An Enhanced DETR Network for Sonar Image Object Detection

Le résumé fourni par la source

This work proposes the more Precise and Faster Detection Transformer (PAF-DETR), a real-time and high-precision sonar image object detector based on the transformer architecture. The method fully considers the challenges posed by weak objects, blurred contours, and real-time requirements in practical underwater detection tasks. By comparing detectors with different architectures, we first analyze the conditions necessary for real-time deployment. To enhance feature representation, we design a feature information aggregation module and embed it into the neck (Path Aggregation Network) of the detection framework, thereby enabling more effective fusion of crucial multi-scale features. Furthermore, to reduce the computational cost of the decoder, PAF-DETR introduces a novel attention execution that selectively filters encoder tokens, thereby improving computational efficiency while preserving detection accuracy. These improvements collectively equip PAF-DETR with superior computational efficiency and real-time performance. Experiments were conducted on two customized sonar data sets, where the proposed PAF-DETR detector achieved 86.7% and 95.4% [in terms of mean Average Precision (mAP)@0.5], outperforming the baseline RT-DETR-L by 2% and 1.3% mAP (mAP@0.5), respectively. In addition, PAF-DETR demonstrated an inference speed of approximately 65 frames per second, slightly surpassing the comparison detector when evaluated on two RTX 3080 Ti GPUs.

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