Research On Face Attribute Recognition Method Based On Multi-Task Learning

As the application field of face recognition technology gradually increases,the technology of face attribute recognition has become one of the hot issues in the field of computer vision.The ways to obtain face images are also gradually diversified,and the images contain rich face information,such as gender,race,age and so on.Face image recognition methods are also widely applied to video surveillance,security,social intercourse and other fields.In practical application,face attribute recognition is also full of challenges due to the influence of uncertain factors such as lighting and angle in the surrounding environment.Compared with traditional methods,deep learning shows better performance in the field of computer vision.As an important part of deep learning,multi-task learning plays a pivotal role in the development of deep learning.In the process of face attribute recognition,multiple attributes need to be recognized simultaneously,and there are different degrees of correlations and differences between attributes,and applying multi-task learning methods to face attribute recognition can help improve the recognition accuracy.In this thesis,the research status of multi-task learning and face attribute recognition at home and abroad is analyzed.On this basis,the face attribute recognition algorithms based on multi-task learning are investigated.The main contents include three aspects: face attribute recognition based on task similarity and attention module,face attribute recognition combining feature fusion and task grouping,and face attribute recognition oriented to parallel sharing.The main research and contributions are as follows:(1)Aiming at the problems that it is difficult to determine the location of branch nodes in the multi-task model of sharing first and then branching,and the features extracted from the shared part have different importance for different tasks,a face attribute recognition method based on task similarity and attention module is proposed.Firstly,Res Net50 was used as feature extraction network for different tasks to construct single-task models and extract the effective feature information of the task.Secondly,the similarity of the single-task model is measured by the centered kernel alignment,and the similarity between the layers of the single-task model is used as the basis for the division of branch nodes to obtain the shared part and branch part of the multi-task model.Finally,the attention module is introduced in the branch part of the model to enable subtasks to obtain more important information from the shared features to improve the performance of the model.Experimental results show that the proposed method is more time-efficient than the method of obtaining branch node position by multiple experiments,and the proposed algorithm can recognize face attributes effectively.(2)Aiming at the problems that the method of dividing different groups according to attribute locations does not consider the inter-attribute correlation degree sufficiently,and most models ignore inter-layer semantic information leading to low recognition accuracy,a face attribute recognition method combining feature fusion and task grouping is proposed.Firstly,different feature fusion modules are added to Slim-CNN to make full use of the semantic information between layers.Secondly,the attribute grouping strategy CKA-SC is proposed to fully consider the correlation degree of the task,and the attribute grouping results are obtained based on the similarity of attributes.Finally,lightweight ECA module and uncertainty weighting strategy are used to improve the model performance.The experimental results show that the proposed method has high recognition accuracy,small number of model parameters,and certain generalization ability.(3)Aiming at the problems of insufficient feature sharing among tasks in the multi-task model and unbalanced positive and negative attribute samples in the existing public dataset,a face attribute recognition method oriented to parallel sharing is proposed.Firstly,Dense Net is used as the backbone network to construct subtask networks for different face attribute recognition tasks respectively,and the subtask networks are connected by NDDR modules and cross-stitch units so that the features in different stages of subtask networks are shared and the parallelized interaction of information between different tasks is realized.Secondly,the weighted loss function is designed to enhance the learning ability of a small number of samples and solve the problem of imbalance between positive and negative attribute samples.Finally,experimental results on relevant datasets show that the proposed algorithm can effectively recognize face attributes,and has higher recognition accuracy than the comparison algorithm.