Research On Multitask Learning For Emotion And Facial Attributes Recognition

As an essential branch of computer vision,facial expression recognition based on deep neural network,as well as gender,age and other facial attributes analysis,has a critical application value in the field of public security,video surveillance and human-computer interaction.Most of the algorithms are single-task recognition of facial expression,age and gender attributes respectively.However,these tasks are related to each other to varying degrees,and the single task model is independent of each other,so it is difficult to achieve collaborative learning between tasks.In order to solve these drawbacks of single task learning,multi-task learning fully excavates the relevant information between tasks,provides additional auxiliary information,and improves the performance of each task.However,at present,most datasets are heterogeneous data designed for a single task.It is difficult for multitask models to learn effective features from those heterogeneous datasets with different distributions.In this paper,the synergy among tasks such as facial expression,gender and age is deeply studied,and a multi-task network is proposed to exploit correlation among these tasks to improve the performance of each task.In order to solve the problem that heterogeneous data is difficult to converge in multi-task network training,this paper leverages pseudo-label to explore inter-task correlation,and uses iterative training strategy to achieve multitask learning on different datasets.This method effectively alleviates the forgetting effect in the training process,and achieves the recognition accuracy of expression,gender and age on multiple datasets.At the same time,the algorithm achieves the same accuracy as single task learning when the data volume decreases to half of the original data,and effectively enhances the data utilization.