Joint Face Attribute Estimation Based On Multi-Task Learning

In recent years,the Internet and its related industries have developed rapidly,and the whole society has entered the era of big data.Getting all kinds of face-related images and video data from the Internet is becoming easier than ever.People's facial images contain a wealth of personal information,including gender,race,age,expression and other personal attributes.This property of face image makes it have a wide range of potential applications in such fields as security,entertainment and social communication.On the other hand,the rapid development of computing power has made deep learning a hot research direction in the last decade.The deep learning-based convolutional neural network performs better than traditional image processing in various computer vision tasks.Multi-task learning also plays an important role in the development of deep learning.Therefore,it is of great academic and applied value to study face attribute joint estimation based on multi-task learning.The specific work of this paper is as follows:First of all,this article from the traditional methods of extracting features based on manual and based on the method of deep learning,estimate properties on the face and multitasking,we deeply analyzed the research status in the field of learning,then the analysis and comparison on the face properties estimation under the background of a particular task,and multitasking learning of different single task.Secondly,this paper proposes an improved multi-task learning network,called limited shared multi-task network,which solves the problem of information flow existing in the current multi-task learning network.The traditional multi-task learning network adopts the structure of shallow network sharing and high-level network branching.The idea is to extract low-level shared semantic features by using the shallow network,and then learn discriminating features for different attributes or attribute groups in each highlevel sub-network.The disadvantage of this structure is that the flow of information between each sub-network in the upper layer tends to stop,which is not conducive to the improvement of performance.Other network structures turn the shared network into a separate sub-network,allowing the flow of information across the network.However,such unrestricted sharing will lead to mutual restriction among subtasks and affect the performance of the network.In this paper,the effect of attribute estimation is effectively improved by reasonably controlling the information flow between sub-networks and deciding the connection mode between sub-networks by the network itself.Finally,combining Inception module and attention mechanism,this paper proposes residual attention module to improve the feature extraction ability and feature location ability of the network.The residual attention module is composed of an improved Inception module and a serial connection of the attention module,in which the attention module takes the form of a residual connection and can use any available attention module.Then,after comparing the performance of two kinds of attention modules,including SE module and CBAM module,the CBAM module with better performance was selected,and the Res-CBAM network was built on the basis of the limited shared multi-task network.