In this paper, we describe our submission to the workshop and challenge on learned image compression (CLIC) hosted at CVPR 2019. Lossy compressed images usually suffer from unpleasant artifacts, especially when the bit-rate is low. In order to improve the image quality without spending extra bit-rate, decoder side quality enhancement becomes necessary. Most approaches focus on spatial information exploration and the quality enhancement is usually only performed on the luminance component, which leads to the neglect of inter-channel correlation. In addition, since compressed images mainly lose the high-frequency components, high-frequency and low-frequency components show different characteristics. Motivated by the characteristics of compressed images, a wavelet transform based 3-stage CNN is proposed in this paper. With the RGB image as input, the proposed network exploits the latent inter-channel correlations and enhances the low-frequency and high-frequency sub-band separately. Both objective and subjective evaluations show the noticeable quality improvements compared to Better Portable Graphics (BPG) and previous approaches.     «

In this paper, we describe our submission to the workshop and challenge on learned image compression (CLIC) hosted at CVPR 2019. Lossy compressed images usually suffer from unpleasant artifacts, especially when the bit-rate is low. In order to improve the image quality without spending extra bit-rate, decoder side quality enhancement becomes necessary. Most approaches focus on spatial information exploration and the quality enhancement is usually only performed on the luminance component, which...     »