Research On Virtual Illumination Sample Generation Method For Face Recognition

As a kind of biometrics recognition technology,face r ecognition has very important value in the intelligent construction of today's society.Thanks to the rapid development of computer hardware and software technology,the research in the field of face recognition is advancing and many problems have been sol ved.Therefore,the application of related technologies in human daily life is becoming more and more extensive.However,nowadays,many face recognition technology applications are implemented on the basis of a large number of training samples.When the number of samples is very limited,the recognition rate will decrease.Especially when each individual in the training database has only one image for training,the recognition rate is significantly reduced and many face recognition algorithms based on multiple samples will no longer be applicable.In some special application scenarios,such as law enforcement agencies tracking suspects,single-sample per-person face recognition has become an urgent problem that must be solved.Based the analysis above,single-sample per-person face recognition is an important research direction in the field of face recognition and has important research value and significance.The content of this dissertation will be focusing on study of the single-sample per-person face recognition under the influence of light,generate virtual illumination samples of the face by the method of sample expansion,verify the effectiveness of the virtual illumination samples through a large number of recognition experiments,and study the single-sample face recognition algorithm.Generating a certain amount of virtual illumination samples from a single frontal face image is of high complexity,involving theories and methods in many fields such as computer vision,computer graphics,digital image processing,etc.The synthesized virtual illumination samples need to bear a high sense of realism.We propose a face re-illumination model based on edge-preserving filtering,which is used to generate face virtual lighting samples to expand the single training sample library from a two-dimensional perspective.A large number of recognition experiments were carried out on three recognized data sets in the research field,which quantitatively verified the effectiveness of virtual lighting sample generation.In order to make the best use of the virtual illumination samples,w e also proposes a face recognition algorithm called multi-illumination discriminant manifold analysis based on illumination pre-judgment which uses the synthesized virtual samples to do recognition tasks.After the image is divided into blocks,they are grouped according to illumination conditions and semantic information,and the image blocks at the same position under the same illumination conditions in the training database(including the generated virtual lighting samples and single training samples)are grouped into one group.Treat each group of image blocks as a manifold,construct the intra-class and inter-class adjacency relations of each group of image blocks,and learn a feature subspace for each group of image blocks.Then use the training samples to establish a light pre-judgment feature space,and use the light pre-judgment space to pre-judge the illumination conditions of the testing samples.In the recognition stage,a k-nearest neighbor classifier is used,combined with the illumination pre-judgment information of the testing sample and the recognition results of all feature subspaces,and a weighted voting is performed to obtain the final recognition result.Recognition experiments were carried out on three recognized data sets in the research field to verify the superiority of the algorithm.