Research Of Face Recognition Based On Semi-supervised Learning

Face recognition as a convenient biometric recognition technology is a popular research topic.In recent years,due to the continuous development and maturity of the technology,face recognition has been widely used in security,finance,daily life and other scenarios,with great social value.In addition,with the rapid development of the Internet,there are large amount of potentially valuable data on the Internet.These data often do not have task-related labels and cannot be used by supervised learning algorithms.Therefore,when data is captured from the Internet,it often needs to be labeled artificially.If the data scale is large,it will consume large amounts of labors.Semi-supervised learning is an algorithm that take advantage of both labeled and unlabeled data.With only part of the data having label,it can improve the performance of the algorithm by using a large number of unlabeled data,which is very suitable for such scenarios.Because the mainstream face recognition algorithms are supervised,this thesis hopes to find a semi-supervised learning method by studying the characteristics of existing face recognition algorithms,and then use unlabeled data to improve the performance of face recognition.The specific work of this thesis is as follows:1.Three key factors for the success of face loss based on angular margin are studied.They are: 1.Let logits only depend on the angle between vectors;2.Use s to control the distribution of p and influence the gradient direction;3.Reduce the target logits value and further adjust the distribution of probability p.2.Based on the three key factors of angular margin loss,a new loss function similar to their performance is proposed,and the importance of these three factors is also verified from the side.3.By analyzing the advantages of regularization loss of existing semi-supervised learning algorithms,a feature-based semi-supervised learning loss is proposed,which does not depend on category information.4.A semi-supervised learning algorithm based on clustering is proposed,and the improvement of face recognition by semi-supervised learning algorithm under ideal conditions is analyzed.