Multi-scale feature reconstruction for unsupervised defect detection and localization

Abstract

In recent years, the rapid development of deep learning makes it more and more widely used in the field of defect detection. Compared with the traditional machine vision methods, the deep learning methods based on Convolutional Neural Networks (CNN) have stronger feature learning abilities and can achieve higher detection accuracy and work efficiency in the field of surface defect detection of industrial products. However, supervised deep learning algorithms require a large amount of labeled data, making it difficult to generalize practically. To this end, we propose an unsupervised defect detection method MSFR-VAE for Multi-Scale Feature Reconstruction-Variational Auto Encoder: It realizes defect detection and localization by reconstructing the deep features of the input image and only needs to be trained on normal samples. Different from the image-based reconstruction, the feature-based reconstruction method can make the model focus more on the key features that can distinguish the normal and defective samples, so as to improve the detection effect. Besides, we use the pre-trained CNN for Multi-Scale feature extraction which is carried out from an image pyramid to detect defects of different sizes. Moreover, in order to make full use of the deep features, we use Variational AutoEncoder (VAE) to learn the feature distribution of normal samples for better reconstruction. Extensive experiments on the challenging and newly proposed MVTec AD dataset show that our method outperforms baselines.