Hierarchical Attention Networks for Multispectral Image Compression

Aiming at the different i mportance a nd p ixel s ize o f feature information at different p ositions o f m ultispectral i mages, t his paper proposes an improved ConvNext architecture for multispectral image compression based on swin transformer, and designs the corresponding decoding network while adding multi-scale channel attention. The multispectral image obtains hierarchically mapped feature information through forward coding network, and then obtains the binary code stream through quantization and entropy coding. The reverse decoding network recovers the original image through hierarchical structure and upsampling, and the rate distortion optimization balances the reconstructed image quality and bit rate in end-to-end overall architecture. The experimental results show that at the same bit rate, the PSNR of the proposed method is better than that of the existing JPEG2000.