LIC_MCM: Learned Image Compression With Mamba-CNN Mixture Architecture for Remote Sensing Images

Abstract

The rapid growth of remote sensing images brings significant challenges to bandwidth and onboard storage resources. This makes remote sensing image compression essential for large-scale storage and subsequent processing. However, achieving a balance between reconstruction quality and model complexity remains a challenging problem for remote sensing image compression. To address this issue, a novel learned image compression model named LIC_MCM is proposed. Specifically, a Mamba-CNN Mixture (MCM) block is introduced into the encoder and decoder to combine the strong local representation ability of convolution with the long-range modeling dependency capability of Mamba, which enables global context modeling with linear computational complexity, thereby improving compression performance without introducing more complexity. In addition, the spatial-channel cascaded context block (SCCCB) is designed for entropy modeling, which improves context modeling and enables more accurate probability estimation, resulting in better compression performance. Experimental results on the GF6 and DIOR datasets demonstrate that the proposed model achieves state-of-the-art rate-distortion performance while maintaining competitive computational cost, showing its effectiveness for remote sensing image compression.