Variable-rate Deep Image Compression with Vision Transformers

Recently, vision transformers have been applied in many computer problems due to its long-range learning ability. However, it has not throughly explored image compression. We propose a patch-based learned compression network by incorporating transformers. The input is divided into patches before feeding the encoder and are reconstructed from decoder form complete image. Different kinds of transformer blocks (TransBlocks) meet various requirements subnetworks. also transformer-based context model (TransContext) facilitate coding based on previously decoded symbols. Since computational complexity attention mechanism quadratic function sequence length, we partition feature tensor different segments conduct each segment save cost. To alleviate artifacts, use overlapping apply an existing deblocking further remove artifacts. At last, residual scheme adopted get performance for variable bit rates. show that our with obtain 0.75dB improvement PSNR at 0.15bpp than prior variable-rate work Kodak dataset. When using strategy, framework keeps good comparable BPG420. For MS-SSIM, higher results BPG444 across range rates (0.021 0.21bpp) other models low