The Design And Implementation Of Face Recognition System In Complex Environment

Face recognition is the most popular biometric technology nowadays.It is widely used in financial payment,security monitoring,entrance guard attendance and other fields.With the rapid development of deep learning and the growing data volume,the face recognition ability of computers has already surpassed the human beings on several benchmarks.However,in practical use,face recognition system still faces complex and diverse challenges such as illumination,living,cross-age,occlusion and so on.Among them,the most core challenge for commercial applications is the requirement of illumination robustness and security.At the same time,even though the recognition performance of face recognition model has been far better than that of human beings in the test set,the higher the accuracy of the model,the lower the speed performance generally.Commercial applications also attach importance to the data security and stability of the system.Edge computing is a long-term trend in the near future.How to optimize,deploy and apply face recognition algorithms on low-power and low-cost processors is a difficult problem for researchers.1.Face Detection:we modified and improved MBLBP feature,proposed improved SE-DMBLBP features,and using Adaboost for model training.At the same time,we designed an algorithm strategy to boost the speed and performance of face detection in various complex illumination conditions real-time;2.Face Anti-spoofing: a Bio-Detection algorithm based on binocular near-infrared + RGB camera is proposed,effective defense against replay attack and print attack;3.Inference Accelerate:According to the existing facial feature extraction CNN model,the feature extraction model can achieve real-time effect in low-power embedded platform such as RK3399 by reconstructing the underlying mathematical library GEMM and quantification means;4.Feature Extraction:Aiming at the difficulty of training on the huge training set: unbearable time cost and difficulties to converge,an attempt based on Softmax Loss improved algorithm is proposed,which effectively accelerates the training process and improves the training efficiency.5.System design: Based on the above arithmetic modules,an embedded automated face recognition system is designed,which has better illumination robustness,strongeranti-counterfeiting ability and faster recognition speed under the condition of accuracy guarantee,in order to meet most of the commercial applications of entrance guard and attendance.