Perceptual Video Content Analysis and Application to HEVC Quantization Refinement

Abstract

In this paper, we propose a set of perceptual features aiming to consistently describe the visual information. The measurement is performed in the complex frequency domain according to human visual system (HVS) mechanisms. The aim is to explore the performance of these features in a video coding scheme. Particularly, we consider the High Efficiency Video Coding (HEV C) standard as it introduces several efficient tools along with new coding structures. The quantization parameter (QP) is an essential factor that affects the coding performance and has a relationship the Lagrangian multiplier. Based on extracted measures, a perceptual factor is proposed to adjust the Lagrangian multiplier and subsequently, the QP is refined over the adjusted value. The achieved BD-rate savings over several resolutions of video sequences, using the Bjontegaard metric, show the promising coding efficiency of the proposed method with regard to an adequate rate-distortion (R-D) compromise. We opted for the Structural SIMilarity (SSIM) metric to carry out a perceptual R-D comparison. The R-D curves demonstrate that the obtained bitrate savings are associated to convenient quality measures, compared to HEVC anchor and a state-of-the-art QP refinement model.