def dual_gradient_energy(img):
'''
Calculate the vertical energy of each pixel
@return: the energy representation of the img
x_gradient = r_x**2 + g_x**2 + b_x**2
y_gradient = r_y**2 + g_y**2 + b_y**2
energy = x_gradient + y_gradient
'''
h, w = img.shape[:2]
energy = [[0 for i in range(0, w)] for j in range(0, h)]
R = img[:, :, 0]
G = img[:, :, 1]
B = img[:, :, 2]
r_y = filter.hsobel(R)
g_y = filter.hsobel(G)
b_y = filter.hsobel(B)
r_x = filter.vsobel(R)
g_x = filter.vsobel(G)
b_x = filter.vsobel(B)
for i in range(0, h):
for j in range(0, w):
y_en = (r_y[i][j]**2) + (g_y[i][j]**2) + (b_y[i][j]**2)
x_en = (r_x[i][j]**2) + (g_x[i][j]**2) + (b_x[i][j]**2)
energy[i][j] = y_en + x_en
return energy
def initialize(self, start, stop, skip):
self._width = self._img.shape[:2][0]
self._height = self._img.shape[:2][1]
size = self._width * self._height
self._weights = [float] * size
self._edgeTo = [int] * size
self._distTo = [float] * size
R = self._img[:, :, 0]
G = self._img[:, :, 1]
B = self._img[:, :, 2]
self._vR = vsobel(R)
self._vG = vsobel(G)
self._vB = vsobel(B)
self._hR = hsobel(R)
self._hG = hsobel(G)
self._hB = hsobel(B)
for v in xrange(0, size, 1):
if ((v >= start) and (v < stop) and (v - start) % skip == 0):
self._distTo[v] = 0
else:
self._distTo[v] = sys.maxint
self._edgeTo[v] = -1
self._weights[v] = self.energy(self.col(v), self.row(v))
def calculate_residue_scaled_gradients(patches,
old_frame,
new_frame,
patch_radius,
w_func=lambda x, y: 1):
#Calculate some gradients:
gradx = hsobel(new_frame)
grady = vsobel(new_frame)
#Calculate the scaled gradients:
diff = np.subtract(old_frame, new_frame)
ex = np.multiply(diff, gradx)
ey = np.multiply(diff, grady)
w = np.array([[None] * (patch_radius * 2 + 1)] * (patch_radius * 2 + 1))
#Calculate the w:
for i in range(-patch_radius, patch_radius + 1):
for j in range(-patch_radius, patch_radius + 1):
w[i + patch_radius, j + patch_radius] = w_func(i, j)
#Get the scaled gradient at each point:
e_list = np.array([None] * len(patches))
for i, (x, y) in enumerate(patches):
e_list[i] = calculate_residue_gradient(ex, ey, x, y, patch_radius, w)
return e_list
def dual_gradient_energy(self, img):
horizontal_gradient = filter.hsobel(img)
horizontal_gradient = np.square(horizontal_gradient)
vertical_gradient = filter.vsobel(img)
vertical_gradient = np.square(vertical_gradient)
energy = horizontal_gradient + vertical_gradient
return energy
def calculate_structure_tensors(frame,
feature_list,
patch_radius,
w_func=lambda x, y: 1):
#Find gradients:
gradx = hsobel(frame)
grady = vsobel(frame)
#Calculate the w:
w = np.array([[None] * (patch_radius * 2 + 1)] * (patch_radius * 2 + 1))
for i in range(-patch_radius, patch_radius + 1):
for j in range(-patch_radius, patch_radius + 1):
w[i + patch_radius, j + patch_radius] = w_func(i, j)
#Precompute values through numpy:
Gxx = np.multiply(gradx, gradx)
Gxy = np.multiply(gradx, grady)
Gyy = np.multiply(grady, grady)
#Calculate the G matrix of every feature:
G_list = np.array([None] * len(feature_list))
for i, (x, y) in enumerate(feature_list):
G_list[i] = calculate_structure_tensor(Gxx, Gxy, Gyy, x, y,
patch_radius, w)
return G_list
def frac_hedges__frac_vedges(self, image):
'''
fraction of all grayscale pixels that lie on vertical and
horizontal edges in the image interior
'''
v = vsobel(image)
h = hsobel(image)
return [v.mean(), h.mean()]
def sobelh(provider):
"""
sobel horizontal
"""
gray = provider.as_gray()
dx = hsobel(gray) # horizontal derivative
return np.mean(dx)
def test_hsobel_horizontal():
"""Horizontal Sobel on an edge should be a horizontal line"""
i, j = np.mgrid[-5:6, -5:6]
image = (i >= 0).astype(float)
result = F.hsobel(image)
# Fudge the eroded points
i[np.abs(j) == 5] = 10000
assert (np.all(result[i == 0] == 1))
assert (np.all(result[np.abs(i) > 1] == 0))
def test_hsobel_horizontal():
"""Horizontal Sobel on an edge should be a horizontal line."""
i, j = np.mgrid[-5:6, -5:6]
image = (i >= 0).astype(float)
result = F.hsobel(image)
# Fudge the eroded points
i[np.abs(j) == 5] = 10000
assert (np.all(result[i == 0] == 1))
assert (np.all(result[np.abs(i) > 1] == 0))
def sobel_features(image):
h = hsobel(image)
v = vsobel(image)
Gx = np.sum(h)
Gy = np.sum(v)
m = math.sqrt(Gx**2 + Gy**2)
g = 0
if(Gx == 0):
g = 0 if Gy == 0 else math.pi
else:
g = math.atan2(Gy,Gx)
return [m, g, Gx, Gy]
def dual_gradient_energy(img):
R = img[:, :, 0]
G = img[:, :, 1]
B = img[:, :, 2]
hR = filter.hsobel(R)
hG = filter.hsobel(G)
hB = filter.hsobel(B)
vR = filter.vsobel(R)
vG = filter.vsobel(G)
vB = filter.vsobel(B)
sumRG = np.add(np.square(hR), np.square(hG))
x_square = np.add(sumRG, np.square(hB))
sumRGv = np.add(np.square(vR), np.square(vG))
y_square = np.add(sumRGv, np.square(vB))
energy = np.add(x_square, y_square)
return energy
def sobeld(provider):
"""
sobel mean direction
"""
gray = provider.as_gray()
dx = hsobel(gray) # horizontal derivative
dy = vsobel(gray) # vertical derivative
dirs = np.arctan2(dy, dx)
return np.mean(dirs)
def percent_vert_horiz_lines(FilledBlobImg, props_area):
v = filter.vsobel(FilledBlobImg)
v_floor = filter.threshold_otsu(v)
ind_v = np.where(v > v_floor)
h = filter.hsobel(FilledBlobImg)
h_floor = filter.threshold_otsu(h)
ind_h = np.where(h > h_floor)
vert_and_horiz = np.zeros(v.shape).astype("bool")
vert_and_horiz[ind_v] = True
vert_and_horiz[ind_h] = True
ind = np.where(vert_and_horiz)[0]
return float(ind.size) / props_area
def process(filename):
imagepath = os.path.join(os.getcwd(), filename)
orig_img = io.imread(filename,True,'pil')
img = orig_img > 0.9 # binary threshold
lines = probabilistic_hough_line(hsobel(img),line_length=200)
for l in lines:
x0, x1 = l[0][0],l[1][0]
y = l[0][1]
for x in range(x0,x1):
img[y+1,x] = 1
img[y,x] = 1
img[y-1,x] = 1
erode_img = erosion(img, square(2))
contours, lengths = compute_contours(erode_img,0.8)
lengths = pd.Series(lengths)
lengths = lengths[lengths > 400]
for i in lengths.index:
contour = contours[i]
box = get_boundingboxes([contour])[0]
x_sum = sum(map(abs, np.gradient(contour[:,1])))
y_sum = sum(map(abs, np.gradient(contour[:,0])))
area = (box[2] - box[0]) * (box[3] - box[1])
plt.plot(contour[:,1],contour[:,0])
contours = [contours[i] for i in lengths.index]
newboxes = set(link_contours(contours))
retboxes = []
for box in newboxes:
minx,miny,maxx,maxy = box
x = (minx, maxx, maxx, minx, minx)
y = (miny, miny, maxy, maxy, miny)
area = (maxx-minx) * (maxy-miny)
if area > 10000:
retboxes.append(box)
plt.plot(x, y, '-b', linewidth=2)
imshow(erode_img)
return retboxes, contours
def process(filename):
imagepath = os.path.join(os.getcwd(), filename)
orig_img = io.imread(filename, True, 'pil')
img = orig_img > 0.9 # binary threshold
lines = probabilistic_hough_line(hsobel(img), line_length=200)
for l in lines:
x0, x1 = l[0][0], l[1][0]
y = l[0][1]
for x in range(x0, x1):
img[y + 1, x] = 1
img[y, x] = 1
img[y - 1, x] = 1
erode_img = erosion(img, square(2))
contours, lengths = compute_contours(erode_img, 0.8)
lengths = pd.Series(lengths)
lengths = lengths[lengths > 400]
for i in lengths.index:
contour = contours[i]
box = get_boundingboxes([contour])[0]
x_sum = sum(map(abs, np.gradient(contour[:, 1])))
y_sum = sum(map(abs, np.gradient(contour[:, 0])))
area = (box[2] - box[0]) * (box[3] - box[1])
plt.plot(contour[:, 1], contour[:, 0])
contours = [contours[i] for i in lengths.index]
newboxes = set(link_contours(contours))
retboxes = []
for box in newboxes:
minx, miny, maxx, maxy = box
x = (minx, maxx, maxx, minx, minx)
y = (miny, miny, maxy, maxy, miny)
area = (maxx - minx) * (maxy - miny)
if area > 10000:
retboxes.append(box)
plt.plot(x, y, '-b', linewidth=2)
imshow(erode_img)
return retboxes, contours
def gradient_map(image):
h = hsobel(image)
v = vsobel(image)
return np.dstack((h,v))
graycheck = False if pix<thresh: if not graycheck: graycheck = True horcount = horcount + 1 else: graycheck = False for pix in vertslice: graycheck = False if pix<thresh: if not graycheck: graycheck = True vertcount = vertcount + 1 else: graycheck = False features['hsobel'] = np.nanmean(filter.hsobel(image[j])) features['vsobel'] = np.nanmean(filter.vsobel(image[j])) features['peaklocalmax'] = np.nanmean(peak_local_max(image[j])) features['felzen'] = np.nanmean(segmentation.felzenszwalb(image[j])) if np.isnan(features['peaklocalmax']): features['peaklocalmax'] = 0.0 if np.isnan(features['felzen']): features['felzen'] = 0.0 hormirror = image[j][:maxPixel/2]-image[j][maxPixel-1:maxPixel/2:-1] vertmirror = image[j][:,:maxPixel/2]-image[j][:,maxPixel-1:maxPixel/2:-1] #__End moved region image[j] = resize(image[j], (maxPixel, maxPixel)) #From http://scikit-image.org/docs/dev/auto_examples/plot_local_binary_pattern.html lbp = local_binary_pattern(image[j], n_points, radius, METHOD) n_bins = lbp.max()+1
def gradient_orientation_map(image):
h = hsobel(image)
v = vsobel(image)
return np.arctan2(h, v)
def test_hsobel_zeros():
"""Horizontal sobel on an array of all zeros"""
result = F.hsobel(np.zeros((10, 10)), np.ones((10, 10), bool))
assert (np.all(result == 0))
def test_hsobel_vertical():
"""Horizontal Sobel on a vertical edge should be zero"""
i, j = np.mgrid[-5:6, -5:6]
image = (j >= 0).astype(float)
result = F.hsobel(image)
assert (np.all(result == 0))
"""
This example illustrates the use of the horizontal Sobel filter, to compute
horizontal gradients.
"""
from skimage import data, filter
import matplotlib.pyplot as plt
text = data.text()
hsobel_text = filter.hsobel(text)
plt.figure(figsize=(12, 3))
plt.subplot(121)
plt.imshow(text, cmap='gray', interpolation='nearest')
plt.axis('off')
plt.subplot(122)
plt.imshow(hsobel_text, cmap='jet', interpolation='nearest')
plt.axis('off')
plt.tight_layout()
plt.show()
def prep():
header = "acantharia_protist_big_center,acantharia_protist_halo,acantharia_protist,amphipods,appendicularian_fritillaridae,appendicularian_s_shape,appendicularian_slight_curve,appendicularian_straight,artifacts_edge,artifacts,chaetognath_non_sagitta,chaetognath_other,chaetognath_sagitta,chordate_type1,copepod_calanoid_eggs,copepod_calanoid_eucalanus,copepod_calanoid_flatheads,copepod_calanoid_frillyAntennae,copepod_calanoid_large_side_antennatucked,copepod_calanoid_large,copepod_calanoid_octomoms,copepod_calanoid_small_longantennae,copepod_calanoid,copepod_cyclopoid_copilia,copepod_cyclopoid_oithona_eggs,copepod_cyclopoid_oithona,copepod_other,crustacean_other,ctenophore_cestid,ctenophore_cydippid_no_tentacles,ctenophore_cydippid_tentacles,ctenophore_lobate,decapods,detritus_blob,detritus_filamentous,detritus_other,diatom_chain_string,diatom_chain_tube,echinoderm_larva_pluteus_brittlestar,echinoderm_larva_pluteus_early,echinoderm_larva_pluteus_typeC,echinoderm_larva_pluteus_urchin,echinoderm_larva_seastar_bipinnaria,echinoderm_larva_seastar_brachiolaria,echinoderm_seacucumber_auricularia_larva,echinopluteus,ephyra,euphausiids_young,euphausiids,fecal_pellet,fish_larvae_deep_body,fish_larvae_leptocephali,fish_larvae_medium_body,fish_larvae_myctophids,fish_larvae_thin_body,fish_larvae_very_thin_body,heteropod,hydromedusae_aglaura,hydromedusae_bell_and_tentacles,hydromedusae_h15,hydromedusae_haliscera_small_sideview,hydromedusae_haliscera,hydromedusae_liriope,hydromedusae_narco_dark,hydromedusae_narco_young,hydromedusae_narcomedusae,hydromedusae_other,hydromedusae_partial_dark,hydromedusae_shapeA_sideview_small,hydromedusae_shapeA,hydromedusae_shapeB,hydromedusae_sideview_big,hydromedusae_solmaris,hydromedusae_solmundella,hydromedusae_typeD_bell_and_tentacles,hydromedusae_typeD,hydromedusae_typeE,hydromedusae_typeF,invertebrate_larvae_other_A,invertebrate_larvae_other_B,jellies_tentacles,polychaete,protist_dark_center,protist_fuzzy_olive,protist_noctiluca,protist_other,protist_star,pteropod_butterfly,pteropod_theco_dev_seq,pteropod_triangle,radiolarian_chain,radiolarian_colony,shrimp_caridean,shrimp_sergestidae,shrimp_zoea,shrimp-like_other,siphonophore_calycophoran_abylidae,siphonophore_calycophoran_rocketship_adult,siphonophore_calycophoran_rocketship_young,siphonophore_calycophoran_sphaeronectes_stem,siphonophore_calycophoran_sphaeronectes_young,siphonophore_calycophoran_sphaeronectes,siphonophore_other_parts,siphonophore_partial,siphonophore_physonect_young,siphonophore_physonect,stomatopod,tornaria_acorn_worm_larvae,trichodesmium_bowtie,trichodesmium_multiple,trichodesmium_puff,trichodesmium_tuft,trochophore_larvae,tunicate_doliolid_nurse,tunicate_doliolid,tunicate_partial,tunicate_salp_chains,tunicate_salp,unknown_blobs_and_smudges,unknown_sticks,unknown_unclassified".split(',')
with open('namesClasses.dat','rb') as f:
namesClasses = cPickle.load(f)
labels = map(lambda s: s.split('\\')[-1], namesClasses)
for i in range(len(namesClasses)):
currentClass = namesClasses[i]
root_class.data_search(currentClass).id_list.append(i)
#get the total test images
#Full set
#fnames = glob.glob(os.path.join("competition_data", "test", "*.jpg"))
#Smaller set
fnames = glob.glob(os.path.join("smaller_comp", "test", "*.jpg"))
numberofTestImages = len(fnames)
X = np.zeros((numberofTestImages, num_features), dtype=float)
#Get filename separate from prefix path
images = map(lambda fileName: fileName.split('\\')[-1], fnames)
i = 0
# report progress for each 1% done
report = [int((j+1)*numberofTestImages/100.) for j in range(100)]
for fileName in fnames:
# Read in the images and create the features
image = imread(fileName, as_grey=True)
image = rotImage(image)
# Added from https://github.com/Newmu/Stroke-Prediction/blob/master/startPredictingGenCode.py
thresh = 0.9*255
# if image.min < 0.75*255:
# img = image < thresh
# else:
# img = image
# if img.sum() != 0:
# imgX,imgY = np.nonzero(img)
# imgW = imgX.max()-imgX.min()
# imgH = imgY.max()-imgY.min()
# if (imgW>1 and imgH>1):
# image = image[imgX.min():imgX.max(),imgY.min():imgY.max()]
# #----------------------------------
cvimage = cv2.imread(fileName)
features = getMinorMajorRatio(image)
#__Begin moved region
#From http://nbviewer.ipython.org/github/kqdtran/AY250-F13/blob/master/hw4/hw4.ipynb
pca = decomposition.PCA(n_components=25)
PCAFeatures = pca.fit_transform(image)
PCAevr = pca.explained_variance_ratio_
for evr in range(len(PCAevr)):
if np.isnan(PCAevr[evr]):
PCAevr[evr] = 0.0
#_____________________________________________________________
corners = getCorners(cvimage,features['orientation'])
horslice = image[:,maxPixel/2]
vertslice = image[maxPixel/2]
#correlation = signal.correlate(image,image)
#horcorrelation = signal.correlate(horslice,horslice)
#vertcorrelation = signal.correlate(vertslice,vertslice)
#crosscorrelation = signal.correlate(horslice,vertslice)
#correlation = correlation/correlation[correlation.shape[0]/2,correlation.shape[0]/2]
#horcorrelation = horcorrelation/horcorrelation[horcorrelation.shape[0]/2]
#vertcorrelation = vertcorrelation/vertcorrelation[vertcorrelation.shape[0]/2]
#crosscorrelation = crosscorrelation/crosscorrelation[horcorrelation.shape[0]/2]
hormean = np.mean(horslice)
horstd = np.std(horslice)
vertmean = np.mean(vertslice)
vertstd = np.std(vertslice)
horcount = vertcount = 0
for pix in horslice:
graycheck = False
if pix<thresh:
if not graycheck:
graycheck = True
horcount = horcount + 1
else:
graycheck = False
for pix in vertslice:
graycheck = False
if pix<thresh:
if not graycheck:
graycheck = True
vertcount = vertcount + 1
else:
graycheck = False
features['hsobel'] = np.nanmean(filter.hsobel(image))
features['vsobel'] = np.nanmean(filter.vsobel(image))
features['peaklocalmax'] = np.nanmean(peak_local_max(image))
features['felzen'] = np.nanmean(segmentation.felzenszwalb(image))
if np.isnan(features['peaklocalmax']):
features['peaklocalmax'] = 0.0
if np.isnan(features['felzen']):
features['felzen'] = 0.0
#hormirror = image[:maxPixel/2]-image[maxPixel-1:maxPixel/2:-1]
#vertmirror = image[:,:maxPixel/2]-image[:,maxPixel-1:maxPixel/2:-1]
image = resize(image, (maxPixel, maxPixel))
#From http://scikit-image.org/docs/dev/auto_examples/plot_local_binary_pattern.html
lbp = local_binary_pattern(image, n_points, radius, METHOD)
n_bins = lbp.max()+1
lbpcounts = np.histogram(lbp,n_bins,normed=True,range=(0, n_bins))[0]
#_____________________________________________________________
#__Moved region was here
#dist_trans = ndimage.distance_transform_edt(image[0])
# Store the rescaled image pixels and the axis ratio
#fd, hog_image = hog(image[0], orientations=8, pixels_per_cell=(2, 2),
# cells_per_block=(1, 1), visualise=True)
#X[i, 0:imageSize] = np.reshape(dist_trans, (1, imageSize))
#X[i, 0:imageSize] = np.reshape(hog_image, (1, imageSize))
# Store the rescaled image pixels and the axis ratio
X[i, 0:imageSize] = np.reshape(image, (1, imageSize))
#X[i, imageSize:imageSize+corr2dsize] = np.reshape(correlation, (1,corr2dsize))
#X[i, imageSize+corr2dsize:imageSize+corr2dsize+corrsize] = np.reshape(horcorrelation, (1,corrsize))
#X[i, imageSize+corr2dsize+corrsize:imageSize+corr2dsize+2*corrsize] = np.reshape(vertcorrelation, (1,corrsize))
#X[i, imageSize+corr2dsize+2*corrsize:imageSize+corr2dsize+3*corrsize] = np.reshape(crosscorrelation, (1,corrsize))
featcount = imageSize+3*corrsize+corr2dsize
for k,v in features.items():
try:
X[i, featcount:featcount+len(v)] = v
featcount = featcount + len(v)
except TypeError, te:
X[i, featcount] = v
featcount = featcount + 1
X[i, featcount:featcount+lbpcounts.size] = lbpcounts
X[i, featcount+lbpcounts.size:featcount+lbpcounts.size+PCAsize] = PCAevr
X[i, featcount+lbpcounts.size+PCAsize] = np.mean(PCAFeatures)
i += 1
if i in report: print np.ceil(i *100.0 / numberofTestImages), "% done"
graycheck = False
if pix < thresh:
if not graycheck:
graycheck = True
horcount = horcount + 1
else:
graycheck = False
for pix in vertslice:
graycheck = False
if pix < thresh:
if not graycheck:
graycheck = True
vertcount = vertcount + 1
else:
graycheck = False
features["hsobel"] = np.nanmean(filter.hsobel(image[j]))
features["vsobel"] = np.nanmean(filter.vsobel(image[j]))
features["peaklocalmax"] = np.nanmean(peak_local_max(image[j]))
features["felzen"] = np.nanmean(segmentation.felzenszwalb(image[j]))
if np.isnan(features["peaklocalmax"]):
features["peaklocalmax"] = 0.0
if np.isnan(features["felzen"]):
features["felzen"] = 0.0
hormirror = image[j][: maxPixel / 2] - image[j][maxPixel - 1 : maxPixel / 2 : -1]
vertmirror = image[j][:, : maxPixel / 2] - image[j][:, maxPixel - 1 : maxPixel / 2 : -1]
# __End moved region
image[j] = resize(image[j], (maxPixel, maxPixel))
# From http://scikit-image.org/docs/dev/auto_examples/plot_local_binary_pattern.html
lbp = local_binary_pattern(image[j], n_points, radius, METHOD)
n_bins = lbp.max() + 1
def test_hsobel_mask():
"""Horizontal Sobel on a masked array should be zero."""
np.random.seed(0)
result = F.hsobel(np.random.uniform(size=(10, 10)),
np.zeros((10, 10), bool))
assert (np.all(result == 0))
def test_hsobel_vertical():
"""Horizontal Sobel on a vertical edge should be zero."""
i, j = np.mgrid[-5:6, -5:6]
image = (j >= 0).astype(float) * np.sqrt(2)
result = F.hsobel(image)
assert_allclose(result, 0, atol=1e-10)
import numpy as np
import skimage.io as io
from skimage import filter
from scipy.signal import argrelextrema
from scipy.signal import medfilt
img = io.imread('../../data/cbct/fp200_x70_3s_5fps_60kV_150uA_00001.tif')
#img=io.imread('../../data/cbct/fp150_x350_3s_5fps_60kV_150uA_00001.tif')
#plt.figure(figsize=(10,2/3*10))
plt.subplot(321)
plt.imshow(img)
##
# Prepare
res = np.abs(filter.gaussian_filter(filter.hsobel(img), 5))
res = np.clip(res, 0.0, 0.2)
plt.subplot(322)
plt.imshow(res)
plt.colorbar()
# Identify the left and right sides of the circles
#vsum=medfilt(np.sum(res,0),7)
vsum = np.sum(res, 0)
idx = argrelextrema(vsum, np.greater)
peaks = vsum[idx]
p1 = np.argmax(peaks)
peaks = np.delete(peaks, p1)
p2 = np.argmax(peaks)
p2 = p2 + 1 if p1 < p2 else p2
p1 = idx[0][p1]
def edge_sobel_h(img2d):
"gray input"
imgsize = float((img2d.shape[0]*img2d.shape[1]))
med_filter = ndimg.median_filter(img2d, size = (5,5))
edges_h = filter.hsobel(med_filter/255.)
return edges_h.sum()/imgsize
def test_hsobel_vertical():
"""Horizontal Sobel on a vertical edge should be zero"""
i, j = np.mgrid[-5:6, -5:6]
image = (j >= 0).astype(float)
result = F.hsobel(image)
assert (np.all(result == 0))
def gradient_orientation_map(image):
h = hsobel(image)
v = vsobel(image)
return np.arctan2(h,v)
def main():
files = []
# Generate training data
i = 0
label = 0
# List of string of class names
namesClasses = list()
features = {}
features = features.copy()
print "Reading images"
# Navigate through the list of directories
for folder in directory_names:
#Get name of class directory separate from prefix path
currentClass = folder.split(os.sep)[-1]
namesClasses.append(currentClass)
root_class.data_search(currentClass).id_list.append(
len(namesClasses) - 1)
for fileNameDir in os.walk(folder):
for fileName in fileNameDir[2]:
# Only read in the images
if fileName[-4:] != ".jpg":
continue
# Read in the images and create the features
nameFileImage = "{0}{1}{2}".format(fileNameDir[0], os.sep,
fileName)
image = []
image.append(imread(nameFileImage, as_grey=True))
features['original_size'] = image[0].size
#image[0] = equalize_hist(image[0])
image[0] = rotImage(image[0])
# Added from https://github.com/Newmu/Stroke-Prediction/blob/master/startPredictingGenCode.py
thresh = 0.9 * 255
# if image.min < 0.75*255:
# img = image < thresh
# else:
# img = image
# if img.sum() != 0:
# imgX,imgY = np.nonzero(img)
# imgW = imgX.max()-imgX.min()
# imgH = imgY.max()-imgY.min()
# if (imgW>1 and imgH>1):
# image = image[imgX.min():imgX.max(),imgY.min():imgY.max()]
#----------------------------------
cvimage = cv2.imread(nameFileImage)
#image[0] = gaussian_filter(image[0],sigma=2)
files.append(nameFileImage)
image.append(np.fliplr(image[0]))
image.append(np.flipud(image[0]))
image.append(np.fliplr(image[2]))
# image.append(np.rot90(image[0]))
# image.append(np.fliplr(image[4]))
# image.append(np.flipud(image[4]))
# image.append(np.fliplr(image[6]))
for j in range(len(image)):
features = getMinorMajorRatio(image[j])
#__Begin moved region
#From http://nbviewer.ipython.org/github/kqdtran/AY250-F13/blob/master/hw4/hw4.ipynb
pca = decomposition.PCA(n_components=25)
PCAFeatures = pca.fit_transform(image[0])
#_____________________________________________________________
corners = getCorners(cvimage, features['orientation'])
horslice = image[j][:, maxPixel / 2]
vertslice = image[j][maxPixel / 2]
#correlation = signal.correlate(image,image)
#horcorrelation = signal.correlate(horslice,horslice)
#vertcorrelation = signal.correlate(vertslice,vertslice)
#crosscorrelation = signal.correlate(horslice,vertslice)
#correlation = correlation/correlation[correlation.shape[0]/2,correlation.shape[0]/2]
#horcorrelation = horcorrelation/horcorrelation[horcorrelation.shape[0]/2]
#vertcorrelation = vertcorrelation/vertcorrelation[vertcorrelation.shape[0]/2]
#crosscorrelation = crosscorrelation/crosscorrelation[horcorrelation.shape[0]/2]
hormean = np.mean(horslice)
horstd = np.std(horslice)
vertmean = np.mean(vertslice)
vertstd = np.std(vertslice)
horcount = vertcount = 0
for pix in horslice:
graycheck = False
if pix < thresh:
if not graycheck:
graycheck = True
horcount = horcount + 1
else:
graycheck = False
for pix in vertslice:
graycheck = False
if pix < thresh:
if not graycheck:
graycheck = True
vertcount = vertcount + 1
else:
graycheck = False
features['hsobel'] = np.nanmean(filter.hsobel(image[j]))
features['vsobel'] = np.nanmean(filter.vsobel(image[j]))
features['peaklocalmax'] = np.nanmean(
peak_local_max(image[j]))
features['felzen'] = np.nanmean(
segmentation.felzenszwalb(image[j]))
if np.isnan(features['peaklocalmax']):
features['peaklocalmax'] = 0.0
if np.isnan(features['felzen']):
features['felzen'] = 0.0
#hormirror = image[j][:maxPixel/2]-image[j][maxPixel-1:maxPixel/2:-1]
#vertmirror = image[j][:,:maxPixel/2]-image[j][:,maxPixel-1:maxPixel/2:-1]
#__End moved region
image[j] = resize(image[j], (maxPixel, maxPixel))
#From http://scikit-image.org/docs/dev/auto_examples/plot_local_binary_pattern.html
lbp = local_binary_pattern(image[j], n_points, radius,
METHOD)
n_bins = lbp.max() + 1
lbpcounts = np.histogram(lbp,
n_bins,
normed=True,
range=(0, n_bins))[0]
#_____________________________________________________________
#__Moved region was here
#dist_trans = ndimage.distance_transform_edt(image[0])
# Store the rescaled image pixels and the axis ratio
#fd, hog_image = hog(image[0], orientations=8, pixels_per_cell=(2, 2),
# cells_per_block=(1, 1), visualise=True)
#X[i*imagesperfile+j, 0:imageSize] = np.reshape(dist_trans, (1, imageSize))
#X[i*imagesperfile+j, 0:imageSize] = np.reshape(hog_image, (1, imageSize))
X[i * imagesperfile + j,
0:imageSize] = np.reshape(image[j], (1, imageSize))
#X[i*imagesperfile+j, imageSize:imageSize+corr2dsize] = np.reshape(correlation, (1,corr2dsize))
#X[i*imagesperfile+j, imageSize+corr2dsize:imageSize+corr2dsize+corrsize] = np.reshape(horcorrelation, (1,corrsize))
#X[i*imagesperfile+j, imageSize+corr2dsize+corrsize:imageSize+corr2dsize+2*corrsize] = np.reshape(vertcorrelation, (1,corrsize))
#X[i*imagesperfile+j, imageSize+corr2dsize+2*corrsize:imageSize+corr2dsize+3*corrsize] = np.reshape(crosscorrelation, (1,corrsize))
featcount = imageSize + 3 * corrsize + corr2dsize
for k, v in features.items():
try:
X[i * imagesperfile + j,
featcount:featcount + len(v)] = v
featcount = featcount + len(v)
except TypeError, te:
X[i * imagesperfile + j, featcount] = v
featcount = featcount + 1
X[i * imagesperfile + j,
featcount:featcount + lbpcounts.size] = lbpcounts
X[i * imagesperfile + j,
featcount + lbpcounts.size:featcount + lbpcounts.size +
PCAsize] = pca.explained_variance_ratio_
X[i * imagesperfile + j, featcount + lbpcounts.size +
PCAsize] = np.mean(PCAFeatures)
# Store the classlabel
y[i * imagesperfile + j] = label
if i * imagesperfile + j in report:
print np.ceil((i * imagesperfile + j) * 100.0 /
(num_rows)), "% done"
i += 1
label += 1
def extract_features(im_cat_path_list):
'''
Extract Features takes a list of 2 element lists:
[catagory of image , path to image]
and extracts 15 features from this image. The output is a list of 3 element lists:
[catagory of image , path to image, [list of features]]
The 15 features are:
1 Product of image pixel dimensions (image size)
2 Mean of Grayscale Image
3 Area of Sobel Edges Normalized By Image Size
4 Area of Sobel Edges Above 2x Mean of Sobel Edges Normalized By Image Size
5 Area of Canny Edges (Sum of booleans) Normalized By Image Size
6 Number of Harris Corners
7 Unique Felzenszwalb Image Segmentation Lines
8 Area of Vertical Sobel Edges Above 2x Mean of Sobel Edges Normalized By Image Size
9 Area of Horizontal Sobel Edges Above 2x Mean of Sobel Edges Normalized By Image Size
10-12 Mean of Red/Green/Blue Channels (if grayscale: mean of the only color channel)
13 Maximum Pixel Value of the Histogram of Oriented Gradients
14 Percent of image that is light versus dark with adaptive thresholding
15-17 Percent of image that is red/green/blue with adaptive thresholding
'''
cat_path_features = []
for im_cat, im_path in im_cat_path_list:
#RAW IMAGE
im_raw = imread(im_path) #image matrix
#RAW IMAGE FLATTENED IF NOT ALREADY FLAT
if len(np.shape(im_raw)) == 3:
im_raw_flat = np.median(im_raw, axis=2)
else:
im_raw_flat = im_raw
#Size of image
im_y = float(im_raw.shape[0])
im_x = float(im_raw.shape[1])
im_size = im_y * im_x
#LIST OF FEATURES
features = []
#FEATURE 1: Product of image pixel dimensions (image size)
features.append(float(im_size))
#FEATURE 2: Mean of Grayscale Image
features.append(im_raw_flat.mean())
#FEATURE 3: Area of Sobel Edges Normalized By Image Size
im_edge_sobel = filter.sobel(im_raw_flat)
features.append(im_edge_sobel.sum() / im_size)
#FEATURE 4: Area of Sobel Edges Above 2x Mean of Sobel Edges Normalized By Image Size
features.append(
float((im_edge_sobel > im_edge_sobel.mean() * 2).sum()) / im_size)
#FEATURE 5: Area of Canny Edges (Sum of booleans) Normalized By Image Size
im_canny = filter.canny(im_raw_flat, sigma=8)
features.append(im_canny.sum().astype(float) / im_size)
#FEATURE 6: Number of Harris Corners
im_corners = feature.corner_peaks(feature.corner_harris(im_raw_flat),
min_distance=5)
features.append(float(len(im_corners)))
#FEATURE 7: Unique Felzenszwalb Image Segmentation Lines
im_raw_float = util.img_as_float(im_raw[::2, ::2])
im_felzen_segments = segmentation.felzenszwalb(im_raw_float,
scale=100,
sigma=0.5,
min_size=50)
features.append(float(len(np.unique(im_felzen_segments))))
#FEATURE 8: Area of Vertical Sobel Edges Above 2x Mean of Sobel Edges Normalized By Image Size
im_edge_vsobel = filter.vsobel(im_raw_flat)
features.append(
float(
(im_edge_vsobel > im_edge_vsobel.mean() * 2).sum()) / im_size)
#FEATURE 9: Area of Horizontal Sobel Edges Above 2x Mean of Sobel Edges Normalized By Image Size
im_edge_hsobel = filter.hsobel(im_raw_flat)
features.append(
float(
(im_edge_hsobel > im_edge_hsobel.mean() * 2).sum()) / im_size)
#FEATURE 10-12: Mean of Red/Green/Blue Channels (if grayscale: mean of the only color channel)
if len(np.shape(im_raw)) == 3:
features.append(im_raw[..., 0].mean())
features.append(im_raw[..., 1].mean())
features.append(im_raw[..., 2].mean())
else:
features.append(im_raw_flat.mean())
features.append(im_raw_flat.mean())
features.append(im_raw_flat.mean())
#FEATURE 13: Maximum Pixel Value of the Histogram of Oriented Gradients
im_fd, im_hog = feature.hog(im_raw_flat,
orientations=8,
pixels_per_cell=(16, 16),
cells_per_block=(1, 1),
visualise=True)
features.append(im_hog.max())
#FEATURE 14: Percent of image that is light versus dark with adaptive thresholding
im_thres_flat = filter.threshold_adaptive(im_raw_flat, 100, 'mean')
features.append(im_thres_flat.sum() / im_size)
#FEATURE 15-17: Percent of image that is red/green/blue with adaptive thresholding
im_thres_red = filter.threshold_adaptive(im_raw[..., 0], 100, 'mean')
im_thres_green = filter.threshold_adaptive(im_raw[..., 1], 100, 'mean')
im_thres_blue = filter.threshold_adaptive(im_raw[..., 2], 100, 'mean')
features.append(im_thres_red.sum() / im_size)
features.append(im_thres_green.sum() / im_size)
features.append(im_thres_blue.sum() / im_size)
#BUILD OUTPUT LIST FOR THIS IMAGE
cat_path_features.append([im_cat, im_path, features])
#CLEAR IMAGE PROC DATA
del im_raw
del im_raw_flat
del im_raw_float
del im_edge_sobel
del im_canny
del im_corners
del im_felzen_segments
del im_edge_vsobel
del im_edge_hsobel
del im_fd
del im_hog
return cat_path_features
def main():
files = []
# Generate training data
i = 0
label = 0
# List of string of class names
namesClasses = list()
features = {}
features = features.copy()
print "Reading images"
# Navigate through the list of directories
for folder in directory_names:
#Get name of class directory separate from prefix path
currentClass = folder.split(os.sep)[-1]
namesClasses.append(currentClass)
root_class.data_search(currentClass).id_list.append(len(namesClasses)-1)
for fileNameDir in os.walk(folder):
for fileName in fileNameDir[2]:
# Only read in the images
if fileName[-4:] != ".jpg":
continue
# Read in the images and create the features
nameFileImage = "{0}{1}{2}".format(fileNameDir[0], os.sep, fileName)
image = []
image.append(imread(nameFileImage, as_grey=True))
features['original_size'] = image[0].size
#image[0] = equalize_hist(image[0])
image[0] = rotImage(image[0])
# Added from https://github.com/Newmu/Stroke-Prediction/blob/master/startPredictingGenCode.py
thresh = 0.9*255
# if image.min < 0.75*255:
# img = image < thresh
# else:
# img = image
# if img.sum() != 0:
# imgX,imgY = np.nonzero(img)
# imgW = imgX.max()-imgX.min()
# imgH = imgY.max()-imgY.min()
# if (imgW>1 and imgH>1):
# image = image[imgX.min():imgX.max(),imgY.min():imgY.max()]
#----------------------------------
cvimage = cv2.imread(nameFileImage)
#image[0] = gaussian_filter(image[0],sigma=2)
files.append(nameFileImage)
image.append(np.fliplr(image[0]))
image.append(np.flipud(image[0]))
image.append(np.fliplr(image[2]))
# image.append(np.rot90(image[0]))
# image.append(np.fliplr(image[4]))
# image.append(np.flipud(image[4]))
# image.append(np.fliplr(image[6]))
for j in range(len(image)):
features = getMinorMajorRatio(image[j])
#__Begin moved region
#From http://nbviewer.ipython.org/github/kqdtran/AY250-F13/blob/master/hw4/hw4.ipynb
pca = decomposition.PCA(n_components=25)
PCAFeatures = pca.fit_transform(image[0])
#_____________________________________________________________
corners = getCorners(cvimage,features['orientation'])
horslice = image[j][:,maxPixel/2]
vertslice = image[j][maxPixel/2]
#correlation = signal.correlate(image,image)
#horcorrelation = signal.correlate(horslice,horslice)
#vertcorrelation = signal.correlate(vertslice,vertslice)
#crosscorrelation = signal.correlate(horslice,vertslice)
#correlation = correlation/correlation[correlation.shape[0]/2,correlation.shape[0]/2]
#horcorrelation = horcorrelation/horcorrelation[horcorrelation.shape[0]/2]
#vertcorrelation = vertcorrelation/vertcorrelation[vertcorrelation.shape[0]/2]
#crosscorrelation = crosscorrelation/crosscorrelation[horcorrelation.shape[0]/2]
hormean = np.mean(horslice)
horstd = np.std(horslice)
vertmean = np.mean(vertslice)
vertstd = np.std(vertslice)
horcount = vertcount = 0
for pix in horslice:
graycheck = False
if pix<thresh:
if not graycheck:
graycheck = True
horcount = horcount + 1
else:
graycheck = False
for pix in vertslice:
graycheck = False
if pix<thresh:
if not graycheck:
graycheck = True
vertcount = vertcount + 1
else:
graycheck = False
features['hsobel'] = np.nanmean(filter.hsobel(image[j]))
features['vsobel'] = np.nanmean(filter.vsobel(image[j]))
features['peaklocalmax'] = np.nanmean(peak_local_max(image[j]))
features['felzen'] = np.nanmean(segmentation.felzenszwalb(image[j]))
if np.isnan(features['peaklocalmax']):
features['peaklocalmax'] = 0.0
if np.isnan(features['felzen']):
features['felzen'] = 0.0
#hormirror = image[j][:maxPixel/2]-image[j][maxPixel-1:maxPixel/2:-1]
#vertmirror = image[j][:,:maxPixel/2]-image[j][:,maxPixel-1:maxPixel/2:-1]
#__End moved region
image[j] = resize(image[j], (maxPixel, maxPixel))
#From http://scikit-image.org/docs/dev/auto_examples/plot_local_binary_pattern.html
lbp = local_binary_pattern(image[j], n_points, radius, METHOD)
n_bins = lbp.max()+1
lbpcounts = np.histogram(lbp,n_bins,normed=True,range=(0, n_bins))[0]
#_____________________________________________________________
#__Moved region was here
#dist_trans = ndimage.distance_transform_edt(image[0])
# Store the rescaled image pixels and the axis ratio
#fd, hog_image = hog(image[0], orientations=8, pixels_per_cell=(2, 2),
# cells_per_block=(1, 1), visualise=True)
#X[i*imagesperfile+j, 0:imageSize] = np.reshape(dist_trans, (1, imageSize))
#X[i*imagesperfile+j, 0:imageSize] = np.reshape(hog_image, (1, imageSize))
X[i*imagesperfile+j, 0:imageSize] = np.reshape(image[j], (1, imageSize))
#X[i*imagesperfile+j, imageSize:imageSize+corr2dsize] = np.reshape(correlation, (1,corr2dsize))
#X[i*imagesperfile+j, imageSize+corr2dsize:imageSize+corr2dsize+corrsize] = np.reshape(horcorrelation, (1,corrsize))
#X[i*imagesperfile+j, imageSize+corr2dsize+corrsize:imageSize+corr2dsize+2*corrsize] = np.reshape(vertcorrelation, (1,corrsize))
#X[i*imagesperfile+j, imageSize+corr2dsize+2*corrsize:imageSize+corr2dsize+3*corrsize] = np.reshape(crosscorrelation, (1,corrsize))
featcount = imageSize+3*corrsize+corr2dsize
for k,v in features.items():
try:
X[i*imagesperfile+j, featcount:featcount+len(v)] = v
featcount = featcount + len(v)
except TypeError, te:
X[i*imagesperfile+j, featcount] = v
featcount = featcount + 1
X[i*imagesperfile+j, featcount:featcount+lbpcounts.size] = lbpcounts
X[i*imagesperfile+j, featcount+lbpcounts.size:featcount+lbpcounts.size+PCAsize] = pca.explained_variance_ratio_
X[i*imagesperfile+j, featcount+lbpcounts.size+PCAsize] = np.mean(PCAFeatures)
# Store the classlabel
y[i*imagesperfile+j] = label
if i*imagesperfile+j in report: print np.ceil((i*imagesperfile+j) *100.0 / (num_rows)), "% done"
i += 1
label += 1
X[i, imageSize + 3 * corrsize + corr2dsize + 11] = perimratio
X[i, imageSize + 3 * corrsize + corr2dsize + 12] = arearatio
X[i, imageSize + 3 * corrsize + corr2dsize + 13] = cornercenter
X[i, imageSize + 3 * corrsize + corr2dsize +
14] = cornercentercoords[0]
X[i, imageSize + 3 * corrsize + corr2dsize +
15] = cornercentercoords[1]
X[i, imageSize + 3 * corrsize + corr2dsize + 16] = cornerstd
X[i,
imageSize + 3 * corrsize + corr2dsize + 17] = cornerstdcoords[0]
X[i,
imageSize + 3 * corrsize + corr2dsize + 18] = cornerstdcoords[1]
X[i, imageSize + 3 * corrsize + corr2dsize + 19] = np.nanmean(
filter.vsobel(image))
X[i, imageSize + 3 * corrsize + corr2dsize + 20] = np.nanmean(
filter.hsobel(image))
X[i, imageSize + 3 * corrsize + corr2dsize + 21] = felzen
X[i, imageSize + 3 * corrsize + corr2dsize + 22] = peaklocalmax
X[i, imageSize + 3 * corrsize + corr2dsize + 23] = lrdiff
X[i, imageSize + 3 * corrsize + corr2dsize + 24] = tbdiff
X[i, imageSize + 3 * corrsize + corr2dsize + 25] = whu1
X[i, imageSize + 3 * corrsize + corr2dsize + 26] = whu2
X[i, imageSize + 3 * corrsize + corr2dsize + 27] = whu3
X[i, imageSize + 3 * corrsize + corr2dsize + 28] = whu12
X[i, imageSize + 3 * corrsize + corr2dsize + 29] = whu13
X[i, imageSize + 3 * corrsize + corr2dsize + 30] = whu23
X[i, imageSize + 3 * corrsize + corr2dsize + 31] = extent
X[i, imageSize + 3 * corrsize + corr2dsize + 32] = minintensity
X[i, imageSize + 3 * corrsize + corr2dsize + 33] = meanintensity
X[i, imageSize + 3 * corrsize + corr2dsize + 34] = maxintensity
X[i, imageSize + 3 * corrsize + corr2dsize + 35] = intensityratio1
def sharpness(image, edge_threshold=0.0001, w=5, t=1.0):
"""
Code implemented by the following article:
Sharpness Estimation for Document and Scene Images,
Kumar, Chen, Doermann
"""
# Median filtering
w_size = (w * 2) + 1
image = util.img_as_float(image)
image_m = util.img_as_float(median(image, mph.square(3)))
# Window functions
def dom_func(window):
# import pdb; pdb.set_trace()
return abs(
abs(window[4] - window[2]) - abs(window[2] - window[0])
)
def contrast_func(window):
# print window
s = 0.0
for i in xrange(0, len(window) - 1):
# print i
s += abs(window[i] - window[i+1])
return s
# Delta DoM in horizontal direction
dom_x_values = generic_filter(image_m, dom_func,
size=(1, 5),
mode='reflect')
# Delta DoM in vertical direction
dom_y_values = generic_filter(image_m, dom_func,
size=(5, 1),
mode='reflect')
dom_x = generic_filter(
dom_x_values, lambda w: sum(w),
size=(1, w_size), mode='reflect')
dom_y = generic_filter(
dom_y_values, lambda w: sum(w),
size=(w_size, 1), mode='reflect')
edges_x = vsobel(image)
# Normalize
edges_x *= (1.0 / edges_x.max())
edges_x_pixels = len(edges_x[edges_x > edge_threshold].ravel())
edges_y = hsobel(image)
# Normalize
edges_y *= (1.0 / edges_y.max())
edges_y_pixels = len(edges_y[edges_y > edge_threshold].ravel())
# Contrast in horizontal direction
contrast_x = generic_filter(image, contrast_func,
size=(1, w_size + 1),
mode='reflect')
# Contrast in vertical direction
contrast_y = generic_filter(image, contrast_func,
size=(w_size + 1, 1),
mode='reflect')
sharpness_x = dom_x / contrast_x
sharpness_y = dom_y / contrast_y
# import pdb; pdb.set_trace()
sharp_x_pixels = len(np.where(
sharpness_x[edges_x > edge_threshold] > t
)[0])
sharp_y_pixels = len(np.where(
sharpness_y[edges_y > edge_threshold] > t
)[0])
# import pdb; pdb.set_trace()
if edges_x_pixels > 0:
rx = (float(sharp_x_pixels) / edges_x_pixels)
else:
rx = 1
if edges_y_pixels > 0:
ry = (float(sharp_y_pixels) / edges_y_pixels)
else:
ry = 1
final_estimate = np.sqrt(
(rx ** 2) + (ry ** 2)
)
return final_estimate
def gradient_map(image):
h = hsobel(image)
v = vsobel(image)
return np.dstack((h, v))
def test_hsobel_zeros():
"""Horizontal sobel on an array of all zeros."""
result = F.hsobel(np.zeros((10, 10)), np.ones((10, 10), bool))
assert (np.all(result == 0))
def test_hsobel_mask():
"""Horizontal Sobel on a masked array should be zero"""
np.random.seed(0)
result = F.hsobel(np.random.uniform(size=(10, 10)), np.zeros((10, 10),
bool))
assert (np.all(result == 0))
def prep():
header = "acantharia_protist_big_center,acantharia_protist_halo,acantharia_protist,amphipods,appendicularian_fritillaridae,appendicularian_s_shape,appendicularian_slight_curve,appendicularian_straight,artifacts_edge,artifacts,chaetognath_non_sagitta,chaetognath_other,chaetognath_sagitta,chordate_type1,copepod_calanoid_eggs,copepod_calanoid_eucalanus,copepod_calanoid_flatheads,copepod_calanoid_frillyAntennae,copepod_calanoid_large_side_antennatucked,copepod_calanoid_large,copepod_calanoid_octomoms,copepod_calanoid_small_longantennae,copepod_calanoid,copepod_cyclopoid_copilia,copepod_cyclopoid_oithona_eggs,copepod_cyclopoid_oithona,copepod_other,crustacean_other,ctenophore_cestid,ctenophore_cydippid_no_tentacles,ctenophore_cydippid_tentacles,ctenophore_lobate,decapods,detritus_blob,detritus_filamentous,detritus_other,diatom_chain_string,diatom_chain_tube,echinoderm_larva_pluteus_brittlestar,echinoderm_larva_pluteus_early,echinoderm_larva_pluteus_typeC,echinoderm_larva_pluteus_urchin,echinoderm_larva_seastar_bipinnaria,echinoderm_larva_seastar_brachiolaria,echinoderm_seacucumber_auricularia_larva,echinopluteus,ephyra,euphausiids_young,euphausiids,fecal_pellet,fish_larvae_deep_body,fish_larvae_leptocephali,fish_larvae_medium_body,fish_larvae_myctophids,fish_larvae_thin_body,fish_larvae_very_thin_body,heteropod,hydromedusae_aglaura,hydromedusae_bell_and_tentacles,hydromedusae_h15,hydromedusae_haliscera_small_sideview,hydromedusae_haliscera,hydromedusae_liriope,hydromedusae_narco_dark,hydromedusae_narco_young,hydromedusae_narcomedusae,hydromedusae_other,hydromedusae_partial_dark,hydromedusae_shapeA_sideview_small,hydromedusae_shapeA,hydromedusae_shapeB,hydromedusae_sideview_big,hydromedusae_solmaris,hydromedusae_solmundella,hydromedusae_typeD_bell_and_tentacles,hydromedusae_typeD,hydromedusae_typeE,hydromedusae_typeF,invertebrate_larvae_other_A,invertebrate_larvae_other_B,jellies_tentacles,polychaete,protist_dark_center,protist_fuzzy_olive,protist_noctiluca,protist_other,protist_star,pteropod_butterfly,pteropod_theco_dev_seq,pteropod_triangle,radiolarian_chain,radiolarian_colony,shrimp_caridean,shrimp_sergestidae,shrimp_zoea,shrimp-like_other,siphonophore_calycophoran_abylidae,siphonophore_calycophoran_rocketship_adult,siphonophore_calycophoran_rocketship_young,siphonophore_calycophoran_sphaeronectes_stem,siphonophore_calycophoran_sphaeronectes_young,siphonophore_calycophoran_sphaeronectes,siphonophore_other_parts,siphonophore_partial,siphonophore_physonect_young,siphonophore_physonect,stomatopod,tornaria_acorn_worm_larvae,trichodesmium_bowtie,trichodesmium_multiple,trichodesmium_puff,trichodesmium_tuft,trochophore_larvae,tunicate_doliolid_nurse,tunicate_doliolid,tunicate_partial,tunicate_salp_chains,tunicate_salp,unknown_blobs_and_smudges,unknown_sticks,unknown_unclassified".split(
',')
with open('namesClasses.dat', 'rb') as f:
namesClasses = cPickle.load(f)
labels = map(lambda s: s.split('\\')[-1], namesClasses)
for i in range(len(namesClasses)):
currentClass = namesClasses[i]
root_class.data_search(currentClass).id_list.append(i)
#get the total test images
#Full set
#fnames = glob.glob(os.path.join("competition_data", "test", "*.jpg"))
#Smaller set
fnames = glob.glob(os.path.join("smaller_comp", "test", "*.jpg"))
numberofTestImages = len(fnames)
X = np.zeros((numberofTestImages, num_features), dtype=float)
#Get filename separate from prefix path
images = map(lambda fileName: fileName.split('\\')[-1], fnames)
i = 0
# report progress for each 1% done
report = [int((j + 1) * numberofTestImages / 100.) for j in range(100)]
for fileName in fnames:
# Read in the images and create the features
image = imread(fileName, as_grey=True)
image = rotImage(image)
# Added from https://github.com/Newmu/Stroke-Prediction/blob/master/startPredictingGenCode.py
thresh = 0.9 * 255
# if image.min < 0.75*255:
# img = image < thresh
# else:
# img = image
# if img.sum() != 0:
# imgX,imgY = np.nonzero(img)
# imgW = imgX.max()-imgX.min()
# imgH = imgY.max()-imgY.min()
# if (imgW>1 and imgH>1):
# image = image[imgX.min():imgX.max(),imgY.min():imgY.max()]
# #----------------------------------
cvimage = cv2.imread(fileName)
features = getMinorMajorRatio(image)
#__Begin moved region
#From http://nbviewer.ipython.org/github/kqdtran/AY250-F13/blob/master/hw4/hw4.ipynb
pca = decomposition.PCA(n_components=25)
PCAFeatures = pca.fit_transform(image)
PCAevr = pca.explained_variance_ratio_
for evr in range(len(PCAevr)):
if np.isnan(PCAevr[evr]):
PCAevr[evr] = 0.0
#_____________________________________________________________
corners = getCorners(cvimage, features['orientation'])
horslice = image[:, maxPixel / 2]
vertslice = image[maxPixel / 2]
#correlation = signal.correlate(image,image)
#horcorrelation = signal.correlate(horslice,horslice)
#vertcorrelation = signal.correlate(vertslice,vertslice)
#crosscorrelation = signal.correlate(horslice,vertslice)
#correlation = correlation/correlation[correlation.shape[0]/2,correlation.shape[0]/2]
#horcorrelation = horcorrelation/horcorrelation[horcorrelation.shape[0]/2]
#vertcorrelation = vertcorrelation/vertcorrelation[vertcorrelation.shape[0]/2]
#crosscorrelation = crosscorrelation/crosscorrelation[horcorrelation.shape[0]/2]
hormean = np.mean(horslice)
horstd = np.std(horslice)
vertmean = np.mean(vertslice)
vertstd = np.std(vertslice)
horcount = vertcount = 0
for pix in horslice:
graycheck = False
if pix < thresh:
if not graycheck:
graycheck = True
horcount = horcount + 1
else:
graycheck = False
for pix in vertslice:
graycheck = False
if pix < thresh:
if not graycheck:
graycheck = True
vertcount = vertcount + 1
else:
graycheck = False
features['hsobel'] = np.nanmean(filter.hsobel(image))
features['vsobel'] = np.nanmean(filter.vsobel(image))
features['peaklocalmax'] = np.nanmean(peak_local_max(image))
features['felzen'] = np.nanmean(segmentation.felzenszwalb(image))
if np.isnan(features['peaklocalmax']):
features['peaklocalmax'] = 0.0
if np.isnan(features['felzen']):
features['felzen'] = 0.0
#hormirror = image[:maxPixel/2]-image[maxPixel-1:maxPixel/2:-1]
#vertmirror = image[:,:maxPixel/2]-image[:,maxPixel-1:maxPixel/2:-1]
image = resize(image, (maxPixel, maxPixel))
#From http://scikit-image.org/docs/dev/auto_examples/plot_local_binary_pattern.html
lbp = local_binary_pattern(image, n_points, radius, METHOD)
n_bins = lbp.max() + 1
lbpcounts = np.histogram(lbp, n_bins, normed=True,
range=(0, n_bins))[0]
#_____________________________________________________________
#__Moved region was here
#dist_trans = ndimage.distance_transform_edt(image[0])
# Store the rescaled image pixels and the axis ratio
#fd, hog_image = hog(image[0], orientations=8, pixels_per_cell=(2, 2),
# cells_per_block=(1, 1), visualise=True)
#X[i, 0:imageSize] = np.reshape(dist_trans, (1, imageSize))
#X[i, 0:imageSize] = np.reshape(hog_image, (1, imageSize))
# Store the rescaled image pixels and the axis ratio
X[i, 0:imageSize] = np.reshape(image, (1, imageSize))
#X[i, imageSize:imageSize+corr2dsize] = np.reshape(correlation, (1,corr2dsize))
#X[i, imageSize+corr2dsize:imageSize+corr2dsize+corrsize] = np.reshape(horcorrelation, (1,corrsize))
#X[i, imageSize+corr2dsize+corrsize:imageSize+corr2dsize+2*corrsize] = np.reshape(vertcorrelation, (1,corrsize))
#X[i, imageSize+corr2dsize+2*corrsize:imageSize+corr2dsize+3*corrsize] = np.reshape(crosscorrelation, (1,corrsize))
featcount = imageSize + 3 * corrsize + corr2dsize
for k, v in features.items():
try:
X[i, featcount:featcount + len(v)] = v
featcount = featcount + len(v)
except TypeError, te:
X[i, featcount] = v
featcount = featcount + 1
X[i, featcount:featcount + lbpcounts.size] = lbpcounts
X[i, featcount + lbpcounts.size:featcount + lbpcounts.size +
PCAsize] = PCAevr
X[i, featcount + lbpcounts.size + PCAsize] = np.mean(PCAFeatures)
i += 1
if i in report: print np.ceil(i * 100.0 / numberofTestImages), "% done"
X[i, imageSize + 5] = hu1
X[i, imageSize + 6] = hu2
X[i, imageSize + 7] = hu3
X[i, imageSize + 8] = hu12
X[i, imageSize + 9] = hu13
X[i, imageSize + 10] = hu23
X[i, imageSize + 11] = perimratio
X[i, imageSize + 12] = arearatio
X[i, imageSize + 13] = cornercenter
X[i, imageSize + 14] = cornercentercoords[0]
X[i, imageSize + 15] = cornercentercoords[1]
X[i, imageSize + 16] = cornerstd
X[i, imageSize + 17] = cornerstdcoords[0]
X[i, imageSize + 18] = cornerstdcoords[1]
X[i, imageSize + 19] = np.nanmean(filter.vsobel(image))
X[i, imageSize + 20] = np.nanmean(filter.hsobel(image))
X[i, imageSize + 21] = felzen
X[i, imageSize + 22] = peaklocalmax
# Store the classlabel
y[i] = label
i += 1
# report progress for each 5% done
report = [int((j + 1) * num_rows / 20.) for j in range(20)]
if i in report: print np.ceil(i * 100.0 / num_rows), "% done"
label += 1
# Loop through the classes two at a time and compare their distributions of the Width/Length Ratio
#Create a DataFrame object to make subsetting the data on the class
df = pd.DataFrame({"class": y[:], "ratio": X[:, num_features - 1]})
X[i, imageSize+3*corrsize+corr2dsize+5] = hu1 X[i, imageSize+3*corrsize+corr2dsize+6] = hu2 X[i, imageSize+3*corrsize+corr2dsize+7] = hu3 X[i, imageSize+3*corrsize+corr2dsize+8] = hu12 X[i, imageSize+3*corrsize+corr2dsize+9] = hu13 X[i, imageSize+3*corrsize+corr2dsize+10] = hu23 X[i, imageSize+3*corrsize+corr2dsize+11] = perimratio X[i, imageSize+3*corrsize+corr2dsize+12] = arearatio X[i, imageSize+3*corrsize+corr2dsize+13] = cornercenter X[i, imageSize+3*corrsize+corr2dsize+14] = cornercentercoords[0] X[i, imageSize+3*corrsize+corr2dsize+15] = cornercentercoords[1] X[i, imageSize+3*corrsize+corr2dsize+16] = cornerstd X[i, imageSize+3*corrsize+corr2dsize+17] = cornerstdcoords[0] X[i, imageSize+3*corrsize+corr2dsize+18] = cornerstdcoords[1] X[i, imageSize+3*corrsize+corr2dsize+19] = np.nanmean(filter.vsobel(image)) X[i, imageSize+3*corrsize+corr2dsize+20] = np.nanmean(filter.hsobel(image)) X[i, imageSize+3*corrsize+corr2dsize+21] = felzen X[i, imageSize+3*corrsize+corr2dsize+22] = peaklocalmax X[i, imageSize+3*corrsize+corr2dsize+23] = lrdiff X[i, imageSize+3*corrsize+corr2dsize+24] = tbdiff X[i, imageSize+3*corrsize+corr2dsize+25] = whu1 X[i, imageSize+3*corrsize+corr2dsize+26] = whu2 X[i, imageSize+3*corrsize+corr2dsize+27] = whu3 X[i, imageSize+3*corrsize+corr2dsize+28] = whu12 X[i, imageSize+3*corrsize+corr2dsize+29] = whu13 X[i, imageSize+3*corrsize+corr2dsize+30] = whu23 X[i, imageSize+3*corrsize+corr2dsize+31] = extent X[i, imageSize+3*corrsize+corr2dsize+32] = minintensity X[i, imageSize+3*corrsize+corr2dsize+33] = meanintensity X[i, imageSize+3*corrsize+corr2dsize+34] = maxintensity X[i, imageSize+3*corrsize+corr2dsize+35] = intensityratio1
def extract_features(im_cat_path_list):
'''
Extract Features takes a list of 2 element lists:
[catagory of image , path to image]
and extracts 15 features from this image. The output is a list of 3 element lists:
[catagory of image , path to image, [list of features]]
The 15 features are:
1 Product of image pixel dimensions (image size)
2 Mean of Grayscale Image
3 Area of Sobel Edges Normalized By Image Size
4 Area of Sobel Edges Above 2x Mean of Sobel Edges Normalized By Image Size
5 Area of Canny Edges (Sum of booleans) Normalized By Image Size
6 Number of Harris Corners
7 Unique Felzenszwalb Image Segmentation Lines
8 Area of Vertical Sobel Edges Above 2x Mean of Sobel Edges Normalized By Image Size
9 Area of Horizontal Sobel Edges Above 2x Mean of Sobel Edges Normalized By Image Size
10-12 Mean of Red/Green/Blue Channels (if grayscale: mean of the only color channel)
13 Maximum Pixel Value of the Histogram of Oriented Gradients
14 Percent of image that is light versus dark with adaptive thresholding
15-17 Percent of image that is red/green/blue with adaptive thresholding
'''
cat_path_features = []
for im_cat, im_path in im_cat_path_list:
#RAW IMAGE
im_raw = imread(im_path) #image matrix
#RAW IMAGE FLATTENED IF NOT ALREADY FLAT
if len(np.shape(im_raw)) == 3:
im_raw_flat = np.median(im_raw, axis=2)
else:
im_raw_flat = im_raw
#Size of image
im_y = float(im_raw.shape[0])
im_x = float(im_raw.shape[1])
im_size = im_y*im_x
#LIST OF FEATURES
features = []
#FEATURE 1: Product of image pixel dimensions (image size)
features.append(float(im_size))
#FEATURE 2: Mean of Grayscale Image
features.append(im_raw_flat.mean())
#FEATURE 3: Area of Sobel Edges Normalized By Image Size
im_edge_sobel = filter.sobel(im_raw_flat)
features.append(im_edge_sobel.sum()/im_size)
#FEATURE 4: Area of Sobel Edges Above 2x Mean of Sobel Edges Normalized By Image Size
features.append(float((im_edge_sobel > im_edge_sobel.mean()*2).sum())/im_size)
#FEATURE 5: Area of Canny Edges (Sum of booleans) Normalized By Image Size
im_canny = filter.canny(im_raw_flat, sigma=8)
features.append(im_canny.sum().astype(float)/im_size)
#FEATURE 6: Number of Harris Corners
im_corners = feature.corner_peaks(feature.corner_harris(im_raw_flat), min_distance=5)
features.append(float(len(im_corners)))
#FEATURE 7: Unique Felzenszwalb Image Segmentation Lines
im_raw_float = util.img_as_float(im_raw[::2, ::2])
im_felzen_segments = segmentation.felzenszwalb(im_raw_float, scale=100, sigma=0.5, min_size=50)
features.append(float(len(np.unique(im_felzen_segments))))
#FEATURE 8: Area of Vertical Sobel Edges Above 2x Mean of Sobel Edges Normalized By Image Size
im_edge_vsobel = filter.vsobel(im_raw_flat)
features.append(float((im_edge_vsobel > im_edge_vsobel.mean()*2).sum())/im_size)
#FEATURE 9: Area of Horizontal Sobel Edges Above 2x Mean of Sobel Edges Normalized By Image Size
im_edge_hsobel = filter.hsobel(im_raw_flat)
features.append(float((im_edge_hsobel > im_edge_hsobel.mean()*2).sum())/im_size)
#FEATURE 10-12: Mean of Red/Green/Blue Channels (if grayscale: mean of the only color channel)
if len(np.shape(im_raw)) == 3:
features.append(im_raw[...,0].mean())
features.append(im_raw[...,1].mean())
features.append(im_raw[...,2].mean())
else:
features.append(im_raw_flat.mean())
features.append(im_raw_flat.mean())
features.append(im_raw_flat.mean())
#FEATURE 13: Maximum Pixel Value of the Histogram of Oriented Gradients
im_fd, im_hog = feature.hog(im_raw_flat, orientations=8, pixels_per_cell=(16 , 16), cells_per_block=(1, 1), visualise=True)
features.append(im_hog.max())
#FEATURE 14: Percent of image that is light versus dark with adaptive thresholding
im_thres_flat = filter.threshold_adaptive(im_raw_flat, 100 , 'mean')
features.append(im_thres_flat.sum()/im_size)
#FEATURE 15-17: Percent of image that is red/green/blue with adaptive thresholding
im_thres_red = filter.threshold_adaptive(im_raw[...,0], 100 , 'mean')
im_thres_green = filter.threshold_adaptive(im_raw[...,1], 100 , 'mean')
im_thres_blue = filter.threshold_adaptive(im_raw[...,2], 100 , 'mean')
features.append(im_thres_red.sum()/im_size)
features.append(im_thres_green.sum()/im_size)
features.append(im_thres_blue.sum()/im_size)
#BUILD OUTPUT LIST FOR THIS IMAGE
cat_path_features.append([im_cat, im_path, features])
#CLEAR IMAGE PROC DATA
del im_raw
del im_raw_flat
del im_raw_float
del im_edge_sobel
del im_canny
del im_corners
del im_felzen_segments
del im_edge_vsobel
del im_edge_hsobel
del im_fd
del im_hog
return cat_path_features