def _equalize_images(distractor_images_dict, targets):
mean_lm_list = []
# calculate mean luminance of each image
# TODO: add contrast equalization
for key in distractor_images_dict:
for im in distractor_images_dict[key]:
data = np.asarray(im)
non_transparent_indices = (data[:, :, 3] != 0)
yiq = rgb2yiq(data[:, :, 0:3] / 255)
mean_lm_list.append(np.mean(yiq[non_transparent_indices, 0]))
for im in targets:
data = np.asarray(im)
non_transparent_indices = (data[:, :, 3] != 0)
yiq = rgb2yiq(data[:, :, 0:3] / 255)
mean_lm_list.append(np.mean(yiq[non_transparent_indices, 0]))
mean_lm = np.mean(mean_lm_list)
for key in distractor_images_dict:
for j in range(len(distractor_images_dict[key])):
new_im = _equalize_luminance(distractor_images_dict[key][j],
mean_lm)
new_im.info = distractor_images_dict[key][j].info
distractor_images_dict[key][j] = new_im
for j in range(len(targets)):
new_im = _equalize_luminance(targets[j], mean_lm)
new_im.info = targets[j].info
targets[j] = new_im
return distractor_images_dict, targets
def test_yuv(self):
rgb = np.array([[[1.0, 1.0, 1.0]]])
assert_array_almost_equal(rgb2yuv(rgb), np.array([[[1, 0, 0]]]))
assert_array_almost_equal(rgb2yiq(rgb), np.array([[[1, 0, 0]]]))
assert_array_almost_equal(rgb2ypbpr(rgb), np.array([[[1, 0, 0]]]))
assert_array_almost_equal(rgb2ycbcr(rgb), np.array([[[235, 128, 128]]]))
rgb = np.array([[[0.0, 1.0, 0.0]]])
assert_array_almost_equal(rgb2yuv(rgb), np.array([[[0.587, -0.28886916, -0.51496512]]]))
assert_array_almost_equal(rgb2yiq(rgb), np.array([[[0.587, -0.27455667, -0.52273617]]]))
assert_array_almost_equal(rgb2ypbpr(rgb), np.array([[[0.587, -0.331264, -0.418688]]]))
assert_array_almost_equal(rgb2ycbcr(rgb), np.array([[[144.553, 53.797, 34.214]]]))
def test_yuv(self):
rgb = np.array([[[1.0, 1.0, 1.0]]])
assert_array_almost_equal(rgb2yuv(rgb), np.array([[[1, 0, 0]]]))
assert_array_almost_equal(rgb2yiq(rgb), np.array([[[1, 0, 0]]]))
assert_array_almost_equal(rgb2ypbpr(rgb), np.array([[[1, 0, 0]]]))
assert_array_almost_equal(rgb2ycbcr(rgb), np.array([[[235, 128, 128]]]))
rgb = np.array([[[0.0, 1.0, 0.0]]])
assert_array_almost_equal(rgb2yuv(rgb), np.array([[[0.587, -0.28886916, -0.51496512]]]))
assert_array_almost_equal(rgb2yiq(rgb), np.array([[[0.587, -0.27455667, -0.52273617]]]))
assert_array_almost_equal(rgb2ypbpr(rgb), np.array([[[0.587, -0.331264, -0.418688]]]))
assert_array_almost_equal(rgb2ycbcr(rgb), np.array([[[144.553, 53.797, 34.214]]]))
def match_luminance(content, style):
content = content / 255
style = style / 255
content = color.rgb2yiq(content)
style = color.rgb2yiq(style)
mean_c = np.mean(content)
mean_s = np.mean(style)
stddev_c = np.std(content)
stddev_s = np.std(style)
style = (stddev_c / stddev_s) * (style - mean_s) + mean_c
style = np.clip(color.yiq2rgb(style), 0, 1) * 255
return style
def test_yuv_roundtrip(self):
img_rgb = img_as_float(self.img_rgb)[::16, ::16]
assert_array_almost_equal(yuv2rgb(rgb2yuv(img_rgb)), img_rgb)
assert_array_almost_equal(yiq2rgb(rgb2yiq(img_rgb)), img_rgb)
assert_array_almost_equal(ypbpr2rgb(rgb2ypbpr(img_rgb)), img_rgb)
assert_array_almost_equal(ycbcr2rgb(rgb2ycbcr(img_rgb)), img_rgb)
assert_array_almost_equal(ydbdr2rgb(rgb2ydbdr(img_rgb)), img_rgb)
def test_yuv_roundtrip(self):
img_rgb = img_as_float(self.img_rgb)[::16, ::16]
assert_array_almost_equal(yuv2rgb(rgb2yuv(img_rgb)), img_rgb)
assert_array_almost_equal(yiq2rgb(rgb2yiq(img_rgb)), img_rgb)
assert_array_almost_equal(ypbpr2rgb(rgb2ypbpr(img_rgb)), img_rgb)
assert_array_almost_equal(ycbcr2rgb(rgb2ycbcr(img_rgb)), img_rgb)
assert_array_almost_equal(ydbdr2rgb(rgb2ydbdr(img_rgb)), img_rgb)
def processingLayer1(self):
"""
Processing for YIQ, LAB, YCBC color models
"""
print(" ")
print("Preprocessing - Layer 1 processing")
widgets = [
progressbar.Percentage(), ' ',
progressbar.Bar(), ' ',
progressbar.ETA(), ' ',
progressbar.AdaptiveETA()
]
bar = progressbar.ProgressBar(widgets=widgets, maxval=3)
bar.start()
self.yiq = rgb2yiq(self.image) ## 1
self.lab = rgb2lab(self.image) ## 2
self.ycb = rgb2ycbcr(self.image) ## 2
bar.update(1)
## morph procedure
binary_yiq = self.morfprocess(self.yiq, 2, 1, 1, 0, self.percent)
binary_lab = self.morfprocess(self.lab, 2, 1, 0, 0, self.percent)
binary_ycb = self.morfprocess(self.ycb, 2, 1, 0, 0, self.percent)
bar.update(2)
## Arragne the information
self.Array_layer1[:, :, 0] = binary_yiq
self.Array_layer1[:, :, 1] = binary_lab
self.Array_layer1[:, :, 2] = binary_ycb
bar.update(3)
## Return the information
return self.Array_layer1
def test_rgb2yiq_conversion(self):
rgb = img_as_float(self.img_rgb)[::16, ::16]
yiq = rgb2yiq(rgb).reshape(-1, 3)
gt = np.array([colorsys.rgb_to_yiq(pt[0], pt[1], pt[2])
for pt in rgb.reshape(-1, 3)]
)
assert_almost_equal(yiq, gt, decimal=2)
def test_rgb2yiq_conversion(self):
rgb = img_as_float(self.img_rgb)[::16, ::16]
yiq = rgb2yiq(rgb).reshape(-1, 3)
gt = np.array([colorsys.rgb_to_yiq(pt[0], pt[1], pt[2])
for pt in rgb.reshape(-1, 3)]
)
assert_almost_equal(yiq, gt, decimal=2)
def dsqm(img1,original,blkSize = 100):
img1 = img1.astype(np.float32) / 256.0
original = original.astype(np.float32) / 256.0
h,w,p = original.shape
imgL = original
imgV = img1
cp = np.zeros((h,w))
offsetX = 2
offsetY = 2
imgLYIQ = rgb2yiq(imgL).astype(np.float32)
imgLY = imgLYIQ[:,:,1]
imgVYIQ = rgb2yiq(imgV).astype(np.float32)
imgVY = imgVYIQ[:,:,1]
brow = blkSize
bcol = brow
blkV = makeBlocks(imgVY, brow, bcol)
blkRows, blkCols = blkV.shape[0:2]
bestMatch = np.full((blkRows,blkCols), {'v':None,'x':None,'y':None})
blkVmatch = np.full((blkRows,blkCols), {})
score = np.zeros(blkV.shape)
for i in xrange(blkCols):
for j in xrange(blkRows):
T = blkV[j,i]
Tx = i * bcol
Ty = j * brow
Bx = (i+1) * bcol
By = (j+1) * brow
img = imgLY[max(0, Ty-offsetY):min(h, By+offsetY),max(0,Tx-offsetX):min(w, Bx+offsetX)]
orig = original[max(0, Ty-offsetY):min(h, By+offsetY),max(0,Tx-offsetX):min(w, Bx+offsetX)]
warped = imgV[max(0, Ty-offsetY):min(h, By+offsetY),max(0,Tx-offsetX):min(w, Bx+offsetX)]
b = imgV[j*bcol:(j+1)*bcol,i*brow:(i+1)*brow]
mp = cv2.matchTemplate(orig, b, cv2.TM_CCORR_NORMED)
#mp = cv2.matchTemplate(img, T, cv2.TM_CCORR_NORMED)
y,x = np.unravel_index(np.argmax(mp),mp.shape)
bestMatch[j,i] = {'v': mp[y,x], 'x': x, 'y': y}
blkVmatch[j,i]['arr'] = img[y:(y+bcol),x:(x+brow)]
a = orig[y:y+bcol,x:x+brow]
x = phasecong(T)
y = phasecong(blkVmatch[j,i]['arr'])
score[j,i] = np.abs(x[0].mean()-y[0].mean())
cp[j*bcol:(j+1)*bcol,i*brow:(i+1)*brow] = (y[0]-x[0])**2*256
return cp.mean(), cp
def extract(self, x):
YIQ = rgb2yiq(x) # From RGB to YIQ
Y = YIQ[:, :, 0]
(hist, _) = np.histogram(Y.ravel(), bins=self.numPointsHistogram)
# normalize the histogram
hist = hist.astype('float')
hist /= (hist.sum() + 1e-7)
return hist
def YIQ(image):
image_yiq = rgb2yiq(image)
image_yiq = np.around(image_yiq * 255, decimals=0)
#print(image_yiq)
#print(type(image_yiq))
image_yiq = image_yiq.astype(np.uint8)
#image_yiq = image_yiq.astype(np.float32)
return image_yiq
def compute_priorities(front, C, D, cote_patch, mask, img, color):
(n, m) = C.shape
P = np.zeros((n, m))
coordonnees = np.stack((np.nonzero(front)[0], np.nonzero(front)[1]),
axis=-1)
if (color):
imgYIQ = rgb2yiq(img)
imgY = imgYIQ[:, :, 0]
raw_gradient = np.gradient(imgY)
gradient = np.stack((raw_gradient[0], raw_gradient[1]), axis=-1)
else:
raw_gradient = np.gradient(img)
gradient = np.stack((raw_gradient[0], raw_gradient[1]), axis=-1)
raw_normal_vectors = compute_n(mask)
normal_vectors = np.stack((raw_normal_vectors[0], raw_normal_vectors[1]),
axis=-1)
element_structurant = np.ones((3, 3))
#temp_mask = scnd.morphology.binary_dilation(mask, element_structurant).astype('float')
old_front_approx = abs(
mask - scnd.morphology.binary_dilation(mask, element_structurant))
for psi_p in coordonnees:
gradient[psi_p[0],
psi_p[1]] = calculate_grad_approx(psi_p[0], psi_p[1],
gradient, old_front_approx)
D[psi_p[0], psi_p[1]] = abs(
np.vdot(
np.array([
gradient[psi_p[0], psi_p[1]][1], gradient[psi_p[0],
psi_p[1]][0]
]), normal_vectors[psi_p[0], psi_p[1]]))
p_x_min = max(0, psi_p[0] - cote_patch // 2)
p_x_max = min(psi_p[0] + cote_patch // 2, n - 1) + 1
p_y_min = max(0, psi_p[1] - cote_patch // 2)
p_y_max = min(psi_p[1] + cote_patch // 2, m - 1) + 1
#confidence = np.sum(C[p_x_min: p_x_max, p_y_min: p_y_max]) / ((p_x_max - p_x_min) * (p_y_max - p_y_min))
norme = (p_x_max - p_x_min) * (p_y_max - p_y_min) # Approximation
sum = 0
for i in range(p_x_min, p_x_max):
for j in range(p_y_min, p_y_max):
if mask[i, j] == 0.0:
sum += C[i, j]
C[psi_p[0], psi_p[1]] = sum / norme
for i in range(n):
for j in range(m):
if front[i, j] == 1.0:
P[i, j] = C[i, j] * D[i, j]
return P, C
def yiq_blend(im1, im2, mask, max_levels, filter_size_im, filter_size_mask):
"""
:param im1: input rgb image to be blended
:param im2: input rgb image to be blended
:param mask: boolean mask containing True and False representing which parts of im1 and im2 should
appear in the resulting im_blend
:param max_levels: max_levels parameter you should use when generating the Gaussian and Laplacian pyramids.
:param filter_size_im: is the size of the Gaussian filter (an odd scalar that represents a squared filter)
which defining the filter used in the construction of the Laplacian pyramids of im1 and im2
:param filter_size_mask: size of the Gaussian filter(an odd scalar that represents a squared filter) which
defining the filter used in the construction of the Gaussian pyramid of mask.
:return: the blended image
"""
y_im1 = rgb2yiq(im1)
y_im2 = rgb2yiq(im2)
channel = 0
y_im2[:, :, channel] = pyramid_blending(y_im1[:, :, channel],
y_im2[:, :,
channel], mask, max_levels,
filter_size_im, filter_size_mask)
blended_i = yiq2rgb(y_im2)
return blended_i
def preprocess(original, marked):
"""
Preprocesses input images: the original grayscale image, and the
version marked with colored scribbles. Converts images to YUV space
and trims based on (???).
"""
# read in images from files
grayscale = cv2.imread(original)
if grayscale is None:
raise Exception(f"Could not read from image file '{original}'")
marked = cv2.imread(marked)
if marked is None:
raise Exception(f"Could not read from image file '{marked}'")
marked = cv2.cvtColor(marked, cv2.COLOR_BGR2RGB)
# scale to float
grayscale = grayscale / 255.
marked = marked / 255.
# isolate colored markings
marks = np.sum(np.abs(grayscale - marked), axis=2) > PRECISION
marks = marks.astype('double')
# convert to YIQ
gray_ntsc = rgb2yiq(grayscale)
marked_ntsc = rgb2yiq(marked)
# create image to be colorized
h, w, c = gray_ntsc.shape
im = np.empty((h, w, c))
im[:, :, 0] = gray_ntsc[:, :, 0]
im[:, :, 1] = marked_ntsc[:, :, 1]
im[:, :, 2] = marked_ntsc[:, :, 2]
return (marks, im)
def extract_superpixel_features(stats, img, adj_matrix, binaryCropMask):
imgHSV = rgb2hsv(img)
imgLAB = rgb2lab(img)
imgYCbCr = rgb2ycbcr(img) / PIXEL_MAX_VAL
imgYIQ = rgb2yiq(img)
# set pixel range to [0,1]
imgRGB = img / PIXEL_MAX_VAL
imgRGBsum = np.sqrt(np.sum(np.square(imgRGB), axis=0))
# broadcasting will make division in 3 channels
imgRGBnorm = np.divide(imgRGB, imgRGBsum)
gradient_maps = get_7_gradient_maps(imgRGB)
minCoverPercentage = 0.5
color_spaces = [imgRGB, imgRGBnorm, imgLAB, imgHSV, imgYCbCr, imgYIQ]
features = []
for i, s in enumerate(stats):
# get features of superpixels inside a mask
if s.area <= 0 or np.mean(binaryCropMask[s.coords[:, 0], s.coords[:, 1]]) <= minCoverPercentage:
continue
feature = []
# put coords of superpixel as a feature
feature.append(s.centroid[0])
feature.append(s.centroid[1])
# get features of 8 neighbors and itself
neighbors = get8neighbors4_superpixel(stats, adj_matrix, i)
neighbors.append(i)
for color_space in color_spaces:
for nei in neighbors:
for channel in range(3):
coords = stats[nei].coords
color_values = color_space[coords[:,
0], coords[:, 1], channel]
feature.append(np.mean(color_values))
feature.append(np.var(color_values))
for gradient_map in gradient_maps:
for nei in neighbors:
coords = stats[nei].coords
values = gradient_map[coords[:, 0], coords[:, 1]]
feature.append(np.mean(values))
feature.append(np.var(values))
features.append(feature)
return features
def _equalize_luminance(im, new_mean):
im_arr = np.asarray(im)
ret_im = np.zeros_like(im_arr, dtype=np.uint8) # asarray is const
ret_im[:, :, 3] = im_arr[:, :, 3]
non_transparent_indices = (im_arr[:, :, 3] != 0)
rgb = im_arr[:, :, 0:3] / 255 # turn to 0-1 image instead of 0-255
yiq = rgb2yiq(rgb)
cur_mean = np.mean(yiq[non_transparent_indices,
0]) # disregard pixels with alpha=0
yiq[non_transparent_indices, 0] -= (cur_mean - new_mean)
rgb = np.round(yiq2rgb(yiq) *
255) # convert back to RGB nad non-normalized image
rgb /= np.max(rgb) # normalize
rgb = np.round(rgb * 255) # back to 0-255
ret_im[:, :, 0:3] = rgb
ret_im = Image.fromarray(ret_im)
return ret_im
def test_yuv_roundtrip(self, channel_axis):
img_rgb = img_as_float(self.img_rgb)[::16, ::16]
img_rgb = np.moveaxis(img_rgb, source=-1, destination=channel_axis)
assert_array_almost_equal(
yuv2rgb(rgb2yuv(img_rgb, channel_axis=channel_axis),
channel_axis=channel_axis), img_rgb)
assert_array_almost_equal(
yiq2rgb(rgb2yiq(img_rgb, channel_axis=channel_axis),
channel_axis=channel_axis), img_rgb)
assert_array_almost_equal(
ypbpr2rgb(rgb2ypbpr(img_rgb, channel_axis=channel_axis),
channel_axis=channel_axis), img_rgb)
assert_array_almost_equal(
ycbcr2rgb(rgb2ycbcr(img_rgb, channel_axis=channel_axis),
channel_axis=channel_axis), img_rgb)
assert_array_almost_equal(
ydbdr2rgb(rgb2ydbdr(img_rgb, channel_axis=channel_axis),
channel_axis=channel_axis), img_rgb)
def histogram_equalize(im_orig):
"""
performs histogram equalization to the image
:param im_orig: an ndimage array
:return: array of equalized image, the original histogram and the cumulative histogram
"""
if len(im_orig.shape) > 2:
# rgb
YIQim = rgb2yiq(im_orig) * 255
hist_orig, bin_edges = np.histogram(YIQim[:, :, 0], 256)
rows, columns, dim = im_orig.shape
cum_hist = np.cumsum(hist_orig)
cum_hist = cum_hist.astype(np.float64)
else:
# grayscale
im_orig *= 255
hist_orig, bin_edges = np.histogram(im_orig, 256)
cum_hist = np.cumsum(hist_orig)
cum_hist = cum_hist.astype(np.float64)
rows, columns = im_orig.shape
tot_pixels = rows * columns
cum_hist = (cum_hist / tot_pixels)
minimum = min(np.nonzero(cum_hist)[0])
maximum = np.nonzero(cum_hist)[0][-1]
minVal = cum_hist[minimum]
maxVal = cum_hist[maximum]
cum_hist = (255 * ((cum_hist - minVal) / (maxVal - minVal)))
cum_hist = np.around(cum_hist)
if len(im_orig.shape) > 2:
im_eq = np.copy(YIQim)
y_values = cum_hist[YIQim[:, :, 0].astype(np.int8)]
im_eq[:, :, 0] = y_values
im_eq = yiq2rgb(im_eq / 255)
else:
im_eq = cum_hist[im_orig.astype(np.int8)]
cum_hist /= 255
cum_hist = np.clip(cum_hist, 0, 1)
return [im_eq, hist_orig, cum_hist]
import cv2
from skimage import color as scl
from matplotlib import pyplot as plt
imgRGB = cv2.cvtColor(cv2.imread("Lenna.jpg", cv2.IMREAD_COLOR),
cv2.COLOR_BGR2RGB)
imgHSV = scl.rgb2hsv(imgRGB)
imgYIQ = scl.rgb2yiq(imgRGB)
plt.subplot(221)
plt.imshow(imgRGB)
plt.title("Original Image")
imgRGB[:, :, 0] = cv2.equalizeHist(imgRGB[:, :, 0])
imgRGB[:, :, 1] = cv2.equalizeHist(imgRGB[:, :, 1])
imgRGB[:, :, 2] = cv2.equalizeHist(imgRGB[:, :, 2])
plt.subplot(222)
plt.imshow(imgRGB)
plt.title("RGB")
# imgHSV[:, :, 2] = cv2.equalizeHist(imgHSV[:, :, 2])
plt.subplot(223)
plt.imshow(scl.hsv2rgb(imgHSV))
plt.title("HSV")
# imgYIQ[:, :, 0] = cv2.normalize(cv2.equalizeHist(cv2.normalize(imgYIQ[:, :, 0], None, 0, 255, cv2.NORM_MINMAX)), None, 0.0, 1.0, cv2.NORM_MINMAX)
# imgYIQ[:, :, 0] = yiq1
plt.subplot(224)
plt.imshow(scl.yiq2rgb(imgYIQ))
print(imgYIQ[:, :, 0].max)
plt.title("YIQ")
def signal_from_ROI(input_fullname):
# Open video
vid = cv2.VideoCapture(input_fullname)
counter = 0
frameNumber = 0
amplified_signal = [0, 0]
length = int(vid.get(cv2.CAP_PROP_FRAME_COUNT))
FPS = int(vid.get(cv2.CAP_PROP_FPS))
print("Frames per second: ", FPS)
temp_cdata = vid.read()
# ROI selection
r = cv2.selectROI(temp_cdata[1])
row_ini = int(r[1])
row_fin = int(r[1] + r[3])
cols_ini = int(r[0])
cols_fin = int(r[0] + r[2])
# Shows interactive image
fig = plt.figure()
ax = fig.add_subplot(111)
Ln, = ax.plot(amplified_signal)
plt.axis([0, 1, 0, 0.1])
ax.set_xlim([0, length + 2])
plt.ion()
plt.show()
# Loop
while vid.isOpened():
print("Procesando: %.1f%%" % (100 * counter / length))
temp_cdata = vid.read()
if temp_cdata[0]:
# creates ROI and convert to RGB
temp_BGR_ROI = temp_cdata[1][row_ini:row_fin, cols_ini:cols_fin, :]
temp_RGB_ROI = cv2.cvtColor(temp_BGR_ROI, cv2.COLOR_BGR2RGB)
currentFrame_ROI = temp_RGB_ROI.astype('float') / 255
# Show ROI in video
cv2.imshow("ROI", temp_BGR_ROI)
# Reference frame
if counter == 2:
RGB_ini_ROI = currentFrame_ROI
# Creates signal
if counter > 2:
diff_ROI_RGB = currentFrame_ROI - RGB_ini_ROI
# diff_ROI_RGB = np.clip(diff_ROI_RGB, a_min=0, a_max=1)
diff_ROI_RGB = np.abs(diff_ROI_RGB)
diff_ROI_YIQ = color.rgb2yiq(diff_ROI_RGB)
mean_Y = np.mean(diff_ROI_YIQ[:, :, 0])
amplified_signal = np.append(amplified_signal, mean_Y)
# shows peaks every 2 seconds
buffer_length = np.round(2 * FPS)
if frameNumber % buffer_length == 0:
# search and draw peaks
peaks, _ = find_peaks(amplified_signal[(
len(amplified_signal) -
buffer_length):len(amplified_signal)],
height=0)
plt.plot(
len(amplified_signal) - buffer_length + peaks,
amplified_signal[len(amplified_signal) -
buffer_length + peaks], "x")
# Update amplified signal in figure
Ln.set_ydata(amplified_signal)
Ln.set_xdata(range(len(amplified_signal)))
plt.pause(0.01)
frameNumber = frameNumber + 1
else:
break
counter = counter + 1
### end of WHILE ####
# Shows remaining peaks in the signal
peaks, _ = find_peaks(
amplified_signal[(len(amplified_signal) -
buffer_length):len(amplified_signal)],
height=0)
plt.plot(
len(amplified_signal) - buffer_length + peaks,
amplified_signal[len(amplified_signal) - buffer_length + peaks], "x")
plt.pause(5.0)
# Release video
vid.release()
def vidmag_fn(input_fullname, parameters):
alpha = parameters['alpha']
lambda_c = parameters['lambda_c']
fl = parameters['fl']
fh = parameters['fh']
samplingRate = parameters['samplingRate']
chromAttenuation = parameters['chromAttenuation']
nlevels = parameters['nlevels']
# Butterworth coefficients
[low_a, low_b] = signal.butter(1, fl / samplingRate, 'low')
[high_a, high_b] = signal.butter(1, fh / samplingRate, 'low')
# output fullname format
input_filename = os.path.splitext(os.path.basename(input_fullname))[0]
output_fullname = c.PROCESSED_VIDEO_DIR+input_filename+'-butter-from-'+str(fl)+'-to-'+str(fh)+'Hz'+\
'-alpha-'+str(alpha)+'-lambda_c-'+str(lambda_c)+\
'-chromAtn-'+str(chromAttenuation)+'.mp4'
input_video = cv2.VideoCapture(input_fullname)
vidHeight = int(input_video.get(cv2.CAP_PROP_FRAME_HEIGHT))
vidWidth = int(input_video.get(cv2.CAP_PROP_FRAME_WIDTH))
fr = int(input_video.get(cv2.CAP_PROP_FPS))
length = int(input_video.get(cv2.CAP_PROP_FRAME_COUNT))
fourcc = cv2.VideoWriter_fourcc(*'MP4V')
output_video = cv2.VideoWriter(output_fullname, fourcc, fr,
(vidWidth, vidHeight), 1)
# First frame
temp_cdata = input_video.read()
rgbframe = temp_cdata[1].astype('float') / 255.0
# get desired sizes used in all the Laplacian pyramid levels
dsizes = np.zeros((nlevels + 1, 2))
for k in range(0, nlevels + 1):
dsizes[k, :] = [
np.floor(rgbframe.shape[0] / (2**k)),
np.floor(rgbframe.shape[1] / (2**k))
]
desired_sizes = tuple(map(tuple, dsizes))
print(desired_sizes)
# first frame processing (initial conditions)
frame = color.rgb2yiq(rgbframe)
lpyr = buildlpyr(frame, nlevels,
desired_sizes) # creates Laplacian pyramid
lowpass1 = lpyr
lowpass2 = lpyr
pyr_prev = lpyr
output_frame = color.yiq2rgb(frame) # yiq color space
output_frame = output_frame * 255
output_video.write(output_frame.astype('uint8'))
# processing remaining frames
counter = 1
while input_video.isOpened():
print("Processing: %.1f%%" % (100 * counter / length))
temp_cdata = input_video.read()
if not (temp_cdata[0]):
break
#from rgb to yiq
rgbframe = temp_cdata[1].astype('float') / 255
frame = color.rgb2yiq(rgbframe)
# Laplacian pyramid (expansion)
lpyr = buildlpyr(frame, nlevels, desired_sizes)
# Temporal filter
lowpass1 = (-high_b[1] * lowpass1 + high_a[0] * lpyr +
high_a[1] * pyr_prev) / high_b[0]
lowpass2 = (-low_b[1] * lowpass2 + low_a[0] * lpyr +
low_a[1] * pyr_prev) / low_b[0]
filtered = lowpass1 - lowpass2
pyr_prev = lpyr
# Amplification
delta = lambda_c / 8 / (1 + alpha)
exaggeration_factor = 2
lambda_ = (vidHeight**2 + vidWidth**2)**0.5 / 3
filtered[0] = np.zeros_like(filtered[0])
filtered[-1] = np.zeros_like(filtered[-1])
for i in range(nlevels - 1, 1, -1):
# equation 14 (paper vidmag, see references)
currAlpha = lambda_ / delta / 8 - 1
currAlpha = currAlpha * exaggeration_factor
# from figure 6 (paper vidmag, see references)
if currAlpha > alpha:
filtered[i] = alpha * filtered[i]
else:
filtered[i] = currAlpha * filtered[i]
lambda_ = lambda_ / 2
# Laplacian pyramid (contraction)
for i in range(nlevels - 1):
aux = cv2.pyrUp(filtered[i],
dstsize=(int(desired_sizes[nlevels - 2 - i][1]),
int(desired_sizes[nlevels - 2 - i][0])))
pyr_contraida = cv2.add(aux, filtered[i + 1])
# Components chrome attenuation
pyr_contraida[:, :, 1] = pyr_contraida[:, :, 1] * chromAttenuation
pyr_contraida[:, :, 2] = pyr_contraida[:, :, 2] * chromAttenuation
# adding contracted pyramid to current frame
output_frame = pyr_contraida + frame
# recovering rgb frame
output_frame = color.yiq2rgb(output_frame)
output_frame = np.clip(output_frame, a_min=0, a_max=1)
output_frame = output_frame * 255
# saving processed frame to output video
output_video.write(output_frame.astype('uint8'))
counter = counter + 1
### end of WHILE ####
# release video
output_video.release()
input_video.release()
return output_fullname
def main(args):
content_image = Image.open(FLAGS.content)
content_image = resize(content_image)
style_image = Image.open(FLAGS.style).resize(content_image.size)
content_image = np.asarray(content_image, dtype=np.float32)
style_image = np.asarray(style_image, dtype=np.float32)
# match luminance between style and content image
style_image = match_luminance(content_image, style_image)
content_image = np.expand_dims(content_image, 0)
style_image = np.expand_dims(style_image, 0)
img_shape = content_image.shape
with tf.name_scope('image'):
random_image = tf.random_normal(mean=1, stddev=.01, shape=img_shape)
image = tf.Variable(initial_value=random_image,
name='image',
dtype=tf.float32)
tf.summary.image('img', tf.clip_by_value(image, 0, 255), max_outputs=1)
# subtract mean
inputs = image - IMAGENET_MEAN
# convert to BGR, because VGG16 was trained on BGR images
channels = tf.unstack(inputs, axis=-1)
inputs = tf.stack([channels[2], channels[1], channels[0]], axis=-1)
_, endpoints = vgg_16(inputs, is_training=False, scope='vgg_16')
saver = tf.train.Saver(var_list=tf.get_collection(
tf.GraphKeys.GLOBAL_VARIABLES, scope='vgg_16'))
style_tensors = [endpoints[l] for l in STYLE_LAYERS]
content_tensors = [endpoints[l] for l in CONTENT_LAYERS]
image_style_tensors = [endpoints[l] for l in STYLE_LAYERS]
image_content_tensors = [endpoints[l] for l in CONTENT_LAYERS]
with tf.Session() as sess:
tf.global_variables_initializer().run()
saver.restore(sess, FLAGS.ckpt_file)
style_features = sess.run(style_tensors,
feed_dict={image: style_image})
content_features = sess.run(content_tensors,
feed_dict={image: content_image})
# define style loss
style_losses = []
for image_layer, style_layer in zip(image_style_tensors, style_features):
_, height, width, channels = image_layer.get_shape().as_list()
size = height * width * channels
# computer gram matrices
image_feats_reshape = tf.reshape(image_layer, [-1, channels])
image_gram = tf.matmul(tf.transpose(image_feats_reshape),
image_feats_reshape) / size
style_feats_reshape = tf.reshape(style_layer, [-1, channels])
style_gram = tf.matmul(tf.transpose(style_feats_reshape),
style_feats_reshape) / size
loss = tf.square(
tf.norm(image_gram - style_gram, ord='fro', axis=(0, 1)))
style_losses.append(loss)
style_loss = STYLE_WEIGHT * tf.add_n(style_losses)
# define content loss
content_losses = []
for image_layer, content_layer in zip(image_content_tensors,
content_features):
_, height, width, channels = image_layer.get_shape().as_list()
size = height * width * channels
loss = tf.nn.l2_loss(image_layer - content_layer) / size
content_losses.append(loss)
content_loss = CONTENT_WEIGHT * tf.add_n(content_losses)
# total variation denoising loss
tvd_loss = TVD_WEIGHT * tf.reduce_sum(tf.image.total_variation(image))
loss = style_loss + content_loss + tvd_loss
global_step = tf.train.get_or_create_global_step()
optim = tf.train.AdamOptimizer(LEARNING_RATE)
train_op = optim.minimize(loss, global_step=global_step, var_list=[image])
with tf.Session() as sess:
tf.global_variables_initializer().run()
saver.restore(sess, FLAGS.ckpt_file)
noise = tf.random_normal(mean=1, stddev=.01, shape=img_shape)
rand_init = tf.clip_by_value(content_image * noise, 0, 255)
image.assign(rand_init).eval()
for step in range(FLAGS.steps):
t0 = time.time()
_, style_loss_val, content_loss_val, tvd_loss_val, loss_val = sess.run(
[train_op, style_loss, content_loss, tvd_loss, loss])
t = time.time() - t0
if step % 10 == 0:
format_str = 'step: {}/{} loss: style: {}, content: {}, tvd: {}, total: {} | time: {:.2f} s/step'
print(
format_str.format(step, FLAGS.steps, style_loss_val,
content_loss_val, tvd_loss_val, loss_val,
t))
img = sess.run(image)[0]
# transfer luminance
img = np.clip(img, 0, 255) / 255
content_image = content_image[0] / 255
result_y = np.expand_dims(color.rgb2yiq(img)[:, :, 0], 2)
content_iq = color.rgb2yiq(content_image)[:, :, 1:]
img = np.dstack((result_y, content_iq))
img = np.clip(color.yiq2rgb(img), 0, 1)
imsave(FLAGS.result_file, img)
def pruebacolor(img):
print("rgb")
figure(0)
imshow(img[:, :, 0])
title("rgb r")
figure(1)
imshow(img[:, :, 1])
title("rgb g")
figure(2)
imshow(img[:, :, 2])
title("rgb b")
yiq = rgb2yiq(img)
print("yiq")
figure(3)
imshow(yiq[:, :, 0])
title("yiq y")
figure(4)
imshow(yiq[:, :, 1])
title("yiq i")
figure(5)
imshow(yiq[:, :, 2])
title("yiq q")
hsv = rgb2hsv(img)
print("hsv")
figure(6)
imshow(hsv[:, :, 0])
title("hsv h")
figure(7)
imshow(hsv[:, :, 1])
title("hsv s")
figure(8)
imshow(hsv[:, :, 2])
title("hsv v")
xyz = rgb2xyz(img)
print("xyz")
figure(9)
imshow(xyz[:, :, 0])
title("xyz x")
figure(10)
imshow(xyz[:, :, 1])
title("xyz y")
figure(11)
imshow(xyz[:, :, 2])
title("xyz z")
lab = rgb2lab(img)
print("lab")
figure(12)
imshow(lab[:, :, 0])
title("lab l")
figure(13)
imshow(lab[:, :, 1])
title("lab a")
figure(14)
imshow(lab[:, :, 2])
title("lab b")
ycbcr = rgb2ycbcr(img)
print("ycbcr")
figure(15)
imshow(ycbcr[:, :, 0])
title("ycbcr y")
figure(16)
imshow(ycbcr[:, :, 1])
title("ycbcr cb")
figure(17)
imshow(ycbcr[:, :, 2])
title("ycbcr cr")
def main(content_image_path, style_image_path, iterations, content_img_height,
style_img_height, tv_weight,
style_weight, content_weight, save_gif, preserve_color, learning_rate,
beta_1, beta_2, epsilon):
"""Performs neural style transfer on a content image and style image."""
model = get_model()
# Load images
content_image = load_and_process_img(content_image_path, content_img_height)
style_image = load_and_process_img(style_image_path, style_img_height)
content_image_yiq = rgb2yiq(
load_img(content_image_path, content_img_height))
# Compute content and style features
style_outputs = model(style_image)
content_outputs = model(content_image)
# Get the style and content feature representations from our model
style_features = [style_layer[0] for style_layer in
style_outputs[:NUM_STYLE_LAYERS]]
content_features = [content_layer[0] for content_layer in
content_outputs[NUM_STYLE_LAYERS:]]
gram_style_features = [gram_matrix(style_feature) for style_feature in
style_features]
# Set initial image
init_image = load_and_process_img(content_image_path, content_img_height,
as_gray=preserve_color)
init_image = tf.Variable(init_image, dtype=tf.float32)
# Create our optimizer
optimizer = tf.optimizers.Adam(learning_rate=learning_rate, beta_1=beta_1,
beta_2=beta_2, epsilon=epsilon)
# Create config dictionary
loss_weights = (style_weight, content_weight, tv_weight)
config = {
'model': model,
'loss_weights': loss_weights,
'init_image': init_image,
'gram_style_features': gram_style_features,
'content_features': content_features
}
images = []
# Optimization loop
for step in range(1, iterations + 1):
start_time = time.time()
grads, loss = compute_grads(config)
optimizer.apply_gradients([(grads, init_image)])
clipped = tf.clip_by_value(init_image, MIN_VALS, MAX_VALS)
init_image.assign(clipped)
img = deprocess_image(init_image.numpy())
if preserve_color:
img = rgb2yiq(img)
img[:, :, 1:] = content_image_yiq[:, :, 1:]
img = yiq2rgb(img)
img = np.clip(img, 0, 1)
img = (img * 255).astype('uint8')
images.append(img)
end_time = time.time()
print('Finished step {} ({:.03} seconds)\nLoss: {}\n'.format(
step, end_time - start_time, loss))
# Save final image
imsave('stylized.jpg', images[-1])
if save_gif:
create_gif(images, 'transformation.gif')
def pixelmatch(img1, img2):
yuv1 = rgb2yiq(img1)
yuv2 = rgb2yiq(img2)
delta2 = np.square(yuv1 - yuv2) # why square?
return delta2 @ [0.5053, 0.299, 0.1957]
for i_b in range(B_prime_pyramid[l].shape[0]):
print("\tRow %d" % i_b)
for j_b in range(B_prime_pyramid[l].shape[1]):
i_a, j_a = best_match(A_pyramid, A_prime_pyramid, B_pyramid,
B_prime_pyramid, s, l, i_b, j_b)
B_prime_pyramid[l][i_b][j_b] = A_prime_pyramid[l][i_a][j_a]
s[(i_b, j_b)] = (i_a, j_a)
return B_prime_pyramid[0]
if __name__ == '__main__':
A = plt.imread(INPUT + A_NAME)
A_prime = plt.imread(INPUT + A_PRIME_NAME)
B = plt.imread(INPUT + B_NAME)
if USE_LUMINANCE:
A, A_prime, B = rgb2yiq(A), rgb2yiq(A_prime), rgb2yiq(B)
transform_func, inverse_transform_func = compute_luminance_transforms(
A, B)
A[:, :, 0] = transform_func(A[:, :, 0])
A_prime[:, :, 0] = transform_func(A_prime[:, :, 0])
B[:, :, 0] = transform_func(B[:, :, 0])
B_prime = create_image_analogy(A, A_prime, B)
if USE_LUMINANCE:
B_prime[:, :, 0] = inverse_transform_func(B_prime[:, :, 0])
B_prime = yiq2rgb(B_prime)
try:
os.makedirs(OUTPUT)
except:
pass
print(img_gray_01)
print("変換後: [0, 255]")
print(img_gray)
io.imshow(img_gray.astype(np.uint8))
# #### グレースケール化②
# 別のグレースケール化方法も試してみましょう。一度RGB空間からYIQ空間へ変換し、Yを利用を利用します。YIQ形式は、グレースケール情報がカラーデータから分離しているため、同じ信号をカラーと白黒の両方で使用可能です。
#
# （※ グレースケール化のアルゴリズムによっては `img_gray_01` と `img_yiq[:, :, 0]` が等しくなりますが、skimageでは異なります）
# In[38]:
img_yiq = color.rgb2yiq(img_true)
img_conb = np.concatenate(
(img_yiq[:, :, 0], img_yiq[:, :, 1], img_yiq[:, :, 2]), axis=1)
io.imshow(img_conb)
# In[33]:
# skimage.color.rgb2gray と比較
img_conb2 = np.concatenate((img_yiq[:, :, 0], img_gray_01), axis=1)
io.imshow(img_conb2)
# In[47]:
def __call__(self, img):
img = np.asarray(img, np.uint8)
img = color.rgb2yiq(img)
return img
def quantize(im_orig, n_quant, n_iter):
"""
a function that performs optimal quantization of a given grayscale or RGB image.
:param im_orig: is the input grayscale or RGB image to be quantized (float64 image with values in [0, 1]).
:param n_quant: is the number of intensities your output im_quant image should have.
:param n_iter: is the maximum number of iterations of the optimization procedure
:return: im_quant - is the quantized output image.
error - is an array with shape (n_iter,) (or less) of the total intensities error for each iteration of the
quantization procedure
"""
error = []
q = np.array([0] * n_quant, dtype=np.float64)
z = [0] * (n_quant + 1)
if len(im_orig.shape) > 2:
# rgb
YIQim = rgb2yiq(im_orig)
hist = np.histogram(YIQim[:, :, 0] * 255, bins=256)[0]
else:
# grayscale
hist = np.histogram(im_orig * 255, bins=256)[0]
cum_hist = np.cumsum(hist)
for i in range(1, n_quant):
z[i] = np.where(cum_hist > (i / n_quant) * cum_hist[255])[0][0] - 1
z[n_quant] = 255
q_1 = 0
q_2 = 0
for i in range(n_iter):
change = False # boolean to see if z changed
# calculate new q
for j in range(n_quant):
for x in range(z[j], z[j + 1] + 1):
q_2 += hist[x]
q_1 += (x * hist[x])
q[j] = int(q_1 / q_2)
q_1 = 0
q_2 = 0
# calculate new z
for j in range(1, n_quant):
z_1 = int((q[j - 1] + q[j]) / 2)
if z_1 != z[j]:
z[j] = z_1
change = True
error.append(compute_error(n_quant, z, q, hist))
if not change:
lut = create_lut(z, q)
if len(im_orig.shape) > 2:
YIQim *= 255
im_quant = np.copy(YIQim)
y_values = lut[YIQim[:, :, 0].astype(np.int8)]
im_quant[:, :, 0] = y_values
im_quant = yiq2rgb(im_quant)
else:
im_orig *= 255
im_quant = lut[im_orig.astype(np.int8)]
return [im_quant, error]
lut = create_lut(z, q)
if len(im_orig.shape) > 2:
YIQim *= 255
im_quant = np.copy(YIQim)
y_values = lut[YIQim[:, :, 0].astype(np.int8)]
im_quant[:, :, 0] = y_values
im_quant = yiq2rgb(im_quant / 255)
else:
im_orig *= 255
im_quant = lut[im_orig.astype(np.int8)]
return [im_quant, error]
def toYCbCr(rgb):
return color.rgb2yiq(rgb)
def generateFeaturesImages(img_rgb, fparams):
f_imgs = dict()
f_imgs['gray'] = color.rgb2gray(img_rgb)
f_imgs['RGB'] = img_rgb
if fparams['LAB']['use']:
f_imgs['LAB'] = color.rgb2lab(img_rgb)
if fparams['HSV']['use']:
f_imgs['HSV'] = color.rgb2hsv(img_rgb)
if fparams['YCbCr']['use']:
f_imgs['YCbCr'] = color.rgb2ycbcr(img_rgb)
if fparams['xyz']['use']:
f_imgs['xyz'] = color.rgb2xyz(img_rgb)
if fparams['yiq']['use']:
f_imgs['yiq'] = color.rgb2yiq(img_rgb)
if fparams['yuv']['use']:
f_imgs['yuv'] = color.rgb2yuv(img_rgb)
if fparams['entropy']['use']:
f_imgs['entropy'] = entropy(img_as_ubyte(f_imgs['gray']), disk(5))
#POR CAUSA DOS COMENTARIOS TESTE
'''
if fparams['texton']['use']: # appende gabor filter responses to extract mean and std
kernels = generateKernels() #generate gabor filters
texton_imgs = np.empty((f_imgs['gray'].shape[0], f_imgs['gray'].shape[1], fparams['texton']['n_kernels']))
for k, kernel in enumerate(kernels):
texton_imgs[:,:,k] = nd.convolve(f_imgs['gray'], kernel)
if fparams['texton']['mean'] or fparams['texton']['std']:
f_imgs['texton'] = texton_imgs
if fparams['texton']['hist_max']:
f_imgs['hist_max'] = np.argmax(texton_imgs, 2)
if fparams['texton']['hist_kmeans']:
data_t = np.empty((f_imgs['gray'].size, fparams['texton']['n_kernels']))
for k in range(fparams['texton']['n_kernels']):
data_t [:, k] = texton_imgs[:,:,k].flatten()
textons = kmeans.predict(data_t)
f_imgs['hist_kmeans'] = textons.reshape(f_imgs['gray'].shape[0], f_imgs['gray'].shape[1])
'''
'''
if fparams['grayD']['use'] or fparams['grayA']['use']:
f_imgs['grayMeans'] = generateGrayMeans(f_imgs['gray'], 6, 6, 1) #gray diff between mean gray of the block and some areas
'''
if fparams['LBP']['use']:
if fparams['LBP']['gray']:
f_imgs['LBP_gray'] = local_binary_pattern(f_imgs['gray'], fparams['LBP']['n_neibor'], fparams['LBP']['radius']) #lbp image
if fparams['LBP']['red']:
f_imgs['LBP_r'] = local_binary_pattern(f_imgs['RGB'][:,:,0], fparams['LBP']['n_neibor'], fparams['LBP']['radius']) #lbp image
if fparams['LBP']['green']:
f_imgs['LBP_g'] = local_binary_pattern(f_imgs['RGB'][:,:,1], fparams['LBP']['n_neibor'], fparams['LBP']['radius']) #lbp image
if fparams['LBP']['blue']:
f_imgs['LBP_b'] = local_binary_pattern(f_imgs['RGB'][:,:,2], fparams['LBP']['n_neibor'], fparams['LBP']['radius']) #lbp image
""" if fparams['VegetationIndex']['use']:
img_nir = img_rgb.astype(float)[:,:,0]
img_g = img_rgb.astype(float)[:,:,1]
img_r = img_rgb.astype(float)[:,:,2]
if fparams['VegetationIndex']['NDVI']:
f_imgs['NDVI'] = ((img_nir-img_r)/(img_nir+img_r))
if fparams['VegetationIndex']['NNIR']:
f_imgs['NNIR'] = ((img_nir)/(img_nir+img_g+img_r))
if fparams['VegetationIndex']['NGREEN']:
f_imgs['NGREEN'] = ((img_g)/(img_nir+img_g+img_r))
if fparams['VegetationIndex']['NRED']:
f_imgs['NRED'] = ((img_r)/(img_nir+img_g+img_r))
if fparams['VegetationIndex']['PVI']:
b,a = np.polyfit(img_r[0,], img_nir[0,], 1) #regression to find values for b and a
f_imgs['PVI'] = ((b*img_nir*img_r)+a)/(np.sqrt(b+1)) """
return f_imgs
import cv2
from skimage import color as scl
from matplotlib import pyplot as plt
img = cv2.cvtColor(cv2.imread("Lenna.jpg", cv2.IMREAD_COLOR),
cv2.COLOR_BGR2RGB)
imgYIQ = scl.rgb2yiq(img)
plt.subplot(221)
plt.imshow(img)
plt.title("Colorful image")
plt.subplot(222)
plt.imshow(imgYIQ[:, :, 0], cmap='gray')
plt.title("Y")
plt.subplot(223)
plt.imshow(imgYIQ[:, :, 1], cmap='gray')
plt.title("I")
plt.subplot(224)
plt.imshow(imgYIQ[:, :, 2], cmap='gray')
plt.title("Q")
plt.subplots_adjust(left=0.1,
bottom=0.1,
right=0.9,
top=0.9,
wspace=0.6,
hspace=0.6)
plt.show()
