Wide Receptive Field and Channel Attention Network for JPEG Compressed Image Deblurring
A motion blurred image stored in the joint photographic experts group (JPEG) image compression format contains both motion blur and JPEG artifacts. Therefore, it is very difficult to restore the original image from a blurred and JPEG-compressed image. To address this problem, this paper proposes two methods: a wide receptive field and channel attention network (WRCAN), and JPEG auto-encoder loss. First, the WRCAN utilizes a large receptive field and considers the interdependencies among channels of a feature map. Second, the proposed JPEG auto-encoder loss enables the WRCAN to learn prior knowledge of JPEG compression artifacts such that the proposed WRCAN can effectively restore the original image from JPEG-compressed images. The proposed methods are evaluated on the JPEG-compressed REDS dataset by participating in the NTIRE 2021 workshop challenges on Image Deblurring Track 2 JPEG artifacts. The WRCAN trained with the proposed loss ranked third with an output of 29.60dB on the REDS test set, indicating that the proposed methods provide state-of-the-art results. The source codes, model, and data are available at https://github.com/dhyeonlee/WRCAN-PyTorch.