Development of Statistics and Convolution as Tools for Image Noise Suppression: Statistical Performance Analysis of Spatial Filters

Systems introduce noise into images. Spatial Filtering is required to reduce or eliminate the noise content and improve peak signal to noise ratio (PSNR). Spatial filtering is based on the assumption that noise has a high spatial frequency and, therefore, can be attenuated by a local operation which makes each pixel's intensity roughly consistent with those of its nearest neighbours. The development of thirty-nine Spatial Filtering Algorithms based on combination of three types of convolution and thirteen types of statistics as tools is presented. Comprehensive study and analysis of performance of Spatial Filtering Algorithms for the suppression of six types of noise in a test bed of thirty-six test images is presented with the aid of statistics. Spatial Filters are found to be low pass filters; the higher the Frequency Estimate of the test image, the lower is the gain of the spatial filter. Modified Center Pixel Based Convolution (MCC) is found to perform better than both Center Pixel Based Convolution with Average on the Edges (ACC) and Center Pixel Based Convolution (CC). For suppression of Gaussian Noise, Localvar Noise and Speckle Noise, Gaussian (G) is the best statistics. Median (MDN) is the best statistics for the suppression of Salt & Pepper Noise while Alpha-Trimmed Mean [4 Pixels Excluded] (AT4P) is the best statistics for the suppression of Poisson Noise and Gaussian Plus Salt & Pepper Noise.