Semantics-guided diffusion prior for dual-branch remote sensing image compression
Year: 2026
Type: article
Abstract: Existing image compression methods do not fully consider the characteristics of remote sensing images and the different visual quality requirements of objects and background, resulting in the problem of structural distortion and texture loss of key objects at low bit rates. To solve this problem, a dual-branch remote sensing image compression method based on object-background decoupling and semantic-guided diffusion prior is proposed in this paper. The remote sensing image is first explicitly separated into object and background regions by a content decoupling module, and then compressed using dual encoding branches. The object compression branch employs a multi-scale residual codec, which cooperatively captures both global geometric structures and local fine textures through parallel global-structure and local-detail sub-branches, ensuring high fidelity in key object regions. The background compression branch incorporates a conditional latent diffusion model guided by scene textual semantics, where textual semantic descriptions of the original images are embedded as fine-grained conditions into the latent space diffusion prior, enabling semantically consistent reconstruction with high perceptual quality at extremely low bitrates. Finally, an iterative convolution-based fusion module is designed to smoothly integrate the compressed and reconstructed object and background regions, yielding the complete reconstructed image. Experimental results demonstrate that the proposed method can significantly improving the fidelity of object regions while maintaining the overall perceptual quality of remote sensing images.
Authors: Fei Xiao, Zhensheng Sun, Junjie Jin
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $1,162
OpenAlex ID: W7214748959