valid['labels']
X_test, y_test = test['features'], test['labels']
# TODO: Number of training examples
n_train = len(y_train)
# TODO: Number of validation examples
n_validation = len(y_valid)
# TODO: Number of testing examples.
n_test = len(y_test)
# TODO: What's the shape of an traffic sign image?
image_shape = X_train[0].shape
# TODO: How many unique classes/labels there are in the dataset.
n_classes = max(np.concatenate((y_train ,y_valid , y_test), axis=0)) + 1
print("Number of training examples =", n_train)
print("Number of testing examples =", n_test)
print("Image data shape =", image_shape)
print("Number of classes =", n_classes)
#DATA analysis
if DISPLAY_DATA_SET_HISTOGRAM:
n, bins, patches = plt.hist(y_train, n_classes, facecolor='blue', alpha=0.5)
plt.show()
random_index = random.randint(0, len(X_train))
rgb_image = X_train[random_index].squeeze()
#convert to grayscale
if image_depth == 1:
X_train = np.sum(X_train/3, axis=3, keepdims=True)
X_valid = np.sum(X_valid/3, axis=3, keepdims=True)
X_test = np.sum(X_test/3, axis=3, keepdims=True)
#Histograms Equalization in OpenCV
if Data_Augmentation_EqualizeHist:
if True:
X_EqualizeHist = np.zeros_like(X_train)
for i in range(X_train.shape[0]):
X_EqualizeHist[i, :, :, 0] = cv2.equalizeHist(X_train[i].astype(np.uint8))
#add_images(X_EqualizeHist)
X_train = X_EqualizeHist
for i in range(X_valid.shape[0]):
X_valid[i, :, :, 0] = cv2.equalizeHist(X_valid[i].astype(np.uint8))
for i in range(X_test.shape[0]):
X_test[i, :, :, 0] = cv2.equalizeHist(X_test[i].astype(np.uint8))
else:
clahe = cv2.createCLAHE(tileGridSize=(2, 2), clipLimit=clipLimit)
X_EqualizeHist = np.zeros_like(X_train)
for i in range(X_train.shape[0]):
X_EqualizeHist[i, :, :, 0] = clahe.apply(X_train[i].astype(np.uint16))
#add_images(X_EqualizeHist)
X_train = X_EqualizeHist
for i in range(X_valid.shape[0]):
X_valid[i, :, :, 0] = clahe.apply(X_valid[i].astype(np.uint16))
for i in range(X_test.shape[0]):
X_test[i, :, :, 0] = clahe.apply(X_test[i].astype(np.uint16))
x = tf.placeholder(tf.float32, (None, 32, 32, image_depth))
y = tf.placeholder(tf.int32, (None, 43))
if Data_Augmentation_Rotate:
print("Data Augmentation %s" %"Rotate")
random_degree = random.uniform(-10, 10)
x_rotate = rotate(X_train, random_degree)
X_train = np.concatenate((X_train, x_rotate), axis=0)
y_train = np.concatenate((y_train, y_train), axis=0)
available_training_data_set += 1
image_gray = X_train[random_index].squeeze()
if Data_Augmentation_Noise:
print("Data Augmentation %s" %"Noise")
x_noise = RandomNoise(X_train/255, mode='pepper')*255
#x_noise = x_noise / 2
image_noise = x_noise[random_index].squeeze()
#image_noise = RandomNoise(image_gray)
plt.figure(figsize=(1,3))
_ ,ax = plt.subplots(1, 3)
ax[0].imshow(rgb_image)
ax[0].set_title("rgb_image", fontsize=20)
ax[1].imshow(image_gray, "gray")
ax[1].set_title("gray", fontsize=20)
ax[2].imshow(image_noise, "gray")
ax[2].set_title("image_noise", fontsize=20)
#print(image_noise)
#print(np.max(image_gray), np.max(image_noise))
plt.show()
X_train = np.concatenate((X_train, x_noise), axis=0)
y_train = np.concatenate((y_train, y_train), axis=0)
available_training_data_set += 1
if Data_Augmentation_Blur:
print("Data Augmentation %s" %"Blur")
x_blur = Blur(X_train, size=3)
image_blur = x_blur[random_index].squeeze()
plt.figure(figsize=(1,3))
_, ax = plt.subplots(1, 3)
ax[0].imshow(rgb_image)
ax[0].set_title("rgb_image", fontsize=20)
ax[1].imshow(image_gray, "gray")
ax[1].set_title("gray", fontsize=20)
ax[2].imshow(image_blur, "gray")
ax[2].set_title("image_blur", fontsize=20)
plt.show()
X_train = np.concatenate((X_train, x_blur), axis=0)
y_train = np.concatenate((y_train, y_train), axis=0)
available_training_data_set += 1
X_train, y_train = shuffle(X_train, y_train)
n_train *=available_training_data_set
# Normalize data
X_train = (X_train - 128)/128
X_valid = (X_valid - 128)/128
X_test = (X_test - 128)/128
encoder = LabelBinarizer()
encoder.fit(y_train)
y_train = encoder.transform(y_train)
y_valid = encoder.transform(y_valid)
y_test = encoder.transform(y_test)
if False:
plt.figure(figsize=(1,1))
plt.imshow(image)
print(y_train_num[random_index])
print(y_train[random_index])
print('Labels One-Hot Encoded')
plt.show()
fc_keep_prob = tf.placeholder(tf.float32)
conv_keep_prob = tf.placeholder(tf.float32)
logits = feed_forward(x, conv_keep_prob, fc_keep_prob, depth=image_depth, conv1_depth=conv1_depth_val, conv2_depth=conv2_depth_val, filter_size=conv_filter_size)
cross_entropy = tf.nn.softmax_cross_entropy_with_logits(labels=y, logits=logits)
loss_operation = tf.reduce_mean(cross_entropy)
optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate )
training_operation = optimizer.minimize(loss_operation)
correct_prediction = tf.equal(tf.argmax(logits, 1), tf.argmax(y, 1))
accuracy_operation = tf.reduce_mean(tf.cast(correct_prediction, tf.float32))
saver = tf.train.Saver()
with tf.Session() as sess:
if Evaluate_NW:
com_sess = sess
com_x = x
#outputFeatureMap(image_gray, conv1)
pass
if CONTINUE == True:
saver.restore(sess, './00_out/2019_01_13-13_52_52/LeNet')
else:
sess.run(tf.global_variables_initializer())
num_examples = int(n_train/BATCH_SIZE)*BATCH_SIZE
if TRAIN:
print("Training...")
for i in range(EPOCHS):
#print(type(y_train))
X_train, y_train = shuffle(X_train, y_train)
for offset in range(0, num_examples, BATCH_SIZE):
end = offset + BATCH_SIZE
if end >= n_train:
end = n_train-1
batch_x, batch_y = X_train[offset:end], y_train[offset:end]
_, l = sess.run([training_operation, loss_operation], feed_dict={x: batch_x, y: batch_y, conv_keep_prob : conv_keep_prob_value, fc_keep_prob : fc_keep_prob_value})
loss_list.append(l)
validation_accuracy = evaluate(X_valid, y_valid)
training_accuracy = evaluate(X_train, y_train)
print("EPOCH {} ...".format(i + 1))
print("Validation Accuracy = {:.3f}".format(validation_accuracy))
print("Training Accuracy = {:.3f}".format(training_accuracy))
acc_list.append(validation_accuracy)
train_acc_list.append(training_accuracy)
print()
sub_folder_name = get_date_time()
fig0 = plt.figure(0)
fig0.clf()
plt.plot(acc_list, label='validation_accuracy')
plt.plot(train_acc_list, label='training_accuracy')
plt.grid(True)
plt.legend()
annot_max(range(len(acc_list)),acc_list)
max_accuracy = max(acc_list)
max_accuracy_epoch = range(len(acc_list))[np.argmax(acc_list)]
print("Max accuracy %.3f at Epoch %d" %(max_accuracy, max_accuracy_epoch))
arr = convert_figure_to_array(fig0)
store_image(arr, "validation_accuracy", out_dir + "/" + sub_folder_name)
fig0 = plt.figure(0)
fig0.clf()
plt.plot(loss_list, label='loss')
plt.grid(True)
plt.legend(loc=2)
annot_min(range(len(loss_list)),loss_list)
arr = convert_figure_to_array(fig0)
store_image(arr, "loss", out_dir + "/" + sub_folder_name)
global test_accuracy
test_accuracy = 0
data_list = ["EPOCHS", "BATCH_SIZE", "learning_rate", "feed_forward.__name__",
"fc_keep_prob_value", "conv_keep_prob_value", "CONTINUE", "image_depth",
"conv1_depth_val", "conv2_depth_val", "conv_filter_size","test_accuracy",
"max_accuracy", "max_accuracy_epoch", "clipLimit",
"acc_list", "train_acc_list", "loss_list"]
file_name = out_dir + "/" + sub_folder_name + "/" + "configs"
save_variables(file_name, data_list)
saver.save(sess, "./" + out_dir + "/" + sub_folder_name + '/' + feed_forward.__name__)
print("Model saved")
if TEST:
print("Testing ...")
test_accuracy = evaluate(X_test, y_test)
print("Testing Accuracy = {:.3f}".format(test_accuracy))
imgArr.append(x_test[i])
pred.append(y_pred[i])
gnd.append(y_test[i])
counter += 1
w = 10
h = 10
fig = plt.figure(figsize=(12, 12))
columns = 4
rows = 1
print("\n\nFailure cases: ")
print("=================")
for i in range(1, columns * rows + 1):
img = imgArr[i]
img = img.reshape((64, 64))
fig.add_subplot(rows, columns, i)
rotated_img = ndimage.rotate(img, 270)
plt.imshow(rotated_img, interpolation='nearest')
plt.title("\nPred: {} Actual o/p: {}".format(pred[i], gnd[i]),
fontsize=11)
plt.axis('off')
plt.show()
# following code displays images of correct prediction.
counter = 0
imgArr = []
pred = []
gnd = []
for i in range(len(y_test)):
if (y_test[i] == y_pred[i] and counter < 10):
imgArr.append(x_test[i])
pred.append(y_pred[i])
cv2.putText(img_new, label, (x, y - 10), font, 2, [255, 0, 0],
d)
if (label == 'red insulator'): #red = màu đỏ
cv2.rectangle(img_new, (x, y), (x + w, y + h), [0, 0, 255], d)
cv2.putText(img_new, label, (x, y - 10), font, 2, [0, 0, 255],
d)
h1 = int(h / 10)
w1 = int(w / 10)
img_crop = img[y - h1:y + h + h1, x - w1:x + w + w1]
#img_crop = cv2.resize(img_crop,(img_crop.shape[1]*5,img_crop.shape[0]*5))
#Xoay ngang ảnh nếu ảnh là ảnh dọc
if (h > w):
img_crop = ndimage.rotate(img_crop, 90)
#Display dùng cv2
cv2.imshow("Image", img_new)
key = cv2.waitKey(1)
cv2.imshow("Image_crop", img_crop)
key = cv2.waitKey(0)
cv2.destroyAllWindows()
'''
#Display dùng plt, hiển thị toàn bộ 1 lần trên 1 khung
# Lưu ý nhớ chuyển hệ màu trước khi dùng
titles = ['Anh goc','Detect',"Crop"]
images = [img, img_new, img_crop]
def test_export_array():
try:
from scipy.ndimage import rotate
except:
rotate = False
pass
namfile = 'freyberg.nam'
model_ws = '../examples/data/freyberg_multilayer_transient/'
m = flopy.modflow.Modflow.load(namfile,
model_ws=model_ws,
verbose=False,
load_only=['DIS', 'BAS6'])
m.sr.rotation = 45.
nodata = -9999
m.sr.export_array(os.path.join(tpth, 'fb.asc'),
m.dis.top.array,
nodata=nodata)
arr = np.loadtxt(os.path.join(tpth, 'fb.asc'), skiprows=6)
m.sr.write_shapefile(os.path.join(tpth, 'grid.shp'))
# check bounds
with open(os.path.join(tpth, 'fb.asc')) as src:
for line in src:
if 'xllcorner' in line.lower():
val = float(line.strip().split()[-1])
# ascii grid origin will differ if it was unrotated
if rotate:
assert np.abs(val - m.sr.bounds[0]) < 1e-6
else:
assert np.abs(val - m.sr.xll) < 1e-6
if 'yllcorner' in line.lower():
val = float(line.strip().split()[-1])
if rotate:
assert np.abs(val - m.sr.bounds[1]) < 1e-6
else:
assert np.abs(val - m.sr.yll) < 1e-6
if 'cellsize' in line.lower():
val = float(line.strip().split()[-1])
rot_cellsize = np.cos(np.radians(
m.sr.rotation)) * m.sr.delr[0] * m.sr.length_multiplier
#assert np.abs(val - rot_cellsize) < 1e-6
break
rotate = False
rasterio = False
if rotate:
rotated = rotate(m.dis.top.array, m.sr.rotation, cval=nodata)
if rotate:
assert rotated.shape == arr.shape
try:
# test GeoTIFF export
import rasterio
m.sr.export_array(os.path.join(tpth, 'fb.tif'),
m.dis.top.array,
nodata=nodata)
with rasterio.open(os.path.join(tpth, 'fb.tif')) as src:
arr = src.read(1)
except:
pass
if rasterio:
assert src.shape == (m.nrow, m.ncol)
assert np.abs(src.bounds[0] - m.sr.bounds[0]) < 1e-6
assert np.abs(src.bounds[1] - m.sr.bounds[1]) < 1e-6
def plotsklearnresult(configini, videoSetting, frameSetting):
config = ConfigParser()
configFile = str(configini)
try:
config.read(configFile)
except MissingSectionHeaderError:
print(
'ERROR: Not a valid project_config file. Please check the project_config.ini path.'
)
csv_dir = config.get('General settings', 'csv_path')
csv_dir_in = os.path.join(csv_dir, "machine_results")
animalsNo = config.getint('General settings', 'animal_no')
projectPath = config.get('General settings', 'project_path')
frames_dir_out = config.get('Frame settings', 'frames_dir_out')
frames_dir_out = os.path.join(frames_dir_out, 'sklearn_results')
if not os.path.exists(frames_dir_out):
os.makedirs(frames_dir_out)
counters_no = config.getint('SML settings', 'No_targets')
vidInfPath = config.get('General settings', 'project_path')
vidInfPath = os.path.join(vidInfPath, 'logs', 'video_info.csv')
vidinfDf = pd.read_csv(vidInfPath)
target_names = []
loopy = 0
Xcols, Ycols, Pcols = getBpNames(configini)
colorList = []
for color in range(len(Xcols)):
r, g, b = (random.randint(0, 255), random.randint(0, 255),
random.randint(0, 255))
colorTuple = (r, g, b)
colorList.append(colorTuple)
filesFound = glob.glob(csv_dir_in + '/*.csv')
print('Processing ' + str(len(filesFound)) + ' videos ...')
########### GET MODEL NAMES ###########
for i in range(counters_no):
currentModelNames = 'target_name_' + str(i + 1)
currentModelNames = config.get('SML settings', currentModelNames)
target_names.append(currentModelNames)
########### FIND PREDICTION COLUMNS ###########
for i in filesFound:
target_counters, target_timers = ([0] * counters_no, [0] * counters_no)
currentVideo = i
loopy += 1
CurrentVideoName = os.path.basename(currentVideo)
if frameSetting == 1:
videoFrameDir = os.path.join(frames_dir_out,
CurrentVideoName.replace('.csv', ''))
if not os.path.exists(videoFrameDir):
os.makedirs(videoFrameDir)
CurrentVideoRow = vidinfDf.loc[vidinfDf['Video'] == str(
CurrentVideoName.replace('.csv', ''))]
try:
fps = int(CurrentVideoRow['fps'])
except TypeError:
print(
'Error: make sure all the videos that are going to be analyzed are represented in the project_folder/logs/video_info.csv file'
)
currentDf = pd.read_csv(currentVideo)
currentDf = currentDf.fillna(0)
currentDf = currentDf.astype(int)
currentDf = currentDf.loc[:,
~currentDf.columns.str.contains('^Unnamed')]
currentDf = currentDf.reset_index()
animalBpHeaderList, animalBpHeaderListY, animalBpHeaderListX = ([], [],
[])
animal1_BPsX, animal1_BPsY = (currentDf[Xcols], currentDf[Ycols])
for i in range(len(animal1_BPsX.columns)):
animalBpHeaderListX.append(animal1_BPsX.columns[i])
animalBpHeaderListY.append(animal1_BPsY.columns[i])
animalBpHeaderList.append(animal1_BPsX.columns[i])
animalBpHeaderList.append(animal1_BPsY.columns[i])
animalBpHeaderListX, animalBpHeaderListY, animalBpHeaderList = ([
x for x in animalBpHeaderListX if "Tail_end" not in x
], [x for x in animalBpHeaderListY if "Tail_end" not in x
], [x for x in animalBpHeaderList if "Tail_end" not in x])
if animalsNo == 2:
animal_1_BpHeaderList = [
s for s in animalBpHeaderList if "_1_" in s
]
animal_2_BpHeaderList = [
s for s in animalBpHeaderList if "_2_" in s
]
if os.path.exists(
os.path.join(projectPath, 'videos',
CurrentVideoName.replace('.csv', '.mp4'))):
videoPathName = os.path.join(
projectPath, 'videos',
CurrentVideoName.replace('.csv', '.mp4'))
elif os.path.exists(
os.path.join(projectPath, 'videos',
CurrentVideoName.replace('.csv', '.avi'))):
videoPathName = os.path.join(
projectPath, 'videos',
CurrentVideoName.replace('.csv', '.avi'))
else:
print('Cannot locate video ' +
str(CurrentVideoName.replace('.csv', '')) +
'in mp4 or avi format')
break
cap = cv2.VideoCapture(videoPathName)
width = int(cap.get(cv2.CAP_PROP_FRAME_WIDTH))
height = int(cap.get(cv2.CAP_PROP_FRAME_HEIGHT))
frames = int(cap.get(cv2.CAP_PROP_FRAME_COUNT))
fourcc = cv2.VideoWriter_fourcc(*'mp4v')
outputFileName = os.path.join(frames_dir_out, CurrentVideoName)
if height < width:
videoHeight, videoWidth = width, height
if height >= width:
videoHeight, videoWidth = height, width
writer = cv2.VideoWriter(outputFileName.replace('.csv', '.mp4'),
fourcc, fps, (videoWidth, videoHeight))
mySpaceScale, myRadius, myResolution, myFontScale = 60, 20, 1500, 1.5
maxResDimension = max(width, height)
circleScale = int(myRadius / (myResolution / maxResDimension))
fontScale = float(myFontScale / (myResolution / maxResDimension))
spacingScale = int(mySpaceScale / (myResolution / maxResDimension))
currRow = 0
while (cap.isOpened()):
ret, frame = cap.read()
if ret == True:
if animalsNo == 1:
currAnimal1 = currentDf.loc[currentDf.index[currRow],
animalBpHeaderList]
currAnimal1 = np.array(currAnimal1).astype(int)
currAnimal1 = np.reshape(currAnimal1, (-1, 2))
M1polyglon_array_hull = cv2.convexHull(
(currAnimal1.astype(int)))
cv2.drawContours(frame,
[M1polyglon_array_hull.astype(int)], 0,
(255, 255, 255), 2)
if animalsNo == 2:
currAnimal1, currAnimal2 = (
currentDf.loc[currentDf.index[currRow],
animal_1_BpHeaderList],
currentDf.loc[currentDf.index[currRow],
animal_2_BpHeaderList])
currAnimal1, currAnimal2 = (
np.array(currAnimal1).astype(int),
np.array(currAnimal2).astype(int))
currAnimal1, currAnimal2 = (np.reshape(
currAnimal1,
(-1, 2)), np.reshape(currAnimal2, (-1, 2)))
M1polyglon_array_hull, M2polyglon_array_hull = (
cv2.convexHull((currAnimal1.astype(int))),
cv2.convexHull((currAnimal2.astype(int))))
cv2.drawContours(frame,
[M1polyglon_array_hull.astype(int)], 0,
(255, 255, 255), 2)
cv2.drawContours(frame,
[M2polyglon_array_hull.astype(int)], 0,
(255, 255, 255), 2)
for cords in range(len(animalBpHeaderListX)):
currXval = animal1_BPsX.loc[animal1_BPsX.index[currRow],
animalBpHeaderListX[cords]]
currYval = animal1_BPsY.loc[animal1_BPsY.index[currRow],
animalBpHeaderListY[cords]]
cv2.circle(frame, (int(currXval), int(currYval)),
circleScale,
colorList[cords],
-1,
lineType=cv2.LINE_AA)
if height < width:
frame = ndimage.rotate(frame, 90)
# draw event timers
for b in range(counters_no):
target_timers[b] = (1 / fps) * target_counters[b]
target_timers[b] = round(target_timers[b], 2)
cv2.putText(frame, str('Timers'),
(10, ((height - height) + spacingScale)),
cv2.FONT_HERSHEY_COMPLEX, fontScale, (0, 255, 0),
2)
addSpacer = 2
for k in range(counters_no):
cv2.putText(frame, (str(target_names[k]) + ' ' +
str(target_timers[k]) + str('s')),
(10,
(height - height) + spacingScale * addSpacer),
cv2.FONT_HERSHEY_SIMPLEX, fontScale,
(0, 0, 255), 2)
addSpacer += 1
cv2.putText(frame, str('ensemble prediction'),
(10, (height - height) + spacingScale * addSpacer),
cv2.FONT_HERSHEY_SIMPLEX, fontScale, (0, 255, 0),
2)
addSpacer += 1
colors = [(2, 166, 249), (47, 255, 173), (0, 165, 255),
(60, 20, 220), (193, 182, 255), (238, 130, 238),
(144, 128, 112), (32, 165, 218), (0, 0, 128),
(209, 206, 0)]
for p in range(counters_no):
TargetVal = int(currentDf.loc[currRow, [target_names[p]]])
if TargetVal == 1:
cv2.putText(
frame, str(target_names[p]),
(10, (height - height) + spacingScale * addSpacer),
cv2.FONT_HERSHEY_TRIPLEX, int(fontScale * 2),
colors[p], 2)
target_counters[p] += 1
addSpacer += 1
if videoSetting == 1:
writer.write(frame)
if frameSetting == 1:
frameName = os.path.join(videoFrameDir,
str(currRow) + '.png')
cv2.imwrite(frameName, frame)
if (videoSetting == 0) and (frameSetting == 0):
print('Error: Please choose video and/or frames.')
break
currRow += 1
print('Frame ' + str(currRow) + '/' + str(frames) +
'. Video ' + str(loopy) + '/' + str(len(filesFound)))
if frame is None:
print('Video ' + str(
os.path.basename(CurrentVideoName.replace('.csv', '.mp4')))
+ ' saved.')
cap.release()
break
def get_process_params(folder, wallcut_file):
"""
Function to get the rotation angle and crop range
Parameters
----------
folder : string
Input folder where the images are located.
Returns
-------
crop_final : 2d array of int32
Array containing the 4 crop parameters at the end.
crop_rot : 2d array of int32
Array containing the 4 crop parameters after the rotation.
angle : float64
Angle of tilt for the images in radians.
name_sliced : string
Name of the images without the number.
img_amount : int
Amount of images in the folder.
idx0 : int
Index of the first image in the folder.
images_exist : boolean
True if there are images in the folder, otherwise false
"""
run_names = wallcut_file[:,0].astype(np.str)
wallcuts = wallcut_file[:,1:].astype(np.float)
# get a list of all the files in the folder
file_list = os.listdir(folder)
# get the amount of images by taking away the labview files
img_amount = len(file_list)-5
# check wether there actually are images in the folder
if(img_amount < 1):
# print error message and return
MEX = 'No images in folder %s' %folder
print(MEX)
return 0, 0, 0, 0, 0, 0, False
# get the name of the first frame
frame0 = file_list[5]
# get the '.' in the file name
indices = [i for i, a in enumerate(frame0) if a == '.']
# extract the index of the first frame
idx0 = int(frame0[indices[0]+1:indices[1]])
# cut of the index of the file to get the sliced name
name_sliced = frame0[:(indices[0])]
# search for the right wallcut
for j in range(0, len(run_names)):
if(name_sliced == run_names[j]):
wallcut_left, wallcut_right = wallcuts[j,:]
# check whether a wallcut was found
if (wallcut_left ==np.nan):
MEX = 'No wallcut was found'
print(MEX)
return 0, 0, 0, 0, 0, 0, False
# calculate the average image
avg = get_avg_image(folder, name_sliced, idx0)
# calculate the angle from the average image
angle = get_angle(avg)
# rotate the image
rotated = ndimage.rotate(avg, angle*180/np.pi)
# get the new image shape
hr, wr = rotated.shape
# crop the edges due to the rotation
crop_rot =(0+int(np.rint(wr*np.tan(angle))), hr-int(np.rint(wr*np.tan(angle))),\
0+int(np.rint(hr*np.tan(angle))), wr-int(np.rint(hr*np.tan(angle))))
cropped = rotated[crop_rot[0]:crop_rot[1],crop_rot[2]:crop_rot[3]]
# get the edges of the rotated image
new_peak_left, new_peak_right = find_edges(cropped)
x_low = get_most_common_element(new_peak_left)
x_high = get_most_common_element(new_peak_right)
# arange coordinates of the final crop
crop_final = (0,cropped.shape[0],x_low+int(wallcut_left),x_high+1-int(wallcut_right))
return crop_final, crop_rot, angle, name_sliced, img_amount, idx0, True
def __init__(self):
# mnist = tf.contrib.learn.datasets.load_dataset("mnist")
# train_labels = np.asarray(mnist.train.labels, dtype=np.int32)
# blocksOfInterestIdx = np.nonzero(train_labels==2)[0]
# self.blocks = mnist.train.images[blocksOfInterestIdx]
self.blockSize = 28 # edge size of one block
self.stride = 28 # num grid cells between adjacent move destinations
# self.stride = 24 # num grid cells between adjacent move destinations
# self.stride = 20 # num grid cells between adjacent move destinations
# self.stride = 14 # num grid cells between adjacent move destinations
# self.stride = 7 # num grid cells between adjacent move destinations
# self.stride = 6 # num grid cells between adjacent move destinations
# self.stride = 3 # num grid cells between adjacent move destinations
# self.numBlocksWide = 8 # number of blocks that can fit end-to-end on one edge of board
self.initStride = self.stride
self.gridSize = 28 * 8
self.num_blocks = 2
self.num_orientations = 4
self.observation_space = spaces.Tuple([
spaces.Box(np.zeros([self.gridSize, self.gridSize, 1]),
np.ones([self.gridSize, self.gridSize, 1])),
spaces.Discrete(2)
])
self.holdingImage = []
self.state = None
self.max_episode = 10
self.gap = 3 # num pixels that need to be clear around a block in order ot move it.
# self.gap = 2 # num pixels that need to be clear around a block in order ot move it.
self.numBlocksInRowGoal = 2
# self.blockType = 'Numerals'
# self.blockType = 'Disks'
self.blockType = 'Blocks'
if self.blockType == 'Numerals': # load MNIST numerals
mnist = tf.contrib.learn.datasets.load_dataset("mnist")
train_labels = np.asarray(mnist.train.labels, dtype=np.int32)
blocksOfInterestIdx = np.nonzero(train_labels == 2)[0]
self.blocks = mnist.train.images[blocksOfInterestIdx]
elif self.blockType == 'Disks': # load random radius disks instead
numBlocks = 10
minRadius = 8
maxRadius = 10
self.blocks = []
for i in range(numBlocks):
X, Y = np.mgrid[0:self.blockSize, 0:self.blockSize]
halfBlock = int(self.blockSize / 2)
dist2 = (X - halfBlock)**2 + (Y - halfBlock)**2
radius = np.random.randint(minRadius, maxRadius)
im = np.int32(dist2 < radius**2)
self.blocks.append(np.reshape(im, int(self.blockSize**2)))
elif self.blockType == 'Blocks': # load random blocks instead
numBlocks = 10
minRadius = 8
maxRadius = 10
self.blocks = []
for i in range(numBlocks):
Y, X = np.mgrid[0:self.blockSize, 0:self.blockSize]
halfBlock = int(self.blockSize / 2)
dist2 = (Y - halfBlock)**2 # vertical rectangle
randNum = np.random.rand()
if randNum < 0.25:
dist2 = dist2 + 0.01
# dist2 = ndimage.rotate(dist2+0.01,45,reshape=False)
elif randNum < 0.5:
# dist2 = dist2+0.01
dist2 = ndimage.rotate(dist2 + 0.01, 45, reshape=False)
elif randNum < 0.75:
# dist2 = dist2+0.01
dist2 = ndimage.rotate(dist2 + 0.01, 90, reshape=False)
# dist2 = ndimage.rotate(dist2+0.01,135,reshape=False)
elif randNum < 1.01:
# dist2 = dist2+0.01
# dist2 = ndimage.rotate(dist2+0.01,90,reshape=False)
dist2 = ndimage.rotate(dist2 + 0.01, 135, reshape=False)
else:
print("error!")
dist2 = dist2 + (dist2 == 0) * np.max(dist2)
radius = np.random.randint(minRadius, maxRadius)
im = np.int32(dist2 < radius**2)
# Create a boundary of zeros around the object. Needed to break
# up adjacent objects.
im[0, :] = 0
im[:, 0] = 0
im[-1, :] = 0
im[:, -1] = 0
self.blocks.append(np.reshape(im, int(self.blockSize**2)))
self.reset()
def balanced_dataset(input_dir, gt_dir):
positive_list = []
negative_list = []
# input = os.listdir(input_dir)
# gt = os.listdir(gt_dir)
input = input_dir
gt = gt_dir
input.sort()
gt.sort()
file_count = 0
for x, y in zip(input, gt):
file_count += 1
print('Executing folder ', file_count)
input_sub_dir = os.path.join(input_dir, x)
gt_sub_dir = os.path.join(gt_dir, y)
input_sub_dir_patches = os.listdir(input_sub_dir)
gt_sub_dir_patches = os.listdir(gt_sub_dir)
input_sub_dir_patches.sort()
gt_sub_dir_patches.sort()
for in_patch, gt_patch in zip(input_sub_dir_patches,
gt_sub_dir_patches):
input_image = os.path.join(input_sub_dir, in_patch)
gt_image = os.path.join(gt_sub_dir, gt_patch)
FLAIR_image_in = nib.load(input_image).get_data()
FLAIR_image_in = np.array(FLAIR_image_in)
FLAIR_image_gt = nib.load(gt_image).get_data()
FLAIR_image_gt = np.array(FLAIR_image_gt)
if FLAIR_image_gt.max() == 1.0:
positive_list.append((FLAIR_image_in, 1.0))
FLAIR_image_in_fliped = FLAIR_image_in[::-1, :, :]
positive_list.append((FLAIR_image_in_fliped, 1.0))
FLAIR_image_in_rotated_1 = sind.rotate(FLAIR_image_in,
-12,
reshape=False)
positive_list.append((FLAIR_image_in_rotated_1, 1.0))
FLAIR_image_in_rotated_2 = sind.rotate(FLAIR_image_in,
12,
reshape=False)
positive_list.append((FLAIR_image_in_rotated_2, 1.0))
FLAIR_image_shifted_1 = np.roll(FLAIR_image_in, 10, 0)
positive_list.append((FLAIR_image_shifted_1, 1.0))
FLAIR_image_shifted_2 = np.roll(FLAIR_image_in, -10, 0)
positive_list.append((FLAIR_image_shifted_2, 1.0))
else:
negative_list.append((FLAIR_image_in, 0.0))
print('Executed all folders')
positive_count = len(positive_list)
negative_list_1 = random.sample(negative_list, positive_count)
balanced_list = positive_list + negative_list_1
print(len(positive_list))
print(len(negative_list_1))
random.shuffle(balanced_list)
print(len(balanced_list))
return balanced_list
transition_probs = np.outer(r_next_prior_uncorr(r[1], r_range, sigma_z),
r_next_prior_uncorr(r[0], r_range, sigma_z))
transition_probs = transition_probs / np.sum(transition_probs)
# Weight the V at next t by the transition probabilities
weighted_vals = V_cube[:, :, -(index - 1)] * transition_probs
# Use mean of above minus opportunity cost in addition to V_d_r to find max V at curr r
V_exp_cube[i, j, -index] = np.sum(weighted_vals) - (c + rho) * t_w * t
V_cube[i, j, -index] = np.max((np.sum(weighted_vals) - t * (c + rho) * t_w, V_d_r))
# Store decision identity in cube
decision_cube[i, j, -index] = np.argmax((np.sum(weighted_vals) - (c + rho) * t_w * t,
V_d_r))
if decision_cube[i, j, -index] == 1:
decision_cube[i, j, -index] = np.argmax(r) + 1
rotmat = rotate(decision_cube, -45, axes=(1, 0, 0))
center_slice = rotmat[int(rotmat.shape[0] / 2), :, :]
# Little movie of decision boundary change
fig = plt.figure(figsize=(6, 9))
ax1 = fig.add_subplot(311)
ax2 = fig.add_subplot(312)
ax3 = fig.add_subplot(313)
ax3.pcolor(center_slice)
ax3.set_xticks([])
ax3.set_yticks([])
ax3.set_xlabel('Time')
ax3.set_ylabel('Estimated Reward Difference')
def anim_update(i):
