| Since the proposal of deep learning,there was a rising enthusiasm in the field of computer vision.Many stubborn obstacles have been easily settled by deep models,such as facial key-points localization for low-quality face image,face recognition in the wild,object classification,etc.There is a generality in these successful deep models.They were all trained in supervised methods,and all need large amount of precisely labeled training data.But it is diffcult to collect so much data,and the cost to label such large amount of data is also great.In this paper,two algorithms are designed to deal with such problem for face recognition.To use large amount of unman-labeled data.An effective weak supervised algorithm is proposed.First,a large weakly labeled face dataset is collected from the Internet using the commercial search engine interface.Because of there exists lots of wrong label data in the dataset,a modification signal is introduced in the loss function.The modification signal can help to amend the wrong label to some extent and make these data beneficial to the model.To make use of the existing data,a transfer learning method is proposed.In real world application,it is difficult to collect a large training set in the specific condition.So a transfer learning method was proposed which first learn good feature from the large amount of existing data and then adapt to the target training set. |
