Research On Transfer Learning Algorithm For Image Classification

Image classification is the basis of image content analysis and understanding.However,in practical image classification applications,due to the illumination,background or acquisition methods of different images,images in the same category have different distribution,and the labeled images which have same distribution with test images for training a classification model is scarce,which is difficult to train classification model effectively based on traditional machine learning theory and method.Transfer learning is trying to use a large number of labeled data that is related to the target tasks to train model for target tasks.The main idea of transfer learning is to transfer the knowledge of the labeled data from related auxiliary fields to target domain,and complete target tasks,or improve the performance of the target tasks.For image classification task,by using the existing relevant annotation image data sets,the annotation information is transferred to assist the unlabeled image data sets to complete the data annotation.In recent years,deep network has made a great breakthrough in the field of image classification.Transfer learning algorithms based on deep learning can effectively learn the cross domain feature representations with good transferability,and improve the performance of target tasks.For existing drawbacks in the existing deep transfer learning models,this thesis addresses the problems of insufficient distribution adaptation between domains,insufficient cross domain generalization ability,negative transfer in multi-source domain transfer learning.The novel contributions are summarized as follows.(1)For addressing the problem of insufficient distribution adaptation between domains,a deep transfer learning algorithm based on adversarial learning and joint distribution adaptation is proposed.Most existing transfer learning algorithms based on deep learning only consider the marginal probability distribution or conditional probability distribution alignment,and fail to align both the marginal and conditional distribution at the same time,which leads to the problem of insufficient distribution adaptation between domains.In our proposed algorithm,a domain discriminator is introduced into the feature space to adapt the marginal distribution between domains.Two classifiers are introduced to adapt the conditional distribution between domains,the experimental results show that aligning both of the marginal and conditional distribution between domains can effectively improve the feature transferability and the classification accuracy on the target data set.(2)For addressing the problem of insufficient cross domain generalization ability,transfer learning algorithms based on semantic adaptation are proposed.Most existing transfer learning algorithms only align the distribution of different domains in feature space.However,there is no guarantee that features of the same category in different domains that are aligned in feature space can be mapped close to each other in semantic space by classifier.which leads to insufficient cross domain generalization ability of transfer model.In this paper,two transfer learning algorithms used for semantic adaptation are proposed.The former is based on the clustering hypothesis,by taking the category centers of source features as the clustering centers of the target features,the pseudo labels of the target samples are assigned as the labels of the corresponding source category center.With the help of the target pseudo labels,the source and target semantic features are aligned with each other in the semantic space by constraining target features mapped into the decision boundary of the source classifier.The latter considers both feature-level adaptation and semantic-level adaptation.For feature-level adaptation,two classifiers are used for reducing the distribution discrepancy between different domains in feature space.For semantic-level adaptation,we constrain the target semantic feature to be mapped close to corresponding source category centers in semantic space.Experimental results show that the proposed transfer learning algorithms used for semantic adaptation can effectively align the features of different domains in the feature space and semantic space,which improve the classification performance of the transfer learning model on the target image data set.(3)Most of the existing transfer learning algorithms focus on the single-source-to-single-target domain adaptation setting.In practical application,there are usually multiple available source domains,and there are distribution discrepancy among different source domains and between each pair source and target domains.Directly combining multiple source domains into one source domain will lead to the problem of negative transfer.This paper extends the theory of transfer learning generalization error under the assumption of single source domain,and proposes the theory of upper bound for multi-source domain generalization error based on the combination of hypothesis classes.Following the theoretic result,we propose a multi-source domain transfer learning algorithm.Experimental results show that the proposed multi-source domain adaptation algorithm can effectively address the problem of negative transfer in multi-source domain adaptation,and the algorithm can use the annotation information of multiple source domains for the training of the transfer model,and improve the classification performance of the model in the target domain dataset.