JPEG Decoding With Nonlinear Inverse Transform Network and Progressive Recurrent Residual Network

JPEG is the most widely used image compression format, especially with the popularity of mobile or portable devices. However, the quality of a decoded JPEG image is usually degraded by compression artifacts such as blocking effect and ringing effect, especially at low bit rates. Recently, some convolutional neural network (CNN) based methods have been designed to solve the above problem. These methods take the problem as post-processing and only add a CNN-based post-processing network after a JPEG decoder to improve the image quality. In this paper, the JPEG decoding with nonlinear inverse transform network and progressive recurrent residual network (dubbed as JDNet) is proposed. JDNet can reconstruct JPEG images of different quality factors (QF) with only one model. In JDNet, first, the CNN-based inverse transform network (iTNet) is proposed to learn the nonlinear mapping from DCT coefficients to their corresponding original pixels, which is against the linear mapping in previous DCT networks. iTNet can reduce error propagation during inverse DCT and obtain more accurate reconstruction. Furthermore, iTNet can be combined with any JPEG post-processing method to improve its performance. Second, the progressive recurrent residual network (PRRN) is proposed for local feature extraction in the designed post-processing network which utilizes local and nonlocal similarities in multi-scale space (LNLMS) to further enhance the decoded image quality. The experimental results show that compared with JPEG, ARCNN, DnCNN and STRRN, the average gains of JDNet are 1.88dB, 0.63dB, 0.35dB, 0.24dB on Live1 dataset, 2.61dB, 1.11dB, 0.69dB and 0.29dB on Urban100 dataset, and 1.91dB, 0.69dB, 0.37dB and 0.13dB on BSD500 dataset, respectively.