Research On Single-sample Low-resolution Face Recognition Method

Face recognition technology has the advantages of non-contact and easy to obtain,so it has been widely used in various fields.However,in unconstrained environment,such as security monitoring,it is more difficult,and the major difficulties are summarized as follows:Firstly,the camera is usually far away from people,so the detected face images have a low resolution.In addition,due to the influence of motion blur and lens distortion,the low-resolution face images degrade seriously,making it difficult to recognize.Secondly,most of the identity classes of the probe samples may belong to the unknown classes outside the gallery set,that is,only a few probe samples belong to the gallery classes.So the open set problem need to be solved.Thirdly,in most cases,it typically contains only a single high-resolution image per person in the gallery set.Therefore,it is also necessary to solve the single-sample face recognition problem.Fourthly,in special cases,the gallery set contains only a single target object and has only a high-resolution image,so it is necessary to solve the problem that the single target single-sample cannot use the inter-class mutual exclusion relationships between different classes of samples.This paper focuses on the above difficult problems of the low-resolution face recognition,and the research goal of this paper is to ensure that algorithm can improve the recall with an precision of nearly 100%.Therefore,the main work and innovation points of this paper are as follows:(1)Low-resolution open-set face recognition,firstly,considering the problem that there are some differences between low-resolution and high-resolution face images features of the same class,for each high-resolution gallery sample,a corresponding low-resolution sample can be found in the probe set,and the high and low resolution samples are jointly used to identify the rest of the probe samples.Secondly,according to the confidence distance obtained by statistics,it determines whether there is a sample of gallery class in the probe set.Finally,constructing face recognition algorithm based on iterative label propagation.After each generation of label propagation,the adaptive thresholds of acceptance and rejection are estimated according to the statistical classification method.The acceptance threshold of the precision approaching 100% is used to estimate the target samples,and the non-target samplesdetermined by rejection thresholds are deleted directly,so that the non-target probe samples reduce the number of times to participate in the identification,thereby improving the algorithm precision.At the process of statistical iterative classification,the deep convolutional neural network is used to obtain the image features,and the label propagation algorithm is used to construct the classifier.(2)Single-target face recognition,because the sample of single-target cannot utilize the mutual exclusion relationships between the samples of different classes,and the number of target class and non-target classes is seriously unbalanced in the probe set.In view of the above problems,this paper proposes a single-sample low-resolution single-target face recognition algorithm.Firstly,the probe samples are divided into subsets,and then the non-target probe samples are respectively obtained in the non-target subsets,and they and the target sample are constructed the multi-target gallery set,thereby converting the single-target face recognition problem into multi-target face recognition problem.(3)The problem of single-target face recognition is time-consuming.In this paper,in the the process of single-target face recognition based on iterative label propagation,firstly,the initial recognition is used for pre-treatment,and the probe samples are divided into two sets:suspected target samples and non-target samples.The cluster number of non-target samples is determined by contour coefficient method,and the samples with the closest distance from the center point of each class are added to the gallery set,so that the single-target problem in this paper is converted into a multi-target problem for solving.Then,perform iterative label propagation,the suspected sample set and the non-target sample set are updated iteratively according to the similarity measurement results of the newly added gallery samples and the suspected sample.The final suspect sample set is the recognition result.In the iteration process,the number of the suspected samples gradually decrease,thus reducing the complexity of each label propagation algorithm,and the recognition efficiency is improved while ensuring a high recognition precision.