Face Detection And Recognition In Classroom Scenarios

With the continuous development and improvement of deep learning technology,face recognition technology,as one of the important branches in the field of target detection,is widely used in many fields,and has achieved good results in terms of recognition accuracy and application efficiency.This paper conducts in-depth study of the relevant literature on face detection and recognition at home and abroad,selects a face detection and recognition model suitable for this paper,and improves the algorithm of the selected RetinaFace face detection network based on the outstanding characteristics of the classroom scene.The face recognition network is modeled and improved,and finally the two are integrated in series to obtain a face detection and recognition system in the classroom scene.The main research contents of this paper are as follows:This paper improves the face detection framework based on RetinaFace.First,deformable convolution is introduced into the backbone network to adapt to the problem of face deformation caused by the different postures of the students in the classroom scene;secondly,according to the face scale characteristics in the classroom scene,the feature pyramid is adjusted.The number of feature layers is adjusted and the preset Anchor is adjusted;and the residual structure is introduced into the context-sensitive module,which not only increases the ability of the network to extract detailed features but also effectively prevents the gradient from disappearing and exploding.Finally,in the face detection model training process,this paper First train the basic weights on the public dataset WIDER FACE,and then perform transfer learning on the dataset in the self-labeled classroom scene to adapt to the characteristics of the classroom scene.Expand it to 1500 sheets to try to cover as many situations as possible in a classroom scene.This algorithm is tested on the self-labeled data set,and the accuracy of face detection is 94.72%,which is 6.03% higher than that before the improvement.In terms of face recognition,this paper adopts the Res Net50 model with high accuracy,excellent performance and fast speed as the basic model of the network,and introduces the SE Block module in the backbone network to increase the feature weight of the target of interest in the network and reduce redundant information.At the same time,the normalization method of the feature extraction network is replaced from batch normalization to group normalization to prevent the network recognition performance from being reduced due to the small batch size;the loss function uses ArcFace Loss to improve the gap between classes and reduce the intraclass gap.After the improvement,the accuracy of face recognition is increased from the original92.11% to 94.16%.The experimental results show that the algorithm has high accuracy and reliability in classroom scenarios,and can be applied to actual face recognition in the scene.