RFD:Detecting SAR Coastal Ship Targets Based On Reducing Feature Decay
Kuan Yang, Shuxing Feng, Wu Xue, Bei Zhu, Wanli Wang, Shuizhou Yu, Hewei Zhang · 2023
Synthetic aperture radar (SAR) image has been widely used in the detection of ship target. Especially with the development of artificial intelligence, machine learing has been applied for ship detection in SAR images. However, it is hard to detect small scale ship in the SAR image using the current methods of SAR ship detection due to the image characteristics such as inshore complex background, geometric distortion, low pixel values, and limited useful information. In response to these problems, we propose a detection method based on feature reducing decay which could decrease the loss of effective information, i.e., reducing feature decay (RFD) module. First, one of the feature extraction models named RFD-ELAN model is suggested to reduce the loss of valuable information caused by multiple convolutions and activation functions. Then, the Refine-IoU loss function is used for loss calculation, which could provide a relatively accurate representation of the difference between prediction bounding box and ground truth bounding box. Finally, we utilize clustering analysis to improve the adaptability of anchor boxes and reduce redundancy in prediction areas. The experimental results show that the detection accuracy of the proposed algorithm has increased by 2.77% and 3.9% respectively, compared with the baseline for mAP_0.5:0.95 in the SIRSDD and HRSID coastal ship dataset. The experimental results show that the detection accuracy of the proposed algorithm is higher than that of baseline for both mAP_0.5:0.95 in the SIRSDD and HRSID coastal ship dataset, which improvement is 2.77% and 5.1% respectively.