A Method for Enhancing the Quality of Compressed Videos Based on 2D Convolution and Aggregating Spatio-Temporal Information

Existing block-based video coding frameworks are often affected by the quantization step size and motion compensation accuracy, resulting in the loss of high-frequency information and compression artifacts. Especially in the case of limited coding resources, the blurring of content edges and obvious compression distortion will have a negative impact on the subjective quality of the video. Therefore, there is an urgent need to build a quality enhancement method to improve the compressed video quality at the receiving end under the same coding resources. This paper proposes a compression video quality enhancement method based on 2D convolution that aggregates spatial and temporal information. Based on the objective facts of analyzing the spatiotemporal correlation and video quality fluctuation, this method constructs a multi-frame input mechanism consisting of the current frame to be enhanced and its adjacent frames; furthermore, it efficiently extracts and integrates the temporal and spatial information features of the input video sequence by utilizing the excellent feature extraction and fusion capabilities of the encoder-decoder structure, achieving implicit alignment. On this basis, an attention mechanism is integrated to more accurately locate and extract key information in the video, thereby more accurately restoring the detail information in the video and improving the performance of the model. In public benchmark tests, our method achieved average ΔPSNR gains of 0.801 dB, 0.796 dB, 0.792 dB, and 0.714 dB on 18 video test sequences with QP = 22, 27, 32, and 37, respectively, outperforming other methods. Compared with the state-of-the-art algorithms, our method achieved speed improvements of 13.2%, 10.5%, and 6.2% for processing videos with resolutions of 832 × 480, 1080 × 720, and 1920 × 1080, respectively. The above results show that our method can improve the compressed video quality at the receiving end under the same coding resources and outperforms other methods in terms of performance.