SJ-PVC: An Efficient Perceptual Video Compression Scheme Based on Adaptive QP and Rate-Distortion Optimization

Perceptual Video Compression (PVC) is a promising approach to enhancing compression efficiency. The Human Visual System (HVS) possesses many important perceptual characteristics, which can be utilized to further enhance encoding efficiency without significantly degrading perceptual quality. This paper addresses the issue that existing video compression methods have not fully leveraged HVS characteristics by proposing a video compression scheme, SJ-PVC, that uses a Just Noticeable Distortion (JND) estimation model based on HVS characteristics. Specifically, we design a structurally simplified network to address the structural redundancy in existing multi-scale feature-based Video Saliency Prediction (VSP) models. This network simplifies the model structure while maintaining high accuracy. Furthermore, we propose an adaptive Quantization Parameter (QP) selection algorithm that classifies each CU based on JND characteristics and saliency maps, allowing for more precise control of QP values, which significantly enhances the overall visual quality of the video. Finally, we introduce a Rate-Distortion Optimization algorithm based on HVS characteristics, which considers visual masking effects and saliency information during the encoding process to select the optimal encoding scheme. Experimental results demonstrate that SJ-PVC improves subjective video quality, significantly reduces bitrate, and shortens encoding time.