Enhanced DETR Structure for Ship Detection in SAR Images

• Published in: Chinese Control Conference pp. 7929 - 7934
• Main Authors: Qin, Weiwei, Chi, Hao, Zhu, Xiaofei, Qin, Qingqiang, Liang, Zhuo
• Format: Conference Proceeding
• Language: English
• Published: Technical Committee on Control Theory, Chinese Association of Automation 28.07.2024
• Subjects: Accuracy; Decoding; DETR; Label Adversarial Training; Measurement; Multi-scale Attention; Radar polarimetry; Synthetic Aperture Radar; Target Detection; Training; Transformer cores; Transformers
• ISSN: 1934-1768
• DOI: 10.23919/CCC63176.2024.10661671

Addressing the challenges posed by slow convergence and the detection of small and faint targets in Synthetic Aperture Radar (SAR) imagery, this research introduces a sophisticated methodology that leverages the DETR architecture. The core innovation involves the integration of a multi-scale attention module into the Transformer encoder framework. This integration enables the efficient utilization of multi-scale image features, thereby enhancing the model's proficiency in discriminating ship targets in SAR imagery. Furthermore, a seamlessly integrated label adversarial training module is introduced into the decoder architecture to address issues such as recurrent target predictions and ambiguities in target preselection. This module plays a pivotal role in expeditiously facilitating the convergence of the model during training iterations. To rigorously validate the proposed algorithm, empirical experiments are conducted on the publicly available High-Resolution SAR Ship Dataset (HRSID). The results demonstrate significant improvements compared to the conventional Faster R-CNN technique. Specifically, the mAP_0.5 is elevated by 12.4%, the mAP_0.5:0.95 witnesses a surge of 13.3%, and the AR metric demonstrates a commendable increment of 14.7% relative to the Faster R-CNN baseline. These promising outcomes underscore the efficacy of the proposed methodology in amplifying target detection accuracy and accelerating convergence in ship detection applications utilizing SAR imagery.