1. Padding
1. Disadvantages of image edge detection with convolution operation:
1) Image edge detection with convolution operation will make the image Shrink (compressed)
After one convolution, the image resolution is: n - f + 1 * n - f + 1 # The original image resolution is n * n; the convolution kernel size is f * f
Image resolution becomes very small after multiple convolutional layers
2) The pixels at the edge of the image are rarely used and lose a lot of information
The solution is to refill the p = padding circle of pixels around the image (all with a gray value of 0 for convenience)
After one convolution, the image resolution is: n + 2p - f + 1 * n + 2p - f + 1
2. How many circles of pixels to fill (two methods)
1)Valid Convolutions
not filled
2)Same Convolutions
Make the output image the same size as the input image
p = (f - 1) / 2 # f is odd (in the field of computer vision, f is usually odd, if f is even, then asymmetric padding on both sides is required)