Research On Few-shot Face Recognition Under Complex Circumstances

In recent years,the demand for biometric technology in social application scenarios such as life,finance,law,and criminal investigation has become higher and higher,and face recognition technology has become the most popular identification technology with the advantages of rich feature information,easy to be collected and high precision.Face recognition systems in real-world application scenarios are mostly few-shot problems,which means each face image category in the face database contains only a single or a few samples,we name this type of problem as Few-shot Face Recognition(FFR).In general,the collection of face images is under uncontrollable natural conditions,in that way,the images often have multiple changes such as illumination,expression,posture,and occlusion.Therefore,the Few-shot Face Recognition Under Complex Circumstances(FFRUCC)is an important task in face recognition field.In this paper,two novel and effective FFR methods are proposed based on the deep analysis and the improvement of some existing algorithms.The main work and improved algorithms are as follows:1)This paper proposes a method called Blocked Dictionary Learning(BDL),which combines Sparse Representation Classification(SRC)and Convolutional Neural Network(CNN),it takes a local-global way to construct a sparse dictionary for solving the FFR problem.First of all,preprocessing is necessary,we make face alignment according to the key points of face images and cut the images into four local regions for sample enhancement.Then more discriminative local and global features have been extracted due to the use of CNN and all of them are utilized to build a sparse dictionary.Next,the loss function(Spars Loss)with sparse constraint and cosine constraint is exploited to optimize the network parameters,which can reduce the intra-class distance as well as expand the inter-class distance.Finally,the Developed Sparse Representation Classification Method(DSRM)makes great efforts for face recognition.The experimental results show that the BDL algorithm achieves a high recognition rate of 92.64%and 91.93%on the AR dataset and the extended Yale B dataset respectively,and it is robust to occlusion changes and facial expression changes.2)An FFR algorithm based on Sliding Block-Generation Adversarial Network(SGAN)is proposed,the sliding window is used to obtain face information and then generate a frontal face image from a side face image.In the end,a generation-recognition system is established for solving FFR problem with posture changes.Firstly,generate virtual samples by using mirror symmetry and axis symmetry.Then,all images need to be cross-aligned in order to ensure the accuracy of the sliding block.Next,the features of face images are turned into several window features while the sliding window passing through the sample images and we regard the window informations as the input of generation adversarial network.Moreover,the frontal images are generated by using the improved system loss function.Finally,we utilize all of frontal images to make classification.The experimental results show that the SGAN algorithm has certain validity for the frontal generation of face images with small poses(deflection angle less than 45 degrees)and improves the performance of FFR with posture changes and illumination changes.