Face Verification Based On Feature Learning

Face recognition is a hot spot in the direction of artificial intelligence recent years,It’s application from a special single scene to a natural changeful scenario based on kinds of hardware products today.Thus,to solve the problems of the real scene is of great significance for the practical application of face recognition and expand practical.Therefore,this paper is dedicated to solve part of natural scenarios problems of face verification.There exisits two aspects of problems when it comes to face verification at natural scenes.Firstly,the reliability of the image quality is not guaranteed in the process of acquisition,the shaking at process of shooting produced slightly blurred or low contrast,it is unable to highlight the pixel of image features.Our paper pay much attention to this two,we adopt the face image reconstruction method based on depth of convolution network method and technology of RAISR to make face image reconstruct.on the other hand,we adopt CLAHE technology which could easily adjust the contrast.As we know,the characteristics of the image depends on the pixels and the relationship between the pixels and the pixels.by this way,this relationship is further strengthen.Secondly,the operation of face image detection and alignment is of great importance,this part of the work plays a crucial role not only in training a optimal model but also in improving the efficiency of the working face verification.In our paper,face detection based on task for alignment of neural network,with the thought of divide and rule,different kinds of face samples send to the different stages of the network,getting rid of the experience of all process.As a result,workload and accuracy are both promoted.In this paper,a method of affine transformation and multi-task alignment network is proposed,the method of using the correction of facial landmark not only standard posture,facial,and accurately obtain the face image area effectively,and obtain the effective area accounted for more than face image,and then by scale transformation can directly input facial feature extraction network,the affine transformation is combining two networks,have played an important role in the two networks.How to obtain a stability and distinct features of face is also a major challenge.In our paper,model training based on the convolutional neural network and adopt the residual network mechanism,the network depth is improved as well as the network performance.In combination with the loss of the cohesion characteristics of the center loss and characteristics and the dispersity characteristics of softmax loss,the two as the The network training supervisory signal obtained a deep network model capable of extracting more stable and distinctive face features.As for how to do verify work and actual testing,we designed a number of experiments in this paper.We analyzed the face features obtained from the LFW test set,and the best set threshold for face similarity is found through the analysis of data and images.In order to improve the accuracy of verification,the idea of using maximum likelihood as the basis of verification was proposed.The feasibility of this scheme was verified through experiments.In order to verify the stability and superiority of the features obtained in this paper,we carried out personalized verification experiments with multiple angles,multiple expressions,and ages.The experimental results verify the reliability of the method.In the end,a simple experimental application platform was also constructed to ensure that future research work should be more integrated with life needs.