SORA DETR: A Remote Sensing Small Object Detection Method Based on Attention Mechanism and Feature Fusion

High-Resolution Remote Sensing Images often exhibit significant variations in object scales, with small objects being dense and having complex backgrounds. To address the issues of missed detection and false detection of small targets in remote sensing images, this paper proposes a small target detection method, SORA-DETR, which is an improvement based on RT-DETR. A lightweight small target feature fusion module, SPD-CSP Omnikernel Fusion Module (SOFM), is introduced. This module integrates low-level image features into the CNN-based Cross-scale Feature Fusion of RT-DETR to enhance the detection capability for small targets. Furthermore, C2f is used to replace the ResNet-18 in RT-DETR, while a Residual Attention Block (RAB) is applied to improve the flow of features through the network and suppress irrelevant features. On the DOTAv1.5 dataset, compared to RT-DETR r18, the mAP50 improved by 2.7% and 1.8%, while the number of parameters decreased by 24%, demonstrating significant improvements in both accuracy and lightweight performance.