Research On Face Recognition Algorithm Based On Super-resolution Reconstruction In Smart Classroom Management System

In recent years,face recognition technology has developed rapidly and has become one of the popular researches in image recognition technology as it has the advantages of easy access and high accuracy compared with other biometric technologies such as fingerprint and voice,and is widely used in public security and cell phone applications.However,in practical applications,most face recognition systems are still influenced by the high resolution requirements of imaging devices and the requirement for recognizers to take pictures under certain lighting conditions.To address the above issues,this paper focuses on the problem that face images taken by cameras in the classroom environment are prone to low resolution and blurred images,resulting in low accuracy of face detection and recognition in small sizes.To improve the accuracy of face recognition in the classroom environment,this paper improves the face detection algorithm based on deep learning and adds a face superresolution reconstruction module for research and improvement,as follows:(1)Improved detection algorithm for small-sized faces based on YOLOv5 s IYOLOv5s-MF.The YOLOv5 s algorithm is improved for problems such as error detection and omission detection easily occurring in small-sized face detection in the classroom environment.By applying FTT structure in the Neck structure of YOLOv5 s model to improve the small-scale face detection accuracy;improving on the positive and negative sample sampling strategy to increase the effective positive samples to enhance the influence of positive face samples in the model;addressing the CIoU Loss function cannot effectively respond to the width and height loss of the prediction frame and the real frame,the Focal-EIoU Loss function is used to enhance the The IYOLOv5s-MF algorithm achieves an average accuracy of 94.90%,92.99%,and 83.55%in different difficulty subsets of the Wider Face dataset,respectively.(2)DRN-based face super-resolution reconstruction algorithm DRNF.In order to improve the effect of low-resolution face images after reconstruction,this paper improves on the basis of DRN algorithm by eliminating the up-and-down flip and 90° flip enhancement of images in model data enhancement,adding RGB channel random transform to enhance the reconstruction performance of the model;adding RAM residual attention module to extract more face The DRN-F algorithm is tested on FFHQ dataset and CASIA-FaceV5 dataset,and achieves the highest scores on image quality evaluation indexes PSNR and SSIM compared with other algorithms.(3)Application of the improved algorithm in this paper for face recognition.In this paper,we apply the IYOLOv5s-MF algorithm for face detection,use the DRN-F face super-resolution reconstruction algorithm to reconstruct the detected small-size face images,and finally use the Arc Face algorithm for face recognition.The experiments show that the improved algorithm achieves 98.5% recognition accuracy in CASIAFaceV5 dataset and 80.8% recognition accuracy in the self-built dataset.