SAR Ship Detection Based on Improved YOLOv7-Tiny
Yang Liu, Xiaoqing Wang
Abstract
Ship detection in Synthetic Aperture Radar (SAR) images is an important remote sensing technology. It plays an essential role in maritime monitoring and warfare. Traditional SAR ship detection methods suffer from the interference of complex backgrounds. The speckle noise and the special imaging system of SAR will deteriorate the performance of traditional methods in SAR image interpretation. Deep learning has recently been widely used in SAR image interpretation while achieving state-of-the-art performance. In this paper, a novel SAR ship detection method based on Improved You Only Look Once version 7-tiny (Improved YOLOv7-tiny) is proposed. Based on the YOLOv7-tiny network, we have made the following improvements. First, the coordinate attention is plugged into the backbone network to ensure the lightweight design. Second, the Spatial Pyramid Pooing (SPP) of the backbone network and the SIoU loss function are modified to improve performance. Eventually, the proposed method has achieved better detection performance than YOLOv7-tiny and YOLOv7 on the HRSID dataset.