2026年“核理安邦”联合博士生学术论坛

A High-fidelity Compression Method for Optical Sensor Satellite Images Based on Detail Information Enhancement

Not scheduled

12m

地址：清华大学校内

口头报告 人工智能 人工智能

Speaker

yuanyuan xiang (浙江工业大学)

Abstract

Aiming at the problems of explosive growth of optical remote sensing satellite image data, limited on-board computing resources, and severe detail distortion at high compression ratios, this paper proposes a high-fidelity compression method for optical satellite images based on detail enhancement. The method introduces an interval sampling module to reduce the input image resolution, which significantly decreases the model computational complexity and meets the requirements of lightweight on-board deployment. Meanwhile, a detail information enhancement module is designed to compensate for the loss of texture and edge during compression and improve the fidelity of reconstructed images. An end-to-end trainable compression framework is constructed based on variational autoencoder, and comparative experiments are carried out on multispectral and panchromatic satellite image datasets. The results show that the proposed method outperforms traditional on-board codecs and existing deep learning compression models in rate-distortion performance, detail reconstruction effect and inference speed. It can realize efficient and high-fidelity transmission of satellite images at high compression ratios, providing an effective technical solution for real-time compression of on-board optical images.

摘要

针对光学遥感卫星图像数据量激增与星上计算资源受限、高压缩比下细节失真严重的问题，本文提出一种基于细节增强的光学卫星图像高保真压缩方法。该方法引入间隔采样模块降低输入图像分辨率，显著减少模型计算复杂度，适配星上轻量化部署需求；同时设计细节信息增强模块，补偿压缩过程中的纹理与边缘损失，提升重建图像保真度。基于变分自编码器构建端到端可训练压缩框架，在多光谱与全色卫星图像数据集上开展对比实验。结果表明，所提方法在率失真性能、细节重建效果与推理速度上均优于传统星上编解码器与现有深度学习压缩模型，可在高压缩比下实现卫星图像的高效、高保真传输，为星载光学图像实时压缩提供有效技术方案。