Convolution

The script runs a 5x5 Convolutional Kernel over the image to recognize horizontal and vertical lines.

Then it draws the horizontal lines red and the vertical ones green.

from PIL import Image
import numpy as np

im = Image.open('../bbtor.jpg')
bw = im.convert('L')  # grayscale

new_size = bw.size[0] // 4, bw.size[1] // 4  # smaller image
small = bw.resize(new_size)

a = np.array(small).astype(np.float64)

# define a 5x5 kernel with preset weights
kernel = np.zeros((5, 5), dtype=np.float64)
kernel[1,:] = -1.0
kernel[2,:] = 2.0
kernel[3,:] = -1.0


def convolve(image, kernel):
    """apply a 5x5 convolutional kernel on the image"""
    feature_map = []
    for y in range(0, image.shape[0] - 4):
        for x in range(0, image.shape[1] - 4):
            dot = (image[y:y+5, x:x+5] * kernel).sum()
            feature_map.append(dot)

    # reshape, cutting of borders
    feature_map = np.array(feature_map, dtype=np.float64)
    new_shape = image.shape[0] - 4, image.shape[1] - 4
    return feature_map.reshape(new_shape)


def scale(fm, threshold=120):
    """scale with a threshold value found by trial and error"""
    fm += threshold
    fm[fm < 0] = 0.0
    fm = 255 * fm / fm.max()
    fm = fm.astype(np.uint8)  # prepare for image
    return fm


# apply the convolution
fm_horizontal = scale(convolve(a, kernel))
fm_vertical = scale(convolve(a, kernel.T))

# set color channels to result
b = np.zeros((fm_horizontal.shape[0], fm_horizontal.shape[1], 3), dtype=np.uint8)
b[:,:,0] = fm_horizontal
b[:,:,1] = fm_vertical

final = Image.fromarray(b, mode='RGB')
final = final.resize((final.size[0]*2, final.size[1]*2)) # increase image size

final.save('convolution.png')