Research And Application Of Low Quality Face Recognition Method Based On Feature Enhancement

Face recognition is one of the important branches in computer vision.With the continuous enhancement of facial recognition model feature representation capabilities and the emergence of large-scale facial datasets,the performance of face recognition based on deep learning has been significantly improved.However,existing face recognition algorithms still have issues such as strong data dependence and poor generalization,making it difficult to recognize low-quality faces.In addition,researchers have proposed methods such as super-resolution reconstruction and learning common feature subspaces for Low resolution face recognition.Although these methods can effectively achieve the extraction and classification of Low resolution facial features,they also face the challenge of difficult deployment in uncontrolled scenes due to the difficulty in obtaining high-resolution images they require.The main problems with low-quality facial images in uncontrolled scenes are as follows:(1)The number of publicly available low-quality facial datasets is very limited,and low-quality facial recognition tasks lack a data foundation?(2)In uncontrolled scenes,the distance between the target and the camera is different,resulting in significant intra class differences in the detected faces and poor inter class distinguishability?(3)The resolution of facial images is low,and complex lighting,camera out of focus,and blurring caused by subject motion can weaken key facial information.In response to the above issues,the main research content of this thesis is as follows:(1)In response to the scarcity of low-quality face datasets,this thesis constructs a face dataset Ivip Class Face based on real teaching scenarios,which includes 3299 images with 28 identity IDs.This provides strong data support for the performance improvement of subsequent face recognition algorithms and the application of classroom attendance systems.(2)Addressing the issues of weak intra class compactness and small inter class differences in low-quality facial images,this thesis designs a face recognition algorithm based on angle triplet loss,which maps the correlation between targets from Euclidean distance to angle measurement,enhancing intra class compactness while increasing inter class differences.At the same time,in order to better distinguish the widely existing similar classes in low-quality facial images,an adaptive margin based on inter class similarity is constructed to enhance the model’s ability to identify similar classes in low-quality facial images.(3)Aiming at the problem of weak activation of key features faced by low-quality faces,this thesis designs an attention feature enhancement algorithm based on self distillation mechanism,uses the self attention module to extract refined shallow features,designs an attention feature enhancement module based on feature segmentation to focus on the global Semantic information of the feature itself,and optimizes the depth features.Under the constraint of L2 loss,distillation between shallow and deep features is achieved to extract more discriminative facial features.(4)Finally,this thesis constructs a classroom scene attendance system based on face recognition.Based on the constructed face dataset Ivip Class Face and the research results of face recognition algorithms,a face recognition system is constructed,and a client operation interface is designed to achieve classroom scene attendance application,verifying the effectiveness of the research results of face recognition algorithms in this thesis.