Research On Transfer Sparse Representation For Facial Expression Recognition

Facial expression recognition has recently received an increasing attention due to its great potentiality in real world applications.Conventional facial expression recognition is often conducted on the assumption that training data and testing data are obtained from the same dataset.Therefore,the training data and testing data follow an identical distribution.However,in reality,the facial expression images are often obtained from varying corpora due to different devices or environments,which will make the training and testing data follow different feature distribution.There are many differences between the training and testing data.In this scenario,traditional methods will lead to poor recognition performance and model generalizability.Based on this,this thesis studies the method of transfer learning combined with sparse representation to solve the challenging cross-domain/cross-database facial expression recognition,which can efficiently improve the performance of facial expression recognition in practical applications.The main contributions are summarized as follows:1)We propose a dual-graph regularized transfer sparse coding method.In order to reduce the distribution divergence,this method constructs a dual-graph to explicitly measure the inter-domain and intra-domain similarity among different datasets,which can not only effectively narrow the distribution gap between source and target databases,but also can preserve the local geometric structural information of features.Then,we combine MMD with dual-graph as a joint distance measurement,in which the global and local distance measurements across domains are considered to effectively reduce the distribution divergence.By utilizing the label information,we introduce a discriminative criterion based on graph embedding strategy,thus the learned sparse representations can have more discriminative power.Finally,our method can learn a common dictionary and transferable sparse representation for source and target data,and obtains satisfactory results for cross-database facial expression recognition.2)We further propose a joint transfer sparse coding and subspace learning method.This method jointly combines subspace learning and transfer learning into the framework of sparse coding,which can obtain transferable sparse representations for source and target data by learning a common low-dimensional projection and dictionary.Meanwhile,this method jointly considers samples alignment and feature distribution alignment strategies,which effectively reduce the distribution divergence between source and target data.Finally,the source and target data can be well interlaced in new sparse codes.This method can learn good sparse representation for cross-corpus facial expression images.Compared with state-of-the-art methods,our method can obtain better recognition performance.