Research On Method For Face Recognition Based On Convolutional Neural Networks And System Implementation

The evolution of artificial intelligence technology has greatly contributed to the development of society,and related products have penetrated into all aspects of people’s daily lives.Computer vision technology,a hot topic in AI research,has also made great strides in recent years.One of the most widely used branches,face recognition,has also made the transition from scientific research to the commercial sector and is being used in a variety of industries.At the same time,various algorithms and new technologies for face recognition have emerged,and the accuracy of recognition is increasing year by year.In practice,however,although most algorithms are highly accurate in recognizing frontal faces,the face images captured in reality are more or less deflected to varying degrees,and deflection generally causes a reduction in recognition accuracy.Also,faces may often be partially obscured by items such as facial masks,sunglasses,hair,hats,and ornaments.These can also result in partial loss of face features,which can affect accuracy.In order to solve the problem of large pose change of human face,an offset network is designed to represent the offset relationship from pose face to face image,and the mapping function from pose face to offset vector is obtained by gradient descent optimization algorithm.Experiments show that the recognition performance of face with pose change is improved compared with Facenet.In order to solve the problem of local occlusion,a weight network is designed to represent the weight distribution of different parts of the face under occlusion,and the weight values of different parts of the face under occlusion are obtained by gradient descent optimization algorithm.Experiments show that the recognition performance of face under occlusion is also improved compared with Facenet.Combined with the research results,a human face recognition system is designed and implemented.The face coding network,offset network and weight network are unified into a network architecture Uni Net,and the model is used to recognize the front face,pose face and occluded face at the same time.At the same time,by using various optimization techniques,the performance of the network is raised.The test shows that on the same test set(especially the dataset mixed with multi pose face data,and face data covered by sunglasses,hats,masks,etc),the recognition accuracy is improved by four percentage points compared with the face recognition model Face Net,and a more significant improvement is obtained compared to some earlier classical models of face recognition such as VGG16.