Self-Supervised Anomaly Detection Algorithm for Industrial Components Based on Multi-scale Feature Fusion
Industrial component anomaly detection is a key issue in industrial production,where the main objective is to detect and identify anomalous components in time to ensure product quality and production efficiency.However,current industrial component anomaly detection algorithms are still extremely challenging,such as the impact of target defects at different scales on the accuracy of the algorithms,the uncer-tainty that all possible anomaly data cannot be exhausted.To solve the above problems,proposed a self-supervised anomaly detection algo-rithm for industrial components based on multi-scale feature fusion.Using Poisson fusion to seamlessly integrate rectangular blocks of different sizes into normal samples to generate anomaly sample label pairs,and proposes an Attention Atrous Spatial Pyramid Pooling(A-ASPP)mod-ule based on a CNN model with encoder-decoder structure,which achieves multi-scale feature extraction of images through Atrous Spatial Pyramid Pooling,and uses channel attention mechanism and spatial attention mechanism to achieve multi-scale feature interaction and focus region weights,and finally locates anomalous regions through the probability map output by the model.The experimental results show that the AUROC metric of this paper's method improves by 9.2%compared to the NSA method for the screw category in the public dataset MvTecAD.The method in this paper achieves an average AUROC of 98.5%on this dataset,superior to NSA methods.
国家科技期刊平台
NSTL
万方数据
