Compressed Video Enhancement with Spatio-Temporal Multi-Feature Extraction

Abstract

Bandwidth limitations require transmitting the original video in a compressed format, leading to video artifacts and quality degradation. Therefore, compressed video enhancement is very important. In recent years, deep learning methods have gained prominence in compressed video enhancement due to their ability to efficiently restore lost details in specific frames by extracting features. Information extraction can be influenced by video artifacts in preceding and subsequent frames, even when techniques like optical flow estimation or quality-based frame selection are employed. To better capture artifact removal details, this paper introduces a spatiotemporal multi-feature extraction method for more accurate video reconstruction. The experimental results show that the peak Signal-To-Noise Ratio (PSNR) of the video enhanced by the method proposed in this paper is about 0.06 dB higher(QP37) on average when compared to the video enhanced by some other methods (e.g., MFQE2.0, MFQE1.0, DS-CNN).