Content adaptive pre-filtering for video compression

Abstract

Bitrate reduction with little to no degradation in visual perception is a long-standing challenge in video coding. This paper targets this challenge by adaptively filtering the content prior to video compression and in the preprocessing stage. This is done by applying a bilateral filter where the filter parameters are selected according to regional content complexity and estimated visual importance besides bitrate and quality requirements. A multi-scale metric based on 2D gradient is employed to determine bandwidth requirements of different regions. A random forest regression model is trained to predict distortion and bit requirements for a block, if it is filtered and encoded at a given quality. The predicted distortion and bit requirements are used to select filter parameters considering a cost function. The proposed approach is applied to both H.264 and HEVC encoders, with different GOP structures. The results show up to 60% bitrate reduction in terms of BD-Rate (about 20% on average) for the attempted test cases with little to no noticeable quality degradation.