This article is devoted to the development of a method for detecting defects on the surface of a product based on anomaly detection methods using a feature extractor based on a convolutional neural network. The method involves the use of machine learning to train classification models based on the obtained features from a layer of a pre-trained U-Net neural network. As part of the study, an autoencoder is trained based on the U-Net model on data that does not contain images of defects. The features obtained from the neural network are classified using classical algorithms for identifying anomalies in the data. This method allows you to localize areas of anomalies in a test data set when only samples without anomalies are available for training. The proposed method not only provides anomaly detection capabilities, but also has high potential for automating quality control processes in various industries, including manufacturing, medicine, and information security. Due to the advantages of unsupervised machine learning models, such as robustness to unknown forms of anomalies, this method can significantly improve the efficiency of quality control and diagnostics, which in turn will reduce costs and increase productivity. It is expected that further research in this area will lead to even more accurate and reliable methods for detecting anomalies, which will contribute to the development of industry and science.
Keywords: U-Net, neural network, classification, anomaly, defect, novelty detection, autoencoder, machine learning, image, product quality, performance