improve image convolve algorithm and add hog feature

digital_image_processing/edge_detection/canny.py

@@ -1,3 +1,6 @@
+"""
+Implementation of canny edge-detection algorithm
+"""
 import cv2
 import numpy as np
 from digital_image_processing.filters.convolve import img_convolve

digital_image_processing/feature/hog.py

@@ -0,0 +1,115 @@
+"""
+Implementation hog(Histogram of Oriented Gradients) feature of image according
+the paper[https://lear.inrialpes.fr/people/triggs/pubs/Dalal-cvpr05.pdf].
+"""
+
+import math
+import cv2
+import numpy as np
+from digital_image_processing.filters.convolve import img_convolve
+
+
+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
+
+
+if __name__ == '__main__':
+    # read original image in gray model
+    img = cv2.imread('../image_data/lena.jpg', 0)
+    # extract hog feature
+    hog_dsp, hog_img = hog_feature(img)
+    cv2.imshow('hog', hog_img.astype(np.uint8))
+    cv2.waitKey(0)

digital_image_processing/filters/convolve.py

@@ -1,39 +1,104 @@
-# @Author  : lightXu
-# @File    : convolve.py
-# @Time    : 2019/7/8 0008 下午 16:13
+"""
+Implementation of image convolve algorithm
+"""
 from cv2 import imread, cvtColor, COLOR_BGR2GRAY, imshow, waitKey
 from numpy import array, zeros, ravel, pad, dot, uint8
+import math
 
 
-def im2col(image, block_size):
-    rows, cols = image.shape
-    dst_height = cols - block_size[1] + 1
-    dst_width = rows - block_size[0] + 1
-    image_array = zeros((dst_height * dst_width, block_size[1] * block_size[0]))
+def im2col(image, block_size, row_stride, col_stride, dst_rows, dst_cols):
+    """
+    :param image: padded image array
+    :param block_size: filter shape tuple
+    :param row_stride: stride in row channels
+    :param col_stride: stride in row channels
+    :param dst_rows: the rows of the filtered input image
+    :param dst_cols: the cols of the filtered input image
+    :return: the reshape array with shape(dst_rows*dst_cols， block_size[1] * block_size[0])
+    """
+
+    image_array = zeros((dst_rows * dst_cols, block_size[1] * block_size[0]))
     row = 0
-    for i in range(0, dst_height):
-        for j in range(0, dst_width):
-            window = ravel(image[i:i + block_size[0], j:j + block_size[1]])
+    for i in range(0, dst_rows):
+        for j in range(0, dst_cols):
+            window = ravel(image[i * row_stride:i * row_stride + block_size[0],
+                           j * col_stride:j * col_stride + block_size[1]])
             image_array[row, :] = window
             row += 1
 
     return image_array
 
 
-def img_convolve(image, filter_kernel):
+def img_convolve(image, kernel, row_stride=1, col_stride=1):
+    """
+    :param image:  input image array
+    :param kernel: filter kernel array
+    :param row_stride: stride in row channels
+    :param col_stride: stride in row channels
+    :return: the filter result
+
+    Example:
+    >>> image = array([[1,2,3,4,5],[2,3,4,5,6], [1,2,3,4,5]])
+    >>> kernel = array([[-1, 0, 1], [-1, 0, 1], [-1, 0, 1]])
+    >>> img_convolve(image, kernel, 1, 1)
+    array([[3., 6., 6., 6., 3.],
+           [3., 6., 6., 6., 3.],
+           [3., 6., 6., 6., 3.]])
+
+    >>> img_convolve(image, kernel, 2, 1)
+    array([[3., 6., 6., 6., 3.],
+           [3., 6., 6., 6., 3.]])
+
+    >>> img_convolve(image, kernel, 2, 2)
+    array([[3., 6., 3.],
+           [3., 6., 3.]])
+
+    >>> kernel_1_3 = array([[-1, 2, -1]])
+    >>> img_convolve(image, kernel_1_3, 1, 1)
+    array([[-1.,  0.,  0.,  0.,  1.],
+           [-1.,  0.,  0.,  0.,  1.],
+           [-1.,  0.,  0.,  0.,  1.]])
+
+    >>> kernel_3_1 = array([[-2], [1], [-2]])
+    >>> img_convolve(image, kernel_3_1, 1, 1)
+    array([[ -5.,  -8., -11., -14., -17.],
+           [ -2.,  -5.,  -8., -11., -14.],
+           [ -5.,  -8., -11., -14., -17.]])
+
+    >>> kernel_3_1 = array([[-2], [1], [-2]])
+    >>> img_convolve(image, kernel_3_1, 2, 1)
+    array([[ -5.,  -8., -11., -14., -17.],
+           [ -5.,  -8., -11., -14., -17.]])
+
+
+    >>> kernel_3_1 = array([[-2], [1], [-2]])
+    >>> img_convolve(image, kernel_3_1, 2, 2)
+    array([[ -5., -11., -17.],
+           [ -5., -11., -17.]])
+    """
+
     height, width = image.shape[0], image.shape[1]
-    k_size = filter_kernel.shape[0]
-    pad_size = k_size//2
-    # Pads image with the edge values of array.
-    image_tmp = pad(image, pad_size, mode='edge')
+    k_size_row, k_size_col = kernel.shape[0], kernel.shape[1]
 
-    # im2col, turn the k_size*k_size pixels into a row and np.vstack all rows
-    image_array = im2col(image_tmp, (k_size, k_size))
+    # "SAME" convolve mode
+    dst_rows = math.ceil(height / row_stride)  # ceil
+    dst_cols = math.ceil(width / col_stride)  # ceil
+
+    pad_h = max((dst_rows - 1) * row_stride + k_size_row - height, 0)
+    pad_top = pad_h // 2  # floor
+    pad_bottom = pad_h - pad_top
+    pad_w = max((dst_cols - 1) * col_stride + k_size_col - width, 0)
+    pad_left = pad_w // 2  # floor
+    pad_right = pad_w - pad_left
+
+    # Pads image with the edge values of array.
+    image_tmp = pad(array=image,
+                    pad_width=((pad_top, pad_bottom), (pad_left, pad_right)),
+                    mode='edge')
 
-    #  turn the kernel into shape(k*k, 1)
-    kernel_array = ravel(filter_kernel)
-    # reshape and get the dst image
-    dst = dot(image_array, kernel_array).reshape(height, width)
+    image_array = im2col(image_tmp, (k_size_row, k_size_col), row_stride, col_stride, dst_rows, dst_cols)
+    kernel_array = ravel(kernel)
+    dst = dot(image_array, kernel_array).reshape(dst_rows, dst_cols)
     return dst
 
 

digital_image_processing/filters/sobel_filter.py

@@ -1,6 +1,6 @@
-# @Author  : lightXu
-# @File    : sobel_filter.py
-# @Time    : 2019/7/8 0008 下午 16:26
+"""
+Implementation of sobel filter algorithm
+"""
 import numpy as np
 from cv2 import imread, cvtColor, COLOR_BGR2GRAY, imshow, waitKey
 from digital_image_processing.filters.convolve import img_convolve