SAR ship detection based on improved YOLOv5 and BiFPN
Chushi Yu, Yoan Shin
Abstract
Synthetic aperture radar (SAR) is an advanced microwave sensor widely used in ocean monitoring, whose operation is not affected by light and weather. Ship targets in SAR images contain characteristically unclear contour information, a complex background, and display strong scattering. Ship detection algorithms based on convolutional neural networks achieved good results, albeit with many missed and false detections. To address this issue, we propose an improved scheme based on YOLOv5, that combines coordinate attention blocks and uses a bidirectional feature pyramid network for better feature fusion. Experimental results obtained with SAR images datasets demonstrate the effectiveness and applicability of the proposed model when applied for ship detection in SAR images. Compared to the original YOLOv5, the detection accuracy of the proposed method was increased from 81.28% to 88.27%, and the mean average precision was increased from 92.57% to 95.02%, which showed significant performance improvement by the proposed method in terms of detection accuracy and speed.