SRS-DETR: DETR for small object detection in remote sensing images

Abstract

In the field of remote sensing image processing, small object detection presents a significant challenge due to the diminutive size of the objects, limitations in image resolution, and the difficulty of recognizing objects within complex backgrounds. To address these challenges, we propose a novel small object detection model for remote sensing images, named SRS-(DETR) DEtection TRansformer. This model is a variant of the DETR. Firstly, we introduce the Deformable Feature Correlation module to mitigate the high computational cost associated with the Multi-head Self-Attention mechanism in the transformer encoder and to enhance the performance of the model in detecting small objects. Secondly, we propose the Dynamic Multi-scale Sequence Fusion (DMSF) module. This module employs a dynamic design to efficiently fuse features at different scales, further improving the ability of the model to detect small objects. Finally, we design an Intersection over Union (IoU) loss function, Dynamic Multi named GWIoU, specifically tailored for small object detection in remote sensing images, to optimize the training process. Experimental results demonstrate that the mAP of SRS-DETR reaches 85.3%, 97.5%, and 96.3% on the LEVIR-Ship, SSDD, and RSOD datasets, respectively. Our model outperforms current popular detection methods for remote sensing images. Overall, the transformer-based architecture of SRS-DETR exhibits superior detection performance and generalization capability. Overall, the transformer-based architecture of SRS-DETR achieves excellent detection performance and sound generalization capability on the three datasets employed in this study.