PFARN: Pyramid Fusion Attention and Refinement Network for Multiscale Ship Detection in SAR Images

Synthetic aperture radar (SAR)-based ship detection is of great value to military surveillance and civil port management. Recently, motivated by the excellent performance of convolutional neural networks (CNNs) for object detection in computer vision, their application to ship detection on SAR images has been attracting attention gradually. However, different from optical images, the SAR images have their own unique characteristics, posing challenges to the direct application of CNNs. First of all, ships from the remote sensing perspective have special proportions, and ship objects in SAR images lack prior multiscale aspect ratio information, which is not conducive to multiscale ship detection. Second, coastal ships sticking to the land have blurred ship edges and positions, which greatly interferes with stable detection. Furthermore, scattered noise at the sea surface can be confused with the tiny ships, possibly causing false alarm. To address the aforementioned problems, this article proposes a novel detection algorithm named pyramid fusion attention and refinement network (PFARN) for SAR ship detection. First, a shape scale convolution is used to improve the focus and extraction of mutiscale-ship features. Second, the feature maps are fused with pyramid fusion attention, which is based on self-attention and Gaussian cross-attention, ensuring alignment of both the semantic and spatial information. Third, a more suitable central loss function is used for the detection head to ensure that the bounding box flexibly adapts to the size and location of the object. Experimental studies show that the proposed PFARN achieves 98.1% average precision (AP) for all and 57.4% AP for tiny ships in the SSDD datasets, and 94.2% AP for all and 68.0% AP for tiny ships in the HRSID datasets, respectively, which is satisfactory for the detection of multiscale ships.