【摘要】 由于合成孔径雷达(SAR)图像的特点，如岸上船舶目标密集排列，容易受到陆地影响，造成SAR图像近岸舰船目标检测检测率低和漏检率高的问题，提出了一种基于YOLOv5的改进BiFPN特征融合模块的目标检测算法。针对海岸边舰船目标容易受到复杂建筑影响造成漏检的问题，在YOLOv5的Backbone中加入CBAM注意力机制，通过注意力机制学习舰船目标的特征，提高主干网络的特征提取能力；使用SIoU作为新的损失函数，重新定义预测框和真实框的关系，实现新的高精度定位；增加为四尺度特征检测，重新定义一个大尺度特征检测层，与此对应将特征融合模块中原有PANet替换成改进的多尺度加权双向特征金字塔(BiFPN)网络结构，实现与检测层相对应的高效特征融合模块。实验结果表明，该算法在公开的HRSID舰船数据集的R和mAP分别为88.2%和94.3%,比原来的YOLOv5算法分别提升了2和2.7百分点，达到了在复杂环境下对小目标和密集目标检测的要求。更多还原

【Abstract】 Due to the characteristics of synthetic aperture radar(SAR) images, such as the dense arrangement of ship targets on the shore, which are easily affected by land, resulting in the low detection rate and high missing rate of ship targets on shore in SAR images, a target detection algorithm based on the improved BiFPN feature fusion module based on YOLOv5 is proposed. The CBAM attention mechanism is added to the Backbone of YOLOv5 to learn the features of ship targets through the attention mechanism and improve the feature extraction capability of the backbone network. Using SIoU as a new loss function, the relationship between the predicted box and the real box is redefined to achieve new high-precision positioning. Added four-scale feature detection, a large-scale feature detection layer is redefined. Correspondingly, the original PANet in the feature fusion module is replaced by Improved multi-scale weighted Bidirectional Feature Pyramid(BiFPN) network structure to achieve an efficient feature fusion module corresponding to the detection layer. The experimental results show that the R and mAP of the algorithm in the public HRSID ship data set have reached 88.2% and 94.3%,respectively, which are 2 and 2.7 percentage points higher than that of the original YOLOv5 algorithm, which have achieved the accuracy of small requirements for object and dense object detection.更多还原