Facial Common Information Recognition System Based On Social Service Robot

With the developing life quality, more and more family or personal service robots have stepped into our life, and normally these robots have the capacity to make simple conversation with people. The problem that how to make robots have more autonomy to communicate more freely and friendly have already drawn more and more attention from numbers of experts. Through a lot of researches and experiments, scientists point out that during the communication process, it is the verbal message, body language and facial information that play the vital role in understanding others,among which the facial information is the most direct and important information to get the opposite information during the communication, thus how to obtain people facial information is very important to the research of human-robot interaction. This thesis focuses on the recognition of common facial information, which is profound meaningful to social service robot.The research about facial information can be classified into individual information research and common information research. Individual information contain a person’s all distinctive and the study of it focuses on recognition of faces, while common information contain the facial common information existing in everyone and the study of it focuses on expression recognition, gender recognition and so on. In this thesis, the deep learning model is used to learn the distinctive expression of people facial common information and recognize facial common information, including mustache existence or not, mouth slightly open or not, wearing lipstick or not, cheekbones is high or not, smiling or not, male or female, young or not, the facial expression and so on. The content of the thesis is consisted of the followings:Applying prominent contribution about the deep learning method in the artificial intelligent field, the facial common attributes learning model is proposed based on the deep convolution neural network. In this thesis, the model studies 7 kinds of property knowledge. Because the sample volume in the database is small, the deep learning model can’t learn effective distinctive expression, thus in order to learn more accurate distinctive expression, the transfer learning method is applied during the model training process, namely the distinctive knowledge of facial recognition research is transferred to the attributes learning training. The transfer learning method makes the facial attributes of model learning become more abundant and accurate, besides with the transfer learning method, the model can get desired recognition performance in the 7 facial property learning tasks.In order to develop the expression recognition success rate, the multi-task facial attributes learning model is proposed to learn related information in facial attributes. In this thesis, the customized labels are added into expression database and attributes database, and the two databases are mixed together to create new training set. With the new training set, the proposed multi-tasks facial property learning model is trained and learns the facial expression and facial attributes, besides the transfer learning method helps the deep learning model learn more effective distinctive expression, thus the expression success rate can reach 91.01%.At last, the facial common information system platform is built and applied on the social service robot. The platform has two main functions:(1)provide algorithm simulation/testing subsystem, and is able to test or simulate all image-based the deep convolution neural network models;(2) provide algorithm deployed subsystem, realize the facial attributes learning model based on multi-task and apply these models to the robot.