Face Recognition Method Based on Unbalanced Label Information Fusion

Abstract

The invention discloses a face recognition algorithm framework based on unbalanced label information fusion, including L1 and L2 two-layer structure. In L1, a supervised learning algorithm is used to train face data and corresponding label information to obtain an initialized face recognition model 1. Finally, unlabeled data is trained using unsupervised methods, and the face data label information and model 1 parameters are alternately optimized, and the final face recognition model 1 is calculated after multiple iterations. L2 is the opposite of L1. It first randomly initializes the parameters of the face recognition model 2, and then conducts unsupervised training to update the model parameters; after that, it inputs labeled data, and uses the supervised learning algorithm to continue training to obtain the final face recognition model 2. Fusion model 1 and model 2 to get the final face recognition model. The invention combines the respective advantages of the supervised learning algorithm and the unsupervised learning algorithm, and gives full play to the role of massive unlabeled data, so that the algorithm can not only have excellent recognition ability in a specific scene, but also adapt to different scenes.

Description

technical field [0001] The invention relates to the field of computer biological feature recognition, in particular to a face recognition method based on unbalanced label information fusion. Background technique [0002] With the gradual maturity of deep learning theory and the massive increase of labeled face data, more and more face recognition algorithms choose to use deep learning, which has greatly improved the performance of face recognition algorithms in recent years. However, such methods face two problems: (1) Most face recognition algorithms based on deep learning require large-scale labeled data, and also require data to be diverse, such as containing multiple races, different poses, multiple sources and different lighting etc. How to obtain face data with label information has become a bottleneck in improving face recognition performance. With the extensive use of sensors, a large amount of data is generated every day, and the calibration of these data will con...

AI Technical Summary

Problems solved by technology

However, such methods face two problems: (1) Most face recognition algorithms based on deep learning require large-scale labeled data, and also require data to be diverse, such as containing multiple races, different poses, multiple sources and different Lighting, etc.

With the extensive use of sensors, a large amount of data will be generated every day, and the calibration of these data will consume a lot of manpower and material resources, and the cost will be huge, resulting in a large amount of unlabeled data that cannot be fully utilized in practical applications.

(2) Under the conditions of existing data scale and computing resources, because the scale and diversity of data used for training are not up to the requirements, over-learning often occurs, resulting in poor generalization ability of the obtained model. Can achieve good performance in some scenarios, once the scene is switched, the performance will drop significantly

However, it is precisely because there is no label information constraint that unsupervised learning methods generally have the problem of poor pertinence, resulting in poor performance in practical applications, so this type of method is still mainly in the research stage.

Images