DFDA-AD: An Approach with Dual Feature Extraction Architecture and Dual Attention Mechanism for Image Anomaly Detection

Detecting and locating unwanted structures or anomalies in the image is one of the important issues in machine vision and industrial inspection. The complexity and variability of data distribution and the lack of labeled data are among the challenges of detecting anomalies in images. In recent years, deep learning methods have provided promising results for solving anomaly detection problems in any data types, especially in images. In this paper, the DFDA-AD architecture, which is an unsupervised approach based on deep learning, is proposed for anomaly detection in industrial images. DFDA-AD consists of dual feature extraction from images by pre-trained DenseNet121 and ResNet50 networks. Two attention mechanisms are improved and developed in this paper, which provide more important feature maps for clustering by K-means algorithm. The evaluation of the model's performance was done on the MVTec AD data set, and the results of the evaluations for anomaly detection and localization were satisfactory compared to several other approaches that have been recently proposed.