Learned Image Compression with Transformer-CNN Mixed Structures and Spatial Checkerboard Context

Learning-based image compression techniques combined with current Transformer models and with checker-board context models have shown the excellent Rate-Distortion performance. However, the mixed structure still has room for optimization in terms of redundancy information and decoding efficiency, while the checkerboard context model has redundancy in capturing correlations between latent represen-tations. To solve these problems, we propose an innovative framework that combines a mixed Transformer-CNN structure with a checkerboard context model. Specifically, we introduce a “Checkerboard Channel-wise Entropy Module” to improve coding efficiency of utilizing contexts through a two-channel decoding method with checkerboard contexts. Then, we propose the “In-slice Odd-even Context”, which improves the handling of spatial redundancy information by adding additional spatial contexts by introducing a checkerboard context model to the original mixed structure with channel contexts and global con-texts. Extensive experimental results demonstrate that our pro-posed method outperforms JPEG, BPG and previous learned image compression on the Kodak dataset.