Research on MTCNN Face Recognition System in Low Computing Power Scenarios

Abstract

This paper analyzes the time-consuming analysis of each cascading network module (PNet module, RNet module and ONet module) in MTCNN, and finds that the time-consuming of PNet module is the highest (about 70%). According to the results of time-consuming analysis, two improved methods are proposed, one is to reduce the number of candidate face frames in input PNet network and the other is to reduce the number of output face frames in PNet network. Then, aiming at the problem that MTCNN algorithm has low detection speed in high-resolution video and cannot meet real-time requirements, a series of optimization such as adjusting the minnisize parameter and PNet threshold in combination with the low computing power application scenarios of the channel bayonet. It is verified in the FDDB face test set and practical application, the detection speed has increased by 70.1% when the detection rate has dropped by only 3.5%, and the improved scheme has achieved good results. Compared with the performance of OpenCV-VJ and SURF face detection algorithms on FDDB, the optimized MTCNN algorithm has better performance. Through the analysis of the detection results of the specific FDDB data set pictures, it is found that the undetected face conditions do not meet the actual application scenarios in this article, which proves that the optimized algorithm has excellent performance in actual applications. The test results reflect that the reproduced and optimized MTCNN face detection algorithm has good robustness to face pose changes, and fully meets the requirements of face recognition systems in low computing power scenarios such as channel bayonet.