def ton_and_color_corrections():
#色调和彩色校正
image=data.astronaut()
h1=color.rgb2hsv(image)
h2=h1.copy()
h1[:,:,1]=h1[:,:,1]*0.5
image1=color.hsv2rgb(h1)
h2[:,:,1]=h2[:,:,1]*0.5+0.5
image2=color.hsv2rgb(h2)
io.imshow(image)
io.imsave('astronaut.png',image)
io.imshow(image1)
io.imsave('astronautlight.png',image1)
io.imshow(image2)
io.imsave('astronautdark.png',image2)
imagered=image.copy()
imagered[:,:,0]=image[:,:,0]*127.0/255+128
io.imsave('astronautred.png',imagered)
imageblue=image.copy()
imageblue[:,:,2]=image[:,:,2]*127.0/255+128
io.imsave('astronautblue.png',imageblue)
imageyellow=image.copy()
imageyellow[:,:,0]=image[:,:,0]*127.0/255+128
imageyellow[:,:,1]=image[:,:,1]*127.0/255+128
io.imsave('astronautyellow.png',imageyellow)
io.imshow(imageyellow)
def classify_and_show(km, img_dir):
'''Use a km object to display both the original and altered version of the image at
img_dir, must be PNG'''
img = rgb2hsv(mpimg.imread(img_dir)[...,:3])
s = img.shape
data = img.reshape((s[0] * s[1], s[2]))
labels = km.classify(data)
plt.imshow(hsv2rgb(img))
plt.figure()
plt.imshow(hsv2rgb(km.means[labels].reshape(s)))
def gaborbank_orientation_vis(d, method='mean', legend=True):
""" Visualise the orientation for each point in the image.
Method 'mean' uses the mean resultant vector over
all orientation filters. Method 'max' takes the orientation
at each pixel to be that of the filter with maximum energy
at that pixel.
Adapted from http://nbviewer.ipython.org/github/gestaltrevision\
/python_for_visres/blob/master/Part7/Part7_Image_Statistics.ipynb
Args:
d: the dict output by gaborbank_convolve.
"""
res = d['res']
e = res.real**2 + res.imag**2 # energy
e = e.sum(axis=3).sum(axis=2) # sum energy over scales and orientations.
if method == 'mean':
ori = gaborbank_mean_orientation(d)
elif method == 'max':
ori = gaborbank_max_orientation(d)
else:
raise ValueError('Unknown method!')
# output values range 0--pi; adjust hues accordingly:
H = ori / np.pi
S = np.ones_like(H)
V = (e - e.min()) / e.max()
HSV = np.dstack((H, S, V))
RGB = hsv2rgb(HSV)
if legend is True:
# Render a hue circle
sz = int(e.shape[0] * 0.1)
r, a = pu.image.axes_polar(sz)
a[a < 0] += np.pi
a /= np.pi
# a = (a - a.min()) / a.max()
a = 1 - a # not sure why I have to flip this, but
# otherwise diagonals are reversed.
mask = (r < 0.9) & (r > 0.3)
hsv_legend = np.dstack((a,
np.ones_like(a, dtype='float'),
mask.astype('float')))
rgb_legend = hsv2rgb(hsv_legend)
RGB[:sz, :sz, :] = rgb_legend[::-1, ::]
return RGB
def get_merged_pic(self, nuclei_color = 0.66, foci_color = 0.33, seeds = False):
'''Return merged pic with foci and nuclei'''
x_max, y_max = self.nuclei.shape
active_cells = self.active_cells()
cell_number = len(active_cells)
if cell_number == 0:
return np.zeros((x_max, y_max, 3), dtype = np.uint8)
merged_pic_peaces = []
nuclei_rgb_koef = hsv2rgb(np.array([nuclei_color, 1., 1.]).reshape((1,1,3))).reshape(3)
foci_rgb_koef = hsv2rgb(np.array([foci_color, 1., 1.]).reshape((1,1,3))).reshape(3)
for cur_cell in active_cells:
if seeds:
foci = cur_cell.foci_seeds
else:
foci = cur_cell.foci_binary
nucleus_only = cur_cell.nucleus - foci
pic_foci_enhanced = 255 - np.floor((255 - cur_cell.pic_foci)*0.6)
pic_foci_enhanced = foci*pic_foci_enhanced
pic_nucleus_only = cur_cell.pic_nucleus*nucleus_only
pic_foci_enhanced_3d = np.dstack((pic_foci_enhanced, pic_foci_enhanced, pic_foci_enhanced))
pic_nucleus_only_3d = np.dstack((pic_nucleus_only, pic_nucleus_only, pic_nucleus_only))
pic_foci_rgb = pic_foci_enhanced_3d*foci_rgb_koef
pic_nucleus_only_rgb = pic_nucleus_only_3d*nuclei_rgb_koef
pic_merged_rgb = np.floor(pic_foci_rgb + pic_nucleus_only_rgb).astype(np.uint8)
merged_pic_peaces.append(peace(pic_merged_rgb, cur_cell.coords))
merged_pic = join_peaces_3d(merged_pic_peaces, x_max, y_max, dtype = np.uint8)
return merged_pic
def RDMcolormap(nCols=256):
# blue-cyan-gray-red-yellow with increasing V (BCGRYincV)
anchorCols = np.array([
[0, 0, 1],
[0, 1, 1],
[.5, .5, .5],
[1, 0, 0],
[1, 1, 0],
])
# skimage rgb2hsv is intended for 3d images (RGB)
# here we add a new axis to our 2d anchorCols to satisfy skimage, and then squeeze
anchorCols_hsv = rgb2hsv(anchorCols[np.newaxis, :]).squeeze()
incVweight = 1
anchorCols_hsv[:, 2] = (1-incVweight)*anchorCols_hsv[:, 2] + \
incVweight*np.linspace(0.5, 1, anchorCols.shape[0]).T
# anchorCols = brightness(anchorCols)
anchorCols = hsv2rgb(anchorCols_hsv[np.newaxis, :]).squeeze()
cols = colorScale(nCols, anchorCols)
return ListedColormap(cols)
def color_augment_image(data):
image = data.transpose(1, 2, 0)
hsv = color.rgb2hsv(image)
# Contrast 2
s_factor1 = numpy.random.uniform(0.25, 4)
s_factor2 = numpy.random.uniform(0.7, 1.4)
s_factor3 = numpy.random.uniform(-0.1, 0.1)
hsv[:, :, 1] = (hsv[:, :, 1] ** s_factor1) * s_factor2 + s_factor3
v_factor1 = numpy.random.uniform(0.25, 4)
v_factor2 = numpy.random.uniform(0.7, 1.4)
v_factor3 = numpy.random.uniform(-0.1, 0.1)
hsv[:, :, 2] = (hsv[:, :, 2] ** v_factor1) * v_factor2 + v_factor3
# Color
h_factor = numpy.random.uniform(-0.1, 0.1)
hsv[:, :, 0] = hsv[:, :, 0] + h_factor
hsv[hsv < 0] = 0.0
hsv[hsv > 1] = 1.0
rgb = color.hsv2rgb(hsv)
data_out = rgb.transpose(2, 0, 1)
return data_out
def _process(self, img):
hsv = rgb2hsv(img)
h = hsv[:, :, 0] + self._adjust
h[h > 1] -= 1
hsv[:, :, 0] = h
img[:, :, :] = hsv2rgb(hsv)
return img
def test_hsv2rgb_conversion(self):
rgb = self.img_rgb.astype("float32")[::16, ::16]
# create HSV image with colorsys
hsv = np.array([colorsys.rgb_to_hsv(pt[0], pt[1], pt[2]) for pt in rgb.reshape(-1, 3)]).reshape(rgb.shape)
# convert back to RGB and compare with original.
# relative precision for RGB -> HSV roundtrip is about 1e-6
assert_almost_equal(rgb, hsv2rgb(hsv), decimal=4)
def stretchImageHue(imrgb): # Image must be stored as 0-1 bound float. If it's 0-255 int, convert if( imrgb.max() > 1 ): imrgb = imrgb*1./255 # Transform to HSV imhsv = rgb2hsv(imrgb) # Find 2-98 percentiles of H histogram (except de-saturated pixels) plt.figure() plt.hist(imhsv[imhsv[:,:,1]>0.1,0].flatten(), bins=360) p2, p98 = np.percentile(imhsv[imhsv[:,:,1]>0.1,0], (2, 98)) print p2, p98 imhsv[:,:,0] = doStretch(imhsv[:,:,0], p2, p98, 0.6, 0.99) plt.figure() plt.hist(imhsv[imhsv[:,:,1]>0.1,0].flatten(), bins=360) imrgb_stretched = hsv2rgb(imhsv) plt.figure() plt.imshow(imrgb) plt.figure() plt.imshow(imrgb_stretched) plt.show()
def hsi_equalize_hist():
image=data.astronaut()
h=color.rgb2hsv(image)
h[:,:,2]=exposure.equalize_hist(h[:,:,2])
image_equal=color.hsv2rgb(h)
io.imshow(image_equal)
io.imsave('astronautequal.png',image_equal)
def saturate(im, amount=1.1): hsvim = skcolor.rgb2hsv(im) hue = np.take(hsvim, 0, axis=2) sat = np.take(hsvim, 1, axis=2) val = np.take(hsvim, 2, axis=2) # sat = sat * amount newhsv = np.dstack((hue, sat, val)) return skcolor.hsv2rgb(newhsv)
def main():
# read the images
image_from = io.imread(name_from) / 256
image_to = io.imread(name_to) / 256
# change to hsv domain (if requested)
if args.use_hsv:
image_from[:] = rgb2hsv(image_from)
image_to[:] = rgb2hsv(image_to)
# get shapes
shape_from = image_from.shape
shape_to = image_to.shape
# flatten
X_from = im2mat(image_from)
X_to = im2mat(image_to)
# number of pixes
n_pixels_from = X_from.shape[0]
n_pixels_to = X_to.shape[0]
# subsample
X_from_ss = X_from[np.random.randint(0, n_pixels_from-1, n_pixels),:]
X_to_ss = X_to[np.random.randint(0, n_pixels_to-1, n_pixels),:]
if save_col_distribution:
import matplotlib.pyplot as plt
import seaborn as sns
sns.set_style('white')
fig, axes = plt.subplots(nrows=2, figsize=(5, 10))
for ax, X in zip(axes, [X_from_ss, X_to_ss]):
ax.scatter(X[:,0], X[:,1], color=X)
if args.use_hsv:
ax.set_xhsvel('hue')
ax.set_yhsvel('value')
else:
ax.set_xhsvel('red')
ax.set_yhsvel('green')
axes[0].set_title('distr. from')
axes[1].set_title('distr. to')
fig.tight_layout()
fig.savefig('color_distributions.png')
# optimal tranportation
ot_color = OptimalTransport(X_to_ss, X_from_ss, lam=lam,
distance_metric=distance_metric)
# model transfer
transfer_model = KNeighborsRegressor(n_neighbors=n_neighbors)
transfer_model.fit(X_to_ss, n_pixels * ot_color.P @ X_from_ss)
X_transfered = transfer_model.predict(X_to)
image_transferd = minmax(mat2im(X_transfered, shape_to))
if args.use_hsv:
image_transferd[:] = hsv2rgb(image_transferd)
io.imsave(name_out, image_transferd)
def _rotate_Scale_fired(self):
"""" _rotate_Scale_fired(self): rotates scale and re-displays image when button is pressed """
max = 255. # this will only work with certain image types...
hsvimage = rgb2hsv([x/max for x in self.image])
hsvimage[:,:,1] = [np.mod(x+0.5,1) for x in hsvimage[:,:,1]]
hsvimage = [np.uint8(x*max) for x in hsv2rgb(hsvimage)]
self.image = hsvimage
self.ax.imshow(hsvimage)
self.figure.canvas.draw()
def colorize(image, hue, saturation=1):
""" Add color of the given hue to an RGB image.
By default, set the saturation to 1 so that the colors pop!
"""
hsv = color.rgb2hsv(image)
hsv[:, :, 1] = saturation
hsv[:, :, 0] = hue
return color.hsv2rgb(hsv)
def _rotate_Hue_fired(self):
"""" _rotate_Hue_fired(self): rotates hue and re-displays image when button is pressed """
max = 255.
hsvimage = rgb2hsv([x/max for x in self.image])
hsvimage[:,:,0] = [np.mod(x+0.5,1) for x in hsvimage[:,:,0]]
hsvimage = [np.uint8(x*max) for x in hsv2rgb(hsvimage)]
self.image = hsvimage
self.ax.imshow(hsvimage)
self.figure.canvas.draw()
def adjust_saturation_hue(image, saturation_factor, hue_delta):
"""Adjust saturation and hue of an RGB image.
Converts to HSV, add an offset to the saturation channel and
converts back to RGB.
"""
img_hsv = color.rgb2hsv(image)
# Adjust saturation and hue channels.
img_hsv[:, :, 1] = np.clip(img_hsv[:, :, 1] * np.abs(saturation_factor), 0.0, 1.0)
img_hsv[:, :, 0] = np.mod(img_hsv[:, :, 0] + 1. + hue_delta, 1.0)
# Back to RGB mod.
return color.hsv2rgb(img_hsv)
def generate(image, width=4096, height=4096): image_width, image_height, image_bands = image.shape left = center(image_width, width) top = center(image_height, height) result = np.zeros((width, height, image_bands), dtype=image.dtype) result[left:left+image_width, top:top+image_height] = image result = rgb2hsv(result) flattened = result.view(hsv_dtype) flattened.shape = -1 # Throws an exception if a copy must be made flattened[np.argsort(flattened, order=['H', 'S', 'V'])] = get_hsv_spectrum() return hsv2rgb(result).view(image.dtype)
def match(img, ref, bins):
hsv = color.rgb2hsv(img)
vals = hsv[:, :, 2].flatten()
vhist, vbins = numpy.histogram(vals, bins, density=True)
vcdf = vhist.cumsum()
a = numpy.interp(vals, vbins[:-1], vcdf)
mapped = ref(a).reshape(hsv[:, :, 2].shape)
hsv[:, :, 2] = mapped
ret_img = color.hsv2rgb(hsv)
return ret_img
def call(self, image, saliency_image):
img_resize = resize(image, saliency_image.shape)
saliency_range = max(0.15, saliency_image.max() - saliency_image.min())
saliency_norm = (saliency_image - saliency_image.min()) / saliency_range
saliency_gamma = adjust_gamma(saliency_norm, gamma=self.gamma)
cmap = matplotlib.cm.get_cmap('viridis')
cmap_hsv = rgb2hsv(cmap(saliency_gamma)[:, :, :3])
hsv = np.stack([
cmap_hsv[:, :, 0],
saliency_gamma,
img_resize
], axis=-1)
return hsv2rgb(hsv)
def make_hsv(magnitude, angle, img=None, alpha=0.5):
"""Convert the result of ``directionality_filter`` to an HSV
image, then convert to RGB."""
magnitude = scale(magnitude)
angle = scale(angle)
h = angle
s = magnitude
v = np.ones(h.shape)
hsv = hsv2rgb(np.dstack([h, s, v]))
if img is None:
return hsv
img = scale(crop_to(img, angle.shape))
result = hsv + gray2rgb(img)
return img_as_float(result)
def updateColorAll(self):
_Vs = list()
for signalType in self.signalColors.keys():
_v = self.signal[signalType][:,:,0]
_v[_v > 100] = 100
_Vs.append(_v/100)
V = sum(_Vs)
S = np.ones(V.shape)
S[V == 0] = 0
H = np.ones(V.shape)
H[V == 0] = 0.5
V = 1 - V
HSV = np.dstack((H, S, V))
self.color = 255*hsv2rgb(HSV).astype(np.uint8, copy=False)
def makeVideoBar(indir='/Volumes/Documents/colbrydi/Documents/DirksWork/chamview/ChamB/'):
sz=''
br=''
for root, dirs, filenames in os.walk(indir):
filenames.sort()
for f in filenames:
if fnmatch.fnmatch(f,'0*.jpeg'):
(ib, sz) = HueBar(os.path.join(root,f),sz);
if br == '':
br = ib
else:
br = np.append(br, ib, axis=1)
bar = color.hsv2rgb(br)
bar = bar * 255
return bar
def embed(self, img, payload, k):
if len(payload) > self.max_payload:
raise ValueError("payload too long")
payload = bytearray(payload) + "\x00" * (self.max_payload - len(payload))
encoded = self.rscodec.encode(payload)
if len(img.shape) == 2:
return self._embed(img, encoded, k)
elif len(img.shape) == 3:
hsv = rgb2hsv(img)
i = 0
hsv[:,:,i] = self._embed(hsv[:,:,i], encoded, k)
return hsv2rgb(hsv)
else:
raise TypeError("img must be a 2d or 3d array")
def preprocess_img(img):
IMG_SIZE = 32
# Histogram normalization in v channel
hsv = color.rgb2hsv(img)
hsv[:, :, 2] = exposure.equalize_hist(hsv[:, :, 2])
img = color.hsv2rgb(hsv)
# central square crop
min_side = min(img.shape[:-1])
centre = img.shape[0]//2, img.shape[1]//2
img = img[centre[0]-min_side//2:centre[0]+min_side//2,
centre[1]-min_side//2:centre[1]+min_side//2,
:]
# rescale to standard size
img = transform.resize(img, (IMG_SIZE, IMG_SIZE))
return img
def hsv_value(image_filter, image, *args, **kwargs):
"""Return color image by applying `image_filter` on HSV-value of `image`.
Note that this function is intended for use with `adapt_rgb`.
Parameters
----------
image_filter : function
Function that filters a gray-scale image.
image : array
Input image. Note that RGBA images are treated as RGB.
"""
# Slice the first three channels so that we remove any alpha channels.
hsv = color.rgb2hsv(image[:, :, :3])
value = hsv[:, :, 2].copy()
value = image_filter(value, *args, **kwargs)
hsv[:, :, 2] = convert(value, hsv.dtype)
return color.hsv2rgb(hsv)
def create_image(model, params, size):
x_dim, y_dim = size
X = np.concatenate(np.array(params), axis=1)
pred = model.predict(X)
img = []
channels = pred.shape[1]
for channel in range(channels):
yp = pred[:, channel]
yp = (yp - yp.min()) / (yp.max()-yp.min())
img.append(yp.reshape(y_dim, x_dim))
img = np.dstack(img)
if channels == 3: img = hsv2rgb(img)
img = (img * 255).astype(np.uint8)
return img
def masked(img, gt, mask, alpha=1):
"""Returns image with GT lung field outlined with red, predicted lung field
filled with blue."""
rows, cols = img.shape
color_mask = np.zeros((rows, cols, 3))
boundary = morphology.dilation(gt, morphology.disk(3)) - gt
color_mask[mask == 1] = [0, 0, 1]
color_mask[boundary == 1] = [1, 0, 0]
img_color = np.dstack((img, img, img))
img_hsv = color.rgb2hsv(img_color)
color_mask_hsv = color.rgb2hsv(color_mask)
img_hsv[..., 0] = color_mask_hsv[..., 0]
img_hsv[..., 1] = color_mask_hsv[..., 1] * alpha
img_masked = color.hsv2rgb(img_hsv)
return img_masked
def array2img(array, colorspace, classification=False):
"""
INPUT:
array: Lab layers in an array of shape=(3, image_x, image_y)
colorspace: the colorspace that the array is in;
''CIELab' for CIELab colorspace
'CIEL*a*b*' for the mapped CIELab colorspace (by function remap_CIELab in NNPreprocessor)
'RGB' for rgb mapped between [0 and 1]
'YCbCr' for YCbCr
'HSV' for HSV
OUTPUT:
Image
"""
# Check if colorspace is properly defined
assert_colorspace(colorspace)
# Convert the image to shape=(image_x, image_y, 3)
image = np.transpose(array,[1,2,0]).astype('float64')
if (colorspace == 'CIEL*a*b*'):
# Convert to CIELab:
image = unmap_CIELab(image)
if classification == True:
# remap L layer
image[:,:,0] *= 1
if ( (colorspace == 'CIELab') or (colorspace == 'CIEL*a*b*') ):
# Convert to rgb:
image = color.lab2rgb(image)
if (colorspace == 'HSV'):
image = color.hsv2rgb(image)
# YCbCr is supported by the PIL Image pkg. so just change the mode that is passed
if colorspace == 'YCbCr':
# Convert to a PIL Image (Should be replaced by the scikit-image equivalent color.rgb2YCbCr in the future)
im = Image.fromarray( np.uint8(image*255.), mode='YCbCr' )
im = im.convert('RGB')
# Put back into a numpy array
image = np.array(im)/255.
# Now the image is definitely in a supported colorspace
return Image.fromarray(np.uint8(image*255.),mode='RGB')
def circle_colony(colony_geometry, colony_score, image_shape, score_min=0.0, score_max=255.0):
# color of circle
hue_start, hue_stop = [0.0, 1/3.0] # from 0 to 120 degrees hue colors
#score_min, score_max = [1.0, 3.0] # colony scores from 1 till 3
color_coeff = ((hue_stop - hue_start)/(score_max - score_min))*(colony_score - score_min)
# create three layers of hsv image (black background) and binary mask
x_max = image_shape[0]
y_max = image_shape[1]
mask_saturation = np.array([[0.0 for i in range(y_max)] for i in range(x_max)])
mask_light = np.array([[0.0 for i in range(y_max)] for i in range(x_max)])
mask_hue = np.array([[0.0 for i in range(y_max)] for i in range(x_max)])
mask_binary = np.array([[False for i in range(y_max)] for i in range(x_max)])
# find circle
x, y, r = colony_geometry
r = r*2
rr, cc = circle_perimeter(x, y, np.round(r).astype(int))
rr_new, cc_new = [], []
for x_c,y_c in zip(rr,cc):
if (x_c >= 0) and (x_c < x_max) and (y_c >= 0) and (y_c < y_max):
rr_new.append(x_c)
cc_new.append(y_c)
# create binary mask with circle
mask_binary[rr_new, cc_new] = True
struct = generate_binary_structure(2, 1)
for i in range(8):
mask_binary = binary_dilation(mask_binary, struct)
# paint circle in hsv layers
mask_saturation[mask_binary] = 1.0
mask_light[mask_binary] = 0.75
mask_hue[mask_binary] = color_coeff
hsv_image = np.dstack([mask_hue, mask_saturation, mask_light])
rgb_image = img_as_ubyte(hsv2rgb(hsv_image))
return rgb_image, mask_binary
def preprocess_img(img):
# Histogram normalization in y
hsv = color.rgb2hsv(img)
hsv[:,:,2] = exposure.equalize_hist(hsv[:,:,2])
img = color.hsv2rgb(hsv)
# central scrop
min_side = min(img.shape[:-1])
centre = img.shape[0]//2, img.shape[1]//2
img = img[centre[0]-min_side//2:centre[0]+min_side//2,
centre[1]-min_side//2:centre[1]+min_side//2,
:]
# rescale to standard size
img = transform.resize(img, (IMG_SIZE, IMG_SIZE))
# roll color axis to axis 0
img = np.rollaxis(img,-1)
return img
def get_preprocessed_state(self):
frame = self.get_state()
if frame is None:
return self.terminal
# Blur, crop, resize
frame = cv2.GaussianBlur(frame, (39, 39), 0, 0)
frame = tf.image.central_crop(frame, 0.5)
frame = tf.image.resize(
frame,
self.resolution,
align_corners=True,
method=tf.image.ResizeMethod.NEAREST_NEIGHBOR).numpy()
# Kmeans clustering
frame = rgb2hsv(frame)
kmeans = KMeans(n_clusters=4).fit(frame.reshape((-1, 3)))
frame = kmeans.cluster_centers_[kmeans.labels_].reshape(frame.shape)
frame = hsv2rgb(frame).astype(np.float32)
# Greyscale
if self.num_channels == 1:
frame = tf.image.rgb_to_grayscale(frame).numpy()
#plt.imshow(frame.reshape((frame.shape[0], frame.shape[1])), cmap="gray")
#plt.show()
return frame
def BeautifyLips(MouthImage, Choice):
'''
return the mouth image after processed with regard to color choice
This will be implemented after Yi Liu implemented his matlab code
and i will transform the code to python as soon as he finished that
'''
hsv_image = color.rgb2hsv(MouthImage)
ratio = 0.25
hsv_image[:, :, 0] = (1 - ratio) * hsv_image[:, :, 0] + ratio * (
cmap[Choice][0] - hsv_image[:, :, 0])
hsv_image[:, :, 1] = (1 - ratio) * hsv_image[:, :, 1] + ratio * (
cmap[Choice][1] - hsv_image[:, :, 1])
hsv_image[:, :, 2] = (1 - ratio) * hsv_image[:, :, 2] + ratio * (
cmap[Choice][2] - hsv_image[:, :, 2])
Mouth = color.hsv2rgb(hsv_image)
#print(Mouth);
print('dalong log : after beautify')
#cv2.imwrite('/home/yuanxl/after_beautify.jpg',255*hsv_image[:,:,::-1])
return np.array(Mouth * 255, dtype=np.uint8)
def RGB2HSV_shift_LAB(I,shift): # shift value0 ~1
# Get Original L in LAB, shift H in HSV
# Get Original LAB
lab_original = color.rgb2lab(I)
l_original = (lab_original[:, :, 0] / 100.0)
# Shift HSV
hsv = color.rgb2hsv(I)
h = ((hsv[:, :, 0] + shift))
s = (hsv[:, :, 1])
v = (hsv[:, :, 2])
hsv2 = color.hsv2rgb(np.dstack([h, s, v]).astype(np.float64))
# Merge (Original LAB, Shifted HSV)
lab = color.rgb2lab(hsv2)
l = l_original
#l = (lab[:, :, 0] / 100.0)
a = (lab[:, :, 1] + 86.1830297444) / (98.2330538631 + 86.1830297444) #* 255.0 # a component ranges from -127 to 127
b = (lab[:, :, 2] + 107.857300207) / (94.4781222765 + 107.857300207) #* 255.0 # b component ranges from -127 to 127
return np.dstack([l, a, b])
def get_raw_data(f, seg, saturation_factor):
serum = quantile_normalize(read_image(f, 'serum_IgG'))
marker = quantile_normalize(read_image(f, 'marker'))
nuclei = quantile_normalize(read_image(f, 'nuclei'))
bg_mask = seg == 0
def subtract_bg(im):
bg = np.median(im[bg_mask])
im -= bg
return im
serum = subtract_bg(serum)
marker = subtract_bg(marker)
nuclei = subtract_bg(nuclei)
raw = np.concatenate([marker[..., None], serum[..., None], nuclei[..., None]], axis=-1)
if saturation_factor > 1:
raw = skc.rgb2hsv(raw)
raw[..., 1] *= saturation_factor
raw = skc.hsv2rgb(raw).clip(0, 1)
return raw, marker
def transform(X: np.ndarray):
""" Randomly manipulate Hue(H), Saturation(S), Luminescence(L) of image. See change above to check how individually them impact the image
"""
# pdb.set_trace()
X = X.astype(np.float32)
hue = np.random.uniform(-0.2, 0.6, 1).round(2)[0]
saturation = np.random.uniform(-0.2, 0.2, 1).round(2)[0]
luminescence = np.random.randint(-20, 90)
# print('hue = {}, saturation= {}, luminescence = '.format(hue , saturation, luminescence))
hsv = color.rgb2hsv(X)
hsv[:, :, 0] += hue
hsv[:, :, 1] += saturation
hsv[:, :, 2] += luminescence
hsv[:, :, 0:2] = hsv[:, :, 0:2].clip(min=0, max=1)
hsv[:, :, 2] = hsv[:, :, 2].clip(min=60, max=250)
X = color.hsv2rgb(hsv)
X = X.astype(np.uint8)
return X
def __call__(self, image):
power = np.random.uniform(self.power[0], self.power[1])
factor = np.random.uniform(self.factor[0], self.factor[1])
addition = np.random.uniform(self.addition[0], self.addition[1])
im_np = np.asarray(image)
#im_np = im_np.transpose(1,0,2) # when we go from pil to numpy array the W and L dimensions are swaped
im_hsv = color.rgb2hsv(im_np)
im_hsv_t_h = [x for x in im_hsv[:, :, 0]]
im_hsv_t_s = [(((x**power) * factor) + addition)
for x in im_hsv[:, :, 1]]
im_hsv_t_v = [(((x**power) * factor) + addition)
for x in im_hsv[:, :, 2]]
im_hsv_t = np.stack((im_hsv_t_h, im_hsv_t_s, im_hsv_t_v), axis=2)
im_rgb = color.hsv2rgb(im_hsv_t)
# Postprocessing to avoid saturating pixels
im_end = im_rgb - im_rgb.min()
im_end = im_end / im_end.max()
return Image.fromarray(np.uint8(im_end * 255))
def overlay_image_with_labels(self, image_data, predicted_labels):
color_mask = np.zeros(
(predicted_labels.shape[0], predicted_labels.shape[1], 3))
color_mask[np.where(predicted_labels == 0)] = [128, 128, 128]
color_mask[np.where(predicted_labels == 1)] = [128, 0, 0]
color_mask[np.where(predicted_labels == 2)] = [128, 128, 0]
color_mask[np.where(predicted_labels == 3)] = [0, 0, 192]
color_mask[np.where(predicted_labels == 4)] = [128, 64, 128]
color_mask[np.where(predicted_labels == 5)] = [64, 0, 128]
alpha = 0.8
# Convert the input image and color mask to Hue Saturation Value (HSV)
# colorspace
img_hsv = color.rgb2hsv(image_data)
color_mask_hsv = color.rgb2hsv(color_mask)
# Replace the hue and saturation of the original image
# with that of the color mask
img_hsv[..., 0] = color_mask_hsv[..., 0]
img_hsv[..., 1] = color_mask_hsv[..., 1] * alpha
img_masked = color.hsv2rgb(img_hsv)
return img_masked
def sobel_edge_detection(img, size):
'''Performs Sobel Edge detection and returns the magnitude and direction of gradients.'''
# Gaussian Smooth
image = gaussian_smooth(img, size, 5)
# Sobel Operator
gx, gy = sobel_operator(image)
# Magnitude
mag = np.sqrt(gx**2 + gy**2)
# Direction
m, n = image.shape
hsv = np.zeros((m, n, 3))
hsv[:, :,
0] = (np.arctan2(gy, gx) + np.pi) / (2 * np.pi) # Hue: Gradient angle
hsv[:, :, 1] = np.ones((m, n)) # Saturation: Maximum, which is 1
hsv[:, :, 2] = (mag - mag.min()) / (mag.max() - mag.min()
) # Value: Gradient magnitude
dir_ = color.hsv2rgb(hsv)
return [mag, dir_]
def hsv_decomposition(image, channel='H'):
"""Decomposes the image into HSV and only returns the channel specified.
Args:
image: numpy array of shape(image_height, image_width, 3).
channel: str specifying the channel. Can be either "H", "S" or "V".
Returns:
out: numpy array of shape(image_height, image_width).
"""
hsv = color.rgb2hsv(image)
out = None
### YOUR CODE HERE
lis = ['H', 'S', 'V']
dic = {x: i for i, x in enumerate(lis)}
mat = np.eye(3)
onl = hsv * mat[dic[channel]]
out = color.hsv2rgb(onl)
### END YOUR CODE
return out
def create_chromakey_image(img_read, green_low_factor, green_high_factor,
saturation, brightness):
hsv = rgb2hsv(img_read)
for pixel_row in hsv:
for pixel_col in pixel_row:
if green_low_factor < pixel_col[0] <= green_high_factor \
and pixel_col[1] > saturation \
and pixel_col[2] > brightness:
pixel_col[0] = 0
pixel_col[1] = 0
pixel_col[2] = 0
else:
pixel_col[0] = 1
pixel_col[1] = 1
pixel_col[2] = 1
img_ndarray = np.array(hsv)
img_rgb = hsv2rgb(img_ndarray)
img_gray = rgb2gray(img_rgb)
img_gray_copy = np.copy(img_gray)
x, y = img_gray_copy.shape
for i in range(0, x):
for j in range(0, y):
pixel = img_gray_copy[i, j]
if pixel > 0:
img_gray_copy[i, j] = 1
else:
img_gray_copy[i, j] = 0
return img_gray_copy
def reconstruction(imagePath, outputPath):
image = io.imread(imagePath)
if (isGrayscale(imagePath)):
image = gray2rgb(image)
elif (isNotHsv(imagePath)):
image = rgb2hsv(image)
img_hsv = image
img_hsv_copy = np.copy(img_hsv)
imgSize = (image.shape)
w = min(imgSize[0], 200)
h = min(imgSize[1], 160)
# flood function returns a mask of flooded pixels
mask = flood(img_hsv[..., 0], (w - 1, h - 1), tolerance=0.016)
# Set pixels of mask to new value for hue channel
img_hsv[mask, 0] = 0.5
# Post-processing in order to improve the result
# Remove white pixels from flag, using saturation channel
mask_postprocessed = np.logical_and(mask, img_hsv_copy[..., 1] > 0.4)
# Remove thin structures with binary opening
# mask_postprocessed = binary_opening(mask_postprocessed,
# np.ones((3, 3)))
# Fill small holes with binary closing
mask_postprocessed = binary_opening(mask_postprocessed, disk(20))
img_hsv_copy[mask_postprocessed, 0] = 0.5
# img_hsv_copy_uint8 = img_as_ubyte(img_hsv_copy)
#is_hsv
if (len(img_hsv_copy.shape) == 3 and img_hsv_copy.shape[2] == 3):
output = hsv2rgb(img_hsv_copy)
else:
output = img_hsv_copy
output_uint8 = img_as_ubyte(output)
imsave('' + outputPath, output_uint8)
# imsave('/home/marekk/workspace/tmp/1234.png', output_uint8)
output_uint8
def convert_warp_to_color(uu: np.ndarray,
vv: np.ndarray,
min_vel_mag: float = MIN_VEL_MAG,
max_vel_mag: float = MAX_VEL_MAG) -> np.ndarray:
""" Convert a warp matrix into a color matrix
:param ndarray uu:
The x-vectors for the flow image
:param ndarray vv:
The y-vectors for the flow image
:param float min_vel_mag:
Minimum vector magnitude to plot
:param float max_vel_mag:
Maximum vector magnitude to plot
:returns:
A color image where angle corresponds to hue and magnitude corresponds
to value on the HSV color scale
"""
# Convert to hsv
mag = np.sqrt(uu**2 + vv**2)
ang = np.arctan2(vv, uu)
hue = (ang % np.pi) / np.pi # hue between 0-1
sat = np.ones_like(hue)
# If no bounds passed, clamp by percentile
if min_vel_mag is None:
min_vel_mag = np.percentile(mag, 5)
if max_vel_mag is None:
max_vel_mag = np.percentile(mag, 95)
# Rescale magnitude to between 0 and 1
value = (mag - min_vel_mag) / (max_vel_mag - min_vel_mag)
value[value < 0] = 0
value[value > 1] = 1
return hsv2rgb(np.stack((hue, sat, value), axis=2))
def deprecated_overlayImg_(img,
mask,
print_color=[5, 119, 72],
linewidth=1,
alpha=0.618):
#img = img_as_float(data.camera())
rows, cols = img.shape[0:2]
# Construct a colour image to superimpose
color_mask = np.zeros((rows, cols, 3))
color_mask[mask == 1] = print_color
color_mask[mask == 0] = img[mask == 0]
# imshow(color_mask)
if len(img.shape) == 2:
img_color = np.dstack((img, img, img))
else:
img_color = img
img_hsv = color.rgb2hsv(img_color)
color_mask_hsv = color.rgb2hsv(color_mask)
img_hsv[..., 0] = color_mask_hsv[..., 0]
img_hsv[..., 1] = color_mask_hsv[..., 1] * alpha
img_masked = color.hsv2rgb(img_hsv)
# Display the output
#f, (ax0, ax1, ax2) = plt.subplots(1, 3,
# subplot_kw={'xticks': [], 'yticks': []})
#ax0.imshow(img, cmap=plt.cm.gray)
#ax1.imshow(color_mask)
#ax2.imshow(img_masked)
#plt.show()
img_masked = np.asarray((img_masked / np.max(img_masked)) * 255,
dtype=np.uint8)
return img_masked
def alter_color(img, predominant_color, min_color, max_color, method=1):
""" alters the color of the image to approximate it to the predominant_color parameter
inputs:
img: downscaled tile image.
predominant_color: the predominant color of a given section of the canvas image (obtained using the mean).
min_color: the minimum color value of a given section of the canvas image.
max_color: the maximum color value of a given section of the canvas image.
method: which color representation method to use in the color processing
1 - RGB: applies a color balancing in the rgb model, so the tile image color moves to the desired color
2 - HSV: use the hue and saturation of the desired color to replace these values in the tile image
In both methods, a normalization using the max and min intensities of the original image is required,
so it can accurately represent the color intensity of the original image.
"""
out_img = np.array(img, copy=True)
if method == 1:
# changes the color channels proportions to the predominant color proportion
out_img[:, :, 0] = out_img[:, :, 0] * predominant_color[0] # R
out_img[:, :, 1] = out_img[:, :, 1] * predominant_color[1] # G
out_img[:, :, 2] = out_img[:, :, 2] * predominant_color[2] # B
elif method == 2:
# transforms to the HSV color representation
out_img = rgb2hsv(out_img)
predominant_color = rgb2hsv(np.reshape(predominant_color, (1, 1, 3)))
out_img[:, :, 0] = (predominant_color[:, :, 0]) # alters hue
out_img[:, :, 1] = (predominant_color[:, :, 1]) # alters saturation
out_img = hsv2rgb(out_img)
# returns to original color interval in rgb to accurately represent the intensity (black/white)
out_img = normalize_image(out_img, new_min=min_color, new_max=max_color)
return np.clip(out_img, 0, 255)
def gen_bra(self, image):
pantie = np.array(image)
# pickup colors
front = pantie[20:100, 30:80, :3] / 255.0
front_shade = pantie[130:150, 0:40, :3] / 255.0
front_color = self.pick_color(front)
front_shade_color = self.pick_color(front_shade)
front_shade_color = rgb2hsv(front_shade_color[None, None])
front_shade_color[0, 0, 1] *= front_shade_color[0, 0, 2] / 0.3
if front_shade_color[0, 0, 1] > 0.7:
front_shade_color[0, 0, 1] *= 0.7
front_shade_color[0, 0, 2] *= front_shade_color[0, 0, 2] / 0.4
front_shade_color = np.clip(hsv2rgb(front_shade_color)[0, 0], 0, 1)
ribbon = pantie[24:32, 15:27, :3] / 255.0
ribbon_color = self.pick_color(ribbon)
# making a center texture
center = pantie[20:170, -200:-15, :3][:, ::-1]
center = resize(center, [2.42, 2.42])
bra_center = np.copy(self.bra_center)
bra_center[225:225 + center.shape[0],
25:25 + center.shape[1], :3] = center * np.float32(
bra_center[225:225 + center.shape[0],
25:25 + center.shape[1], :3] > 0)
bra = self.bra[:, :, :3] * front_color
bra_shade = (self.bra_shade[:, :, -1])[:, :, None] * front_shade_color
bra_component = self.bra_component[:, :, :3] * ribbon_color
# overlaying layers
bra = alpha_brend(bra_center[:, :, :3], bra[:, :, :3],
bra_center[:, :, 0] > 0.1)
bra = alpha_brend(bra_component, bra, self.bra_component_mask)
bra = alpha_brend(bra_shade, bra, self.bra_shade_alpha)
bra = np.dstack((bra, self.bra[:, :, 0] > 0.8))
return Image.fromarray(np.uint8(np.clip(bra, 0, 1) * 255))
def quantizeImage(model, rgb_image, foreground_label_image):
# Convert image to HSV space because cluster centers are in HSV space
hsv_image = color.rgb2hsv(rgb_image)
# Make a background mask
in_foreground = foreground_label_image != 0
foreground_pixels = hsv_image[in_foreground, :]
# Assign each foreground pixel to the nearest cluster center
cluster_labels = model.predict(foreground_pixels)
quantized_hsv = model.cluster_centers_[cluster_labels]
# Make an image displaying the segmentation
cluster_label_image = np.zeros(foreground_label_image.shape, dtype=int)
cluster_label_image[in_foreground] = cluster_labels + 1
# Make a quantized image
quantized_hsv_image = np.zeros_like(hsv_image)
quantized_hsv_image[in_foreground] = quantized_hsv
# Convert from HSV to RGB
quantized_rgb_image = color.hsv2rgb(quantized_hsv_image)
return quantized_rgb_image, cluster_label_image
def overlay_grey_and_color(Xgrey, Xcolor):
""""
overlay a grayscale image with some color (e.g. segmentation masks)
note that this IGNORES the magnitude of the color image, essentially it becomes a MASK
:param Xgrey:
:param Xcolor:
:return:
"""
from skimage import color, io, img_as_float
import numpy as np
import matplotlib.pyplot as plt
alpha = 0.6
img = img_as_float(Xgrey)
rows, cols = img.shape
color_mask = Xcolor
# Construct RGB version of grey-level image
img_color = np.dstack((img, img, img))
# Convert the input image and color mask to Hue Saturation Value (HSV)
# colorspace
img_hsv = color.rgb2hsv(img_color)
color_mask_hsv = color.rgb2hsv(color_mask)
# Replace the hue and saturation of the original image
# with that of the color mask
img_hsv[..., 0] = color_mask_hsv[..., 0]
img_hsv[..., 1] = color_mask_hsv[..., 1] * alpha
img_masked = color.hsv2rgb(img_hsv)
return img_masked
def overlay_on_gray(oimg,
overlayGREEN=None,
overlayRED=None,
overlayBLUE=None,
alpha=0.6):
rows, cols = oimg.shape
color_mask = np.zeros((rows, cols, 3))
if overlayGREEN is not None:
color_mask[:, :, 1] = overlayGREEN
if overlayRED is not None:
color_mask[:, :, 2] = overlayRED
if overlayBLUE is not None:
color_mask[:, :, 0] = overlayBLUE
img_color = np.dstack((oimg, oimg, oimg))
img_hsv = color.rgb2hsv(img_color)
color_mask_hsv = color.rgb2hsv(color_mask)
img_hsv[..., 0] = color_mask_hsv[..., 0]
img_hsv[..., 1] = color_mask_hsv[..., 1] * alpha
img_masked = color.hsv2rgb(img_hsv)
return img_masked, color_mask
def match_background_hsv(foreground_img, background_img):
"""use hsv histogram matching to match the foreground contrast more closely to background
:param foreground_img: ndimage 4 channel array
:param background_img: ndimage
:return: PIL RGBA image
"""
foreground_img = as_ndarray(foreground_img)
background_img = as_ndarray(background_img)
n_bins = 255
foreground_img_hsv = color.rgb2hsv(foreground_img[:, :, :3])
background_img_hsv = color.rgb2hsv(background_img[:, :, :3])
foreground_alpha = np.expand_dims(foreground_img[:, :, 3], 2)
foreground_img_hsva = np.concatenate(
(foreground_img_hsv, foreground_alpha), axis=2)
matched = match_channels(foreground_img_hsva, background_img_hsv, n_bins)
matched_rgb = color.hsv2rgb(matched[:, :, :3]) * 255
matched[:, :, :3] = matched_rgb[:, :, :3]
matched = Image.fromarray(matched.astype('uint8'))
return matched
def create_image(model, x, y, r, z):
'''
create an image for the given latent vector z
'''
# create input vector
Z = np.repeat(z, x.shape[0]).reshape((-1,x.shape[0]))
X = np.concatenate([x, y, r, Z.T], axis=1)
pred = model.predict(X)
img = []
for k in range(pred.shape[1]):
yp = pred[:, k]
# if k == pred.shape[1]-1:
# yp = np.sin(yp)
yp = (yp - yp.min()) / (yp.max()-yp.min())
img.append(yp.reshape(y_dim, x_dim))
img = np.dstack(img)
if img.shape[-1] == 3:
from skimage.color import hsv2rgb
img = hsv2rgb(img)
return (img*255).astype(np.uint8)
def hsv_method(img, n_clusters, n_colors):
img = rgb2hsv(img)
# make img array for kmeans
X = img.reshape(-1, 3)
n_clusters = n_clusters if n_clusters > n_colors else n_colors
km = MiniBatchKMeans(n_clusters=n_clusters,
init='k-means++',
n_init=3,
random_state=0)
labels = km.fit_predict(X)
cluster_centers = km.cluster_centers_
bincount = np.bincount(labels)
# find index of data point closest to each center
closest, _ = pairwise_distances_argmin_min(cluster_centers, X)
# sort colors by frequency
dominants = 1. * X[closest][np.argsort(bincount,
axis=0)[::-1]][:n_colors]
colors = [hsv2rgb([[x]])[0][0] for x in dominants]
return colors
def Filter_syn(I, mode, w, x, y, z, start_time):
# Preprocessing
this_time = time.time()
elapsed_time = this_time - start_time
if mode == "random":
w = ((elapsed_time * 1000) % 1000) * (1.5) / 1000 # above 1
if w > 1.3:
w = 1.3
if w < 0.4:
w = 0.4
# Change Saturation
hsv = color.rgb2hsv(I)
h = (hsv[:, :, 0] / 360.0) * 255.0
s = (hsv[:, :, 1] / 100.0) * 255.0 * (w)
v = (hsv[:, :, 2] / 100.0) * 255.0
r = (h / 255.0) * 360.0
g = (s / 255.0) * 100.0
b = (v / 255.0) * 100.0
rgb = color.hsv2rgb(numpy.dstack([r, g, b]).astype(numpy.float64))
# Change Color
rgb[:, :, 0] = rgb[:, :, 0] * x
rgb[:, :, 1] = rgb[:, :, 1] * y
rgb[:, :, 2] = rgb[:, :, 2] * z
# Post-processing
rgb[:, :, 0] = numpy.clip(rgb[:, :, 0], 0, 1)
rgb[:, :, 1] = numpy.clip(rgb[:, :, 1], 0, 1)
rgb[:, :, 2] = numpy.clip(rgb[:, :, 2], 0, 1)
return rgb
def satura(s, i):
img = io.imread(i)
hsv = color.rgb2hsv(img)
rows, cols, dim = hsv.shape
if s > 1:
s = 1
if s < 0:
s = 0
for i in range(rows):
for j in range(cols):
if s >= 0.5:
hsv[i, j, 1] += (1 - hsv[i, j, 1]) * (s - 0.5) / 0.5
else:
hsv[i, j, 1] = hsv[i, j, 1] * s / 0.5
new_img = color.hsv2rgb(hsv)
plt.figure(5)
plt.title('a' + '')
plt.imshow(img)
plt.axis('off')
plt.figure(6)
plt.title('b Change the satura of "a" to %f' % s)
plt.imshow(new_img)
plt.axis('off')
def add_class_colour(self, image, pred, label):
# load the image and add a mask in the colour of the corresponding predicted label, with transparency alpha.
# then resize the image to make it an appropriate plot point size and append it to the image list.
img = img_as_float(image)[0]
rows, cols = img.shape
color_mask = np.zeros((rows, cols, 3))
color_mask[0:rows, 0:cols] = Visualization.colours[pred]
img_color = np.dstack((img, img, img))
img_hsv = color.rgb2hsv(img_color)
color_mask_hsv = color.rgb2hsv(color_mask)
img_hsv[..., 0] = color_mask_hsv[..., 0]
img_hsv[..., 1] = color_mask_hsv[..., 1] * Visualization.alpha
img_masked = color.hsv2rgb(img_hsv)
img_small = resize(img_masked, (40, 40),
anti_aliasing=True,
mode='constant')
#img_small.clip(min=0, out=img_small)
if pred != label:
img_small[:2, :] = [0.7, 0, 0]
img_small[38:, :] = [0.7, 0, 0]
img_small[:, :2] = [0.7, 0, 0]
img_small[:, 38:] = [0.7, 0, 0]
img_small.clip(min=0, max=1, out=img_small)
Visualization.plot_image.append(img_small)
def gen_components(self, image):
pantie = np.array(image)
# pickup colors
front = pantie[20:100, 30:80, :3] / 255.0
front_shade = pantie[130:150, 0:40, :3] / 255.0
front_color = self.pick_color(front)
front_shade_color = self.pick_color(front_shade)
front_shade_color = rgb2hsv(front_shade_color[None, None])
front_shade_color[0, 0, 1] *= front_shade_color[0, 0, 2] / 0.3
if front_shade_color[0, 0, 1] > 0.7:
front_shade_color[0, 0, 1] *= 0.7
front_shade_color[0, 0, 2] *= front_shade_color[0, 0, 2] / 0.4
front_shade_color = np.clip(hsv2rgb(front_shade_color)[0, 0], 0, 1)
components = np.copy(self.components)
components = self.components[:, :, :3] * front_color
components_shade = (
self.components_shade[:, :, -1])[:, :, None] * front_shade_color
# overlaying layers
components = alpha_brend(components_shade, components,
self.components_shade[:, :, -1])
components = np.dstack((components, self.components[:, :, -1]))
return Image.fromarray(np.uint8(np.clip(components, 0, 1) * 255))
def make_overlay(lac, label, void=0, saturation=0.7, blend=0.6, colors=None):
""" Make overlay image of lac and label """
if not isinstance(void, (list, tuple)):
void = [void]
rgb_colors = [hex2rgb(h) for h in colors]
img = normalize_ndarray(lac)
# Construct a colour image for the labels
color_mask = np.dstack([img] * 3)
# replace labels with color
labelindecies = [i for i in range(np.max(label) + 1) if i not in void]
for i in labelindecies:
color_mask[label == i] = rgb_colors[i]
# Convert the color mask to Hue Saturation Value (HSV)
color_mask_hsv = color.rgb2hsv(color_mask)
# decrease the saturation by saturation
# blend the value with LAC
color_mask_hsv[..., 1] *= saturation
color_mask_hsv[..., 2] = (1.0 - blend) * img + color_mask_hsv[..., 2] * blend
return color.hsv2rgb(color_mask_hsv)
def flow_to_color(flow, mask=None, max_flow=None):
"""Converts flow to 3-channel color image.
Args:
flow: tensor of shape [num_batch, 2, height, width].
mask: flow validity mask of shape [num_batch, 1, height, width].
"""
n = 8
B, _, H, W = flow.size()
mask = torch.ones(B, 1, H, W, dtype=flow.dtype, device=flow.device) \
if mask is None else mask
flow_u, flow_v = torch.split(flow, 1, dim=1)
if max_flow is not None:
max_flow = torch.max(torch.tensor(max_flow), torch.tensor(1.0))
else:
max_flow = torch.max(torch.abs(flow * mask))
mag = torch.pow(torch.sum(torch.pow(flow, 2), dim=1, keepdim=True), 0.5)
angle = torch.atan2(flow_v, flow_u)
im_h = torch.fmod(angle / (2 * np.pi) + 1.0, 1.0)
im_s = torch.clamp(mag * n / max_flow, 0., 1.)
im_v = torch.clamp(n - im_s, 0., 1.)
im_hsv = torch.cat((im_h, im_s, im_v), dim=1)
im_hsv = im_hsv.permute(0, 2, 3, 1)
im_rgb = np.empty((B, H, W, 3))
for i in range(B):
im_rgb[i, :, :, :] = hsv2rgb(im_hsv[i, :, :, :].cpu().numpy())
return torch.tensor(im_rgb, dtype=im_hsv.dtype).permute(0, 3, 1, 2)
def preprocess_img(self, img):
# save non processed image as local variable
self.nonprocessed_image = img
# Histogram normalization in y
hsv = color.rgb2hsv(img)
hsv[:, :, 2] = exposure.equalize_hist(
hsv[:, :, 2]
) # equalize_hist(image, nbins=256, mask=None) : return image array as HSV : (Hue (degrees), saturation (%), value (%))
img = color.hsv2rgb(hsv)
if (self.debug):
io.imsave("1_image_after_equalizehist.ppm", img)
# central scrop
min_side = min(img.shape[:-1])
centre = img.shape[0] // 2, img.shape[1] // 2
img = img[centre[0] - min_side // 2:centre[0] + min_side // 2,
centre[1] - min_side // 2:centre[1] + min_side // 2, :]
if (self.debug):
io.imsave("2_image_after_centralscrop.ppm", img)
# rescale to standard size
img = transform.resize(img, (IMG_SIZE, IMG_SIZE))
if (self.debug):
io.imsave("3_image_after_resize.ppm", img)
# roll color axis (RGB) to axis 0 : NEW IMAGE : (RGB, X, Y) instead of (X, Y, RGB)
img = np.rollaxis(img, -1)
# add one "id" axis at the beginning of the image. Not useful with only one image but required in the trained model
img = np.expand_dims(img, axis=0)
# save processed image in local variable
self.processed_img = img
return img
pass
def overlay_cam(img, cam, alpha, show=True, outFile=None):
if len(cam) > 0:
cam_heatmap = get_heatmap(cam)
img = img.astype(float)
img = img - np.min(img)
img = img / np.max(img)
# Construct RGB version of grey-level image
img_color = np.dstack((img, img, img))
# Convert the input image and color mask to Hue Saturation Value (HSV)
# colorspace
img_hsv = color.rgb2hsv(img_color)
color_mask_hsv = color.rgb2hsv(cam_heatmap)
# Replace the hue and saturation of the original image
# with that of the color mask
img_hsv[..., 0] = color_mask_hsv[..., 0]
img_hsv[..., 1] = color_mask_hsv[..., 1] * alpha
img_masked = color.hsv2rgb(img_hsv)
if show:
fig = plt.figure(figsize=(12, 16))
ax = fig.add_subplot(222)
plt.imshow(img_masked)
ax.axis('off')
ax.set_title('Grad-CAM')
plt.show()
if outFile:
plt.savefig(outFile)
return img_masked
def get_overlayed_image(x, c, gray_factor_bg=0.3,alpha = 0.5):
'''
For an image x and a relevance vector c, overlay the image with the
relevance vector to visualise the influence of the image pixels.
From: https://github.com/lmzintgraf/DeepVis-PredDiff/blob/master/utils_visualise.py
'''
imDim = x.shape[0]
if np.ndim(c) == 1:
c = c.reshape((imDim, imDim))
if np.ndim(x) == 2: # this happens with the MNIST Data
x = 1 - np.dstack((x, x, x)) * gray_factor_bg # make it a bit grayish
if np.ndim(x) == 3: # this is what happens with cifar data
x = color.rgb2gray(x)
x = 1 - (1 - x) * 0.5
x = np.dstack((x, x, x))
# Construct a colour image to superimpose
im = plt.imshow(c, cmap=cm.seismic, vmin=-np.max(np.abs(c)), vmax=np.max(np.abs(c)), interpolation='nearest')
color_mask = im.to_rgba(c)[:, :, [0, 1, 2]]
# Convert the input image and color mask to Hue Saturation Value (HSV) colorspace
img_hsv = color.rgb2hsv(x)
color_mask_hsv = color.rgb2hsv(color_mask)
# Replace the hue and saturation of the original image
# with that of the color mask
img_hsv[..., 0] = color_mask_hsv[..., 0]
img_hsv[..., 1] = color_mask_hsv[..., 1] * alpha
img_masked = color.hsv2rgb(img_hsv)
return img_masked
