融合注意力机制与全局信息的SAR舰船目标检测算法

SAR Ship Target Detection Algorithm Combining Attention Mechanism and Global Information

为提高合成孔径雷达图像中舰船目标的检测精度,基于YOLOv8算法框架提出了一种融合感受野注意力机制与全局信息的目标检测算法RA-YOLO(Receptive-field Attention YOLO)。为了增强对重要信息的感知能力,在骨干网络中引入感受野注意力卷积;同时,在空间金字塔池化结构中加入全局最大池化层和全局平均池化层,融入背景信息和边缘信息以提高检测精度;最后,引入内部交并比损失函数提升网络的泛化能力,以提高回归速度和预测框的准确度。实验结果表明:在公开数据集上与YOLOv8算法相比,在参数量和运算速度几乎不变的情况下,所设计算法的平均精度提高了3.0个百分点;综合性能优于其他经典目标检测算法。

In order to improve the detection accuracy of ship targets in synthetic aperture radar(SAR)images,a target detection algorithm RA-YOLO integrating the receptive field attention mechanism and global information was proposed based on YOLOv8 algorithm framework.In order to enhance the important information perception ability,the Receptive-Field Attention convolution was introduced into the backbone network.At the same time,a global maximum pooling layer and a global average pooling layer were added to the spatial pyramid pooling structure to integrate background information and edge information,by which the detection accuracy was improved.Finally,the Inner-IoU loss function was introduced to improve the generalization ability of the network.Thereby,the regression speed and the accuracy of the prediction frame were improved.Results of experiments on open data set showed that when the number of parameters and operation speed are almost unchanged,the average accuracy of the proposed algorithm is 3%higher that that of YOLOv8 algorithm,and the comprehensive performance is better than other classical target detection algorithms.