Research On Head Pose Recognition And Face Recognition In Classroom Scenes

This thesis uses deep learning methods to analyze classroom monitoring videos,analyze students’ learning status and behaviors by recognizing their head posture,and provide accurate data support for instructors to improve teaching strategies and enhance teaching quality.To solve multiple problems in classroom scenarios,this thesis investigates three aspects of face and head detection,head pose recognition and face recognition.(1)Face and head detection: Head and face detection are the basic techniques for head pose recognition and face recognition,respectively,so this thesis uses the YOLOv5 s network for student face and head detection.For the problem of small-scale faces and YOLOv5 s network misdetection in classroom scenes,this thesis designs the CSP＿dense module to enhance the network’s ability to extract multi-scale face features.Secondly,for the problem of YOLOv5 s network miss detection,the coordinate attention mechanism is introduced into the CSP＿dense module to improve the network’s ability to obtain coordinate information.Experimental results show the effectiveness of the YOLOv5 s network for head detection,and the improved YOLOv5 s network performs better on both homemade and public datasets.(2)Head posture recognition: This thesis transforms the task of estimating students’ head posture in classroom scenes into an image classification task based on in-depth learning and proposes an improved head posture recognition method based on the Efficient Net V2-S network.First,this thesis uses the ECA-Net module to replace the SENet module in the MBConv module to reduce the number of network parameters and improve the network recognition accuracy.Then it proposes the new-CBAM module and introduces it before the fully connected layer for enhancing the ability of the network feature extraction and the connection between channels,at the same time,introducing label smoothing in the cross-entropy loss function to prevent the network from overfitting.Finally,experimental results on homemade and public datasets show that the improved Efficient Net V2-S network effectively improves head pose accuracy.(3)Face recognition: This thesis proposes a face recognition method for students based on the Mobile Face Net network.For the problem of multi-pose and small faces,this thesis designs the S-bottleneck module to replace the original bottleneck layer to enhance the information exchange between channels and the network’s ability to extract features.The proposed method is experimentally validated to be more advantageous for face recognition tasks.Finally,this thesis achieves the simultaneous recognition of students’ head posture and face.