Remote Sensing Object Detection Based on Detail-Semantic Decoupling and Multiscale Coordinate-Guided Semantic Enhancement

Abstract

Due to the severe scale variation of targets in remote sensing images, the dense distribution of objects, and the fact that many small targets occupy only a very limited number of pixels, existing detection methods are prone to losing shallow details and suffering from insufficient low-level semantic representation during repeated downsampling. To address this issue, this article proposes DSH-DETR, a remote sensing object detection method based on RT-DETRv2 that incorporates detail-semantic decoupling and multiscale position-aware semantic enhancement. First, in the encoder, shallow features are split into a detail path and a semantic path, so as to alleviate the representation conflict of shared shallow features between detail preservation and semantic modeling. Then, a multiscale position-aware semantic enhancement module (MPSEM) is designed to improve the representation capability of low-level semantic features in complex remote sensing scenes through directional-aware convolution, multiscale context modeling, adaptive semantic interactive fusion, and coordinate attention (CA) calibration. Finally, a high-resolution semantic reconstruction module (HRSM) is introduced to generate additional high-resolution detection features, which, together with the original multiscale semantic features, form a four-scale detection framework, thereby enhancing the model’s ability to perceive small targets and targets in complex backgrounds. Experimental results show that the proposed method achieves better detection performance than the baseline RT-DETR on the VisDrone2019-DET, RSOD, USOD, and RS-STOD datasets. In particular, it performs better on small-object detection metrics and under high intersection over union (IoU) thresholds, demonstrating its effectiveness in complex remote sensing scenarios.