Padding is the addition of artificial data at the borders of the original data, so that a convolution kernel can operate at the borders (of the original data). For example: in order for a convolution of a 2D feature map with a 3x3 kernel to produce an output of the same size as the input (assuming stride of 1), the input feature map needs to be padded with one extra line of pixels on each side (top, bottom, left, and right). The values at the extended borders are often zero, in which case the process is known as zero-padding.