Research On Image Classification Method Based On Adversarial Domain Adaptive Transfer Learning

With the advent of the era of big data,traditional machine learning methods can no longer solve the increasingly updated task requirements in the field of computer vision.Since the deep network was proposed,it has gained more and more recognition in various fields.However,training a deep network requires sufficient data,which is often accompanied by huge costs and requires the training set and the test set to be the same distribution,which reduces the reuse rate of the model.In recent years,transfer learning has become a popular method to solve these problems.It can realize cross-domain tasks,transfer models trained in a large number of labeled data sets to similar or related fields,and allow the distribution of training set and test set to be different,which can fully solve the problem of low model reuse rate.The main content of this paper is the single-source transfer and partial transfer learning in the field of image classification.The main contents are as follows:Firstly,this paper studies the problem of unsupervised domain adaptation,and proposes an adversarial domain adaptation method based on multiple domain discriminators.The feature extractor is responsible for extracting sample features,and the domain discriminator is responsible for identifying which domain the sample comes from.The purpose of the two is opposite.Through adversarial training,it can extract common features as much as possible to improve classification performance.And this article divides the domain discriminator into multiple,each category corresponds to a domain discriminator,each domain discriminator is only responsible for distinguishing the samples belonging to the category,so that when distinguishing,adding the label information,aligning the source domain and the target Conditional distribution of the domain.Under the guidance of the unsupervised domain adaptation problem standard,experiments are carried out on several public datasets and analyzed from multiple angles to illustrate the effectiveness of the method.Secondly,this paper combines the maximum mean distance with the adversarial domain adaptation method.On the one hand,the feature extractor obtains domain invariant features through the confrontation training with the domain discriminator,and aligns the edge distribution of the source domain and the target domain;On the other hand,this paper combines the maximum mean distance with the feature extractor to reduce the difference between the source domain and the target domain,and achieve the alignment of the conditional distribution.Train the classifier under the guidance of the source domain supervision information and use it in the target domain.Experiments on multiple datasets are compared with some classic methods,and the experimental results are analyzed from multiple angles.It shows that this method can be well applied to the problem of unsupervised domain adaptation.Finally,this paper studies partial transfer learning,and proposes a partial domain adaptation method based on adversarial.Partial transfer learning means that the label space of the target domain is contained in the source domain,and the biggest challenge is the negative transfer problem.First of all,the core is the confrontation training between the feature extractor and the domain discriminator.Furthermore,the non-adversarial domain discriminator is added to obtain the similarity of the source domain of the sample and quantify the transferability of the sample.and each sample in the source domain can obtain the corresponding weight,to reduce the impact of negative migration,promote positive migration,and improve the accuracy of the classifier.Experiments on multiple public data sets of transfer learning are compared with common transfer learning methods and some transfer learning methods,and the experimental results are analyzed from multiple angles.It shows that the method can be well applied to the target task.