基于特征重建的无监督木材图像异常检测  

Feature reconstruction for unsupervised wood image anomaly detection

作　　者：耿磊[1,3] 张文跃 肖志涛[1,3] 王雯[1,3] 李晓捷[1,3] GENG Lei;ZHANG Wen-yue;XIAO Zhi-tao;WANG Wen;LI Xiao-jie(School of Life Sciences,Tiangong University,Tianjin 300387,China;School of Electronics and Information Engineering,Tiangong University,Tianjin 300387,China;Tianjin Key Laboratory of Optoelectronic Detection Technology and Systems,Tiangong University,Tianjin 300387,China)

机构地区：[1]天津工业大学生命科学学院,天津300387 [2]天津工业大学电子与信息工程学院,天津300387 [3]天津工业大学天津市光电检测技术与系统重点实验室,天津300387

出　　处：《计算机工程与设计》2024年第6期1829-1835,共7页Computer Engineering and Design

基　　金：天津市科技计划基金项目(20YDTPJC00110)。

摘　　要：为有效解决目前木材图像异常边缘区域检测精度不高的问题,提出一种基于特征重建的无监督异常检测模型FRNet。设计多层级特征提取器为图像子区域生成多个空间上下文特征表示;多尺度特征生成器将多层特征融合为一幅具有多尺度特征表达的特征图;设计具有跳跃连接的卷积自编码器,通过补充下采样时丢失的细节信息重建特征图,根据重建误差定位异常区域。在构建的木材异常数据集上进行实验,其结果表明,FRNet取得了最好的异常检测性能。To address the problem of low detection accuracy of anomaly edge regions in current wood image anomaly detection,an unsupervised anomaly detection model FRNet based on feature reconstruction was proposed.A multi-level feature extractor was designed to generate multiple spatial context feature representations for image sub-regions.Multi-level features were fused through a multi-scale feature generator into a feature map with multi-scale feature representations.A convolutional autoencoder with skip connections was designed to reconstruct the feature map by supplementing the detail information lost during down sampling,and the anomalous regions were located according to the reconstruction error.Experimental results on the constructed wood anomaly datasets show that FRNet achieves the best anomaly detection performance.

关 键 词：异常检测 无监督学习 特征重建 预训练网络 深度卷积自编码器 木材图像 多尺度特征 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]