Enhanced Remote Sensing Image Compression Method Using Large Network with Sparse Extracting Strategy

Abstract

Deep neural networks based on hyper-encoders play a critical role in estimating prior distributions in remote sensing image compression issues. However, most of the existing encoding methods suffer from a problem on the hyper-encoding side, namely the mismatch of extraction ability with the encoder. This ability bias results in likelihood features that fail to extract sufficient information from latent representations. To solve this problem, the feature extraction capabilities of the hyper-encoder are enhanced to better estimate the Gaussian likelihood of the latent representation in end-to-end network optimization. Specifically, residual blocks and a parameter estimation module are incorporated to balance the performance of the encoder and the hyper-encoder. Furthermore, it is observed that the well-trained compression model tends to generate a fixed pattern of latent representations. Therefore, we incorporate a nonlocal cross-channel graph (NCG) on the backside of the encoder. Specifically, it aggregates features between similar latent representations in a graphical manner to further enhance the side information extraction capability of the hyper-encoder. Considering the computational cost, a sparse graph strategy is further developed to dynamically select the most relevant latent representations for aggregation operations, which greatly reduces the computational effort. The proposed algorithm is named nonlocal cross-channel efficient graph (NCEG). A long-dependent residual network is selected as the backbone, and a sparse attention module is inserted into the encoder/decoder side to enhance the perceptual field of the network. The experimental results on two evaluation datasets demonstrate that the proposed method achieves satisfactory results compared to other learning-based methods.