Research On Vision-Based Domain Adaptation Methods

The domain adaptaion learning method was proposed to solve the main problem of machine learning,that is,how to use the existing data set with sufficient labeled samples to help the relevant target data set with only a few labeled samples to learn an effective model.Domain adaption learning aims to transfer the discriminatory information of the source domain to the target domain where information is scarce to help the learning of the target domain model.However,the existing domain-adaption learning solutions based on visual classification mostly focus on visual understanding or feature learning,which may result in the existence of the so-called semantic gap between low-level features and high-level semantics.In addition,how to effectively distinguish the useful information in the prior model is another unsolved problem.These problems also exist in the method of using visual data to identify depression.These methods have a common assumption:the training and test samples are sampled from the same domain,and their feature space is the same as the marginal probability distribution.However,when collecting visual data,there will be many uncertain changes due to some external environmental factors,making these methods may lack inductive capabilities,which will hinder their real-world scenarios.Therefore,this paper proposes a robust latent domain adaptive depression recognition（LDADR）framework by jointly exploring facial appearance and dynamic feature representation,and using limited information in the target domain to effectively combine source domain information.Specifically,LDADR uses facial appearance features to find a latent space with amount of information to minimize the distribution difference between the source and target domains,and encodes the common components of the source and target domain classifier models as low-rank regularization item achieves the purpose of information sharing.Finally,the obtained two classifiers are effectively integrated with the classifier based on the facial dynamic feature training in the target domain to form an efficient target classification model.In this paper,LDADR and several representative methods are used to evaluate the performance on three depression data sets.The results show that LDADR has better performance in almost all learning tasks.On the basis of the above,the paper then proposes a new multi-source adaption learning framework:multi-source adaption classification framework with feature selection（MSACCF）.The biggest difference between this framework and LDADR is that it can utilize the knowledge of multi-source domains at the same time,not just the knowledge transfer from one source domain to the target domain.The framework integrates feature selection and latent space learning into a unified framework for joint learning.Specifically,MSACCF learns multi-source domain classification models by projecting feature data from multi-source domains into different latent spaces,then discriminately integrate multiple classification results,which are obtained from the target data on multi-source domain classifiers,and are used to learn the target classification model.In addition,the framework also uses L_2,1-norm sparse regression instead of the traditional least squares regression based on L₂-norm to improve the robustness of the framework.Finally,experimental analysis is performed on multiple public data sets to prove that the proposed framework has stronger performance than existing related methods.Finally,the research contents of this paper are summarized and prospected.