Multispectral Image Compression Based on Prediction Network

Multispectral images have rich spatial and spectral information which contain great application superiority. Therefore, effective compression of multispectral images is crucial. This paper proposes an end-to-end network architecture based on prediction networks to complete multispectral image compression tasks. Specifically, the feature extraction module can extract spatial and spectral information effectively and reduce information redundancy. The prediction module is able to predict the original image and obtain the residual one according to the reference spectral image and the extracted features. All modules are jointly optimized by a single loss function. The experimental results show that proposed compression framework outperforms conventional methods, including JPEG2000 and 3D-SPIHT.