Research On Domain Adaptation Algorithm Based On Generative Adversarial Networks

In recent years,with the improvement of computer computing power and data accumulation,deep learning algorithms relying on a large number of labeled samples have achieved great success in the fields of computer vision,natural language processing,speech recognition,etc.,and their performance exceeds traditional manual design features and shallow learning algorithms.However,the success of deep learning depends on a large amount of labeled data,and the labeled data has the disadvantages of low timeliness and high cost,so when a new task comes,a deep learning model cannot be quickly deployed.In response to this problem,researchers proposed a training strategy based on transfer learning,where domain adaptation is one of the branches of transfer learning research,the purpose is to solve the problem when the training set(source domain)and test set(target domain)have different feature distributions but the same task.The latest research shows that Generative Adversarial Networks(GAN)can align the features of the source and target domains and realize the transfer of knowledge.However,in adversarial training,most methods do not consider the sample category information sufficiently,so that the generated features lose part of the category information,which leads to a decrease in classification accuracy.In order to solve this problem,this research proposes an Improved Feature's Capability of Carrying Category-specific Information for Adversarial Domain Adaptation(IFCADA)algorithm.The algorithm consists of a pre-trained feature classifier and a feature alignment framework composed of two sets of GANs.First,in the pre-training phase,the encoder-decoder-classifier structure is trained using the source domain label data,and the classification information of the source domain data and the pixel-level information carried by the data itself are extracted as completely as possible to improve the discrimination and representation capabilities of the features.Secondly,in the feature alignment stage,the first group of GANs are trained to generate features that are aligned with the features of the source domain,and the second group of GANs are used to supplement the classification information and pixel-level information of the source domain data lost by the generated target domain features.Finally,the classifier of the source domain is used to classify the adapted features of the target domain.This method alleviates the problem of losing the specific classification information of the data during the feature alignment stage,and thus can improve the classification accuracy.The method has been fully experimentally verified on the standard datasets Digits and Office-31,and the average accuracy is 96.5% and 95.1%,respectively.The experimental results are better than the current mainstream domain adaptation algorithms.