IIRNet: Infinite impulse response inspired network for compressed video quality enhancement

Abstract

Deep learning approaches have demonstrated notable advancements in a compressed video quality enhancement (CVQE) task. Nevertheless, they inadequately account for the loss of correlation detail information between features caused by the large motion in the video sequence and fail to comprehensively address the impact of diverse compression artifacts on the inaccurate capture of inter-frame dependent information. To address these issues, we find that the process of enhancing the input videos to restore high-quality video in the CVQE task can be analogized to the process of enhancing the input signals to emphasize useful signals in the digital signal processing (DSP) domain. In this paper, we present an infinite impulse response inspired network (IIRNet) for CVQE. Specifically, we propose to employ a design scheme inspired by the five-point difference format (FDF) of IIR and develop a multilevel 3D-FDF cycle module to obtain rich detailed feature information at multiple levels to maintain a strong correlation between features. Furthermore, we devise a cross-domain point fusion attention module to extract dissimilar information from the frequency domain of adjacent frames by non-local dissimilarity technique, especially in the region prone to artifacts, for efficient processing. The proposed IIRNet is evaluated on the public MFQE 2.0 dataset, demonstrating effectiveness in enhancing the quality of compressed videos by removing compression artifacts and reducing noise. Compared with the state-of-the-art (SOTA) methods, the proposed method achieves better performance in both objective metrics and visual quality.