A Nonlinear Loop Filter Model for Video Coding

Hybrid coding schemes employing motion compensated prediction in the time domain and transform coding in the spatial domain have been quite popular for codecs operating in the 64 KBPS - 2 MBPS range. In order to reduce high frequency artifacts in troduced by the motion compensation and quantization noise in the feedback loop, a loop filter is commonly applied to the contents of the frame store in the prediction loop. Mostly, linear filters giving rise to blurred reconstructed images have been employed. In this paper we present some new nonlinear and locally adaptive filters for use as loop filters. These filters remove the accumulation of high frequency artifacts and introduce considerable less blurring than the standard CCITT H.261 loop filter. The proposed filters smooth small intensity variations in the intensity function, leave monotonic image segments unfiltered and are based on a sound theory for iterative filtering. Simulation results show that the application of the presented filters as loop filter for hybrid video coders gives rise to improved decoded images both in terms of visual quality and peak-to-peak signal to noise ratio (SNR).