FS-DETR: Fast and accurate object detection in sonar images

To address the challenges in underwater object detection based on sonar images, including the prevalence of small objects, poor image details, and high real-time processing demands, we propose FS-DETR (Detection Transformer for Sonar), a real-time, high-accuracy detector based on the DETR architecture. The detector is built upon the classical RT-DETR framework and leverages the attention mechanism within the encoder-decoder structure to efficiently aggregate global information, alleviating the issue of insufficient detection accuracy for small objects in sonar images. To mitigate the degradation of feature details due to noise, we design a Noise Suppression and Feature Aggregation Module (NSFAM) that effectively suppresses redundant noise and reinforces the representation of critical information, improving the efficiency of feature utilization. Additionally, we introduce a more efficient attention implementation, enabling object queries in the decoder to quickly and accurately extract useful information from memory, accelerating model convergence and improving inference speed to meet the real-time requirements of underwater tasks. Extensive experiments on a customized sonar dataset show that FS-DETR achieves an mAP of 86.7%, outperforming the baseline RT-DETR-L by 2.0% in mAP. Furthermore, the method achieves 65 FPS on dual RTX 3080 Ti GPUs, demonstrating superior real-time performance over existing approaches.