The Study Of Facial Expression Recognition Based On Transfer Learning

In recent years,artificial intelligence has increasingly become people's focus of attention.Among them,emotional recognition is an important area in the development of artificial intelligence,and facial expression recognition is the main component of emotional expression.Traditional machine learning based on the two most basic assumptions.One is that training and testing set meet independent identical distributed and the other is that training set samples must be enough.But in the practical application of face expression recognition,because the illumination and view of samples are different,samples don't meet the above conditions in many cases.In order to address the practical problems of face expression recognition and extend the limitation of traditional machine learning,this paper first proposed a transfer learning algorithm based on sparse coding(Transfer Learning Based on Sparse Coding,TLbSC)applied in facial expression recognition.The learning algorithms using sparse coding got common basic,called knowledge,in the image and the learned knowledge was transferred to the test samples and training samples to get their feature codes.This paper used SVM as the classification model and trained the characteristics after the transfer of knowledge.A classifier for cross-domain expression recognition was obtained.Comparing TLbSC and other dimensionality reduction methods in the face expression sets,it showed that TLbSC had good recognition rate for cross-domain facial expression recognition in dimension reduction.Based on the TLbSC algorithm,this paper presented the multi source transfer learning based on Particle Swarm Optimization algorithm,which addressed the problem of inadequate data samples used for training in the cross domain.On the basis of single source transfer learning,the supplementary data was increased to act as source domain,which could increase the training data samples.In this paper,after transfer the knowledge by sparse coding,the decision-making level used the selective integration algorithm for multiple classifiers integration and used the simplified Particle Swarm algorithm for optimization.Two data sets of three facial data were selected as training samples of the algorithm.The algorithm was used for cross-domain face expression recognition.Experimental results showed that the proposed multi source migration learning had a good recognition effect on the multi domain problem.The algorithms presented in this paper based on the actual application environment of facial expression recognition and test samples were usually not labeled.Therefore,the single source and multi-source transfer learning algorithms proposed in this paper belonged to the transductive transfer algorithm.Experimental results showed that the two algorithms had better recognition effect in cross-domain facial expression recognition.