Test-Time Optimization for Post-Processing of Compressed Videos

Hongil Kim, Changwoo Han, Donghyun Kim, Sung-Chang Lim, Seung-Won Jung

Abstract: Lossy compression is widely used for video compression, but it often introduces compression artifacts that degrade the visual quality of compressed videos. Consequently, numerous deep learning-based methods have been developed to post-process compressed videos. However, previous post-processing models often encounter difficulties when there is a domain gap between the training and test datasets. Test-time optimization (TTO), a technique that finetunes the model during the test stage, has been considered an effective solution to address the domain gap problem. In this paper, we introduce a novel TTO method specialized for compression artifacts reduction. Specifically, we propose using image pairs available on the decoder-side, i.e., the images before and after the adaptive loop filtering of the versatile video coding standard, as input and target of TTO such that the post-processing model can be adapted to the characteristics of test data. Experimental results on several baseline models and test datasets demonstrate the effectiveness of the proposed method in post-processing compressed videos.