A Spatial-Frequence Semantic-Detail Network for Remote Sensing Object Detection

Le résumé fourni par la source

Remote sensing target detection is a critical technology in fields such as intelligent transportation, resource exploration, and environmental protection. However, challenges in computer vision arise due to issues in remote sensing imagery, including relatively weak feature extraction capabilities, insufficient feature representation, high difficulty in target localization, and large model parameter counts. To address these challenges, this paper proposes a Spatial-Frequence Semantic-Detail network (SSD-Net) for remote sensing object detection. SSD-Net comprises two plug-and-play modules (SFD and MMFE) and a feature fusion framework (SDFF). These components collectively reduce feature loss during extraction, enhance local perception and localization capabilities, suppress background noise interference, strengthen shape perception of remote sensing targets, and improve the correlation between semantic and feature information, thereby enabling target detection against complex backgrounds. We validated the model’s effectiveness on three public remote sensing datasets (DIOR, TGRS-HRRSD, and RSOD), and achieving accuracies of 88.9%, 97.1%, and 99.3% (in terms of mAP50), respectively. These results surpass several benchmark models and state-of-the-art (SOTA) approaches.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.