基于影响空间与YOLOv3的古建筑检测方法  

An Ancient Architectural Detection Method Combining Influence Space with YOLOv3 Network

作　　者：胡立华[1] 王敏敏 刘爱琴[1] 张素兰[1] HU Li-hua;WANG Min-min;LIU Ai-qin;ZHANG Su-lan(School of Computer Science and Technology,Taiyuan University of Science and Technology,Taiyuan 030024,China;Foshan Science and Automation Intelligence Technology Corporations,Foshan 528010,China)

机构地区：[1]太原科技大学计算机科学与技术学院,山西太原030024 [2]佛山市科自智能系统技术有限公司,广东佛山528010

出　　处：《计算机技术与发展》2022年第12期185-193,共9页Computer Technology and Development

基　　金：国家自然科学面上基金(61873264);佛山市科技创新项目(2017IT100022)。

摘　　要：物体检测是计算机视觉领域的一个关键内容,主要研究如何在静态图像或动态视频流中快速、准确地识别及定位出其中的目标。基于图像的古建筑检测可用于古建筑三维重建、智慧旅游等领域,具有重要的研究意义和实际应用价值。然而,受到古建筑样式、形状、花纹及纹理质地等影响,目前的物体检测器存在检测精度低和定位不准的问题。针对上述问题,基于YOLOv3网络模型,结合密度聚类和距离聚类思想,设计了一种基于RNN-DBSCAN+k-means的古建筑检测方法。该方法首先结合影响空间思想,采用RNN-DBSCAN算法对已标注的古建筑图像聚类,生成聚类结果集;其次从聚类结果集中选取最优的k个结果作为k-means的初始聚类中心;然后将这k个聚类中心作为聚类初始值,结合k-means算法得出聚类结果,并作为YOLOv3网络的先验框;最后以voc数据集(20类)和古建筑数据集为对象,验证了算法的有效性。针对古建筑数据集,算法检出率提高了0.33%;而在voc数据集单类检测中,算法检出率提高了0.04%~0.84%。Object detection is a key content in the field of computer vision,which mainly studies how to quickly and accurately identify and locate objects in static images or dynamic video streams.Image-based detection of ancient architectures can be used in 3D reconstruction of ancient architectures,intelligent tourism and other fields,which has important research significance and practical application value.However,affected by the ancient architectures’style,shape,pattern and texture,the object detectors have the problems of poor detection accuracy and inaccurate positioning at present.In order to solve those troubles,a new detective method of ancient architectures based on YOLOv3 network model is proposed.The method named RNN-DBSCAN+k-means combines density clustering idea with distance clustering idea.In this method,combined with the idea of influence space,RNN-DBSCAN algorithm is used to cluster labeled images of ancient architectures and generate the clustered result set firstly.Secondly,the optimal k results are selected from the clustered result set as the initial clustering centers of k-means.Thirdly,those clustered centers are taken as the initial values of clustering,and the clustered results are obtained by through k-means algorithm,which are used as anchors of YOLOv3 network.Finally,voc datasets(20 categories)and ancient architectural datasets are used to verify the effectiveness of the algorithm.The detection rate of the proposed algorithm is improved by 0.33%in the datasets of ancient architectures,while in voc datasets,it is increased by 0.04%~0.84%.

关 键 词：影响空间 YOLOv3 古建筑检测 交并比 检出率 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]