Toward Better Accuracy-Efficiency Tradeoffs for Oriented SAR Ship Object Detection

Abstract

In oriented synthetic aperture radar (SAR) ship detection task, convolutional neural network based detectors have dramatically improved the detection performance, but enormous parameters make it difficult to realize model lightweighting. Recently, DETR and its variants have demonstrated excellent performance in object detection task, while model construction through linear layers has great potential in terms of model lightweighting. However, DETR-based models are rarely applied to oriented object detection task, while the network structure relies on manual experience and cannot be designed automatically. In this article, we propose a novel neural architecture search based lightweight detector in polar coordinate system with DETR as search space for oriented SAR ship detection, where oriented bounding boxes are encoded and decoded in polar coordinate system to cope with boundary discontinuity problems, and the weight entanglement strategy is adopted to realize automatic and lightweight design of DETR. Meanwhile, we design an oriented multiscale attention to alleviate the problem of sampling a large amount of background due to offset learning. Furthermore, we introduce a downsampling feedforward network to significantly reduce network floating point operations. Finally, we transplant FPDDet head as auxiliary head to improve encoder potential ship feature learning and decoder cross-attention learning. Experimental results show that our models not only achieve DETR lightweighting and real-time detection, but also improve detection performance. Our base models achieve state-of-the-art performance on both RSSDD and RSDD datasets compared to previous best models, with 1.36% and 2.28% improvement in mAP with 32.67 and 32.14 GFLOPs, respectively.