Zhang Xiaoshuai ; Yang Wenhan ; Hu Yueyu ; Liu Jiaying

Abstract

JPEG is one of the most commonly used standards among lossy image compressionmethods. However, JPEG compression inevitably introduces various kinds ofartifacts, especially at high compression rates, which could greatly affect theQuality of Experience (QoE). Recently, convolutional neural network (CNN) basedmethods have shown excellent performance for removing the JPEG artifacts. Lotsof efforts have been made to deepen the CNNs and extract deeper features, whilerelatively few works pay attention to the receptive field of the network. Inthis paper, we illustrate that the quality of output images can besignificantly improved by enlarging the receptive fields in many cases. Onestep further, we propose a Dual-domain Multi-scale CNN (DMCNN) to take fulladvantage of redundancies on both the pixel and DCT domains. Experiments showthat DMCNN sets a new state-of-the-art for the task of JPEG artifact removal.