A Fast Threshold Neural Network for Ship Detection in Large-Scene SAR Images

Multiscale ship detection in large-scene offshore synthetic aperture radar (SAR) images is of great significance in civil and military fields, such as maritime management and wartime reconnaissance. Methods based on deep learning apply a deep neural network to extract multiscale information from SAR images, which improves detection performance. However, deep neural networks are computationally complex, and even with GPU acceleration, the timeliness of ship detection in large-scene SAR images is still constrained. Methods based on threshold segmentation, in contrast, are efficient and straightforward, but they are less robust and need to be adjusted with complex and changing scenes. This article combines two methods and proposes a lightweight framework based on a threshold neural network (TNN) to achieve fast detection. Specifically, the TNN is carefully designed to extract the grayscale features of the SAR image, which predicts the optimal detection threshold within the sliding window and separates the targets adaptively. In addition, a false alarm rejection network is used to discriminate candidate targets and improve detection accuracy. Experiments are carried out on the public SSDD offshore dataset and the FUSAR-Ship-Detection dataset. The results show that the proposed framework performs 14.43% better than the Multi-CFAR for the SSDD offshore dataset and 7.36% better for the FUSAR-Ship-Detection dataset when using F1 as the metric. Furthermore, the floating point operations of the proposed framework are only 1/240 of those of YOLO-v4 with comparable performance.