Face Attribute Classification Based On Deep Learning

Face attributes are middle-level representation of a person's identity information,usually containing attribute information such as the age,gender,glasses,and hat of the person.Due to it's obvious and short-term characteristics of semantic features,face attributes become an important feature to describe a pedestrian identity.It has great advantages in face verification,video surveillance,face retrieval,social media and other fields.The existing face attribute recognition methods mainly include two types,one of which is a single task learning method,and each model is individually trained for each task.Such a method can perform separate data tuning for each attribute with re-sampling strategy.The method can essentially solve the problem of data imbalance in multitasking learning.But this kind of approach ignores the relationship among individual attributes.The other is multi-task learning,which uses attribute relationships,low-level feature sharing,and high-level features for group learning.This method uses manual grouping to assist network modeling attribute correlation to improve algorithm performance,but this method ignores the specificity among tasks,while training multiple attribute classification,it is difficult to guarantee the convergence of each task.In this dissertation,the relevance and specificity of face attributes are considered comprehensively,and some modifications have been made based on the existing work:(1)For the problem of face attribute relevance,the existing method is to manually determine the related attributes to group and explore the influence of different grouping methods on the results.This paper proposes a spatial LSTM network by modifying the existing network structure to predict face attributes.By scanning the face attribute feature map to construct the feature sequence,the LSTM network is used to deal with the advantages of sequence dependence.The end-to-end modeling of the correlation among face attributes,eliminating the manual grouping operation,and verifying the validity of the network on the mainstream face attribute data set.(2)Among attributes of face attribute recognition,due to data distribution,attribute scale and other reasons,the convergence difficulty among the attribute is different,and the existing methods do not distinguish each task.Aiming at the problem of face attribute specificity,this thesis changes the traditional multi-task loss function,and proposes a multi-task face attribute recognition method based on adaptive weight.By designing adaptive weight parameters,the network can adaptively train different difficulty tasks.To enable the network to achieve complex task mining while improving the generalization ability of simple task.The algorithms proposed in this paper have been tested on the mainstream face attribute dataset and carried out a series of visual display analysis,hoping to bring some deeper understanding to the reader to inspire more future work.