面向工业目标检测的密集控制阀零件数据集  

Dense control valve parts dataset for industrial object detection

作　　者：王琳毅 白静[1,3] 李彦梅 李文静 WANG Linyi;BAI Jing;LI Yanmei;LI Wenjing(Department of Computer Science and Engineering,North Minzu University,Yinchuan 750021,China;Mount Liupan Laboratory,Yinchuan 750021,China;Key Laboratory of Image and Graphics Intelligent Processing of State Ethnic Affairs Commission,North Minzu University,Yinchuan 750021,China)

机构地区：[1]北方民族大学计算机科学与工程学院,宁夏银川750021 [2]六盘山实验室,宁夏银川750021 [3]北方民族大学图像图形智能处理国家民委重点实验室,宁夏银川750021

出　　处：《光学精密工程》2024年第8期1241-1251,共11页Optics and Precision Engineering

基　　金：国家自然科学基金项目(No.62162001);宁夏自然科学基金项目(No.2022AAC02041);宁夏优秀人才支持计划项目;北方民族大学研究生创新项目(No.YCX23150)。

摘　　要：工业生产中的自动化智能化离不开自动目标检测,而高准确性的自动目标检测则依赖于与实际场景相适应的数据集。本文针对工业实际场景,发布了一个密集控制阀零件数据集,命名为PD4CV(Part Detection for Control Valve)2023。该数据集的图像全部来源于控制阀生产车间,图像采集完成后,首先对数据集图片进行预处理操作,接着对数据集图片中的零件目标进行标注,然后再对数据集图片进行训练集、验证集以及测试集的划分。PD4CV2023数据集共涵盖9类零件,包括510张工盘图像和15 015个零件样本,平均每张图像含有约29个零件样本。与现有的目标检测数据集相比,该数据集具有零件摆放密集、遮挡,零件尺寸差异大,部分零件外形相似,零件样本数量不均衡等特点。最后,在不同类型数据集上的预训练对比实验表明,一般场景数据集、特定工业场景数据集只适用于一般和特定任务,而代表实际控制阀生产工况下的PD4CV2023数据集,可用于控制阀零件目标检测,其具有其特殊性和不可替代性;一系列目标检测算法在该数据集上的综合对比则验证了PD4CV2023数据集在一般性目标检测、多尺度目标检测、小规模、不均衡数据下目标检测中的有效性。PD4CV2023数据集可用于面向工业的目标检测的相关研究。Automated intelligence in industrial production is inseparable from automatic object detection,and high-accuracy automatic object detection relies on datasets adapted to the actual scene.This article published a dense control valve parts dataset for industrial practical scenarios,named PD4CV(Part Detec⁃tion for Control Valve)2023.The image of this dataset came from the control valve production work⁃shop,and after the image collection was completed,it underwent steps such as dataset preprocessing,dataset annotation,and dataset partitioning.The images of this dataset were all from the control valve pro⁃duction workshop.After the image collection was completed,the dataset images were first preprocessed,followed by labeling the part targets in the dataset images.Then,the dataset images were divided into training,validation,and testing sets.The PD4CV2023 dataset covered a total of 9 types of parts,includ⁃ing 510 workstation images and 15015 part samples,with an average of approximately 29 part samples per image.Compared with the existing object detection datasets,this dataset had the characteristics of dense placement and occlusion of parts,large size differences of parts,similar shapes of some parts,and unbalanced number of parts samples.Finally,pre training comparative experiments on different types of datasets show that general scenario datasets and specific industrial scenario datasets are only suitable for general and specific tasks,while the PD4CV2023 dataset,which represents the actual production condi⁃tions of control valves,can be used for target detection of control valve parts,and has its particularity and irreplaceability;a comprehensive comparison of a series of algorithms on this dataset verifies the effective⁃ness of PD4CV2023 dataset in general object detection,multi-scale object detection,and object detection under small-scale and imbalanced data.The PD4CV2023 dataset can be used for research on industrial ori⁃ented object detection algorithms.

关 键 词：深度学习 工业目标检测 数据集 控制阀零件 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术] TG95[自动化与计算机技术—计算机科学与技术]