Unsupervised Deep Learning-based End-to-end Network for Anomaly Detection and Localization

Abstract

These days there is great demand for automatizing a visual inspection process in industrial companies since it is a tedious and time-consuming task. Recent progress in deep convolutional neural networks allowed to automatize visual inspection procedure. However, currently available supervised learning methods require large amount of labeled data, while the unsupervised learning techniques suffer from lack of accuracy. To address these problems, we propose a deep learning-based unsupervised learning method that exhibits fast and precise performance. The proposed unsupervised learning method based pseudo-labeling algorithm using graph Laplacian matrix that allows transferring computationally expensive autoencoder problem to classification task, the proposed system benefits from very fast convergence ability and significantly outperforms currently available deep learning-based AVI methods. In the conducted experiments using real-life fabric image datasets, the proposed method outperformed the currently available methods in terms of speed and accuracy.