def sobel_filter(image):
kernel_x = np.array([[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]])
kernel_y = np.array([[1, 2, 1], [0, 0, 0], [-1, -2, -1]])
dst_x = img_convolve(image, kernel_x)
dst_y = img_convolve(image, kernel_y)
dst = np.sqrt((np.square(dst_x)) + (np.square(dst_y))).astype(np.uint8)
degree = np.arctan2(dst_y, dst_x)
return dst, degree
def sobel_filter(image):
kernel_x = np.array([[-1, 0, 1], [-2, 0, 2], [-1, 0, 1]])
kernel_y = np.array([[1, 2, 1], [0, 0, 0], [-1, -2, -1]])
dst_x = np.abs(img_convolve(image, kernel_x))
dst_y = np.abs(img_convolve(image, kernel_y))
# modify the pix within [0, 255]
dst_x = dst_x * 255 / np.max(dst_x)
dst_y = dst_y * 255 / np.max(dst_y)
dst_xy = np.sqrt((np.square(dst_x)) + (np.square(dst_y)))
dst_xy = dst_xy * 255 / np.max(dst_xy)
dst = dst_xy.astype(np.uint8)
theta = np.arctan2(dst_y, dst_x)
return dst, theta
def canny(image, threshold_low=15, threshold_high=30, weak=128, strong=255):
image_row, image_col = image.shape[0], image.shape[1]
# gaussian_filter
gaussian_out = img_convolve(image, gen_gaussian_kernel(9, sigma=1.4))
# get the gradient and degree by sobel_filter
sobel_grad, sobel_theta = sobel_filter(gaussian_out)
gradient_direction = np.rad2deg(sobel_theta)
gradient_direction += PI
dst = np.zeros((image_row, image_col))
"""
Non-maximum suppression. If the edge strength of the current pixel is the largest
compared to the other pixels in the mask with the same direction, the value will be
preserved. Otherwise, the value will be suppressed.
"""
for row in range(1, image_row - 1):
for col in range(1, image_col - 1):
direction = gradient_direction[row, col]
if (
0 <= direction < 22.5
or 15 * PI / 8 <= direction <= 2 * PI
or 7 * PI / 8 <= direction <= 9 * PI / 8
):
W = sobel_grad[row, col - 1]
E = sobel_grad[row, col + 1]
if sobel_grad[row, col] >= W and sobel_grad[row, col] >= E:
dst[row, col] = sobel_grad[row, col]
elif (PI / 8 <= direction < 3 * PI / 8) or (
9 * PI / 8 <= direction < 11 * PI / 8
):
SW = sobel_grad[row + 1, col - 1]
NE = sobel_grad[row - 1, col + 1]
if sobel_grad[row, col] >= SW and sobel_grad[row, col] >= NE:
dst[row, col] = sobel_grad[row, col]
elif (3 * PI / 8 <= direction < 5 * PI / 8) or (
11 * PI / 8 <= direction < 13 * PI / 8
):
N = sobel_grad[row - 1, col]
S = sobel_grad[row + 1, col]
if sobel_grad[row, col] >= N and sobel_grad[row, col] >= S:
dst[row, col] = sobel_grad[row, col]
elif (5 * PI / 8 <= direction < 7 * PI / 8) or (
13 * PI / 8 <= direction < 15 * PI / 8
):
NW = sobel_grad[row - 1, col - 1]
SE = sobel_grad[row + 1, col + 1]
if sobel_grad[row, col] >= NW and sobel_grad[row, col] >= SE:
dst[row, col] = sobel_grad[row, col]
"""
High-Low threshold detection. If an edge pixel’s gradient value is higher
than the high threshold value, it is marked as a strong edge pixel. If an
edge pixel’s gradient value is smaller than the high threshold value and
larger than the low threshold value, it is marked as a weak edge pixel. If
an edge pixel's value is smaller than the low threshold value, it will be
suppressed.
"""
if dst[row, col] >= threshold_high:
dst[row, col] = strong
elif dst[row, col] <= threshold_low:
dst[row, col] = 0
else:
dst[row, col] = weak
"""
Edge tracking. Usually a weak edge pixel caused from true edges will be connected
to a strong edge pixel while noise responses are unconnected. As long as there is
one strong edge pixel that is involved in its 8-connected neighborhood, that weak
edge point can be identified as one that should be preserved.
"""
for row in range(1, image_row):
for col in range(1, image_col):
if dst[row, col] == weak:
if 255 in (
dst[row, col + 1],
dst[row, col - 1],
dst[row - 1, col],
dst[row + 1, col],
dst[row - 1, col - 1],
dst[row + 1, col - 1],
dst[row - 1, col + 1],
dst[row + 1, col + 1],
):
dst[row, col] = strong
else:
dst[row, col] = 0
return dst
def test_convolve_filter():
# laplace diagonals
Laplace = array([[0.25, 0.5, 0.25], [0.5, -3, 0.5], [0.25, 0.5, 0.25]])
res = conv.img_convolve(gray, Laplace).astype(uint8)
assert res.any()
def hog_feature(image, gamma=0.4, cell_size=6, bin_size=18, block_size=4):
rows, cols = image.shape[0], image.shape[1]
"""
Gamma normalization, however, the author point out this step can be omitted in HOG descriptor computation.
"""
norm = (image + 0.5) / 255
norm = np.power(norm, 1 / gamma)
img_norm = 255 * norm - 0.5
# Get gradient and angle, the kernel below perform better than others(3x3 Sobel).
kernel_x = np.array([[-1, 0, 1]])
kernel_y = np.array([[-1], [0], [1]])
dst_x = img_convolve(img_norm, kernel_x)
dst_y = img_convolve(img_norm, kernel_y)
dst_xy = np.sqrt((np.square(dst_x)) + (np.square(dst_y)))
dst_xy = dst_xy * 255 / np.max(dst_xy)
gradient_magnitude_global = dst_xy
# Get the angles and convert them in range (0, 360)
theta = np.arctan2(dst_y, dst_x)
gradient_angle_global = np.rad2deg(theta) # range(-180, 180)
gradient_angle_global = np.where(gradient_angle_global < 0,
gradient_angle_global + 360,
gradient_angle_global) # range(0, 360)
"""
Orientation binning, The second step of calculation is creating the cell histograms. Each pixel within the cell
casts a weighted vote for an orientation-based histogram channel based on the values found in the gradient
computation. In tests, the gradient magnitude itself generally produces the best results.
"""
angle_unit = 360 / bin_size
cell_gradient_mtx = np.zeros(
(rows // cell_size, cols // cell_size, bin_size))
for i in range(cell_gradient_mtx.shape[0]):
for j in range(cell_gradient_mtx.shape[1]):
pixes_grad_per_cell = gradient_magnitude_global[i *
cell_size:(i + 1) *
cell_size, j *
cell_size:(j + 1) *
cell_size]
pixes_angle_per_cell = gradient_angle_global[i *
cell_size:(i + 1) *
cell_size, j *
cell_size:(j + 1) *
cell_size]
orientation_centers = [0] * bin_size
for cell_i in range(pixes_grad_per_cell.shape[0]):
for cell_j in range(pixes_grad_per_cell.shape[1]):
gradient_strength = pixes_grad_per_cell[cell_i][cell_j]
gradient_angle = pixes_angle_per_cell[cell_i][cell_j]
bin_index = int(gradient_angle / angle_unit)
if gradient_angle == 360:
bin_index = 0
orientation_centers[bin_index] += gradient_strength
cell_gradient_mtx[i][j] = orientation_centers
"""
Descriptor blocks. Grouping cells into larger spatial blocks and contrast normalizing each
block separately. The final descriptor is then the vector of all components of the normalized cell
responses from all of the blocks in the detection window.
"""
hog_descriptor_mtx = []
out_rows, out_cols = cell_gradient_mtx.shape[
1] - block_size + 1, cell_gradient_mtx.shape[0] - block_size + 1
for i in range(0, out_rows):
for j in range(0, out_cols):
block_vector = np.ravel(cell_gradient_mtx[i:i + block_size,
j:j + block_size, :])
# Block L2 normalization
eps = 1e-5
block_vector = block_vector / np.sqrt(
np.sum(block_vector**2) + eps**2)
hog_descriptor_mtx.append(block_vector)
# showing hog
hog_dst_image = np.zeros([rows, cols])
cell_gradient = cell_gradient_mtx
cell_width = cell_size // 2
max_mag = np.array(cell_gradient).max()
for x in range(cell_gradient.shape[0]):
for y in range(cell_gradient.shape[1]):
cell_grad = cell_gradient[x][y]
cell_grad /= max_mag
angle = 0
angle_gap = angle_unit
for magnitude in cell_grad:
angle_radian = math.radians(angle)
x1 = int(x * cell_size +
magnitude * cell_width * math.cos(angle_radian))
y1 = int(y * cell_size +
magnitude * cell_width * math.sin(angle_radian))
x2 = int(x * cell_size -
magnitude * cell_width * math.cos(angle_radian))
y2 = int(y * cell_size -
magnitude * cell_width * math.sin(angle_radian))
strength = int(255 * math.sqrt(magnitude))
cv2.line(hog_dst_image, (y1, x1), (y2, x2), strength)
angle += angle_gap
return hog_descriptor_mtx, hog_dst_image
