内河航运密集小目标船舶的图像检测方法  

Image Detection Method for Dense Small Target Ships in Inland Waterway Navigation

作　　者：吴志华 钟铭恩 邓智颖 吴航星 谭佳威 周美华 WU Zhihua;ZHONG Ming’en;DENG Zhiying;WU Hangxing;TAN Jiawei;ZHOU Meihua(School of Mechanical and Automotive Engineering,Xiamen University of Technology,Xiamen 361024,China;School of Opto-Electronic and Communication Engineering,Xiamen University of Technology,Xiamen 361024,China;School of Medical Informatics,Wannan Medical College,Wuhu 241002,China)

机构地区：[1]厦门理工学院机械与汽车工程学院,福建厦门361024 [2]厦门理工学院光电与通信工程学院,福建厦门361024 [3]皖南医学院医学信息学院,安徽芜湖241002

出　　处：《厦门理工学院学报》2024年第1期58-66,共9页Journal of Xiamen University of Technology

基　　金：福建省自然科学基金项目(2019J01859)。

摘　　要：针对小目标船舶在密集性场景中检测的漏检率和误检率较高的问题,基于深度卷积神经网络,提出一种船舶检测方法。该方法选用YOLOv7模型作为算法参考框架进行改进,设计自适应锚框匹配算法,利用CIoU距离度量方式重构K-means聚类算法,以更好地适应船舶数据集中物体的尺寸和比例分布;增加一个针对小尺度目标的细粒度检测头,并使用分散注意力机制对网络模型结构进行重新设计;采用负片图像增强技术来扩充数据样本以获得更多训练样本。实验结果表明:与YOLOv7原模型相比,本文算法在内河航运船舶检测任务中的查准率和查全率分别达到92.8%和88.9%,分别提高10.9%和23.6%;mAP50达到92.6%,提高23.4%;FPS指标下降11.4;模型大小为47.1 MB,在PC端上单张图片耗时需要32.26 ms,实现了小目标船舶的高效检测。A ship detection method based on deep convolutional neural network is proposed to solve the problem of high miss rate and high false alarm rate of small target ships in dense scenes.The YOLOv7 model is selected as the reference frame of the algorithm to improve the method.Firstly,we design an adaptive anchor boxes matching algorithm that reconfigures the K-means clustering algorithm using the CIoU distance metric to better match the size and scale distribution of objects in our ship dataset.Secondly,we add a fine-grained detection head for small-scale targets and redesign the network model structure using the shuffle attention(SA)mechanism.Lastly,we utilize a negative image enhancement technique to expand the data samples and obtain more training examples.Experimental results show that the algorithm achieves an accuracy and recall of 92.8%and 88.9%,respectively,for the inland waterway shipping ship detection task,representing an improvement of 10.9%and 23.6%,respectively.The mAP50 value also improves by 23.4%,reaching 92.6%,while the FPS index decreases by 11.4.Our model size is 47.1 MB,and the time required for a single image on the PC side is 32.26 ms,achieving efficient detection of small target ships.

关 键 词：内河航运 船舶检测 YOLOv7算法模型 图像增强 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]