Channel-Stationary Entropy Model for Multispectral Image Compression

Multispectral image have a large number of complex features. The existing network models have good compression performance for multispectral image, but the encoding and decoding time is long. In order to extract latent statistical features to guide arithmetic coding and save compression time as much as possible, a multispectral image compression framework based on channel-stationary is proposed. The multi-spectral image first go through the feature extraction network to get the feature maps, they can be down-sampling and compressed and then input to the arithmetic encoder and entropy priori module respectively. The weight features obtained by the entropy prior module feed the aid into the image features of the autoregressive module for entropy coding. The auto-regressive module can realize decoding of images, thus saving the time required for encoding and decoding. In addition, the balance between image reconstruction quality and bit rate is achieved by using distortion optimization technology. Then the image is reconstructed by inverse quantization, up-sampling and deconvolution module. The experimental results show that the method is better than JPEG2000 under the condition of similar bit rate, the compression time is shorter, and the network time complexity of the method is higher.