Abstract: Using traditional fully connected Convolutional Neural Network (CNN) for face recognition and verification task has the problems of long training time and low recognition rate. Through using cascaded layer (nonlinear) filters to take omnidirectional anti-noise face recognition and verification in railway security checking, this paper proposed a method of face recognition and verification with deep learning cascaded filters. Compared with the optimal fully connected CNN, the application of nonlinear threshold function in cascade filter can effectively improve the filtering recognition accuracy and shorten the recognition time. The experimental results show that this filter structure can be cascaded to form a multi-layer filtering system, and the performance can be improved about 8% in recognition rate and three times better in efficiency. Finally, the paper tested the performance of two-layer cascaded filters for the application of face recognition and verification, which provided theoretical support for identification verification in railway security checking.