Faster Linear Convolution Using Careful Zero Padding
When computing a linear convolution between an image I and a kernel K, the convention is to use FFT, pad with zeroes (also called zero filling) and use the convolution theorem to do this efficiently.…