def shuffle(im, lb, num, rotate= False):
map = {}
rows=cols=num
blk_size=im.shape[0]//rows
img_blks=view_as_blocks(im,block_shape=(blk_size,blk_size,3)).reshape((-1,blk_size,blk_size,3))
lbl_blks=view_as_blocks(lb,block_shape=(blk_size,blk_size)).reshape((-1,blk_size,blk_size))
img_shuff=np.zeros((im.shape[0],im.shape[1],3),dtype=np.uint8)
lbl_shuff=np.zeros((lb.shape[0],lb.shape[1]),dtype=np.uint8)
lbl_rotn=np.zeros((lb.shape[0]//blk_size,lb.shape[1]//blk_size),dtype=np.uint8)
a=np.arange(rows*rows, dtype=np.uint8)
b=np.random.permutation(a)
map = {k:v for k,v in zip(a,b)}
print ("Key Map:-\n" + str(map))
for i in range(0,rows):
for j in range(0,cols):
x,y = i*blk_size, j*blk_size
if(rotate):
rot_val=random.randrange(0,4)
lbl_rotn[i,j]=rot_val
img_shuff[x:x+blk_size, y:y+blk_size] = np.rot90(img_blks[map[i*rows + j]],rot_val)
lbl_shuff[x:x+blk_size, y:y+blk_size] = lbl_blks[map[i*rows + j]]
else:
img_shuff[x:x+blk_size, y:y+blk_size] = img_blks[map[i*rows + j]]
lbl_shuff[x:x+blk_size, y:y+blk_size] = lbl_blks[map[i*rows + j]]
return img_shuff,lbl_shuff,lbl_rotn
def shuffle_image(im, num):
map = {}
rows = cols = num
blk_size = im.shape[0] // rows
img_blks = view_as_blocks(im, block_shape=(blk_size, blk_size, 3)).reshape(
(-1, blk_size, blk_size, 3))
print("img_blks.shape: ", img_blks.shape)
img_shuff = np.zeros((im.shape[0], im.shape[1], 3), dtype=np.uint8)
print("img_shuff.shape: ", img_shuff.shape)
a = np.arange(rows * rows, dtype=np.uint8)
b = np.random.permutation(a)
map = {k: v for k, v in zip(a, b)}
print("Key Map:-\n" + str(map))
for i in range(0, rows):
for j in range(0, cols):
x, y = i * blk_size, j * blk_size
img_shuff[x:x + blk_size,
y:y + blk_size] = img_blks[map[i * rows + j]]
shuf_img_blks = view_as_blocks(img_shuff,
block_shape=(blk_size, blk_size,
3)).reshape((-1, blk_size,
blk_size, 3))
print("shuf_img_blks.shape: ", shuf_img_blks.shape)
return img_shuff, img_blks, map, blk_size, shuf_img_blks
def hog_descriptor(patch, pixels_per_cell=(8, 8)):
"""
Generating hog descriptor by the following steps:
1. Compute the gradient image in x and y directions (already done for you)
2. Compute gradient histograms for each cell
3. Flatten block of histograms into a 1D feature vector
Here, we treat the entire patch of histograms as our block
4. Normalize flattened block
Normalization makes the descriptor more robust to lighting variations
Args:
patch: grayscale image patch of shape (H, W)
pixels_per_cell: size of a cell with shape (M, N)
Returns:
block: 1D patch descriptor array of shape ((H*W*n_bins)/(M*N))
"""
assert (patch.shape[0] % pixels_per_cell[0] == 0),\
'Heights of patch and cell do not match'
assert (patch.shape[1] % pixels_per_cell[1] == 0),\
'Widths of patch and cell do not match'
n_bins = 9
degrees_per_bin = 180 // n_bins
Gx = filters.sobel_v(patch)
Gy = filters.sobel_h(patch)
# Unsigned gradients
G = np.sqrt(Gx**2 + Gy**2)
theta = (np.arctan2(Gy, Gx) * 180 / np.pi).astype(int) % 180
# Group entries of G and theta into cells of shape pixels_per_cell, (M, N)
# G_cells.shape = theta_cells.shape = (H//M, W//N)
# G_cells[0, 0].shape = theta_cells[0, 0].shape = (M, N)
G_cells = view_as_blocks(G, block_shape=pixels_per_cell)
theta_cells = view_as_blocks(theta, block_shape=pixels_per_cell)
rows = G_cells.shape[0]
cols = G_cells.shape[1]
# For each cell, keep track of gradient histrogram of size n_bins
cells = np.zeros((rows, cols, n_bins))
# Compute histogram per cell
for i in range(rows):
for j in range(cols):
G_patch = G_cells[i, j]
theta_patch = theta_cells[i, j]
for m in range(pixels_per_cell[0]):
for n in range(pixels_per_cell[1]):
bin_idx = int(theta_patch[m, n] / degrees_per_bin)
# print(theta_patch[m, n], bin_idx)
cells[i, j, bin_idx] += G_patch[m, n]
block = cells.flatten()
block = block / np.linalg.norm(block)
return block
def hog_descriptor(patch, pixels_per_cell=(8, 8)):
"""
Generating hog descriptor by the following steps:
1. Compute the gradient image in x and y directions (already done for you)
2. Compute gradient histograms for each cell
3. Flatten block of histograms into a 1D feature vector
Here, we treat the entire patch of histograms as our block
4. Normalize flattened block
Normalization makes the descriptor more robust to lighting variations
Args:
patch: grayscale image patch of shape (H, W)
pixels_per_cell: size of a cell with shape (M, N)
Returns:
block: 1D patch descriptor array of shape ((H*W*n_bins)/(M*N))
"""
assert (patch.shape[0] % pixels_per_cell[0] == 0),\
'Heights of patch and cell do not match'
assert (patch.shape[1] % pixels_per_cell[1] == 0),\
'Widths of patch and cell do not match'
n_bins = 9
degrees_per_bin = 180 // n_bins
Gx = filters.sobel_v(patch)
Gy = filters.sobel_h(patch)
# Unsigned gradients
G = np.sqrt(Gx**2 + Gy**2)
theta = (np.arctan2(Gy, Gx) * 180 / np.pi) % 180
# Group entries of G and theta into cells of shape pixels_per_cell, (M, N)
# G_cells.shape = theta_cells.shape = (H//M, W//N)
# G_cells[0, 0].shape = theta_cells[0, 0].shape = (M, N)
G_cells = view_as_blocks(G, block_shape=pixels_per_cell)
theta_cells = view_as_blocks(theta, block_shape=pixels_per_cell)
rows = G_cells.shape[0]
cols = G_cells.shape[1]
# For each cell, keep track of gradient histrogram of size n_bins
cells = np.zeros((rows, cols, n_bins))
# Compute histogram per cell
# YOUR CODE HERE
for i in range(rows):
for j in range(cols):
cell_hist = np.histogram(theta_cells[i, j].flatten(),
bins=n_bins,
range=(0, 180),
weights=G_cells[i, j].flatten())[0]
cells[i, j, :] = cell_hist
# normalize the HoG
cells = (cells - np.mean(cells)) / np.std(cells, ddof=1)
block = cells.flatten()
# YOUR CODE HERE
return block
def hog_descriptor(patch, pixels_per_cell=(8, 8)):
"""
Generating hog descriptor by the following steps:
1. compute the gradient image in x and y (already done for you)
2. compute gradient histograms
3. normalize across block
4. flattening block into a feature vector
Args:
patch: grayscale image patch of shape (h, w)
pixels_per_cell: size of a cell with shape (m, n)
Returns:
block: 1D array of shape ((h*w*n_bins)/(m*n))
"""
assert (patch.shape[0] % pixels_per_cell[0] == 0),\
'Heights of patch and cell do not match'
assert (patch.shape[1] % pixels_per_cell[1] == 0),\
'Widths of patch and cell do not match'
n_bins = 9
degrees_per_bin = 180 // n_bins
Gx = filters.sobel_v(patch)
Gy = filters.sobel_h(patch)
# Unsigned gradients
G = np.sqrt(Gx**2 + Gy**2)
theta = (np.arctan2(Gy, Gx) * 180 / np.pi) % 180
G_cells = view_as_blocks(G, block_shape=pixels_per_cell)
theta_cells = view_as_blocks(theta, block_shape=pixels_per_cell)
rows = G_cells.shape[0]
cols = G_cells.shape[1]
cells = np.zeros((rows, cols, n_bins))
o = pixels_per_cell[0]
p = pixels_per_cell[1]
b = np.zeros((n_bins))
x_theta = np.zeros((o, p))
y_G = np.zeros((o, p))
block = np.zeros((patch.shape[0] * patch.shape[1] * n_bins / o / p))
block = []
for i in range(0, rows):
for j in range(0, cols):
x_theta = flat(theta_cells[i][j])
y_G = flat(G_cells[i][j])
x_theta = (x_theta / 20).astype(int)
for m in range(0, o * p):
b[x_theta[m]] += 1 * y_G[m]
pass
block.append(b)
block = np.array(block)
block = simple_descriptor(block)
# YOUR CODE HERE
return block
def hog_descriptor(patch, pixels_per_cell=(8, 8)):
"""
Generating hog descriptor by the following steps:
1. compute the gradient image in x and y (already done for you)
2. compute gradient histograms
3. normalize across block
4. flattening block into a feature vector
Args:
patch: grayscale image patch of shape (h, w)
pixels_per_cell: size of a cell with shape (m, n)
Returns:
block: 1D array of shape ((h*w*n_bins)/(m*n))
"""
assert (patch.shape[0] % pixels_per_cell[0] == 0),\
'Heights of patch and cell do not match'
assert (patch.shape[1] % pixels_per_cell[1] == 0),\
'Widths of patch and cell do not match'
n_bins = 9
degrees_per_bin = 180 // n_bins
Gx = filters.sobel_v(patch)
Gy = filters.sobel_h(patch)
# Unsigned gradients
G = np.sqrt(Gx**2 + Gy**2)
theta = (np.arctan2(Gy, Gx) * 180 / np.pi) % 180
G_cells = view_as_blocks(G, block_shape=pixels_per_cell)
theta_cells = view_as_blocks(theta, block_shape=pixels_per_cell)
rows = G_cells.shape[0]
cols = G_cells.shape[1]
cells = np.zeros((rows, cols, n_bins))
# Compute histogram per cell
### YOUR CODE HERE
for i in range(rows):
for j in range(cols):
for i1 in range(pixels_per_cell[0]):
for j1 in range(pixels_per_cell[1]):
Gt = G_cells[i, j, i1, j1]
Tt = theta_cells[i, j, i1, j1]
bin = int(Tt // degrees_per_bin) % 9
cells[i, j, bin] += Gt
mean = np.mean(cells)
std = np.std(cells)
if std == 0:
std = 1
cells = (cells - mean) / std
block = cells.flatten()
pass
### YOUR CODE HERE
return block
def hog_descriptor(patch, pixels_per_cell=(8, 8)):
"""
Generating hog descriptor by the following steps:
1. compute the gradient image in x and y (already done for you)
2. compute gradient histograms
3. normalize across block
4. flattening block into a feature vector
Args:
patch: grayscale image patch of shape (h, w)
pixels_per_cell: size of a cell with shape (m, n)
Returns:
block: 1D array of shape ((h*w*n_bins)/(m*n))
"""
assert (patch.shape[0] % pixels_per_cell[0] == 0),\
'Heights of patch and cell do not match'
assert (patch.shape[1] % pixels_per_cell[1] == 0),\
'Widths of patch and cell do not match'
n_bins = 9
degrees_per_bin = 180 // n_bins
Gx = filters.sobel_v(patch)
Gy = filters.sobel_h(patch)
# Unsigned gradients
G = np.sqrt(Gx**2 + Gy**2)
theta = (np.arctan2(Gy, Gx) * 180 / np.pi) % 180
G_cells = view_as_blocks(G, block_shape=pixels_per_cell)
theta_cells = view_as_blocks(theta, block_shape=pixels_per_cell)
rows = G_cells.shape[0]
cols = G_cells.shape[1]
cells = np.zeros((rows, cols, n_bins))
# Compute histogram per cell
# YOUR CODE HERE
for i in range(rows):
for j in range(cols):
idxs = theta_cells[i, j] // degrees_per_bin
idxs[idxs == 9] = 8
idxs = idxs.astype(int)
for m in range(G_cells.shape[2]):
for n in range(G_cells.shape[3]):
cells[i, j, idxs[m, n]] += G_cells[i, j, m, n]
# idxs = idxs.astype(int)
# cells[i, j, idxs] += G_cells[i, j, idxs]
# Nomalize across block
cells = (cells - np.mean(cells)) / np.std(cells)
block = np.ravel(cells)
# YOUR CODE HERE
return block
def hog_descriptor(patch, pixels_per_cell=(8, 8)):
"""
Generating hog descriptor by the following steps:
1. compute the gradient image in x and y (already done for you)
2. compute gradient histograms
3. normalize across block
4. flattening block into a feature vector
Args:
patch: grayscale image patch of shape (h, w)
pixels_per_cell: size of a cell with shape (m, n)
Returns:
block: 1D array of shape ((h*w*n_bins)/(m*n))
"""
assert (patch.shape[0] % pixels_per_cell[0] == 0),\
'Heights of patch and cell do not match'
assert (patch.shape[1] % pixels_per_cell[1] == 0),\
'Widths of patch and cell do not match'
n_bins = 9
degrees_per_bin = 180 // n_bins
Gx = filters.sobel_v(patch)
Gy = filters.sobel_h(patch)
# Unsigned gradients
G = np.sqrt(Gx**2 + Gy**2)
theta = (np.arctan2(Gy, Gx) * 180 / np.pi) % 180
G_cells = view_as_blocks(G, block_shape=pixels_per_cell)
theta_cells = view_as_blocks(theta, block_shape=pixels_per_cell)
rows = G_cells.shape[0]
cols = G_cells.shape[1]
cells = np.zeros((rows, cols, n_bins))
# Compute histogram per cell
### YOUR CODE HERE
for i in range(G_cells.shape[0]):
for j in range(G_cells.shape[1]):
for k in range(G_cells.shape[2]):
for l in range(G_cells.shape[3]):
#if int(theta_cells[i, j, k, l] / degrees_per_bin) == 9:
index = int(theta_cells[i, j, k, l] / degrees_per_bin) - 1
#else:
# index = int(theta_cells[i, j, k, l] / degrees_per_bin)
cells[i, j, index] += G_cells[i, j, k, l]
block = cells.flatten()
block = block / np.linalg.norm(block, ord=2)
# block = (block - np.min(block)) / (np.max(block) - np.min(block))
### YOUR CODE HERE
return block
def hog_descriptor(patch, pixels_per_cell=(8, 8)):
"""
Generating hog descriptor by the following steps:
1. Compute the gradient image in x and y directions (already done for you)
2. Compute gradient histograms for each cell
3. Flatten block of histograms into a 1D feature vector
Here, we treat the entire patch of histograms as our block
4. Normalize flattened block
Normalization makes the descriptor more robust to lighting variations
Args:
patch: grayscale image patch of shape (H, W)
pixels_per_cell: size of a cell with shape (M, N)
Returns:
block: 1D patch descriptor array of shape ((H*W*n_bins)/(M*N))
"""
assert (patch.shape[0] % pixels_per_cell[0] == 0),\
'Heights of patch and cell do not match'
assert (patch.shape[1] % pixels_per_cell[1] == 0),\
'Widths of patch and cell do not match'
n_bins = 9
degrees_per_bin = 180 // n_bins
Gx = filters.sobel_v(patch)
Gy = filters.sobel_h(patch)
G = np.sqrt(Gx**2 + Gy**2)
theta = (np.arctan2(Gy, Gx) * 180 / np.pi) % 180
g_C = view_as_blocks(G, block_shape=pixels_per_cell)
t_C = view_as_blocks(theta, block_shape=pixels_per_cell)
r = g_C.shape[0]
c = g_C.shape[1]
cells = np.zeros((r, r, n_bins))
### YOUR CODE HERE
for i in range(r):
for j in range(c):
for z in range(pixels_per_cell[0]):
for t in range(pixels_per_cell[1]):
d_lower = t_C[i][j][z][t] % degrees_per_bin
division_ratio = float(d_lower) / degrees_per_bin
lBN = (int(np.trunc(
t_C[i][j][z][t] / degrees_per_bin))) % 9
uBN = (lBN + 1) % 9
cells[i][j][uBN] += (division_ratio * g_C[i][j][z][t])
cells[i][j][lBN] += ((1 - division_ratio) *
g_C[i][j][z][t])
cells[i][j] = cells[i][j] / np.linalg.norm(cells[i][j])
block = cells.flatten()
### YOUR CODE HERE
return block
def hog_descriptor(patch, pixels_per_cell=(8, 8)):
"""
Generating hog descriptor by the following steps:
1. compute the gradient image in x and y (already done for you)
2. compute gradient histograms
3. normalize across block
4. flattening block into a feature vector
Args:
patch: grayscale image patch of shape (h, w)
pixels_per_cell: size of a cell with shape (m, n)
Returns:
block: 1D array of shape ((h*w*n_bins)/(m*n))
"""
assert (patch.shape[0] % pixels_per_cell[0] == 0),\
'Heights of patch and cell do not match'
assert (patch.shape[1] % pixels_per_cell[1] == 0),\
'Widths of patch and cell do not match'
n_bins = 9
degrees_per_bin = 180 // n_bins
Gx = filters.sobel_v(patch)
Gy = filters.sobel_h(patch)
# Unsigned gradients
G = np.sqrt(Gx**2 + Gy**2)
theta = (np.arctan2(Gy, Gx) * 180 / np.pi) % 180
G_cells = view_as_blocks(G, block_shape=pixels_per_cell)
theta_cells = view_as_blocks(theta, block_shape=pixels_per_cell)
rows = G_cells.shape[0]
cols = G_cells.shape[1]
cells = np.zeros((rows, cols, n_bins))
# Compute histogram per cell
### YOUR CODE HERE
m = pixels_per_cell[0]
n = pixels_per_cell[1]
for i in np.arange(rows):
for j in np.arange(cols):
G_cells[i][j] = G[m * i:m * (i + 1), n * j:n * (j + 1)]
theta_cells[i][j] = theta[m * i:m * (i + 1), n * j:n * (j + 1)]
G_cells_v = G_cells[i][j].flatten()
theta_cells_v = ((theta_cells[i][j] // 20) %
9).flatten().astype(int)
for k in np.arange(n_bins):
cells[i][j][k] = G_cells_v[theta_cells_v == k].sum()
cells[i][j] = (cells[i][j] - np.mean(cells[i][j])) / np.std(
cells[i][j])
block = cells.flatten()
### YOUR CODE HERE
return block
def hog_descriptor(patch, pixels_per_cell=(8,8)):
"""
Generating hog descriptor by the following steps:
1. compute the gradient image in x and y (already done for you)
2. compute gradient histograms
3. normalize across block
4. flattening block into a feature vector
Args:
patch: grayscale image patch of shape (h, w)
pixels_per_cell: size of a cell with shape (m, n)
Returns:
block: 1D array of shape ((h*w*n_bins)/(m*n))
"""
assert (patch.shape[0] % pixels_per_cell[0] == 0),\
'Heights of patch and cell do not match'
assert (patch.shape[1] % pixels_per_cell[1] == 0),\
'Widths of patch and cell do not match'
n_bins = 9
degrees_per_bin = 180 // n_bins
Gx = filters.sobel_v(patch)
Gy = filters.sobel_h(patch)
# Unsigned gradients
G = np.sqrt(Gx**2 + Gy**2)
theta = (np.arctan2(Gy, Gx) * 180 / np.pi + 1e-5) % 180
G_cells = view_as_blocks(G, block_shape=pixels_per_cell)
theta_cells = view_as_blocks(theta, block_shape=pixels_per_cell)
rows = G_cells.shape[0]
cols = G_cells.shape[1]
cells = np.zeros((rows, cols, n_bins))
# Compute histogram per cell
### YOUR CODE HERE
for r in range(rows):
for c in range(cols):
bin_id = np.int32((theta_cells[r, c] // degrees_per_bin).flatten())
G_flat = G_cells[r, c].flatten()
# print (theta_cells, bin_id)
for i,b in enumerate(bin_id):
cells[r, c, b] += G_flat[i]
block = cells.flatten()
block /= np.linalg.norm(block)
# block /= np.sum(np.abs(block))
# variance = np.sum(np.var(cells, axis = (0,1)))
# block = block / np.sqrt(variance)
### YOUR CODE HERE
return block
def hog_descriptor(patch, pixels_per_cell=(8,8)):
"""
Generating hog descriptor by the following steps:
1. compute the gradient image in x and y (already done for you)
2. compute gradient histograms
3. normalize across block
4. flattening block into a feature vector
Args:
patch: grayscale image patch of shape (h, w)
pixels_per_cell: size of a cell with shape (m, n)
Returns:
block: 1D array of shape ((h*w*n_bins)/(m*n))
"""
assert (patch.shape[0] % pixels_per_cell[0] == 0),\
'Heights of patch and cell do not match'
assert (patch.shape[1] % pixels_per_cell[1] == 0),\
'Widths of patch and cell do not match'
n_bins = 9
degrees_per_bin = 180 // n_bins
Gx = filters.sobel_v(patch)
Gy = filters.sobel_h(patch)
# Unsigned gradients
G = np.sqrt(Gx**2 + Gy**2)
theta = (np.arctan2(Gy, Gx) * 180 / np.pi) % 180
G_cells = view_as_blocks(G, block_shape=pixels_per_cell)
theta_cells = view_as_blocks(theta, block_shape=pixels_per_cell)
rows = G_cells.shape[0]
cols = G_cells.shape[1]
cells = np.zeros((rows, cols, n_bins))
# Compute histogram per cell
for r in range(rows):
for c in range(cols):
G = G_cells[r,c]
theta = theta_cells[r,c]
bin_num = theta//degrees_per_bin
for i in range(pixels_per_cell[0]):
for j in range(pixels_per_cell[1]):
cells[r, c, int(bin_num[i, j])%9] += G[i, j]
flat = cells.flatten()
block = flat/np.linalg.norm(flat)
return block
def hog_descriptor(patch, pixels_per_cell=(8, 8)):
"""
Generating hog descriptor by the following steps:
1. compute the gradient image in x and y (already done for you)
2. compute gradient histograms
3. normalize across block
4. flattening block into a feature vector
Args:
patch: grayscale image patch of shape (h, w)
pixels_per_cell: size of a cell with shape (m, n)
Returns:
block: 1D array of shape ((h*w*n_bins)/(m*n))
"""
assert (patch.shape[0] % pixels_per_cell[0] == 0),\
'Heights of patch and cell do not match'
assert (patch.shape[1] % pixels_per_cell[1] == 0),\
'Widths of patch and cell do not match'
n_bins = 9
degrees_per_bin = 180 // n_bins
Gx = filters.sobel_v(patch)
Gy = filters.sobel_h(patch)
# Unsigned gradients
G = np.sqrt(Gx**2 + Gy**2)
theta = (np.arctan2(Gy, Gx) * 180 / np.pi) % 180
G_cells = view_as_blocks(
G,
block_shape=pixels_per_cell) #返回一个4维矩阵，表示一共有m行n列个cell,每个cell的维度是x行y列
theta_cells = view_as_blocks(theta, block_shape=pixels_per_cell)
rows = G_cells.shape[0]
cols = G_cells.shape[1]
cells = np.zeros((rows, cols, n_bins))
# Compute histogram per cell
for i in range(0, rows):
for j in range(0, cols):
for n_bin in range(0, n_bins):
cells[i, j, n_bin] = np.sum(
G_cells[i, j] *
((theta_cells[i, j] >= n_bin * degrees_per_bin) &
(theta_cells[i, j] < (n_bin + 1) * degrees_per_bin)))
block = cells.flatten()
block = block - block.mean()
if block.std() < 1e-5:
pass #方差为0就不动它了
else:
block /= block.std()
return block
def hog_descriptor(patch, pixels_per_cell=(8,8)):
"""
Generating hog descriptor by the following steps:
1. compute the gradient image in x and y (already done for you)
2. compute gradient histograms
3. normalize across block
4. flattening block into a feature vector
Args:
patch: grayscale image patch of shape (h, w)
pixels_per_cell: size of a cell with shape (m, n)
Returns:
block: 1D array of shape ((h*w*n_bins)/(m*n))
"""
assert (patch.shape[0] % pixels_per_cell[0] == 0),\
'Heights of patch and cell do not match'
assert (patch.shape[1] % pixels_per_cell[1] == 0),\
'Widths of patch and cell do not match'
n_bins = 9
degrees_per_bin = 180 // n_bins
Gx = filters.sobel_v(patch)
Gy = filters.sobel_h(patch)
# Unsigned gradients
G = np.sqrt(Gx**2 + Gy**2)
theta = (np.arctan2(Gy, Gx) * 180 / np.pi) % 180
G_cells = view_as_blocks(G, block_shape=pixels_per_cell)
theta_cells = view_as_blocks(theta, block_shape=pixels_per_cell)
rows = G_cells.shape[0]
cols = G_cells.shape[1]
cells = np.zeros((rows, cols, n_bins))
block = []
# Compute histogram per cell
### YOUR CODE HERE
for ci in range(rows):
for cj in range(cols):
cell = G_cells[ci,cj]
cellTheta = theta_cells[ci,cj]
for pi in range(cell.shape[0]):
for pj in range(cell.shape[1]):
tbin = (cellTheta[pi,pj] // degrees_per_bin) % n_bins
cells[ci,cj,int(tbin)] += cell[pi,pj]
cells[ci,cj,:] = cells[ci,cj,:] / np.sum(cells[ci,cj,:])
block = cells.flatten()
### YOUR CODE HERE
return block
def hog_descriptor(patch, pixels_per_cell=(8, 8)):
"""
Generating hog descriptor by the following steps:
1. compute the gradient image in x and y (already done for you)
2. compute gradient histograms
3. normalize across block
4. flattening block into a feature vector
Args:
patch: grayscale image patch of shape (h, w)
pixels_per_cell: size of a cell with shape (m, n)
Returns:
block: 1D array of shape ((h*w*n_bins)/(m*n))
"""
assert (patch.shape[0] % pixels_per_cell[0] == 0),\
'Heights of patch and cell do not match'
assert (patch.shape[1] % pixels_per_cell[1] == 0),\
'Widths of patch and cell do not match'
n_bins = 9
degrees_per_bin = 180 // n_bins
Gx = filters.sobel_v(patch)
Gy = filters.sobel_h(patch)
# Unsigned gradients
G = np.sqrt(Gx**2 + Gy**2)
theta = (np.arctan2(Gy, Gx) * 180 / np.pi) % 180
G_cells = view_as_blocks(G, block_shape=pixels_per_cell)
theta_cells = view_as_blocks(theta, block_shape=pixels_per_cell)
rows = G_cells.shape[0]
cols = G_cells.shape[1]
cells = np.zeros((rows, cols, n_bins))
# Compute histogram per cell
# YOUR CODE HERE
M = G_cells.shape[2]
N = G_cells.shape[3]
for i in range(rows):
for j in range(cols):
for m in range(M):
for n in range(N):
deg = int((theta_cells[i, j, m, n] + 180) * n_bins) // 360
if (deg == 9):
deg = 0
cells[i, j, deg] += G_cells[i, j, m, n]
block = np.reshape(cells, newshape=-1)
# YOUR CODE HERE
return block
def __load__(self, id_name):
image_path = os.path.join(self.path, id_name,
"lattice_light_sheet") + ".tif"
mask_path = os.path.join(self.path, id_name, "truth") + ".tif"
latticeMovieImage = skimage.external.tifffile.imread(image_path)
latticeMovieImage = latticeMovieImage[:self.image_size, :self.
image_size, :self.image_size]
#Standardizing globally
#image_mean = np.mean(latticeMovieImage, axis=(0, 1, 2), keepdims=True)
#image_std = np.std(latticeMovieImage, axis=(0, 1, 2), keepdims=True)
#latticeMovieImage = (latticeMovieImage - image_mean) / image_std
lattice_patches = view_as_blocks(latticeMovieImage,
block_shape=(self.patch_size,
self.patch_size,
self.patch_size))
lattice_patches = lattice_patches.reshape(
int((self.image_size / self.patch_size)**3), self.patch_size,
self.patch_size, self.patch_size)
lattice_patches = np.expand_dims(lattice_patches, axis=-1)
mask_image = skimage.external.tifffile.imread(mask_path)
mask_image = mask_image[:self.image_size, :self.image_size, :self.
image_size]
#mask_image = np.expand_dims(mask_image, axis=-1)
mask = np.zeros((self.image_size, self.image_size, self.image_size))
mask = np.maximum(mask, mask_image)
#TODO Check if view_as_blocks gives all possible blocks
mask_patches = view_as_blocks(mask,
block_shape=(self.patch_size,
self.patch_size,
self.patch_size))
mask_patches = mask_patches.reshape(
int((self.image_size / self.patch_size)**3), self.patch_size,
self.patch_size, self.patch_size)
mask_patches = np.expand_dims(mask_patches, axis=-1)
#lattice_patches = lattice_patches/(2/3*65535.0)
#BOTTOM LINE COMMENTED FOR PSNR5 Data. UNCOMMENT FOR FUTURE USE and use 255.0
mask_patches = mask_patches / 255.0
weight_patches = np.zeros((self.patch_size**3, 2))
weight_patches[:, 0] = 0.005
weight_patches[:, 1] = 0.995
weight_patches = np.squeeze(np.sum(weight_patches, axis=-1))
#print(np.count_nonzero(mask == 1.0)/1000000.0)
return lattice_patches, mask_patches, weight_patches
def hog_descriptor(patch, pixels_per_cell=(8, 8)):
"""
Generating hog descriptor by the following steps:
1. compute the gradient image in x and y (already done for you)
2. compute gradient histograms
3. normalize across block
4. flattening block into a feature vector
Args:
patch: grayscale image patch of shape (h, w)
pixels_per_cell: size of a cell with shape (m, n)
Returns:
block: 1D array of shape ((h*w*n_bins)/(m*n))
"""
assert (patch.shape[0] % pixels_per_cell[0] == 0),\
'Heights of patch and cell do not match'
assert (patch.shape[1] % pixels_per_cell[1] == 0),\
'Widths of patch and cell do not match'
n_bins = 9
degrees_per_bin = 180 // n_bins
Gx = filters.sobel_v(patch)
Gy = filters.sobel_h(patch)
# Unsigned gradients
G = np.sqrt(Gx**2 + Gy**2)
theta = (np.arctan2(Gy, Gx) * 180 / np.pi) % 180
G_cells = view_as_blocks(G, block_shape=pixels_per_cell)
theta_cells = view_as_blocks(theta, block_shape=pixels_per_cell)
rows = G_cells.shape[0]
cols = G_cells.shape[1]
cells = np.zeros((rows, cols, n_bins))
# Compute histogram per cell
### YOUR CODE HERE
for i in range(rows):
for j in range(cols):
hist, _ = np.histogram(a=theta_cells[i, j],
bins=n_bins,
range=(0, 180),
weights=G_cells[i, j])
cells[i, j] = hist
# follow `simple_descriptor`, describe the patch by normalizing the image values
# into a standard normal distribution (having mean of 0 and standard deviation of 1)
cells = (cells - np.mean(cells)) / np.std(cells)
block = cells.flatten()
### YOUR CODE HERE
return block
def hog_descriptor(patch, pixels_per_cell=(8, 8)):
"""
Generating hog descriptor by the following steps:
1. compute the gradient image in x and y (already done for you)
2. compute gradient histograms
3. normalize across block
4. flattening block into a feature vector
Args:
patch: grayscale image patch of shape (h, w)
pixels_per_cell: size of a cell with shape (m, n)
Returns:
block: 1D array of shape ((h*w*n_bins)/(m*n))
"""
assert (patch.shape[0] % pixels_per_cell[0] == 0),\
'Heights of patch and cell do not match'
assert (patch.shape[1] % pixels_per_cell[1] == 0),\
'Widths of patch and cell do not match'
n_bins = 9
degrees_per_bin = 180 // n_bins
Gx = filters.sobel_v(patch)
Gy = filters.sobel_h(patch)
# Unsigned gradients
G = np.sqrt(Gx**2 + Gy**2)
theta = (np.arctan2(Gy, Gx) * 180 / np.pi) % 180
G_cells = view_as_blocks(G, block_shape=pixels_per_cell)
print(G_cells)
theta_cells = view_as_blocks(theta, block_shape=pixels_per_cell)
rows = G_cells.shape[0]
cols = G_cells.shape[1]
cells = np.zeros((rows, cols, n_bins))
# Compute histogram per cell
### YOUR CODE HERE
pass
### YOUR CODE HERE
return block
# matches.remove([362, 56])# cheo dai tren
# matches.remove([404,304])#cheo dai duoi
def hog_descriptor(patch, pixels_per_cell=(8, 8)):
"""
Generating hog descriptor by the following steps:
1. compute the gradient image in x and y (already done for you)
2. compute gradient histograms
3. normalize across block
4. flattening block into a feature vector
Args:
patch: grayscale image patch of shape (h, w)
pixels_per_cell: size of a cell with shape (m, n)
Returns:
block: 1D array of shape ((h*w*n_bins)/(m*n))
"""
assert (patch.shape[0] % pixels_per_cell[0] == 0),\
'Heights of patch and cell do not match'
assert (patch.shape[1] % pixels_per_cell[1] == 0),\
'Widths of patch and cell do not match'
n_bins = 9
degrees_per_bin = 180 // n_bins
Gx = filters.sobel_v(patch)
Gy = filters.sobel_h(patch)
# Unsigned gradients
G = np.sqrt(Gx**2 + Gy**2)
theta = (np.arctan2(Gy, Gx) * 180 / np.pi) % 180
G_cells = view_as_blocks(G, block_shape=pixels_per_cell)
theta_cells = view_as_blocks(theta, block_shape=pixels_per_cell)
rows = G_cells.shape[0]
cols = G_cells.shape[1]
cells = np.zeros((rows, cols, n_bins))
# Compute histogram per cell
Gr = G.reshape((rows, cols, pixels_per_cell[0] * pixels_per_cell[1]))
cells = np.apply_along_axis(
lambda x: np.histogram(x, bins=n_bins, density=True)[0],
axis=2,
arr=Gr)
block = cells.flatten()
block = block / np.sum(block)
return block
def hog_descriptor(patch, pixels_per_cell=(8, 8)):
"""
Generating hog descriptor by the following steps:
1. Compute the gradient image in x and y directions (already done for you)
2. Compute gradient histograms for each cell
3. Flatten 3D matrix of histograms into a 1D feature vector.
4. Normalize flattened histogram feature vector by L2 norm
Normalization makes the descriptor more robust to lighting variations
Args:
patch: grayscale image patch of shape (H, W)
pixels_per_cell: size of a cell with shape (M, N)
Returns:
block: 1D patch descriptor array of shape ((H*W*n_bins)/(M*N))
"""
assert (patch.shape[0] % pixels_per_cell[0] == 0),\
'Heights of patch and cell do not match'
assert (patch.shape[1] % pixels_per_cell[1] == 0),\
'Widths of patch and cell do not match'
n_bins = 9
degrees_per_bin = 180 // n_bins
Gx = filters.sobel_v(patch)
Gy = filters.sobel_h(patch)
# Unsigned gradients
G = np.sqrt(Gx**2 + Gy**2)
theta = (np.arctan2(Gy, Gx) * 180 / np.pi) % 180
# Group entries of G and theta into cells of shape pixels_per_cell, (M, N)
# G_cells.shape = theta_cells.shape = (H//M, W//N)
# G_cells[0, 0].shape = theta_cells[0, 0].shape = (M, N)
G_cells = view_as_blocks(G, block_shape=pixels_per_cell)
theta_cells = view_as_blocks(theta, block_shape=pixels_per_cell)
rows = G_cells.shape[0]
cols = G_cells.shape[1]
# For each cell, keep track of gradient histrogram of size n_bins
hists = np.zeros((rows, cols, n_bins))
# Compute histogram per cell
### YOUR CODE HERE
pass
### YOUR CODE HERE
return block
def load_data_make_jpeg(folder):
entrynumber = 0
list = glob.glob(folder)
for entry in list:
entrynumber += 1
img_size = (256, 256, 3)
img_new = io.imread(entry)
shape = img_new.shape
height = shape[0] // 256
height = height * 256
width = shape[1] // 256
width = width * 256
img_new = img_new[:height, :width, :3]
img_new_w = view_as_blocks(img_new, img_size)
img_new_w = np.uint8(img_new_w)
r = 0
for i in range(img_new_w.shape[0]):
for j in range(img_new_w.shape[1]):
A = np.zeros((img_size[0], img_size[1], 3))
A[:, :, :] = img_new_w[i, j, :, :]
A = np.uint8(A)
imageio.imwrite(
'/home/diego/Desktop/Output/' + str(entrynumber) + '-' +
str(r) + '.png', A)
r += 1
def test_view_as_blocks_2D_array():
A = np.arange(4 * 4).reshape(4, 4)
B = view_as_blocks(A, (2, 2))
assert_equal(B[0, 1], np.array([[2, 3],
[6, 7]]))
assert_equal(B[1, 0, 1, 1], 13)
def downscale_all(img, mask, edges, downscale):
"""
Downscales all the inputs by a given scale.
:param img: A sparse flow map - h x w x 2
:param mask: A binary mask - h x w x 1
:param edges:An edges map - h x w x 1
:param downscale: Downscaling factor.
:return: the downscaled versions of the inputs
"""
from skimage.util.shape import view_as_blocks
from skimage.transform import rescale
img[:, :, 0][mask == -1] = np.nan
img[:, :, 1][mask == -1] = np.nan
img = img[:(img.shape[0] -
(img.shape[0] % downscale)), :(img.shape[1] -
(img.shape[1] % downscale)), :]
blocks = view_as_blocks(img, (downscale, downscale, 2))
img = np.nanmean(blocks, axis=(-2, -3, -4))
mask = np.ones_like(img)
mask[np.isnan(img)] = -1
mask = mask[:, :, 0]
img[np.isnan(img)] = 0
if edges is not None:
edges = edges[:(edges.shape[0] - (edges.shape[0] % downscale)), :(
edges.shape[1] - (edges.shape[1] % downscale))]
edges = rescale(edges, 1 / float(downscale), preserve_range=True)
return img, mask, edges
def get_patch(image_data, patch_shape):
# image_data.shape: (300, 240, 416, 3)
n_image = image_data.shape[0]
# block_hight: 15, block_width: 26
(block_hight, block_width) = get_block_shape(image_data.shape, patch_shape)
# n_patch: 15 * 16 -> 390
n_patch = block_hight * block_width
# Initialize patch_data (390*300, 16, 16, 3)
patch_data = np.empty([n_patch * n_image] + list(patch_shape),
dtype=np.float32)
# block_reshape: (390, 16, 16, 3)
block_reshape = [n_patch] + list(patch_shape)
for i in range(n_image):
# Get each patch block of a image, and organize the patch block to (390, 16, 16, 3)
block = view_as_blocks(image_data[i],
block_shape=patch_shape).reshape(block_reshape)
# Add patch from a patch block to the patch_data one by one
for j in range(n_patch):
patch_data[i * n_patch + j] = block[j]
return patch_data
def downscale_all(sparse_flow, mask, edges, downscale):
"""
Downscales all the inputs by a given scale.
:param sparse_flow: A sparse flow map - h x w x 2
:param mask: A binary mask - h x w x 1
:param edges:An edges map - h x w x 1
:param downscale: Downscaling factor.
:return: the downscaled versions of the inputs
"""
sparse_flow[:, :, 0][mask == -1] = np.nan
sparse_flow[:, :, 1][mask == -1] = np.nan
sparse_flow = sparse_flow[:(sparse_flow.shape[0] -
(sparse_flow.shape[0] % downscale)), :(
sparse_flow.shape[1] -
(sparse_flow.shape[1] % downscale)), :]
blocks = view_as_blocks(sparse_flow, (downscale, downscale, 2))
sparse_flow = np.nanmean(blocks, axis=(-2, -3, -4))
mask = np.ones_like(sparse_flow)
mask[np.isnan(sparse_flow)] = -1
mask = mask[:, :, 0]
sparse_flow[np.isnan(sparse_flow)] = 0
if edges is not None:
edges = edges[:(edges.shape[0] - (edges.shape[0] % downscale)), :(
edges.shape[1] - (edges.shape[1] % downscale))]
edges = rescale(edges, 1 / float(downscale), preserve_range=True)
return sparse_flow, mask, edges
def shuffle(im, inverse=False):
# Configure block size, rows and columns
blk_size = 56
rows = np.uint8(img.shape[0] / blk_size)
cols = np.uint8(img.shape[1] / blk_size)
# Create a block view on image
img_blks = view_as_blocks(im,
block_shape=(blk_size, blk_size, 3)).squeeze()
img_shuff = np.zeros((img.shape[0], img.shape[1], 3), dtype=np.uint8)
# Secret key maps
map = {0: 2, 1: 0, 2: 3, 3: 1}
inv_map = {v: k for k, v in map.items()}
# Perform block shuffling
for i in range(0, rows):
for j in range(0, cols):
x, y = i * blk_size, j * blk_size
if (inverse):
img_shuff[x:x + blk_size,
y:y + blk_size] = img_blks[inv_map[i], inv_map[j]]
else:
img_shuff[x:x + blk_size, y:y + blk_size] = img_blks[map[i],
map[j]]
return img_shuff
def shuffle(im, num, rotate=False):
im = torch2numpy(im.permute(1, 2, 0))
rows = cols = num
blk_size = im.shape[0] // rows
img_blocks = view_as_blocks(im,
block_shape=(blk_size, blk_size, 3)).reshape(
(-1, blk_size, blk_size, 3))
img_shuffle = np.zeros((im.shape[0], im.shape[1], 3))
a = np.arange(rows * rows, dtype=np.uint8)
b = np.random.permutation(a)
map = {k: v for k, v in zip(a, b)}
for i in range(0, rows):
for j in range(0, cols):
x, y = i * blk_size, j * blk_size
if rotate:
rot_val = random.randrange(0, 4)
img_shuffle[x:x + blk_size, y:y + blk_size] = np.rot90(
img_blocks[map[i * rows + j]], rot_val)
else:
img_shuffle[x:x + blk_size,
y:y + blk_size] = img_blocks[map[i * rows + j]]
img_shuffle = torch.FloatTensor(img_shuffle).permute(0, 1, 2)
return img_shuffle
def setupPatches(self, dims):
self.dims = dims
self.im = mpimg.imread(self.inputfile)
self.block_shape = (self.dims[0], self.dims[1], self.im.shape[2]) #height, width
margin=np.mod(self.im.shape,self.block_shape)
self.im_crop = self.im[:(self.im.shape-margin)[0],:(self.im.shape-margin)[1],:(self.im.shape-margin)[2]]
self.view = view_as_blocks(self.im_crop, self.block_shape)
def load_data_make_jpeg():
img_size = (256, 256, 3)
img_new = io.imread(
'/home/diego/MPFI/Images/041719 JCFFRIL 07 Wt3 profile 13 Spiny top montage color for Diego.tif'
)
#img_ref = io.imread('/home/diego/MPFI/Images/CycleGAN_onlyGreen/Unlabels/onlygreen_2.png')
img_new = img_new[:9984, :16128, :3]
#img_ref = img_ref[:7424,:5632,:3]
img_new_w = view_as_blocks(img_new, img_size)
img_new_w = np.uint8(img_new_w)
#img_ref_w = view_as_blocks(img_ref, img_size)
#img_ref_w = np.uint8(img_ref_w)
r = 0
for i in range(img_new_w.shape[0]):
for j in range(img_new_w.shape[1]):
A = np.zeros((img_size[0], img_size[1], 3))
#B = np.zeros((img_size[0], img_size[1], 3))
A[:, :, :] = img_new_w[i, j, :, :]
#B[:,:,:] = img_ref_w[i,j,:,:]
A = np.uint8(A)
#B = np.uint8(B)
imageio.imwrite(
'/home/diego/MPFI/Datasets/profile13_numbered/' + str(r) +
'.png', A)
#imageio.imwrite('/home/diego/MPFI/Datasets/OnlyGreen/trainB/'+ str(r) + '.png', B)
r += 1
def load_data_make_jpeg(folder):
list = glob.glob(folder)
for entry in list:
img_size = (256, 256, 3)
img_new = io.imread(entry)
img_new = (img_new / 256).astype('uint8')
shape = img_new.shape
height = shape[0] // 256
height256 = height * 256
width = shape[1] // 256
width256 = width * 256
img_new = img_new[:height256, :width256, :3]
img_new_w = view_as_blocks(img_new, img_size)
img_new_w = np.uint8(img_new_w)
imageio.imwrite('/home/diego/Desktop/Output_Final/' + 'CroppedVersion' + '.png', img_new)
r = 0
for i in range(img_new_w.shape[0]):
for j in range(img_new_w.shape[1]):
A = np.zeros((img_size[0], img_size[1], 3))
A[:, :, :] = img_new_w[i, j, :, :]
A = np.uint8(A)
imageio.imwrite('/home/diego/Desktop/Output/' + str(r) + '.png', A)
r += 1
return width, height
def test_view_as_blocks_3D_array():
A = np.arange(4 * 4 * 6).reshape(4, 4, 6)
B = view_as_blocks(A, (1, 2, 2))
assert_equal(B.shape, (4, 2, 3, 1, 2, 2))
assert_equal(B[2:, 0, 2],
np.array([[[[52, 53], [58, 59]]], [[[76, 77], [82, 83]]]]))
def test_view_as_blocks_3D_array():
A = np.arange(4 * 4 * 6).reshape(4, 4, 6)
B = view_as_blocks(A, (1, 2, 2))
assert_equal(B.shape, (4, 2, 3, 1, 2, 2))
assert_equal(B[2:, 0, 2], np.array([[[[52, 53],
[58, 59]]],
[[[76, 77],
[82, 83]]]]))
def small_spectrogram_max_pooling(spec):
"""
Using code adapated from:
http://scikit-image.org/docs/dev/auto_examples/plot_view_as_blocks.html
"""
spec = force_spectrogram_length(spec, 384)
im_norm = (spec - spec.mean()) / spec.var()
view = view_as_blocks(im_norm, (32, 32))
flatten_view = view.reshape(view.shape[0], view.shape[1], -1)
return np.max(flatten_view, axis=2).flatten()
def getFeature():
cap = cv2.VideoCapture('/home/xingzhong/Videos/heat.mkv')
fn = 0
ret, iframe = cap.read()
H, W, _ = iframe.shape
gmask = globalMask(iframe)
nznsLeft = deque(maxlen=25)
nznsCenter = deque(maxlen=25)
nznsRight = deque(maxlen=25)
features = []
while(cap.isOpened()):
ret, frame = cap.read()
if not ret:
break
# if fn >= 100000: break
fn += 1
frameHSV = cv2.cvtColor(frame, cv2.COLOR_BGR2HSV)
ret, threshColor = cv2.threshold(
frameHSV[:, :, 0], 30, 179, cv2.THRESH_BINARY_INV)
#threshColorMask = cv2.bitwise_and(frame, frame, mask = threshColor)
#import ipdb; ipdb.set_trace()
#imgs = view_as_blocks(threshColorMask, block_shape = (H/9, W/16, 3)).reshape(-1, H/9, W/16, 3)
imgs = view_as_blocks(
threshColor, block_shape=(H / 9, W / 16)).reshape(-1, H / 9, W / 16)
nzns = map(lambda x: np.count_nonzero(x) / float(x.size), imgs)
#import ipdb; ipdb.set_trace()
#nznLeft = 3 * np.count_nonzero(threshColorMask[:, :W/3]) / float(threshColorMask.size)
#nznCenter = 3 * np.count_nonzero(threshColorMask[:, W/3:2*W/3]) / float(threshColorMask.size)
#nznRight = 3 * np.count_nonzero(threshColorMask[:, -W/3:]) / float(threshColorMask.size)
# nznsLeft.append(nznLeft)
# nznsCenter.append(nznCenter)
# nznsRight.append(nznRight)
features.append(nzns)
if fn % 100 == 0:
cv2.putText(frame, "#f %d" %
fn, (10, 30), FONT, 1, (255, 255, 255), 1, cv2.LINE_AA)
#cv2.putText(threshColorMask, "%d"%(int(100*nznLeft)) ,(W/6, 100), FONT, 1,(255,255,255),1,cv2.LINE_AA)
#cv2.putText(threshColorMask, "%d"%(int(100*nznCenter)) ,(W/2, 100), FONT, 1,(255,255,255),1,cv2.LINE_AA)
#cv2.putText(threshColorMask, "%d"%(int(100*nznRight)) ,(5*W/6, 100), FONT, 1,(255,255,255),1,cv2.LINE_AA)
mask = np.array(nzns).reshape(9, 16)
cv2.imshow('mask', mask)
cv2.imshow('frame', frame)
if cv2.waitKey(1) & 0xFF == ord('q'):
break
pickle.dump(features, open("segment_features_full.p", "wb"))
cap.release()
cv2.destroyAllWindows()
def _downsample(array, factors, sum=True):
"""Performs downsampling with integer factors.
Parameters
----------
array : ndarray
Input n-dimensional array.
factors: tuple
Tuple containing downsampling factor along each axis.
sum : bool
If True, downsampled element is the sum of its corresponding
constituent elements in the input array. Default is True.
Returns
-------
array : ndarray
Downsampled array with same number of dimensions as that of input
array.
"""
pad_size = []
if len(factors) != array.ndim:
raise ValueError("'factors' must have the same length "
"as 'array.shape'")
else:
for i in range(len(factors)):
if array.shape[i] % factors[i] != 0:
pad_size.append(factors[i] - (array.shape[i] % factors[i]))
else:
pad_size.append(0)
for i in range(len(pad_size)):
array = _pad_asymmetric_zeros(array, pad_size[i], i)
out = view_as_blocks(array, factors)
block_shape = out.shape
if sum:
for i in range(len(block_shape) // 2):
out = out.sum(-1)
else:
for i in range(len(block_shape) // 2):
out = out.mean(-1)
return out
def frequency_max_pooling(spec, normalise=True):
"""
Using code adapated from:
http://scikit-image.org/docs/dev/auto_examples/plot_view_as_blocks.html
"""
if normalise:
im_norm = (spec - spec.mean()) / spec.var()
else:
im_norm = spec
view = view_as_blocks(im_norm, (8, spec.shape[1]))
flatten_view = view.reshape(view.shape[0], view.shape[1], -1)
A = np.max(flatten_view, axis=2).flatten()
B = np.var(flatten_view, axis=2).flatten()
C = np.mean(flatten_view, axis=2).flatten()
return np.hstack((A, B, C))
def get_subvolume(self, box_zyx, scale=0):
"""
Extract the subvolume, specified in new (scaled) coordinates from the
original volume service, then scale result accordingly before returning it.
"""
true_scale = scale + self.scale_delta
if true_scale in self.original_volume_service.available_scales:
# The original source already has the data at the necessary scale.
return self.original_volume_service.get_subvolume( box_zyx, true_scale )
# Start with the closest scale we've got
base_scales = np.array(self.original_volume_service.available_scales)
i_best = np.abs(base_scales - true_scale).argmin()
best_base_scale = base_scales[i_best]
delta_from_best = true_scale - best_base_scale
if delta_from_best > 0:
orig_box_zyx = box_zyx * 2**delta_from_best
orig_data = self.original_volume_service.get_subvolume(orig_box_zyx, best_base_scale)
if self.dtype == np.uint64:
# Assume that uint64 means labels.
downsampled_data, _ = downsample_labels_3d( orig_data, 2**self.scale_delta )
else:
downsampled_data = downsample_raw( orig_data, self.scale_delta )[-1]
return downsampled_data
else:
upsample_factor = int(2**-delta_from_best)
orig_box_zyx = downsample_box(box_zyx, np.array(3*(upsample_factor,)))
orig_data = self.original_volume_service.get_subvolume(orig_box_zyx, best_base_scale)
orig_shape = np.array(orig_data.shape)
upsampled_data = np.empty( orig_shape * upsample_factor, dtype=self.dtype )
v = view_as_blocks(upsampled_data, 3*(upsample_factor,))
v[:] = orig_data[:,:,:,None, None, None]
relative_box = box_zyx - upsample_factor*orig_box_zyx[0]
requested_data = upsampled_data[box_to_slicing(*relative_box)]
# Force contiguous so caller doesn't have to worry about it.
return np.asarray(requested_data, order='C')
def demo_upsample_nearest():
a = sp.misc.lena() / 1.
a = sp.misc.lena() / 1.
a.shape = a.shape[:2] + (1,)
print a.shape
#print np.tile(a, 2).shape
#a = np.dstack((a, -a))
N = 96
a = np.tile(a, N)
a[:,:,95] = -a[:,:,95]
#r = np.tile(a, (2, 2, 1))
#np.kron(a, np.ones((2,2,1))).shape
# -- loop
#a = a[:, :, 0].reshape(256, 256, 1)
#r = np.empty((1024, 1024, 1))
#r[0::2, 0::2] = a
#r[0::2, 1::2] = a
#r[1::2, 0::2] = a
#r[1::2, 1::2] = a
# -- block view
r = np.empty((1024, 1024, N))
b = view_as_blocks(r, (2, 2, 1))
print b.shape
a2 = a.reshape(a.shape + (1, 1, 1))
#a[:, :, :, np.newaxis, np.newaxis, np.newaxis]
b[:] = a2
#import IPython; ipshell = IPython.embed; ipshell(banner1='ipshell')
#b2 = b.swapaxes(1, 3).reshape(r.shape)
b2 = b.transpose((0, 3, 1, 4, 2, 5)).reshape(r.shape)
pl.matshow(b2[:,:,0])
pl.matshow(b2[:,:,1])
pl.matshow(b2[:,:,95])
pl.show()
def extract_blocks(img_gray):
patch_sizes = [128, 256]
result = []
for size in patch_sizes:
blocks = view_as_blocks(img_gray, (size, size))
for row in range(blocks.shape[0]):
for col in range(blocks.shape[1]):
block = blocks[row, col]
pred = rc.predict(block)
if pred == None:
continue
pred_prob = rc.predict_prob(block)[0]
top1 = numpy.argsort(pred_prob)[-1:][0]
top1_prob = pred_prob[top1]
tops = numpy.argsort(pred_prob)[-5:]
tops = tops[::-1]
result.append((top1_prob, pred[0], row, col, size, block))
#print "Size", size, "Prediction:", pred, "Argmax:", numpy.argmax(pred_prob), "Class:", classes[numpy.argmax(pred_prob)]
#for idx, top in enumerate(tops):
# print "", idx+1, ": ", classes[top], " : ", pred_prob[top]
#print "="*80
return result
if 'zz_blank' in fname:
continue
out_fname = fname + ext
if path.exists(out_fname):
continue
arr = misc.imread(fname, flatten=True)
arr = misc.imresize(arr, in_shape).astype('float32')
arr -= arr.min()
arr /= arr.max()
farr = model.transform(
arr,
pad_apron=pad_apron, interleave_stride=interleave_stride
)
if reduce is not None:
assert block is not None
farr_b = view_as_blocks(farr, block + (1,))
farr_br = farr_b.reshape(farr_b.shape[:3] + (-1,))
if reduce == 'mean':
farr_brm = farr_br.mean(-1)
elif reduce == 'min':
farr_brm = farr_br.min(-1)
elif reduce == 'max':
farr_brm = farr_br.max(-1)
elif reduce == 'median':
farr_brm = np.median(farr_br, axis=-1)
else:
raise ValueError("'%s' reduce not understood" % reduce)
farr = farr_brm
print farr.shape, out_fname
np.save(out_fname, farr)
end = time.time()
def test_view_as_blocks_block_not_a_tuple():
A = np.arange(10)
with testing.raises(TypeError):
view_as_blocks(A, [5])
def test_view_as_blocks_negative_shape():
A = np.arange(10)
with testing.raises(ValueError):
view_as_blocks(A, (-2,))
def test_view_as_blocks_block_too_large():
A = np.arange(10)
with testing.raises(ValueError):
view_as_blocks(A, (11,))
def test_view_as_blocks_1D_array_wrong_block_shape():
A = np.arange(10)
with testing.raises(ValueError):
view_as_blocks(A, (3,))
def test_view_as_blocks_1D_array():
A = np.arange(10)
B = view_as_blocks(A, (5,))
assert_equal(B, np.array([[0, 1, 2, 3, 4], [5, 6, 7, 8, 9]]))
def test_view_as_blocks_block_not_a_tuple():
A = np.arange(10)
view_as_blocks(A, [5])
def test_view_as_blocks_wrong_block_dimension():
A = np.arange(10)
view_as_blocks(A, (2, 2))
def test_view_as_blocks_1D_array_wrong_block_shape():
A = np.arange(10)
view_as_blocks(A, (3,))
def test_view_as_blocks_block_too_large():
A = np.arange(10)
view_as_blocks(A, (11,))
def test_view_as_blocks_negative_shape():
A = np.arange(10)
view_as_blocks(A, (-2,))
def test_view_as_blocks_wrong_block_dimension():
A = np.arange(10)
with testing.raises(ValueError):
view_as_blocks(A, (2, 2))
def main():
print "Load data"
sourceHeight = 'height_masked_final.asc'
sourceBiomass = 'ndvi_masked_final.asc'
sourceCoverageModel = 'testCov.txt'
sourceHeightModel = 'heightModel.txt'
sourceLCC = "LCC.asc"
heightGrid = numpy.loadtxt(sourceHeight, skiprows=6)
heightGrid = heightGrid[500:1050, 300:750]
# heightGrid = heightGrid[640:685,455:500]
# heightGrid = heightGrid[860:890,600:700]
rgb = (heightGrid - numpy.min(heightGrid)) / (numpy.max(heightGrid) - numpy.min(heightGrid))
rgb *= 255
heightRGBA = numpy.zeros((heightGrid.shape[0], heightGrid.shape[1], 3), dtype=numpy.uint8)
heightRGBA[:, :, 0:3] = rgb[:, :, numpy.newaxis]
# misc.imsave('heightMap_Paulinapolder.png',heightRGBA)
ndviGrid = numpy.loadtxt(sourceBiomass, skiprows=6)
ndviGrid = ndviGrid[500:1050, 300:750]
# ndviGrid = ndviGrid[640:685,455:500]
# ndviGrid = ndviGrid[860:890,600:700]
rgb = (ndviGrid - numpy.min(ndviGrid)) / (numpy.max(ndviGrid) - numpy.min(ndviGrid))
rgb *= 255
ndviRGBA = numpy.zeros((ndviGrid.shape[0], ndviGrid.shape[1], 3), dtype=numpy.uint8)
ndviRGBA[:, :, 0:3] = rgb[:, :, numpy.newaxis]
misc.imsave('ndviMap_Paulinapolder.png', ndviRGBA)
lccGrid = numpy.loadtxt(sourceLCC, skiprows=6)
lccGrid = lccGrid[500:1050, 300:750]
heightModelGrid = numpy.loadtxt(sourceHeightModel)
coverageModelGrid = numpy.loadtxt(sourceCoverageModel)
PaulinaPolder = False
NDVI = True
LCC = False
if PaulinaPolder:
if NDVI and LCC:
vegetationMask = ndviGrid > 0
elif NDVI:
vegetationMask = ndviGrid > 0.08 # 0.02 demo
# figure()
# tempveg = numpy.zeros((ndviGrid.shape[0],ndviGrid.shape[1],3))
# tempveg[:,:,:] = (ndviGrid[:,:,numpy.newaxis]+1) / 2.0;
# imshow(tempveg)
# show()
# figure()
# tempveg = numpy.zeros((vegetationMask.shape[0],vegetationMask.shape[1],3))
# tempveg[:,:,:] = vegetationMask[:,:,numpy.newaxis];
# imshow(tempveg)
# show()
# figure()
# tempveg = numpy.zeros((vegetationMask.shape[0],vegetationMask.shape[1],3))
# tempveg[:,:,:] = heightGrid[:,:,numpy.newaxis] / numpy.max(heightGrid);
# imshow(tempveg)
# show()
elif LCC:
vegetationMask = lccGrid > 0
heightValues = heightGrid[vegetationMask]
baseValues = ndviGrid[vegetationMask]
lccValues = lccGrid[vegetationMask]
lengthX, lengthY = heightGrid.shape
nTypes = 7
area = "NDVI"
lXTemp = lengthX
lYTemp = lengthY
if lengthX % 2 == 1:
lXTemp += 1
if lengthY % 2 == 1:
lYTemp += 1
vegetationMaskExtended = np.zeros((lXTemp, lYTemp), dtype=bool)
vegetationMaskExtended[0:lengthX, 0:lengthY] = vegetationMask
res = 2.0 # 2.0
wangGridLengthX = np.ceil(lengthX / res)
wangGridLengthY = np.ceil(lengthY / res)
xWangIndices, yWangIndices = numpy.indices((wangGridLengthX, wangGridLengthY))
blocks = view_as_blocks(vegetationMaskExtended, block_shape=(int(res), int(res)))
blocks = blocks.reshape(wangGridLengthX, wangGridLengthY, res * res)
blocks_summed = np.sum(blocks, axis=2)
wangVegetationMask = blocks_summed > 0
print "wvm", wangVegetationMask.shape
print "vm", vegetationMask.shape
else:
vegetationMask = coverageModelGrid > 0
heightValues = heightModelGrid[vegetationMask] * 1
baseValues = coverageModelGrid[vegetationMask] * .01
lengthX, lengthY = heightModelGrid.shape
nTypes = 3
heightValues += numpy.fabs(numpy.min(heightValues))
minHeight = numpy.min(heightValues)
maxHeight = numpy.max(heightValues)
heightValues = (heightValues - minHeight) / (maxHeight - minHeight)
rgb = (heightModelGrid - numpy.min(heightModelGrid)) / (numpy.max(heightModelGrid) - numpy.min(heightModelGrid))
rgb *= 255
heightRGBA = numpy.zeros((heightModelGrid.shape[0], heightModelGrid.shape[1], 3), dtype=numpy.uint8)
heightRGBA[:, :, 0:3] = rgb[:, :, numpy.newaxis]
# misc.imsave('heightMap_Ecomodel.png',heightRGBA)
rgb = (coverageModelGrid - numpy.min(coverageModelGrid)) / (
numpy.max(coverageModelGrid) - numpy.min(coverageModelGrid))
rgb *= 255
coverageRGBA = numpy.zeros((coverageModelGrid.shape[0], coverageModelGrid.shape[1], 3), dtype=numpy.uint8)
coverageRGBA[:, :, 0:3] = rgb[:, :, numpy.newaxis]
# misc.imsave('coverageMap_Ecomodel.png',coverageRGBA)
heightValues *= 100
area = "MODEL"
tileIDs = numpy.arange(0, heightValues.shape[0])
tileIDsGrid = numpy.zeros((lengthX, lengthY))
tileIDsGrid[vegetationMask] = tileIDs
wangRes = 1
res = 1
wangLengthX = lengthX * res
wangLengthY = lengthY * res
wangVegetationMask = numpy.zeros((wangLengthX, wangLengthY), dtype=bool)
xWangIndices, yWangIndices = numpy.indices((wangLengthX, wangLengthY))
xw = numpy.trunc(xWangIndices / res).astype(int)
yw = numpy.trunc(yWangIndices / res).astype(int)
wangVegetationMask[xWangIndices, yWangIndices] = vegetationMask[xw, yw]
print wangVegetationMask.shape
nTiles = heightValues.shape[0]
plantTypeLevels = numpy.zeros((nTypes))
if PaulinaPolder:
if LCC:
area2 = "LCC"
elif NDVI:
area2 = "NDVI"
heightMatrix = getHeightMatrix("NDVI")
coverageMatrix = getCoverageMatrix(area2) * .01
hurstMatrix = getHurstMatrix("NDVI")
else:
heightMatrix = getHeightMatrix("MODEL")
coverageMatrix = getCoverageMatrix("MODEL") * .01
hurstMatrix = getHurstMatrix("MODEL")
unit = 1.0
heightBins = numpy.zeros((2, heightValues.shape[0]))
heightBins[:, :] = (heightValues / unit)[numpy.newaxis, :]
heightBins = numpy.trunc(heightBins)
heightBins[1, :] += 1
heightBins *= unit
heightBins = numpy.unique(heightBins)
# print heightBins
heightCoverage = numpy.zeros((nTypes, heightBins.size))
heightHurst = numpy.zeros((nTypes, heightBins.size))
for i in range(0, nTypes):
fc = interpolate.interp1d(heightMatrix, coverageMatrix[i], kind='slinear', bounds_error=False, fill_value=0)
fh = interpolate.interp1d(heightMatrix, hurstMatrix[i], kind='slinear', bounds_error=False, fill_value=0)
heightCoverage[i] = fc(heightBins)
heightHurst[i] = fh(heightBins)
print "Declare data"
coveragePerTile = numpy.zeros((nTypes, nTiles), dtype=float)
hurstPerTile = numpy.zeros((nTypes, nTiles), dtype=float)
constraintsPerTile = numpy.zeros((nTypes, nTiles), dtype=float)
compositionPerTile = numpy.zeros((nTypes, nTiles), dtype=float)
ndviPerTile = numpy.ones((nTypes, nTiles), dtype=float)
print "Get data"
for i in range(0, nTypes):
fc = interpolate.interp1d(heightBins, heightCoverage[i], kind='slinear', bounds_error=False, fill_value=0)
fh = interpolate.interp1d(heightBins, heightHurst[i], kind='slinear', bounds_error=False, fill_value=0)
coveragePerTile[i] = fc(heightValues)
hurstPerTile[i] = fh(heightValues)
if NDVI and LCC:
ndviPerTile[i] = calculateCovNDVI(area, i, baseValues)
lccTiles = calculateCovLCC(i, lccValues)
constraintsPerTile[i] = numpy.minimum(ndviPerTile[i], lccTiles)
compositionPerTile[i] = numpy.minimum(coveragePerTile[i], constraintsPerTile[i])
elif NDVI:
ndviPerTile[i] = calculateCovNDVI(area, i, baseValues)
constraintsPerTile[i] = ndviPerTile[i]
compositionPerTile[i] = numpy.minimum(coveragePerTile[i], constraintsPerTile[i])
elif LCC:
constraintsPerTile[i] = calculateCovLCC(i, lccValues)
compositionPerTile[i] = numpy.minimum(coveragePerTile[i], constraintsPerTile[i])
if LCC:
if NDVI:
tempCov = numpy.minimum(ndviPerTile, coveragePerTile)
else:
tempCov = numpy.minimum(ndviPerTile, coveragePerTile) # coveragePerTile.copy()#numpy.ones((nTypes,nTiles))
lccGroups = 4
noCoveragePerTile = 1 - numpy.sum(tempCov, axis=0)
noCoveragePerTile[noCoveragePerTile < 0] = 0
for i in range(0, lccGroups):
lccMask = numpy.where(lccValues == i + 1)[0]
plantsLCC = getLCCTypes(i + 1)
tcon = constraintsPerTile[:, lccMask]
tcon = tcon[plantsLCC, :]
totCovlcc = numpy.sum(tcon, axis=0)
for j in range(0, plantsLCC.shape[0]):
compositionPerTile[plantsLCC[j], lccMask] /= (totCovlcc + noCoveragePerTile[lccMask])
elif NDVI:
totCov = numpy.sum(compositionPerTile, axis=0)
noCoveragePerTile = 1 - totCov
noCoveragePerTile[noCoveragePerTile < 0] = 0
compositionPerTile /= totCov + noCoveragePerTile
xWang = xWangIndices[wangVegetationMask]
yWang = yWangIndices[wangVegetationMask]
nWangTiles = xWang.shape[0]
print "Start Point Generation"
if PaulinaPolder:
tileIDs = numpy.arange(0, nTiles)
tileIDsGrid = numpy.zeros((lengthX, lengthY))
tileIDsGrid[vegetationMask] = tileIDs
dist = numpy.array([.4]) / res # .4 standard # .8 medium# 1.6 low
points = wangtiles.generatePoints_cornerbased(wangGridLengthX, wangGridLengthY, nWangTiles, xWang, yWang, dist,
1, 4)
points[:, 0:2] *= res
trunckedPoints = np.trunc(points[:, 0:2]).astype(int)
cull = vegetationMaskExtended[trunckedPoints[:, 0], trunckedPoints[:, 1]]
points = points[cull, :]
trunckedPoints = trunckedPoints[cull, :]
points[:, 3] = tileIDsGrid[trunckedPoints[:, 0], trunckedPoints[:, 1]]
else:
dist = numpy.array([.33]) / wangRes
points = wangtiles.generatePoints_cornerbased(wangLengthX, wangLengthY, nWangTiles, xWang, yWang, dist, 1, 4)
points[:, 0:2] *= wangRes
trunckedPoints = np.trunc(points[:, 0:2] / (res * wangRes)).astype(int)
points[:, 3] = tileIDsGrid[trunckedPoints[:, 0], trunckedPoints[:, 1]]
figure(constrain_navigation=True, constrain_ratio=True, antialiasing=True)
# imshow(numpy.rot90(tempveg,1))
plot(points[:, 0], points[:, 1], color='w', primitive_type='POINTS', marker='.', marker_size=6)
show()
totalPoints = points.shape[0]
print totalPoints
print totalPoints / (nTiles * 1.0);
hurstPoints = hurstPerTile[:, points[:, 3].astype(int)]
coveragePoints = compositionPerTile[:, points[:, 3].astype(int)]
heightPoints = heightValues[points[:, 3].astype(int)]
nocoveragePoints = noCoveragePerTile[points[:, 3].astype(int)]
basePoints = constraintsPerTile[:, points[:, 3].astype(int)]
print "Generate MultiFractal Noise"
scaleFactor = math.sqrt(totalPoints / (nTiles * 1.0)) * 0.5
cNoisePoints = fractalnoise.generateFractalNoise(points[:, 0:2] * scaleFactor, nTypes, hurstPoints, hurstPoints)
avgFreqT = numpy.average(hurstPoints, axis=1)
meanFreq = numpy.average(avgFreqT)
sd = numpy.fabs(avgFreqT - meanFreq)
order = numpy.argsort(sd)[::-1]
cNoisePoints2 = cNoisePoints.copy()
maxHurst = numpy.max(cNoisePoints2, axis=1)
minHurst = numpy.min(cNoisePoints2, axis=1)
noiseTypes = (cNoisePoints2 - minHurst[:, numpy.newaxis]) / (
maxHurst[:, numpy.newaxis] - minHurst[:, numpy.newaxis])
for i in range(0, nTypes):
noiseColors = numpy.zeros((points.shape[0], 3))
noiseColors += noiseTypes[i, :, numpy.newaxis]
figure(constrain_navigation=True, constrain_ratio=True, antialiasing=True)
plot(points[:, 0], points[:, 1], primitive_type='POINTS', color=noiseColors, marker='.', marker_size=6)
show()
print "Start classification process"
pointType = PlantSpeciesGeneration.speciesGeneration(cNoisePoints, coveragePoints.copy(), heightPoints, basePoints, points, heightCoverage,
nocoveragePoints.copy(), unit, order, plantTypeLevels)
# pointType = classify_final(cNoisePoints,coveragePoints, heightPoints,diffHeight,points,False)
print "visualize and writing"
resultMask = pointType > 0;
fractalValues = cNoisePoints.copy();
resultTotPoints = numpy.sum(resultMask);
maxfract = numpy.max(fractalValues, axis=1);
fractalValues /= maxfract[:, numpy.newaxis];
selectedFractalValues = fractalValues[:, (pointType - 1)];
results = numpy.zeros((resultTotPoints), dtype=[('x', numpy.float64), ('y', numpy.float64), ('t', numpy.int64)])
results['x'] = points[resultMask, 0]
results['y'] = points[resultMask, 1]
results['t'] = pointType[resultMask]
# results['f'] = selectedFractalValues[resultMask];
pointColors = getColor(pointType[resultMask])
# numpy.savetxt('locations_ecomodel.txt',results, delimiter=" ", fmt="%s")
tot2 = numpy.sum(coveragePoints, axis=0) + nocoveragePoints
tot2[tot2 == 0] = 1
coveragePoints /= tot2
tot = float(points.shape[0])
print numpy.sum(nocoveragePoints) / tot, numpy.sum(pointType == 0) / tot
print numpy.sum(coveragePoints[0, :]) / tot, numpy.sum(pointType == 1) / tot
print numpy.sum(coveragePoints[1, :]) / tot, numpy.sum(pointType == 2) / tot
print numpy.sum(coveragePoints[2, :]) / tot, numpy.sum(pointType == 3) / tot
if PaulinaPolder:
print numpy.sum(coveragePoints[3, :]) / tot, numpy.sum(pointType == 4) / tot
print numpy.sum(coveragePoints[4, :]) / tot, numpy.sum(pointType == 5) / tot
print numpy.sum(coveragePoints[5, :]) / tot, numpy.sum(pointType == 6) / tot
print numpy.sum(coveragePoints[6, :]) / tot, numpy.sum(pointType == 7) / tot
if PaulinaPolder:
interval = int(20)
bins = 31
else:
interval = int(5)
bins = 21
coverageStatistics = numpy.zeros((nTypes, 2, bins))
for t in range(0, nTypes):
for i in range(0, bins):
b1 = i * interval
b2 = i * interval + interval
m1 = heightPoints >= b1
m2 = heightPoints < b2
m = numpy.logical_and(m1, m2)
tot = float(numpy.sum(m))
if tot == 0:
tot = 1
coverageStatistics[t, 0, i] = numpy.sum(coveragePoints[t, m]) / tot
coverageStatistics[t, 1, i] = numpy.sum(pointType[m] == t + 1) / tot
print coverageStatistics[t, :]
if PaulinaPolder:
figure(constrain_navigation=True, constrain_ratio=True, antialiasing=True)
plot(results['x'], results['y'], primitive_type='POINTS', color=pointColors, marker='.', marker_size=6)
show()
else:
figure(constrain_navigation=True, constrain_ratio=False, antialiasing=True)
plot(results['x'], results['y'], primitive_type='POINTS', color=pointColors, marker='.', marker_size=5)
show()
import matplotlib.cm as cm
from skimage import data
from skimage import color
from skimage.util.shape import view_as_blocks
# -- get `astronaut` from skimage.data in grayscale
l = color.rgb2gray(data.astronaut())
# -- size of blocks
block_shape = (4, 4)
# -- see `astronaut` as a matrix of blocks (of shape
# `block_shape`)
view = view_as_blocks(l, block_shape)
# -- collapse the last two dimensions in one
flatten_view = view.reshape(view.shape[0], view.shape[1], -1)
# -- resampling `astronaut` by taking either the `mean`,
# the `max` or the `median` value of each blocks.
mean_view = np.mean(flatten_view, axis=2)
max_view = np.max(flatten_view, axis=2)
median_view = np.median(flatten_view, axis=2)
# -- display resampled images
fig, axes = plt.subplots(2, 2, figsize=(8, 8), sharex=True, sharey=True)
ax0, ax1, ax2, ax3 = axes.ravel()
ax0.set_title("Original rescaled with\n spline interpolation (order=3)")
