改进Shufflenetv2_YOLOv5的轻量级SAR图像舰船目标实时检测  

Inqjroved Shufflenetv2_YOLOv5 lightweight SAR image ship object real-time detection

作　　者：卢鹏[1] 曹阳 邹国良[1] 王振华[1] 郑宗生[1] LU Peng;CAO Yang;ZOU Guoliang;WANG Zhenhua;ZHENG Zongsheng(College of Information,Shanghai Ocean University,Shanghai 201306,China)

机构地区：[1]上海海洋大学信息学院,上海201306

出　　处：《海洋测绘》2023年第1期58-62,82,共6页Hydrographic Surveying and Charting

基　　金：上海市地方能力建设项目(19050502100);上海海洋大学科技发展专项(A2-2006-20-200211)。

摘　　要：针对当前SAR图像舰船目标检测实时性不高、准确率较低等问题,提出了一种改进Shufflenetv2_YOLOv5的轻量级目标检测模型。首先对YOLOv5网络的主干网络进行改进,在对比了MobileNetv2、Shufflenetv2两种轻量级网络的效果之后,构建了Shufflenetv2_YOLOv5网络。引入Transformer结构和双重注意力机制模块对Shufflenetv2_YOLOv5网络进行改进,以增强舰船目标的特征表达。使用FReLU替换原YOLOv5的激活函数,进一步提高网络的性能。同时,提出一种新的损失函数,增强舰船目标的定位效果。实验结果表明,本文算法在SSDD数据集上的准确率为80.2%,FPS为193帧/s,在保证精度的情况下,实现了SAR图像舰船目标实时检测。The current SAR image ship object detection has the shortcomings of low real-time and low accuracy,an improved Shufflenetv2_YOLOv5 lightweight object detection model is proposed.Firstly,this paper improves the backbone network on the basis of YOLOv5,and constructs Shufflenetv2_YOLOv5 network after comparing the effects of MobileNetv2 and Shufflenetv2 lightweight networks.Then,the transformer structure and the dual attention mechanism are introduced to improve Shufflenetv2_YOLOv5 network to enhance the feature expression of ship targets.The activation function of YOLOv5 is replaced by FReLU(Funnel Rectified Linear Unit)to further improve the performance of the model.Meanwhile,a new loss function is proposed to enhance the positioning effect of ship targets.Experimental results show that the accuracy of the proposed algorithm on SSDD datasets is 80.2%,the FPS is 193 frames per second,and the real-time ship detection in SAR images is realized under the condition of ensuring the accuracy.

关 键 词：舰船目标检测 合成孔径雷达 轻量级卷积神经网络 深度学习 注意力机制 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]