End-to-End Learnt Image Compression via Non-Local Attention Optimization and Improved Context Modeling

This paper proposes a novel Non-Local Attention Optimized Deep Image Compression (NLAIC) framework, which is built on top of the popular variational auto-encoder (VAE) structure. Our NLAIC framework embeds non-local operations in encoders and decoders for both image latent feature probability information (known as hyperprior) to capture local global correlations, apply attention mechanism generate masks that are used weigh features hyperprior, implicitly adapt bit allocation different based their importance. Furthermore, hyperpriors spatial-channel neighbors improve entropy coding. The proposed model outperforms existing methods Kodak dataset, including learned (e.g., Balle2019, Balle2018) conventional BPG, JPEG2000, JPEG) compression methods, PSNR MS-SSIM distortion metrics.