Detecting features at the edge of images

• when applying convolution
• there is a difference in the number of times that each pixel in the image is used when we determine the feature map.
• This difference is largest at the corners of the image
• $\therefore$ difficult to detect features at edges of the image

Zero-padding

• include some imaginary pixels at the border of image.
• The amount of zero-padding is usually set to 0.5*(kernel_size – 1)