A Lightweight YOLOv4 Based SAR Image Ship Detection
Liqun Zhou, Jin-Chun Piao
Abstract
In recent years, deep learning technology has made significant achievements in SAR image ship detection. However, many models often occupy too much memory and cannot achieve real-time detection in the micro platform. To solve this problem, this paper proposes a lightweight YOLOv4 ship detection method. First, the backbone uses the lightweight network MobileNetv2. Then, the standard convolution in PANet is replaced by depth-wise separable convolution, which can greatly reduce the number of parameters. In addition, this method embeds coordinate attention in MobileNetv2 to encodes for each channel. Compared with original YOLOv4 experimental results show that, the number of parameters is reduced by 40%, and the mAP reaches 95.5%.