Research On Face Recognition Based On Objective Space Construction

Face recognition is a widely used identity authentication technology in daily life and has sophisticated applications in the fields of public safety and quick payment.The main application of face recognition is identity verification,which includes facial information extraction and face matching in the database.Face recognition is one of the most challenging tasks when deployed in unconstrained environments due to the high variability.Some of these variations include low-resolution,occlusions,aging,poses,and illumination conditions.In recent years,with the fast development of deep learning,face recognition has also made significant breakthroughs,in which the loss function is an essential factor in improving the performance of the model.There are two main lines of research include metric learning and multi-class classifier.In this thesis,the aims of optimization the above methods is analyzed,and those methods construct a highly discriminative objective space.Based on those analyses,the Sample Self-Assemble algorithm is proposed.It constructs discriminative objective space by assembling the same class point to the corresponding class center.The main work of this thesis is as follows:(1)This thesis reviews the face recognition algorithms based on convolutional neural networks and analyzes the algorithms in the two research lines of metric learning and multi-classifiers in detail.Metric learning aims to learn an embedding space that the maximum intra-class distance is greater than the minimum inter-class distance by optimized model directly.The multi-class classifier is to incorporate margins in classifier loss function to maximize class separability.Thus they construct discriminative objective space indirectly.By comparing the optimization purposes of each loss function,this thesis proposes the assumption of constructing a generalized objective space.(2)This thesis proposes the Sample Self-Assemble algorithm based on the hypotheses,which include the resultant force loss function,the random triplet generator,and the model analyzer.Combined with the proposed loss function and specific training data,the algorithm can make the samples of each class constrained in their respective hypersphere,and different classes also have a large boundary.The model analyzer verifies the convergence status of the model.(3)The proposed algorithm is used to train the handwritten digit recognition model,and the necessary process of constructing the objective space is presented.Then the way of determining the class center and the method of dataset cleaning are proposed.According to the distribution of class samples and the characteristics of the proposed algorithm,a multi-stage incremental learning method is proposed in face recognition experiment.As the amount of training data increases,the performance of the model improves linearly.The training model is tested in three benchmark datasets and gets the competitive result.The current flaws of the algorithm are obtained in the experimental analysis,and corresponding improvement measures were proposed.