Research On Low Resolution Face Recognition Technology Based On Deep Learning

With the gradual popularity of the computer vision industry,face recognition technology has been widely used in daily life,especially in this year's epidemic prevention work,the face recognition and temperature measurement integrated machine has played an important role.Although existing face recognition algorithms can accurately recognize high-resolution images under restricted conditions,environmental and human factors in actual scenes affect the image quality,resulting in a decrease in recognition accuracy.The dissertation applies deep learning to solve face detection,image reconstruction,feature extraction and face recognition in low-resolution scenes.The main tasks are as follows:(1)As an important module of the face recognition system,face detection needs to identify the face frame from the tested image and provide the source image for the subsequent recognition algorithm.In practical applications,factors such as face scale and lighting conditions will affect the detection accuracy.The dissertation makes improvements on the MTCNN model,including the application of the most advanced non-maximum suppression algorithm Soft-NMS,the proposed synthesis of difficult samples to fill the training set,and misjudgment algorithm for face detection.Experiments have proved that the algorithm can accurately identify the face area in low-resolution images,and the detection time is relatively low.(2)We propose a low-resolution face recognition algorithm based on deep learning.First,use SRCNN to reconstruct low-resolution images,then extract facial features through the VGG model,and apply principal components analysis(PCA)filters the features,finally the face recognition is completed by the classifier.Improve the performance of the model by optimizing the loss function,and evaluate the proposed algorithm on the two down-sampled public data sets FERET and LFW,and the recognition accuracy has been improved to a certain extent.(3)Fusion of the above-mentioned algorithms,modular implementation of low resolution face recognition system,and low-resolution face images captured by an external camera in a real scene are collected to evaluate the system.The improved MTCNN model is used to complete face detection,SRCNN is used to reconstruct low-resolution face images,Res Net is used for feature extraction,and the classifier is used for feature matching.The experimental results proves that proposed method can recognize face in low-resolution scenes,guarantees accuracy and efficiency,and has certain application value.