Research And Application Of Face Attribute Recognition Based On Deep Learning

The face attribute classification task is to extract features from a given face image and perform multi-label classification of specific attributes.Face attributes are intuitive seman-tic features understandable by humans,such as eyes,beards,wrinkles,etc.Expressions are also subdivisions of face attributes,such as smiles,anger,etc.From this,it seems that face attributes express the semantic level of face features Very important.The existing face at-tribute recognition methods include two types.One type is to separately train two classifiers for each attribute,and finally integrate multiple classifiers to complete multi-attribute recog-nition.This method has good controllability,but does not consider the attributes.The rela-tionship between them;there is also a class of multi-task learning,feature sharing,and mod-eling the relationship between attributes through grouping learning of semantic features.Obviously this end-to-end learning method is more in line with human logic,but it also faces two important The first problem is how to group attributes reasonably,and the second is that the imbalance of the data set has a great impact on training.In response to these problems,this paper mainly does the following work:First,the graph convolution module is introduced to model the relationship of attributes.After the facial features are extracted by the convolutional neural network,the features are first con-strained to the dimension of the attribute category,and then used The graph neural network fits the transfer process of attribute features in non-Euclidean space.Each node of the graph corresponds to different attributes,and the edges of the graph correspond to the connections between attributes.The attribute classification problem is converted into the classification of nodes on the graph Problem;Second,considering the imbalance of the data set,a dynamic weighting method of the loss function is proposed.The loss of each type of attribute corre-sponds to a weight.The weight is updated based on statistical information after each round of training.Positive samples Categories with less distribution will get higher weight.This article also applies the above model to the media field.Currently,the video play-back platforms on the market all have dynamic image interception functions,but they all need to be manually intercepted by users.In this paper,the facial expression recognition model trained by the above method is used to intercept the facial expression fragments in the neighborhood of the time node selected by the user,and at the same time generate the expression descriptor,and match the matched copy from the constructed copy library to finally synthesize the expression package Recommend to users.