Research And Implementation On Key Issues Of Face Detection And Recognition Based On Deep Learning

Face is widely used in identification as a stable biological feature with noticeable individual differences.Benefit from the development of Convolutional Neural Networks(CNN),The performance of face detection and face recognition has been greatly improved.So far Face Recognition technology has been applied in automatic ticket gates system,mobile phone login,payment certifications in the financial system and so on.However,we found in the actual project development that the current multi-task face detection algorithm in complex background has a high false discovery rate,and low recognition accuracy rate for Low-Resolution(LR)face.Not only that,We can further improve the accuracy rate and robustness of CNN through structural optimization.We analyzed the causes of the above problems,improved the performance of the face detection and recognition algorithm,designed and implemented the "Person-ID" Audit Svstem.The main research work of this thesis is as follows:(1)Face Detection:To solve the problem of high false discovery rate of Multi-Task Convolutional Neural Network(MTCNN)in the complex background,We improved the MTCNN network structure and introduced Focal Loss to MTCNN,which focuses on mining hard samples,we call it Focal MTCNN.The experimental results show that the performance of Focal MTCNN proposed in this thesis is better than that of MTCNN.(2)Face Recognition:To improve the performance of FR,We used DropBlock to ameliorate ResNet-34,adopted Margin Based Loss to improve FR accuracy rate.The experimental results show that ResNet-34 with DropBlock has 8.8%lower error rate than ResNet-34 without DropBlock on Labeled Faces in the Wild(LFW).To Compare with ResNet-50,the speed increases by 21.62%when the accuracy rate decreases by only 0.11%.(3)Face Screening:To Solve the problem that Focal MTCNN misdetects hard negative examples similar to real face.We adopted AdaBoost+Haar-like Classifier to screen false detected face.The experimental results show that after the face screening process,hard negative examples are basically deleted.(4)Super-Resolution:To Solve the problem that low recognition accuracy rate caused by Low Resolution(LR)of ID photo,we made the Face Super-Resolution(SR)Dataset,and trained face SR model based on Enhanced Super-Resolution Generative Adversarial Networks(ESRGAN)to enhance the face.The experimental results show that the recognition accuracy rate of LR faces enhanced by SR performed better than that of prediction-based methods.(5)System Design and Implementation:We designed and implemented the;"Person-ID" Audit System,which has been released as an interface in CTRIP to verify the local tour guides.