Frequency-Dependent Perceptual Quantisation for Visually Lossless Compression Applications

TL;DR

This paper introduces a Frequency-Dependent Perceptual Quantisation (FDPQ) method for HEVC that selectively coarsens quantisation of less perceptually relevant high-frequency coefficients, achieving significant bitrate reductions while maintaining perceptual quality.

Contribution

The novel FDPQ technique leverages MTF characteristics to improve perceptual coding efficiency in HEVC by adaptively quantising transform coefficients based on perceptual relevance.

Findings

Achieves up to 41% bitrate reduction compared to RDOQ.

Produces visually lossless video at given QP levels.

Maintains perceptual quality despite coarser quantisation of high-frequency coefficients.

Abstract

The default quantisation algorithms in the state-of-the-art High Efficiency Video Coding (HEVC) standard, namely Uniform Reconstruction Quantisation (URQ) and Rate-Distortion Optimised Quantisation (RDOQ), do not take into account the perceptual relevance of individual transform coefficients. In this paper, a Frequency-Dependent Perceptual Quantisation (FDPQ) technique for HEVC is proposed. FDPQ exploits the well-established Modulation Transfer Function (MTF) characteristics of the linear transformation basis functions by taking into account the Euclidean distance of an AC transform coefficient from the DC coefficient. As such, in luma and chroma Cb and Cr Transform Blocks (TBs), FDPQ quantises more coarsely the least perceptually relevant transform coefficients (i.e., the high frequency AC coefficients). Conversely, FDPQ preserves the integrity of the DC coefficient and the very low frequency AC coefficients. Compared with RDOQ, which is the most widely used transform coefficient-level quantisation technique in video coding, FDPQ successfully achieves bitrate reductions of up to 41%. Furthermore, the subjective evaluations confirm that the FDPQ-coded video data is perceptually indistinguishable (i.e., visually lossless) from the raw video data for a given Quantisation Parameter (QP).