DC-PaDiM: Fast Anomaly Detection based on Effective Feature Dimension-Compression

Visual inspection is one of extremely important tasks in the manufacturing system. Recently, in order to realize high performance, several promising anomaly detection algorithms based on machine learning techniques have been proposed. In particular, methods utilizing pre-trained convolutional neural networks (CNNs) as feature extractors have demonstrated high effectiveness. The well-known PaDiM algorithm represents normal patch embeddings obtained from pre-trained CNNs as the multivariate normal distribution. It demonstrates high performance in the evaluation using the AUC metrics which is generally used in this research field. But, from the viewpoint of practicality, the processing speed is not enough. The major reason of high computational cost is that the compression of feature-dimension is based on random selection, and its efficiency is low. Randomly selected dimensions do not guarantee that they can effectively contribute to high AUC. Such dimensions may include sub dimension-sets which are not effective. Here, we propose a new method of dimension compression by selecting only effective dimensions to improve AUC metrics. As a result of this idea, we can extremely improve the speed of the original PaDiM while maintaining its performance. In this study, we propose a method to optimize dimensional axis selection using the genetic algorithm to maximize AUC for a small number of samples. As a result, only a small number of dimensions effective for anomaly detection are selected. When we have experimented the PaDiM based on proposed dimension compression, the mean of processing time during testing has been 1.23 seconds on the texture classes of the MVTecAD. This result shows a reduction of 40.57 seconds in processing time, compared to the original PaDiM based on 550 randomly selected dimensions.