Robustness Study Of Unsupervised Anomaly Detection Methods

With the increasing demand for product quality and the continuous application of artificial intelligence in the industrial field,the importance of industrial defect detection is growing.Due to the lack of defect samples and the need to deal with various types of corruption during the testing phase,unsupervised methods are often used to train models,and their robustness needs to be ensured.Although some of the latest unsupervised anomaly detection models achieve high accuracy,there is a lack of research on the robustness of the models.Currently,there are two problems in the field of unsupervised defect detection:(1)the lack of robustness evaluation methods for unsupervised defect detection models,which makes it difficult to evaluate and improve the robustness of existing methods and ensure production safety;(2)the lack of methods to enhance the robustness of unsupervised defect detection models.Even if the model performs well on the training set,it can still make mistakes and miss or misjudge defects when facing various non-ideal conditions and environmental changes in real production environments.This thesis addresses these two problems in two ways:For the first problem,this thesis proposes a robustness benchmark,including a dataset and evaluation metrics for unsupervised defect detection,and explores the robustness of existing models and factors that affect their robustness.First,a robustness evaluation dataset is constructed by applying common image corruption to normal images.The dataset includes 8 types of corruption,each with 5 levels of severity,to simulate unexpected corruption that the model may encounter in production environments.Then,the ratio of the average accuracy of the model on different types and severities of corrupted images to the accuracy on corrupted images is calculated as the evaluation metric for robustness.Based on the dataset and evaluation metrics,the robustness of mainstream unsupervised defect detection models is evaluated,and the strengths and weaknesses of different unsupervised defect detection paradigms in terms of robustness are analyzed.Finally,the factors that affect the robustness of the model are explored by conducting ablation studies on two existing methods with the most research value.For the second problem,this thesis proposes a method based on feature semantic enhancement and feature alignment to improve the robustness of existing unsupervised defect models to common image corruption.First,multi-scale features are extracted from the images with the pre-trained model,and more semantically feature representations are obtained by modulating features from different semantic levels.It is worth noting that the feature semantic enhancement is inspired by a research conclusion on the first question of this thesis,which is that features with stronger semantics have higher robustness.Then,image patch features are extracted from this feature representation,and the image patch features of normal samples in the training set are sampled and saved as normal feature memory bank.Next,the feature alignment method is used to optimize the feature representation of the test image by using the normal features memory bank to reduce the feature shift caused by corruption.Finally,the optimized feature representation can be used for the subsequent process of existing unsupervised defect detection methods and enhance their robustness.In the experiments,the robustness benchmark proposed in this thesis is adopted as the evaluation method.Experimental results show that after integrating this method with existing methods,the performance of existing models under various types of corruption is improved,which proves that this method can effectively enhance the robustness of existing unsupervised defect detection models.