基于YOLOX网络的SAR图像舰船目标检测

针对SAR图像舰船目标尺寸大小不一、舰船分布密集、背景复杂等问题，本文提出一种改进YOLOX网络并用于SAR图像舰船目标检测。该网络包括主干特征提取网络、加强特征提取网络、解耦头、预测框优化及损失计算等4个部分。与常规YOLOX网络相比，本文作了如下改进：首先，在主干特征提取网络中，3个基础特征层之后都添加了CA模块；在加强特征提取网络中，两处下采样之后也都添加了CA模块。以强化对SAR图像中重要区域的特征提取。其次，在框回归损失函数中，引入CIOU替代IOU，以更好地利用预测框和真实框之间的相对位置信息和形状信息，提升预测框回归精度。本文基于AIR-SARSHIP-2.0数据集进行了大量的舰船目标检测实验，并选择了Faster-RCNN、YOLOv3和常规YOLOX等3种网络与本文的改进YOLOX网络进行对比。实验结果表明，本文的改进YOLOX网络整体性能优于其他3种对比网络，有更少的虚警和漏警、更高的检测精度。

Ship Target Detection in SAR Images Based on YOLOX

According to the problems of different sizes of ship targets, dense distribution of ships and complex background in SAR images, this paper proposes an improved YOLOX network for ship target detection in SAR images. The network includes four parts: backbone feature extraction network, enhanced feature extraction network, decoupled head, prediction frame optimization and loss calculation. Compared with the conventional YOLOX, the improved YOLOX has following modifications: First, in the backbone feature extraction network, three CA modules are respectively added after the three basic feature layers; In the enhanced feature extraction network, each of the two down?sampling modules is added with a CA module to enhance the feature extraction of important regions in SAR images. Second, in the box regression loss function, CIOU is introduced to replace IOU to make better use of shape information and relative position information between the predicted box and the real box, and improve the accuracy of the predicted box regression. Based on the AIR?SARSHIP?2.0, a large number of ship target detection experiments are finished, and three networks ? Faster?RCNN, YOLOv3 and conventional YOLOX are selected, to compare with the improved YOLOX. The experimental results show that the overall performance of the improved YOLOX network in this paper is better than the other three compared networks, which not only has less false alarms and missed alarms, but also has higher detection accuracy.