基于FastFlow的复杂纹理图像异常检测  

FastFlow-based anomaly detection for complex texture images

作　　者：张果 夏小东 付茂栗 魏麟 ZHANG Guo;XIA Xiaodong;FU Maoli;WEI Lin(Chengdu Institute of Computer Application,Chinese Academy of Sciences,Chengdu Sichuan 610041,China;School of Computer Science and Technology,University of Chinese Academy of Sciences,Beijing 100049,China;Shenzhen CBPM-KEXIN Banking Technology Company Limited,Shenzhen Guangdong 518206,China)

机构地区：[1]中国科学院成都计算机应用研究所,成都610041 [2]中国科学院大学计算机科学与技术学院,北京100049 [3]深圳市中钞科信金融科技有限公司,广东深圳518206

出　　处：《计算机应用》2024年第S01期256-261,共6页journal of Computer Applications

摘　　要：为了提高FastFlow的缺陷检测准确率和分割能力,分析FastFlow在复杂图像异常检测上性能下降的原因,提出一种基于FastFlow的改进模型SAFlow(Self-Attention-Flow)。为了解决异常检测准确率低的问题,首先,采用ResNet18为特征提取器,FastFlow为主干网络,均值方差阈值分割作为分割方法;其次,添加一条特征信息添加分支,包含特征金字塔(FP)模块与自注意力(SA)机制模块,旨在给予网络更多的特征信息参考;最后,采用K-sigma动态阈值作为生成的热力图分割阈值,使分割结果更准确。实验结果表明,与FastFlow模型相比,SAFlow在4个复杂纹理图像数据集上的Pixel-AUC(Area Under Receiver Operating Characteristic Curve)精度分别提升了4.98、3.05、1.44和3.88个百分点;在总数据集上,Pixel-AUC平均精度提升了3.34个百分点。所提网络结构使样本数据最后形成的正态分布更具有代表性、更集中,分布中异常负对数似然得分更高,异常检测准确率更高,解决了复杂纹理图像中异常缺陷的检测问题难点。To improve the defect detection accuracy and segmentation ability of FastFlow,the reasons for the performance degradation of FastFlow on complex image anomaly detection was analyzed,and an improved model SAFlow(Self-Attention-Flow)based on FastFlow was proposed.To solve the problem of low accuracy of anomaly detection,firstly,ResNet18 was adopted as the feature extractor,FastFlow was used as the backbone network,and mean-variance threshold segmentation was used as the segmentation method.Then,a feature information adding branch was added,which contained the Feature Pyramid(FP)module and the Self-Attention mechanism(SA)module,and was designed to give the network more references to the feature information,and finally,K-sigma dynamic thresholds were used as the generated heat map segmentation thresholds to make the segmentation results more accurate.The experimental results show that SAFlow improves the Pixel-AUC(Area Under Receiver Operating Characteristic Curve)accuracy by 4.98,3.05,1.44 and 3.88 percentage points on the four complex texture image datasets,and the average Pixel-AUC accuracy on the total dataset is improved by 3.34 percentage points compared with the FastFlow model.The proposed network structure makes the final normal distribution formed by the sample data more representative and concentrated,with higher abnormal negative log-likelihood score in the distribution,higher accuracy of anomaly detection,and solves the difficult problem of detecting abnormal defects in complex texture images.

关 键 词：深度学习 FastFlow 自注意力 复杂纹理图像 特征金字塔 

分 类 号：TP399[自动化与计算机技术—计算机应用技术]