ORS-Net: A Ship Detector for SAR Image Based on Oriented R-CNN
Hongning Liu, Rubo Zhang, Mingjie Xie, Pengming Feng, Guangjun He, Guokai Xu · 2023
Synthetic aperture radar (SAR) is of significance in the field of marine monitoring. SAR collects reflection imaging of marine ship targets to radar beams and records a variety of information such as radiation phase, vibration frequency, and reflection intensity of the target area. Due to the complex sea surface environment, traditional vision methods cannot effectively identify weak ship targets. Deep learning methods, especially two-stage detectors, e.g. Oriented Region Convolution Neural Network (R-CNN) have given solutions to address such challenges, which have shown certain advantages when facing the problem of complex ocean scenes and weak ship targets. However, the presence of a large amount of coherent noise and speckle noise in SAR images makes the two-stage deep learning model approach still deficient in SAR image processing tasks, which leads to the degradation of typical model detection accuracy. In this paper, an enhanced Oriented R-CNN detector, namely ORS-Net, is proposed to balance between high efficiency and high performance, by augmenting both the backbone network and feature fusion network, and employing various optimization strategies in the training and inference stages. To improve the model's ability to extract feature information from SAR image ship targets, two backbone feature extraction network structures are optimized and improved. One is Dilated ResNet improved based on dilated convolution, and the other is Deep Swin-Transformer based on unstructured model tuning. In addition, the FPN structure is extended by embedding dilated convolution modules in the feature fusion network to ensure efficient feature fusion. Finally, the model is optimized in various aspects during the training and inference phases to further improve its performance. It is experimentally demonstrated that ORS-NET achieves 53.79 mAP, which is better than the current mainstream general-purpose rotating object detectors. ORS-Net can perform the SAR image ship object detection task with high quality.