[Python] Use Python's Scipy to convolve a color image with user supplied kernel

I have noted that scipy.ndimage.convolve / scipy.signal.convolve (scipy.ndimage.correlate/ scipy.signal.correlate) gives unexpected (and maybe wrong results) when done on color images with user supplied kernels. However, it works fine for 2d (gray) images. Am I missing something? Kindly help. The following example illustrates this:

import numpy as np, skimage as sk, matplotlib.pyplot as plt, imageio, PIL, scipy
# convolving 3d input with 3d kernel directly in scipy
input_image = np.random.randint(0, 100, (5,5,3)).astype('float32') # input image
kernel = (np.random.randint(0, 100, (3,3,3))).astype('float32') # kernel
kernel[..., 0] /= kernel[..., 0].sum()
kernel[..., 1] /= kernel[..., 1].sum()
kernel[..., 2] /= kernel[..., 2].sum()
output = scipy.signal.convolve(input_image , kernel, 'same')
output = np.clip(output, 0, 255).astype('uint8')

# convolving 3d input with 3d kernel using nested loops (my ground truth)
input_image_padded = np.pad(input_image, 1) [...,1:-1]
s = np.zeros(input_image.shape, 'float32')
for i in range(3):
for x in range(input_image.shape[0]):
for y in range(input_image.shape[1]):
s[x,y,i] = np.sum(input_image_padded[x:x+kernel.shape[0], y:y+kernel.shape[1],i]* kernel[..., i][::-1,::-1])
s = np.clip(s, 0, 255).astype('uint8')

# Check results
print(np.equal(output, s).all())
print(output)
print(s)

# splitting the image and kernel into their constituent 2d images and kernels then convolving
#them appropriately

r1 = scipy.ndimage.convolve(input_image[..., 0], kernel[..., 0], mode =
'constant').astype('uint8')
r2 = scipy.ndimage.convolve(input_image[..., 1], kernel[..., 1], mode =
'constant').astype('uint8')
r3 = scipy.ndimage.convolve(input_image[..., 2], kernel[..., 2], mode =
'constant').astype('uint8')
r = np.dstack([r1,r2,r3])

# Check results
print(np.equal(r, s).all())
print(r)
print(s)

Continue reading...