def data_generator_landmarks(dataCSV,
selection,
batchSize,
pimgSize=256,
augment=True):
aug = Augmenter()
numSamples = len(dataCSV)
pathImgs = np.array(dataCSV['img'])
pathKps = np.array(dataCSV['keypoints'])
idxRange = np.array(range(numSamples))
while True:
rndIdx = np.random.permutation(idxRange)[:batchSize]
dataX = None
dataY = None
for ii, iidx in enumerate(rndIdx):
pimg = pathImgs[iidx]
pkp = pathKps[iidx]
timg = read_img(pimg, pimgSize).astype(np.uint8)
tkp = read_keypoints_synth(pkp, selection) * 256
timg, tkp = utils.fixbb(timg, tkp)
timg = utils.background_blend(timg)
tkp = tkp / 256
if np.random.rand() < 0.5: # we always augment flips
timg, tkp = utils.fliplr(timg, tkp)
tkp = (tkp * 256).astype(np.int32)
timg = utils.to_uint8(timg)
if dataX is None:
dataX = np.zeros([batchSize] + list(timg.shape),
dtype=np.uint8)
dataY = np.zeros([batchSize] + list(tkp.shape), dtype=np.int32)
dataX[ii] = timg
dataY[ii] = tkp
# augment and flatten dataY
if augment:
keypoints = [
ia.KeypointsOnImage.from_coords_array(ptsi, (256, 256))
for ptsi in dataY
]
seq_geom = aug.seq_geom.to_deterministic()
seq_color = aug.seq_color.to_deterministic()
dataX, keypoints = list(
seq_geom.augment_batches(batches=[dataX, keypoints],
background=False))
dataX, = list(
seq_color.augment_batches(batches=[dataX], background=False))
dataX = dataX.astype(np.float32) / 127.5 - 1.0
dataY = np.array([kp.get_coords_array() for kp in keypoints])
else:
dataX = dataX.astype(np.float32) / 127.5 - 1.0
yield dataX, dataY.reshape((batchSize, -1)).astype(np.float32) / 256
def data_generator_segmentation(dataCSV,
numCls,
batchSize=4,
pimgSize=256,
augment=False):
aug = Augmenter()
numSamples = len(dataCSV)
pathImgs = np.array(dataCSV['img'])
pathMsks = np.array(dataCSV['mask'])
idxRange = np.array(range(numSamples))
while True:
rndIdx = np.random.permutation(idxRange)[:batchSize]
dataX = None
dataY = None
for ii, iidx in enumerate(rndIdx):
pimg = pathImgs[iidx]
pmsk = pathMsks[iidx]
timg = read_img(pimg, pimgSize)
timg = utils.background_blend(timg)
timg = utils.to_uint8(timg)
tmsk = read_msk(pmsk, pimgSize)[..., :3]
if dataX is None:
dataX = np.zeros([batchSize] + list(timg.shape),
dtype=np.uint8)
dataY = np.zeros([batchSize] + list(tmsk.shape),
dtype=np.uint8)
dataX[ii] = timg
dataY[ii] = tmsk
# augment and flatten dataY
if augment:
seq_geom = aug.seq_geom.to_deterministic()
seq_color = aug.seq_color.to_deterministic()
#seq_geom.augment_batches()
dataX, dataY = list(
seq_geom.augment_batches(batches=[dataX, dataY],
background=False))
dataX, = list(
seq_color.augment_batches(batches=[dataX], background=False))
dataY = labels_from_colored(dataY)
dataY = np_utils.to_categorical(dataY, numCls).reshape(
batchSize, pimgSize * pimgSize, numCls)
dataX = dataX.astype(np.float32) / 127.5 - 1.0
else:
dataY = labels_from_colored(dataY)
dataY = np_utils.to_categorical(dataY, numCls).reshape(
batchSize, pimgSize * pimgSize, numCls)
dataX = dataX.astype(np.float32) / 127.5 - 1.0
yield dataX, dataY
def data_generator_landmarks(dataCSV, selection, batchSize, pimgSize=256, augment=True):
aug = Augmenter()
numSamples = len(dataCSV)
pathImgs = np.array(dataCSV['img'])
pathKps = np.array(dataCSV['keypoints'])
idxRange = np.array(range(numSamples))
while True:
rndIdx = np.random.permutation(idxRange)[:batchSize]
dataX = None
dataY = None
for ii, iidx in enumerate(rndIdx):
pimg = pathImgs[iidx]
pkp = pathKps[iidx]
timg = read_img(pimg, pimgSize).astype(np.uint8)
tkp = read_keypoints_synth(pkp, selection)*256
timg, tkp = utils.fixbb(timg, tkp)
timg = utils.background_blend(timg)
tkp = tkp/256
if np.random.rand() < 0.5: # we always augment flips
timg, tkp = utils.fliplr(timg, tkp)
tkp = (tkp*256).astype(np.int32)
timg = utils.to_uint8(timg)
if dataX is None:
dataX = np.zeros([batchSize] + list(timg.shape), dtype=np.uint8)
dataY = np.zeros([batchSize] + list(tkp.shape), dtype=np.int32)
dataX[ii] = timg
dataY[ii] = tkp
# augment and flatten dataY
if augment:
keypoints = [ia.KeypointsOnImage.from_coords_array(ptsi, (256, 256)) for ptsi in dataY]
seq_geom = aug.seq_geom.to_deterministic()
seq_color = aug.seq_color.to_deterministic()
dataX, keypoints = list(seq_geom.augment_batches(batches=[dataX, keypoints], background=False))
dataX, = list(seq_color.augment_batches(batches=[dataX], background=False))
dataX = dataX.astype(np.float32) / 127.5 - 1.0
dataY = np.array([kp.get_coords_array() for kp in keypoints])
else:
dataX = dataX.astype(np.float32) / 127.5 - 1.0
yield dataX, dataY.reshape((batchSize,-1)).astype(np.float32)/256
def data_generator_bgfg(dataCSV, batchSize=4, pimgSize=256, augment=False):
numCls=2
aug = Augmenter()
numSamples = len(dataCSV)
pathImgs = np.array(dataCSV['img'])
pathMsks = np.array(dataCSV['mask'])
idxRange = np.array(range(numSamples))
while True:
rndIdx = np.random.permutation(idxRange)[:batchSize]
dataX = None
dataY = None
for ii, iidx in enumerate(rndIdx):
pimg = pathImgs[iidx]
pmsk = pathMsks[iidx]
timg = read_img(pimg, pimgSize)
timg = utils.background_blend(timg)
timg = utils.to_uint8(timg)
tmsk = read_msk(pmsk, pimgSize)[...,:3]
if dataX is None:
dataX = np.zeros([batchSize] + list(timg.shape), dtype=np.uint8)
dataY = np.zeros([batchSize] + list(tmsk.shape), dtype=np.uint8)
dataX[ii] = timg
dataY[ii] = tmsk
# augment and flatten dataY
if augment:
seq_geom = aug.seq_geom.to_deterministic()
seq_color = aug.seq_color.to_deterministic()
#seq_geom.augment_batches()
dataX, dataY = list(seq_geom.augment_batches(batches=[dataX, dataY], background=False))
dataX, = list(seq_color.augment_batches(batches=[dataX], background=False))
dataY = labels_bg_fg(dataY)
dataY = np_utils.to_categorical(dataY, numCls).reshape(batchSize, pimgSize*pimgSize, numCls)
dataX = dataX.astype(np.float32)/127.5 - 1.0
else:
dataY = labels_bg_fg(dataY)
dataY = np_utils.to_categorical(dataY, numCls).reshape(batchSize, pimgSize * pimgSize, numCls)
dataX = dataX.astype(np.float32) / 127.5 - 1.0
yield dataX, dataY
for row, col in seed_points:
segmented[row, col] = True
neighbourhood(im, segmented, row, col, T)
return segmented
### END YOUR CODE HERE ###
if __name__ == "__main__":
# DO NOT CHANGE
im = utils.read_image("defective-weld.png")
seed_points = [ # (row, column)
[254, 138], # Seed point 1
[253, 296], # Seed point 2
[233, 436], # Seed point 3
[232, 417], # Seed point 4
]
intensity_threshold = 50
segmented_image = region_growing(im, seed_points, intensity_threshold)
assert im.shape == segmented_image.shape, \
"Expected image shape ({}) to be same as thresholded image shape ({})".format(
im.shape, segmented_image.shape)
assert segmented_image.dtype == np.bool, \
"Expected thresholded image dtype to be np.bool. Was: {}".format(
segmented_image.dtype)
segmented_image = utils.to_uint8(segmented_image)
utils.save_im("defective-weld-segmented.png", segmented_image)
args:
im: np.ndarray of shape (H, W) with boolean values (dtype=np.bool)
return:
(np.ndarray) of shape (H, W). dtype=np.bool
"""
### START YOUR CODE HERE ### (You can change anything inside this block)
# You can also define other helper functions
structuring_element = np.array([[1, 1, 1], [1, 1, 1], [1, 1, 1]],
dtype=bool)
eroded_im = binary_erosion(im, selem=structuring_element)
boundary = np.bitwise_xor(im, eroded_im)
return boundary
### END YOUR CODE HERE ###
if __name__ == "__main__":
im = utils.read_image("lincoln.png")
binary_image = (im != 0)
boundary = extract_boundary(binary_image)
assert im.shape == boundary.shape, \
"Expected image shape ({}) to be same as resulting image shape ({})".format(
im.shape, boundary.shape)
assert boundary.dtype == np.bool, \
"Expected resulting image dtype to be np.bool. Was: {}".format(
boundary.dtype)
boundary = utils.to_uint8(boundary)
utils.save_im("lincoln-boundary.png", boundary)
Returns:
im: np.array of shape [H, W]
"""
laplacian = np.array([[0, -1, 0], [-1, 4, -1], [0, -1, 0]])
### START YOUR CODE HERE ### (You can change anything inside this block)
# Convolve img with laplacian kernel
convolved_im = convolve_im(im, laplacian, verbose=True)
# Use equation 6.
resulting_im = np.add(im, (im * convolved_im))
# Limit values between [0,1]
im = np.add(resulting_im, resulting_im.min())
im = np.multiply(im, 1 / im.max())
### END YOUR CODE HERE ###
return im
if __name__ == "__main__":
# DO NOT CHANGE
im = skimage.data.moon()
im = utils.uint8_to_float(im)
sharpen_im = sharpen(im)
sharpen_im = utils.to_uint8(sharpen_im)
im = utils.to_uint8(im)
# Concatenate the image, such that we get
# the original on the left side, and the sharpened on the right side
im = np.concatenate((im, sharpen_im), axis=1)
utils.save_im("moon_sharpened.png", im)
A function that computes the distance to the closest boundary pixel.
args:
im: np.ndarray of shape (H, W) with boolean values (dtype=np.bool)
return:
(np.ndarray) of shape (H, W). dtype=np.int32
"""
# START YOUR CODE HERE ### (You can change anything inside this block)
# You can also define other helper functions
assert im.dtype == np.bool
structuring_element = np.array([[1, 1, 1], [1, 1, 1], [1, 1, 1]],
dtype=bool)
result = im.astype(np.int32)
return result
### END YOUR CODE HERE ###
if __name__ == "__main__":
im = utils.read_image("noisy.png")
binary_image = (im != 0)
noise_free_image = remove_noise(binary_image)
distance = distance_transform(noise_free_image)
assert im.shape == distance.shape, "Expected image shape ({}) to be same as resulting image shape ({})".format(
im.shape, distance.shape)
assert distance.dtype == np.int32, "Expected resulting image dtype to be np.int32. Was: {}".format(
distance.dtype)
distance = utils.to_uint8(distance)
utils.save_im("noisy-distance.png", distance)
(np.ndarray) of shape (H, W). dtype=np.bool
"""
### START YOUR CODE HERE ### (You can change anything inside this block)
# You can also define other helper functions
sh = disk(10)
binary_closing(im, selem=sh, out = im)
binary_opening(im, selem=sh, out = im)
# im = binary_erosion(im, selem=sh)
# im = binary_dilation(im, selem=sh)
return im
### END YOUR CODE HERE ###
if __name__ == "__main__":
# DO NOT CHANGE
im = utils.read_image("noisy.png")
binary_image = (im != 0)
noise_free_image = remove_noise(binary_image)
assert im.shape == noise_free_image.shape, \
"Expected image shape ({}) to be same as resulting image shape ({})".format(
im.shape, noise_free_image.shape)
assert noise_free_image.dtype == np.bool, \
"Expected resulting image dtype to be np.bool. Was: {}".format(
noise_free_image.dtype)
noise_free_image = utils.to_uint8(noise_free_image)
utils.save_im("noisy-filtered.png", noise_free_image)
def imread(filename, as_gray = False):
image = skio.imread(filename, as_gray = as_gray)
if image.dtype == np.float64 :
image = utils.to_uint8(image)
return image
(np.ndarray) of shape (H, W). dtype=np.bool
"""
# START YOUR CODE HERE ### (You can change anything inside this block)
# You can also define other helper functions
structuring_element = np.array([[1, 1, 1], [1, 1, 1], [1, 1, 1]],
dtype=bool)
result = im.copy()
return result
### END YOUR CODE HERE ###
if __name__ == "__main__":
im = utils.read_image("balls-with-reflections.png")
binary_image = im != 0
starting_points = [ # (row, column)
[51, 64], [44, 180], [35, 365], [156, 94], [141, 264], [138, 467],
[198, 180], [229, 413], [294, 103], [302, 230], [368, 388], [352, 489],
[454, 57], [457, 236], [469, 400], [489, 506]
]
num_iterations = 30
result = fill_holes(binary_image, starting_points, num_iterations)
assert im.shape == result.shape, "Expected image shape ({}) to be same as resulting image shape ({})".format(
im.shape, result.shape)
assert result.dtype == np.bool, "Expected resulting image dtype to be np.bool. Was: {}".format(
result.dtype)
result = utils.to_uint8(result)
utils.save_im("balls-with-reflections-filled.png", result)
import utils
import matplotlib.pyplot as plt
import utils
if __name__ == '__main__':
#filename = '../images/color/rgb_cube.png'
#filename = '../images/color/fichas.jpg'
filename = '/home/jsaavedr/Documents/Docencia/2020/CC5508/tareas/tarea3_2020/tarea3/im3.jpg'
image = pai_io.imread(filename)
print('shape: {}'.format(image.shape))
im_red = image[:, :, 0]
im_green = image[:, :, 1]
im_blue = image[:, :, 2]
#yellow
im_yellow = im_red.astype(np.float64) + im_green.astype(
np.float64) - im_blue.astype(np.float64)
im_yellow = utils.to_uint8(im_yellow / (255.0 + 255.0))
##showing image
fig, ax = plt.subplots(2, 3)
ax[0, 1].imshow(image)
ax[0, 1].set_title('image')
ax[1, 0].imshow(im_red, cmap='gray')
ax[1, 0].set_title('Red')
ax[1, 1].imshow(im_green, cmap='gray')
ax[1, 1].set_title('Green')
ax[1, 2].imshow(im_yellow, cmap='gray')
ax[1, 2].set_title('Yellow')
for i in range(2):
for j in range(3):
ax[i, j].set_axis_off()
plt.show()
