Video applications use different types of filters extensively for multiple reasons: filter noise, manipulate motion blur, enhance color and contrast, edge detection, creative effects, etc. At its core, a convolution video filter carries out some form of data average around a pixel, which redefines the amount and type of correlation any pixel has to its surrounding area. Such filtering is carried out for all the pixels in a video frame.
A matrix of coefficients defines a convolution filter. The convolution operation is essentially a sum of products performed on a pixel set (a frame/image sub-matrix centered around a given pixel) and a coefficients matrix. The following figure illustrates how convolution is calculated for a pixel; it is highlighted in yellow. Here the filter has a coefficient matrix that is 3x3 in size. The figure also displays how the whole output image is generated during the filtering process. The index of the output pixel being generated is the index of the input pixel highlighted in yellow that is being filtered. In algorithmic terms, the process of filtering consists of:
Selecting an input pixel as highlighted in yellow in the following figure
Extracting a sub-matrix whose size is the same as filter coefficients
Calculating an element-wise sum-of-product of the extracted sub-matrix and coefficients matrix
Placing the sum-of-product as an output pixel in an output image/frame on the same index as the input pixel
