Lightweight Wavelet-Based Network for JPEG Artifacts Removal

Abstract

In recent years, deep learning-based methods have made remarkable progress in removing blocking artifacts caused by JPEG compression. However, most of these methods directly learn the mapping between compressed images and corresponding clear versions in pixel domain by designing complex network structures and consuming huge computing power, which limits the application on mobile devices. To address this issue, we propose a lightweight wavelet-based network for JPEG compression artifact removal. Specifically, we observe that the signal removal operation in JPEG will introduce diverse distortion for different sub-frequency spectrums. Based on this, we propose a divide-and-conquer strategy to learn the mapping on each sub-frequency spectrum with a lightweight sub-network to reduce network parameters and save computation power. Furthermore, we speed up the inference time by integrating reparameterization technology. The comparison results with the state-of-the-art methods on mobile device demonstrate that our method achieves comparable deblocking performance with ${\times }$100 less computationally complex and ${\times }$50 faster inference time.