点云与图像融合的无监督异常检测算法研究  

Research on unsupervised anomaly detection algorithm for point cloud and image fusion

作　　者：谢宏兴 林珊玲 林志贤[1,2] 郭太良 林坚普 吕珊红 XIE Hong-xing;LIN Shan-ling;LIN Zhi-xian;GUO Tai-liang;LIN Jian-pu;LV Shan-hong(School of Advanced Manufacturing,Fuzhou University,Quanzhou 362251,China;Fujian Science and Technology Innovation Laboratory for Photoelectric Information,Fuzhou 350116,China)

机构地区：[1]福州大学先进制造学院,福建泉州362251 [2]中国福建光电信息科学与技术创新实验室,福建福州350116

出　　处：《激光与红外》2024年第10期1642-1648,共7页Laser & Infrared

基　　金：国家重点研发资助项目(No.2021YFB3600603);福建省自然科学基金资助项目(No.2020J01468)资助。

摘　　要：针对多模态工业检测中高维特征之间存在干扰,导致检测率不理想的问题,提出了一种基于标准化流的多模态工业异常检测方法。首先提取图像的3D点云的深度信息将其作为第4个通道添加到RGB图像中,生成融合后的RGBD图像,然后使用预训练的特征提取网络提取融合后的图像特征,最后使用特征训练得到一个用于异常检测的标准化流模型。实验结果表明,异常检测模型在MVTec 3D-AD数据集的平均Pixel AUROC达到95.8%,平均AUPRO达到86.2%。相较于其他模型分别提升了2.6%和9.1%。A multimodal industrial anomaly detection method based on normalizing flow is proposed to address the issue of interference between high-dimensional features in multimodal industrial detection,resulting in unsatisfactory detection rates.Firstly,the depth information of the 3D point cloud of the image is extracted and added to the RGB image as the fourth channel to generate the fused RGBD image.Then,the fused image features are extracted using a pretrained feature extraction network.Finally,a normalizing flow model for anomaly detection is obtained using feature training.The experimental results show that the anomaly detection model achieves an average Pixel AUROC of 95.8%and an average AUPRO of 86.2%on the MVTec 3D-AD dataset,which is an improvement of 2.6%and 9.1%,respectively,compared to other models.

关 键 词：工业异常检测 多模态 点云 标准化流 

分 类 号：TP391[自动化与计算机技术—计算机应用技术] TN247[自动化与计算机技术—计算机科学与技术]