def colorization(img, marked_img, k):
H, W, C = img.shape
# convert to YUV color space
img_yuv = rgb2yuv(img)
marked_yuv = rgb2yuv(marked_img)
Y = np.array(img_yuv[:, :, 0])
# find the color position and update sparse matrix
diff = np.where(marked_yuv != img_yuv)
colored_coord_idx = np.unique(diff[0] * W + diff[1])
weight_matrix = build_weights_matrix(Y, colored_coord_idx, k)
# compute the least-square solution, by computing the pseudo-inverse
LU = splu(weight_matrix)
b1 = (marked_yuv[:, :, 1]).flatten()
b2 = (marked_yuv[:, :, 2]).flatten()
U = LU.solve(b1)
V = LU.solve(b2)
sol = np.zeros_like(img)
sol[:, :, 0] = Y
sol[:, :, 1] = U.reshape((H, W))
sol[:, :, 2] = V.reshape((H, W))
return sol
def ssim(ims1, ims2, scale):
mean_ssim = 0.0
for im1, im2 in zip(ims1, ims2):
im1 = color.rgb2yuv(im1[scale:-scale, scale:-scale])[..., 0]
im2 = color.rgb2yuv(im2[scale:-scale, scale:-scale])[..., 0]
mean_ssim += measure.compare_ssim(im1, im2) / len(ims1)
return mean_ssim
def SSIM(img1, img2):
[h, w, c] = img1.shape
if c > 2:
img1 = color.rgb2yuv(img1)
img1 = img1[:, :, 0]
img2 = color.rgb2yuv(img2)
img2 = img2[:, :, 0]
score = ssim(img1, img2)
return score
def rmse(ims1, ims2, scale):
mean_rmse = 0.0
for im1, im2 in zip(ims1, ims2):
im1 = color.rgb2yuv(im1[scale:-scale, scale:-scale])[..., 0]
im2 = color.rgb2yuv(im2[scale:-scale, scale:-scale])[..., 0]
mean_rmse += np.sum(np.square(im2 - im1)) / len(ims1)
mean_rmse = np.sqrt(mean_rmse)
return mean_rmse
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 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 gen_pencil_drawing(img, kernel_size, stroke_width=0, num_of_directions=8, smooth_kernel="gauss",
gradient_method=0, rgb=False, w_group=0, pencil_texture_path="", stroke_darkness=1, tone_darkness=1):
if not rgb:
# Grayscale image:
im = img
else:
# RGB image:
yuv_img = color.rgb2yuv(img)
im = yuv_img[:,:,0]
# Generate the Stroke Map:
S = gen_stroke_map(im, kernel_size, stroke_width=stroke_width, num_of_directions=num_of_directions,
smooth_kernel=smooth_kernel, gradient_method=gradient_method)
S = np.power(S, stroke_darkness)
# Generate the Tone Map:
J = gen_tone_map(im, w_group=w_group)
# Read the pencil texture:
if not pencil_texture_path:
pencil_texture = io.imread('./pencils/pencil0.jpg', as_gray=True)
else:
pencil_texture = io.imread(pencil_texture_path, as_gray=True)
# Generate the Pencil Texture Map:
T = gen_pencil_texture(im, pencil_texture, J)
T = np.power(T, tone_darkness)
# The final Y channel:
R = np.multiply(S, T)
if not rgb:
return R
else:
yuv_img[:,:,0] = R
return exposure.rescale_intensity(color.yuv2rgb(yuv_img), in_range=(0, 1))
def loadSeq(self, cam,person):
seqImgs = self.getSeqImgFiles(os.path.join(self.opt.dir_RGB,cam,person))
seqLen = len(seqImgs)
for i in range(0,seqLen):
filename = seqImgs[i]
img = sio.imread(os.path.join(self.opt.dir_RGB,cam,person,filename)) * 1.0
img = rgb2yuv(resize(img, (64, 48), mode='reflect'))
imgof = sio.imread(os.path.join(self.opt.dir_OF,cam,person,filename)) * 1.0
imgof = resize(imgof, (64, 48), mode='reflect')
if i == 0:
s = img.shape
imagePixelData = torch.zeros((seqLen, 5, s[0], s[1]),dtype=torch.float)
for c in range(0, 3):
d = torch.from_numpy(img[:, :, c])
m = torch.mean(d, axis=(0, 1))
v = torch.std(d, axis=(0, 1))
d = d - m
d = d / v
imagePixelData[i, c, :, :] = d
for c in range(0, 2):
d = torch.from_numpy(imgof[:, :, c])
m = torch.mean(d, axis=(0, 1))
v = torch.std(d, axis=(0, 1))
d = d - m
d = d / v
imagePixelData[i, c + 3, :, :] = d
return imagePixelData
def preprocess_image(x):
if __GLOBAL_PARAMS__['NORMALIZATION']:
x = (x - 128.0) / 128.0
if __GLOBAL_PARAMS__['YUV']:
x = np.array([rgb2yuv(x.reshape((1,180,180,3)))])
x = x.reshape((3,180,180))
return x
def convert(colorspace, image):
# Normalize data.
image = image.astype('float32') / 255
#conver to HSL
if colorspace == "HSL":
image = convertHSL(image)
#conver to HSV
if colorspace == "HSV":
image = color.rgb2hsv(image)
#conver to XYZ
if colorspace == "XYZ":
image = color.rgb2xyz(image)
#conver to LUV
if colorspace == "LUV":
image = color.rgb2luv(image)
#conver to LAB
if colorspace == "LAB":
image = color.rgb2lab(image)
#conver to YUV
if colorspace == "YUV":
image = color.rgb2yuv(image)
print("finish")
from scipy.misc import imsave
imsave('test.png', image)
return image
def moments(colln_imgs):
'''
Function to compute color moments of list of images
Input: List of length N, where N is the number of images
Output: Numpy Array of size (N, M), where N is the number of images and M = 1728
'''
win_h, win_w = 100, 100
colln_mnts = []
for img in colln_imgs:
img = rgb2yuv(img)
img_h, img_w = img.shape[0], img.shape[1]
mean, std_dev, skwn, mnts = [], [], [], []
for h in range(0, img_h-win_h+1, win_h):
for w in range(0, img_w-win_w+1, win_w):
win = img[h:h+win_h, w:w+win_w, :]
mean.append([np.mean(win[:,:,0]), np.mean(win[:,:,1]), np.mean(win[:,:,2])])
std_dev.append([np.std(win[:,:,0]), np.std(win[:,:,1]), np.std(win[:,:,2])])
skwn.append([skew(win[:,:,0], axis=None), skew(win[:,:,1], axis=None), skew(win[:,:,2], axis=None)])
mnts += [mean, std_dev, skwn]
colln_mnts.append(mnts)
colln_mnts = np.array(colln_mnts)
w,x,y,z = colln_mnts.shape
colln_mnts = colln_mnts.reshape((w,x*y*z))
return colln_mnts
def loadSequenceImages(self, cameraDir, opticalflowDir, filesList):
nImgs = len(filesList)
for i, file in enumerate(filesList):
filename = cameraDir + '/' + file
filenameOF = opticalflowDir + '/' + file
img = sio.imread(filename) * 1.0
img = rgb2yuv(resize(img, (64, 48), mode='reflect'))
imgof = sio.imread(filenameOF) * 1.0
imgof = resize(imgof, (64, 48), mode='reflect')
if i == 0:
s = img.shape
imagePixelData = np.zeros((nImgs, 5, s[0], s[1]),
dtype=np.float64)
for c in xrange(0, 3):
d = img[:, :, c]
d = d
m = np.mean(d, axis=(0, 1))
v = np.std(d, axis=(0, 1))
d = d - m
d = d / v
imagePixelData[i, c, :, :] = d
for c in xrange(0, 2):
d = imgof[:, :, c]
d = d
m = np.mean(d, axis=(0, 1))
v = np.std(d, axis=(0, 1))
d = d - m
d = d / v
imagePixelData[i, c + 3, :, :] = d
return imagePixelData
def preproc(data, normalize=False, flip=False, mean_image=None, outType='YUV'):
data_size = data.shape[0]
img_size = int(data.shape[1] / 3)
if normalize:
if mean_image is None:
mean_image = np.mean(data)
mean_image = mean_image / np.float32(255)
data = (data - mean_image) / np.float32(255)
data_RGB = np.dstack((data[:, :img_size], data[:, img_size:2 * img_size], data[:, 2 * img_size:]))
data_RGB = data_RGB.reshape((data_size, int(np.sqrt(img_size)), int(np.sqrt(img_size)), 3))
if flip:
data_RGB = data_RGB[0:data_size, :, :, :]
data_RGB_flip = data_RGB[:, :, :, ::-1]
data_RGB = np.concatenate((data_RGB, data_RGB_flip), axis=0)
if outType == 'YUV':
data_out = color.rgb2yuv(data_RGB)
return data_out, data_RGB # returns YUV as 4D tensor and RGB as 4D tensor
elif outType == 'LAB':
data_out = color.rgb2lab(data_RGB)
data_gray = color.rgb2gray(data_RGB)[:, :, :, None]
return data_out, data_gray # returns LAB and grayscale as 4D tensor
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 find_setups(init_img, show=False):
"""
Finds the 6 frames in an image by using the YUV colorspace
Args:
init_img: filename of first image
show: boolean flag for plotting of identified frames
Returns:
List of frame corners coordinates
"""
setups = []
img = mpimg.imread(f'{init_img}')
img = color.rgb2yuv(img)[:, :, 1]
img = filters.gaussian(img, 10)
t = filters.threshold_otsu(img)
img = img > t
label = measure.label(img)
regions = measure.regionprops(label)
avg = []
# detection might add small non-panel areas and these must be removed
for r in regions:
avg.append(r.area)
avg = np.array(avg).mean()
for r in regions:
if r.area > 0.9*avg:
setups.append([r.bbox[0]+10, r.bbox[2]-10, r.bbox[1]+10, r.bbox[3]-10])
if show:
fig, ax = plt.subplots(figsize=(6, 6))
ax.imshow(img[r.bbox[0]+10:r.bbox[2]-10, r.bbox[1]+10:r.bbox[3]-10])
ax.grid(False)
plt.show()
return setups
def convert_from_rgb2yuv(x):
device = x.device
x = x.cpu()
x = x.permute(0, 2, 3, 1).numpy()
x = rgb2yuv(x)
x = torch.from_numpy(x).float().permute(0, 3, 1, 2)
x = x.to(device)
return x
def preprocess_image(image):
# image = tf.image.decode_png(image, channels=3)
# image = tf.image.resize(image, [FLAGS.height, FLAGS.width])
# image_float = tf.image.convert_image_dtype(image, tf.float32)
# image_yuv = tf.image.rgb_to_yuv(image_float)
image = resize(image, (FLAGS.height, FLAGS.width))
image_yuv = rgb2yuv(image)
return image_yuv
def __getitem__(self, index):
img = Image.open(self.files[index])
yuv = rgb2yuv(img)
y = yuv[..., 0] - 0.5
u_t = yuv[..., 1] / 0.43601035
v_t = yuv[..., 2] / 0.61497538
return torch.Tensor(np.expand_dims(y, axis=0)), torch.Tensor(
np.stack([u_t, v_t], axis=0))
def __init__(self, img, criterion=None, min_nrows=4, min_ncols=4):
self.img = img
self.criterion = criterion if criterion is not None else lambda img: img.max() - img.min() > 0.25
self.root = None
self.ychan = rgb2yuv(img)[:, :, 0]
self.ychan = (self.ychan - self.ychan.min()) / (self.ychan.max() - self.ychan.min())
self.min_nrows = min_nrows
self.min_ncols = min_ncols
def load_extra_data(outType='YUV'):
names = np.array(glob.glob('extra/*.jpg'))
files = np.array([imread(f) for f in names])
if outType == 'YUV':
return color.rgb2yuv(files), files
elif outType == 'LAB':
return color.rgb2lab(files), color.rgb2gray(files)[:, :, :, None]
def scale_image(x):
# Extract the Y component after converting to YUV
x_yuv = rgb2yuv(x)
x_y = x_yuv[:, :, 0]
# Scale to input size for convnet
temp = resize(x_y, (84, 84), interpolation=INTER_LINEAR)
scipy.misc.toimage(temp).save('outfile.jpg')
return resize(x_y, (84, 84), interpolation=INTER_LINEAR)
def preprocess(self, input):
input = np.asarray(input)
if time.time() - self.time > self.fps:
self.time = time.time()
self.vis.image(input.transpose((2, 0, 1)), win=0)
input = resize(input, (self.args.im_size, self.args.im_size), mode='reflect')
input = rgb2yuv(input)[:,:,0]
return input
def get_data(vimagedata):
from sklearn.feature_extraction.image import extract_patches
vdataname = vimagedata.pop()
data = img_as_float(imread(vdataname))
data = rgb2yuv(data)[:, :, 0:1]
data_ik = extract_patches(data, (high_dim, high_dim, 1),
(high_dim // 2, high_dim // 2,
1)).reshape([-1] +
list((high_dim, high_dim, 1)))
return data_ik
def pencil_draw_color(I, texture):
if len(I.shape) == 2:
return pencil_draw(I, texture)
else:
I_yuv = color.rgb2yuv(I)
Y_gray = pencil_draw(I_yuv[:, :, 0], texture)
I_yuv[:, :, 0] = Y_gray
I_after = color.yuv2rgb(I_yuv)
I_after = np.maximum(I_after, 0)
I_after = np.minimum(I_after, 1)
#I_ruv[:, :, 0] = 0.5
return I_after
def alterYUV(img):
img = color.rgb2yuv(img)
params = []
#Y
img[:, :, 0] += randUnifC(-0.05, 0.05, params=params)
#U
img[:, :, 1] += randUnifC(-0.02, 0.02, params=params)
#V
img[:, :, 2] += randUnifC(-0.02, 0.02, params=params)
# U & V channels can have negative values; clip only Y
img[:, :, 0] = np.clip(img[:, :, 0], 0, 1.0)
img = color.yuv2rgb(img)
img = np.clip(img, 0, 1.0)
return img
def inference(G, in_path, out_path):
p = Image.open(in_path).convert('RGB')
W, H = p.size
dest_yuv = rgb2yuv(p)
dest_img = np.expand_dims(np.expand_dims(dest_yuv[..., 0], axis=0), axis=0)
p.thumbnail((args.res, args.res))
img_yuv = rgb2yuv(p)
#H,W,_ = img_yuv.shape
print('size:' + str((p.size)))
infimg = np.expand_dims(np.expand_dims(img_yuv[..., 0], axis=0), axis=0)
img_variable = Variable(torch.Tensor(infimg - 0.5))
if args.gpu >= 0:
img_variable = img_variable.cuda(args.gpu)
res = G(img_variable)
uv = res.cpu().detach().numpy()
uv[:, 0, :, :] *= 0.436
uv[:, 1, :, :] *= 0.615
(_, _, H1, W1) = uv.shape
print('size out:' + str((W1, H1)))
uv = zoom(uv, (1, 1, H / H1, W / W1))
yuv = np.concatenate([dest_img, uv], axis=1)[0]
rgb = yuv2rgb(yuv.transpose(1, 2, 0))
cv2.imwrite(out_path, (rgb.clip(min=0, max=1) * 256)[:, :, [2, 1, 0]])
def dir_data_generator(dir, batch_size, data_range=(0, 0), outType='YUV'):
names = np.array(glob.glob(dir + '/*.jpg'))
if data_range != (0, 0):
names = names[data_range[0]:data_range[1]]
batch_count = len(names) // batch_size
while True:
for i in range(0, batch_count):
files = np.array([imread(f) for f in names[i * batch_size:i * batch_size + batch_size]])
if outType == 'YUV':
yield color.rgb2yuv(files), files
elif outType == 'LAB':
yield color.rgb2lab(files), color.rgb2gray(files)[:, :, :, None]
def hog_feature(img, feature_vectore=True):
"""Return the HOG features based on YUV channels of the given image"""
img = img_as_ubyte(color.rgb2yuv(img))
# parameters to tune
orient = 9
pix_per_cell = 8
cell_per_block = 2
features = [feature.hog(img[:, :, channel], orientations=orient, pixels_per_cell=(pix_per_cell, pix_per_cell),
cells_per_block=(cell_per_block, cell_per_block), feature_vector=feature_vectore)
for channel in range(3)]
if feature_vectore:
return np.concatenate(features)
else:
return features
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 convolver_rgb(image, kernel, iterations=1):
img_yuv = rgb2yuv(image)
img_yuv[:, :, 0] = multi_convolver(img_yuv[:, :, 0], kernel, iterations)
final_image = yuv2rgb(img_yuv)
fig, ax = plt.subplots(1, 2, figsize=(17, 10))
ax[0].imshow(image)
ax[0].set_title(f'Original', fontsize=20)
ax[1].imshow(final_image)
ax[1].set_title(f'YUV Adjusted, Iterations = {iterations}', fontsize=20)
[axi.set_axis_off() for axi in ax.ravel()]
fig.tight_layout()
return final_image
def generate_layers_simple(config, back_list, objects_all, num_objects_list):
max_objects = max(num_objects_list)
num_objects = np.random.choice(num_objects_list)
idx_back = np.random.randint(0, len(back_list))
back = back_list[idx_back]
image_shape = back.shape
obj_layers = np.zeros((max_objects, *image_shape), dtype=np.uint8)
occ_masks = np.zeros((num_objects, *image_shape[:-1]), dtype=np.float)
indices = np.random.randint(len(objects_all), size=num_objects)
scales = np.random.uniform(config['range_scale'][0],
config['range_scale'][1],
size=num_objects)
obj_images = [
rescale_image(objects_all[idx][0], scale)
for idx, scale in zip(indices, scales)
]
if 'range_intensity' in config:
y_scales = np.random.uniform(config['range_intensity'][0],
config['range_intensity'][1],
size=num_objects)
yuv_images = [rgb2yuv(image[..., :-1]) for image in obj_images]
for image, y_scale in zip(yuv_images, y_scales):
image[..., 0] *= y_scale
rgb_images = [(np.clip(yuv2rgb(image), 0, 1) * 255).astype(np.uint8)
for image in yuv_images]
obj_images = [
np.concatenate([rgb, obj[..., -1:]], axis=-1)
for rgb, obj in zip(rgb_images, obj_images)
]
obj_classes = np.array([objects_all[idx][1] for idx in indices])
for idx, image in enumerate(obj_images):
row1 = np.random.randint(image_shape[0] - image.shape[0] + 1)
row2 = row1 + image.shape[0]
col1 = np.random.randint(image_shape[1] - image.shape[1] + 1)
col2 = col1 + image.shape[1]
obj_layers[idx, row1:row2, col1:col2] = image
occ_masks[idx, row1:row2,
col1:col2] = 1 if config['overlap_bbox'] else image[...,
-1] / 255
layers = np.concatenate([back[None], obj_layers])
layers = np.rollaxis(layers, -1, -3)
classes = np.full(max_objects + 1, -1, dtype=np.int8)
classes[1:obj_classes.shape[0] + 1] = obj_classes
return layers, classes, occ_masks