Research On Attendance System Based On Dynamic Face Recognition

Face recognition technology that uses human facial features for identity verification has received widespread attention in recent years.Face recognition technology based on deep learning has been widely used in the time.Attendance system implementing dynamic face recognition in time and solving the problem of attendance are time-consuming and limited to the environment,has important practical application value.This paper focuses on the problem of dynamic face recognition in the time and attendance system.The difficulty lies in the need for real-time and fast face recognition,and how to solve the effects of motion,lighting and other factors on the collected face images.In response to the above problems,the main research contents of this article are as follows:1.In order to solve the problem of low face recognition rate caused by low quality of face images captured in unsatisfactory scenes such as motion blur and poor lighting,add the face image quality evaluation process.For input continuous face pictures,use the face image quality assessment Algorithm to process.We input the quality score of the face image fused with quality factors such as lighting and clarity,and effectively filter the face image with lower quality to improve the recognition efficiency.In addition,the MTCNN face detection algorithm is also optimized,and feature pyramids and other optimizations are integrated in the MTCNN neural network,which effectively reduces the missed and false detections of face detection in complex scenes.2.In order to solve the problem of time-consuming and the demand for high algorithm speed,this paper proposes to use a lightweight network optimized in model size and speed as the basic network,and merge it with a high-precision face recognition model.Fusion greatly reduces the amount of calculation,and uses the joint loss function of the central loss function to replace the triple loss function that is difficult to converge during training.Experimental results show that the MobileNetV2 model can achieve a maximum accuracy of 97.4%,and the speed can also meet real-time requirements.3.Aiming at the problems of the performance degradation of the face recognition model with good results on the public test test set in the attendance scenario and the change in the number of attendance registrations,the Siamese face recognition model is used to transfer and learn face features to enhance the face in the actual attendance scenario.