Unsupervised Anomaly Detection And Localization In Industrial Images

With the development of Industry 4.0,intelligent manufacturing is receiving increasing attention.As a key component of intelligent manufacturing,industrial image defect detection helps to improve production efficiency and reduce industrial costs.Unlike traditional image anomaly detection,industrial defect detection not only needs to perform image-level anomaly judgment but also needs to locate the specific anomaly area,involving two tasks: anomaly detection and anomaly localization.Due to the difficulties in obtaining abnormal data and the high cost of labeling in actual industrial scenarios,current research works use unsupervised learning algorithms,which construct detection models based only on normal images.This thesis focuses on the study of unsupervised anomaly detection and localization algorithms for industrial images,utilizing relevant theories and technologies of deep learning and unsupervised learning,and designs three different algorithm models.The effectiveness of the algorithms is demonstrated through experiments on two real industrial defect detection datasets,MVTec AD and BTAD.The specific research work and main contributions of this thesis are as follows:(1)Inspired by semantic segmentation,the problem is modeled as a pixel category assignment task,and a model,Pixel AD,based on pixel-level feature clustering learning is proposed.In the training phase,normal pixel features are clustered to form clusters as pseudo categories,assigning pixel-level pseudo category labels to training samples for pixel-level classification learning.As the abnormal area pixels differ from normal patterns,the model cannot assign high confidence to them for category assignment.Therefore,the algorithm calculates pixel-level anomaly scores using the model confidence to perform anomaly detection and localization.The experiments show that the algorithm can accurately detect abnormal areas of different types and sizes,and is suitable for different industrial application scenarios.(2)To address the problem of low inference efficiency of OOD(out-of-distribution)-based detection methods,an anomaly detection network,Proto AD,based on feature prototype learning is proposed.This algorithm does not require training and directly uses pretrained feature extractors to obtain normal pixel features and perform L2 normalization.Then,this algorithm use non-parametric clustering automatically finds feature prototypes,and construct an end-to-end anomaly detection and localization network by attaching an L2 normalization,a 1×1 convolution layer,a channel maximum pooling operation,and a subtraction operation after the feature extractor,with the prototype as the kernel parameter of the 1 × 1 convolution layer.The experiments show that the algorithm achieves competitive detection performance comparable to the state-of-the-art methods,significantly improving the inference speed of the OOD-based method.(3)To improve the performance of inpainting-based detection methods,a novel mask auto-encoder model,Mae AD,is proposed,consisting of asymmetric Transformer encoderdecoder module and a lightweight convolutional U-Net refinement module.The Transformer is good at capturing long-distance information and modeling high-level semantic structures,while the CNN module is good at repairing local visual details.Hierarchical feature alignment loss is also used,which constrains the consistency between the reconstructed image and the original image in the feature space during training,and participates in the anomaly score calculation in the testing phase.The experiments show that the algorithm achieves the best performance among such methods and narrows the performance gap between reconstruction-based and OOD-based methods.