UAVRS-CompNet: An End-to-End Compression Network for UAV Remote Sensing Images

Unmanned Aerial Vehicle (UAV) remote sensing images are characterized by high resolution, complex textures, and multi-scale ground object features, making their compression and reconstruction more challenging compared to natural images. To enhance the compression performance for UAV remote sensing imagery, this paper proposes an end-to-end deep learning-based compression network named UAVRS-CompNet. At its core, the model employs a Multi-scale Residual Bidirectional Pyramid Backbone (MRBPB) to replace traditional Generalized Divisive Normalization (GDN) and its inverse transformation IGDN modules, thereby strengthening feature extraction and reconstruction capabilities. Additionally, a Dual-Branch Synergistic Attention Module (DSAM) is introduced to achieve cooperative feature enhancement in both spatial and frequency domains, improving the model's adaptive feature extraction and representation capacity for complex textures and multi-scale objects. A dedicated UAV remote sensing image dataset containing diverse ground object types is constructed for comprehensive experimental evaluation. Experimental results demonstrate that the proposed model outperforms several state-of-the-art learning-based compression methods across various bitrates, achieving the best performance in both PSNR and MS-SSIM metrics. Furthermore, visual comparisons reveal that UAVRS-CompNet more effectively preserves ground object details and texture information, leading to higher-quality image reconstruction. The proposed method provides a viable and superior solution for efficient compression and high-fidelity reconstruction of UAV remote sensing imagery.