Incorporating content sensitivity into standard video compression algorithms

Abstract

Typical implementations of standard video compression algorithms use variance-based perceptual masking to adaptively vary the degree of compression noise with video content; the idea being that the human visual system is more sensitive to errors in low variance regions of the image. On a similar note, quality assessment algorithms employ similar spatial-frequency adaptive measures to judge compressed video quality. Such simplistic analysis leads to problems when we encounter clean edges in video (for example, text) which although classified as high frequency areas, are still very sensitive to compression noise. Such discrimination between good and bad texture is particularly important for low bit rate, internet video transmission. In this paper we will propose a novel, computationally simple scheme which further distinguishes between clean and random edges. By comparing the proposed algorithm with other standard texture classification schemes in literature, we will show that we obtain a substantial improvement in discriminatory ability, albeit at a much lower computational cost. Further more, we will show that the proposed scheme can be applied on a per macro-block basis, thus making it particularly suitable to drive standard, block based video compression algorithms such as H.261/3 or MPEG-1/2 which are quite popular.