Research And Application Of Face Attribute Recognition Based On Deep Learning

With the rapid development of multimedia and the widespread application of image and video acquisition equipment,researchers are paying more and more attention to the related issues in the field of computer vision.With the continuous development of network communication technology,information transmission rate and network storage capacity are constantly improving.In addition to the easy acquisition of face information,the number of face images in the Internet has exploded,which provides a large number of data samples for related research topics,such as face recognition and face detection.Researchers are also developing ways to make the most of this data,both to improve the quality of existing topics and to explore new areas of research.In the era of artificial intelligence,recognition based on biological information such as face punch-in,phone unlock and face payment is becoming an increasingly indispensable part of people's life.In order to meet the strong interest of modern people in face research,face attribute recognition,which is more detailed than face recognition,came into being.Face attribute recognition aims to identify accurate and rich detailed description based on given face image.Face attributes are varied,including age,gender,race and various physical characteristics.Face attribute recognition has a wide range of applications in monitoring security,entertainment,criminal investigation,finance and other aspects.At present,most of the work in this direction only supports single attribute recognition.Even if it is multi-attribute recognition,most of the recognition is carried out in the way of multi-model parallel learning,which will not only cause low learning efficiency,but also cause a huge amount of model parameters and computation.In addition,with the urgent need of machine learning implementation and the widespread popularity of smart phone devices,it is an inevitable trend to transplant machine learning algorithm model to mobile devices.Therefore,while improving the accuracy of the model,the model must also be simplified.Based on this,this article aims to use the powerful learning ability of CNN to build a simple but efficient model and complete the joint learning of multi-attribute.In view of the main problems existing in the current work,the main work and contributions made by this article are as follows:1.Based on deep learning,design an efficient neural network to carry out face multi-attribute joint learning,including age,gender,race and appearance rating.For a variety of face attributes,this article designs a network to realize multi-attribute sharing features at the shallower layer and learning unique features independently at the higher layer.Sharing shallow features improves the generalization ability of network,and the independent learning of high-level features improves the model accuracy.2.Simplify model.Reduce the number of parameters and computation by using the sharing feature principle of different attributes.In addition,in order to reduce unnecessary computation caused by redundant information of face,face detection technology is used to cut out the face before attribute recognition.Secondly,in order to facilitate the migration to the mobile terminal,this article seeks the most suitable face size to balance of recognition accuracy and the reduction of computation,so as to minimize the recognition time without affecting the recognition accuracy and improve the user experience.Finally,I made an android App and debugged it on a real machine to test its feasibility.3.In view of the main problems existing in the current face attribute data sets,such as unbalanced category distribution and groundtruth labels thoes cannot meet the requirements of multi-task learning,a solution is proposed.In order to balance the preference of loss function to different categories,the method of setting the weight among different classes is adopted for the categories with great differences in data set distribution.For the problem that attributes needed by the multi-task are distributed in different data sets,this article proposes an incremental multi-source data learning method to recognize multiple attributes in a single model.