基于改进YOLO方法的海上风电场入侵船舶识别  

Ships invading recognition to offshore wind farms based on improved YOLO method

作　　者：高亚娴 张敏 杨璐雅 GAO Yaxian;ZHANG Min;YANG Luya(School of Information Engineering,Shaanxi Xueqian Normal University,Xi’an 710100,China)

机构地区：[1]陕西学前师范学院信息工程学院,西安710100

出　　处：《国外电子测量技术》2025年第1期119-125,共7页Foreign Electronic Measurement Technology

基　　金：2023年教育部人文社科基金项目(23YJA870016)。

摘　　要：为了解决闯入海上风电场的船舶发现难,传统的基于船舶自动识别系统(Automatic Identification System,AIS)系统的识别方法精度低、实时性差等问题,提出一种利用改进YOLO模型从视频中识别入侵船舶的方法。采用均匀融合方法将Transformer和YOLO网络相结合,将主干输出连接到Transformer编码器,将多头注意力Transformer解码器输出连接到全连接层,解决了Seq2Seq问题;基于Transformer多头注意力方法模拟人类注意力机制,计算基于内容的向量序列的凸组合。选取更适合回归的YOLOv5泄漏整流线性单元LReLU(Leaky ReLU)作为损失函数,使用二元交叉熵损失(Binary Cross-Entropy Loss,BCE-loss)函数作为YOLOv5模型的损失分类函数。利用集成学习方法对改进YOLO模型进行预训练,并利用自建数据集进行了训练并实现模型最优。结果表明输入图像在512×512分辨率下具有最佳精度,其精度为83.10%。分类交叉熵(Categorical cross-entropy)函数是YOLO模型中损失函数的最佳选择。在海上风电场监控视频进行的船舶识别实验中,结果显示,采用改进的YOLO方法,在极端天气条件下,对海上风电场附近拍摄的监控图片的识别准确率能达到90%;而在晴好天气下,准确率达到98%。有效解决了海上风电场入侵船舶的识别难题。In order to solve the problem of difficulty in detecting ships entering offshore wind farms,traditional identification methods based on AIS(Automatic Identification System)systems have low accuracy and poor realtime performance.A method is proposed to use an improved YOLO model to identify invading ships from videos.By using a uniform fusion method to combine Transformer and YOLO networks,the backbone output is connected to the Transformer encoder,and the multi head attention Transformer decoder output is connected to the fully connected layer,solving the Seq2Seq problem;Simulate human attention mechanism based on Transformer multi head attention method and calculate convex combinations of content-based vector sequences.Select the more suitable YOLOv5 leakage rectification linear unit LReLU(Leaky ReLU)as the loss function,and use the BCE-loss(Binary Cross-Entropy Loss)function as the loss classification function for the YOLOv5 model.The improved YOLO model was pre trained using ensemble learning methods,and the model was trained using a self built dataset to achieve optimal performance.The results indicate that the input image has the best accuracy at a resolution of 512×512,with an accuracy of 83.10%.The categorical cross-entropy function is the optimal choice for the loss function in the YOLO model.In the ship recognition experiment conducted on the monitoring video of offshore wind farms,the results showed that using the improved YOLO method,the recognition accuracy of monitoring pictures taken near the offshore wind farm could reach 90%under extreme weather conditions;On sunny days,the accuracy rate reaches 98%.Effectively solved the problem of identifying invading ships in offshore wind farms.

关 键 词：海上风电场 船舶识别 神经网络 深度学习 YOLO 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]