As a data-driven technology, face recognition based on deep learning demands high-quality images as input.

In order to filter out facal images with abnormalities caused by various factors in the scene of railway face recognition entering and exiting stations, and improve the accuracy of facial recognition, this paper studied the feature distribution of normal facial images, through knowledge transfer, and proposed a anomaly detection algorithm for facial images that did not require modeling abnormal samples. The AUROC of this algorithm for detecting facial images with abnormalities was 0.979. The experimental results show that the algorithm has a high degree of freedom in the combination of computational accuracy and operating costs, and can meet the algorithm adaptation in different scenarios and hardware conditions. It provides technical support for optimizing the input link of passenger face recognition and improving the passenger face recognition rate in various scenarios.