An End-to-End Multispectral Image Compression Network Based on Weighted Channels

Abstract

For the spectral correlation of multispectral images, a novel compression framework, named an end-to-end multispectral image compression network based on weighted channels, is proposed. The framework consists of a forward coding network, a quantizer and an inverse decoding network. The multispectral images are fed into the forward coding network, which is composed of the residual block based on weighted channels, to obtain intermediate features based on weighted channels. The intermediate features are quantized and encoded by the quantizer and the entropy encoder to obtain the compressed code stream. The compressed code stream passes through the entropy decoder and the inverse decoding network, which is symmetric to the forward coding network, to reconstruct multispectral images. The results validate that the proposed network shows better performance compared to JPEG2000 and JPEG.