Zheng Bolun ; Sun Rui ; Tian Xiang ; Chen Yaowu

Abstract

Recent studies have used deep residual convolutional neural networks (CNNs)for JPEG compression artifact reduction. This study proposes a scalable CNNcalled S-Net. Our approach effectively adjusts the network scale dynamically ina multitask system for real-time operation with little performance loss. Itoffers a simple and direct technique to evaluate the performance gains obtainedwith increasing network depth, and it is helpful for removing redundant networklayers to maximize the network efficiency. We implement our architecture usingthe Keras framework with the TensorFlow backend on an NVIDIA K80 GPU server. Wetrain our models on the DIV2K dataset and evaluate their performance on publicbenchmark datasets. To validate the generality and universality of the proposedmethod, we created and utilized a new dataset, called WIN143, forover-processed images evaluation. Experimental results indicate that ourproposed approach outperforms other CNN-based methods and achievesstate-of-the-art performance.