Research On Image Classification Based On Unsupervised Domain Adaptation

Methods based on deep learning and neural networks can classify images quickly and accurately,and have a wide range of application scenarios in the real world.However,with the advent of the era of data explosion,the diversification of image sources and the various acquisition methods lead to the different distributions of image features,which bring the problem of domain shift.In addition,during training,the target images may not be labeled due to the lack of prior knowledge and the high labeling cost.To address the above problems,unsupervised domain adaptation methods can transfer knowledge from the source domain to the target domain,so as to classify images under domain shift.The existing researches combine unsupervised domain adaptation methods with existing vision models,which are successfully applied to various image classification tasks.Currently,most of the existing unsupervised domain adaptation methods assume that the label sets of the source domain and the target domain are identical.However,in practical applications,the label sets of the two domains are likely inconsistent and even the data of the source domain is unavailable.Therefore,the main topic of the research is to align the feature distributions of source images and target images.In this case,the labeled source knowledge can be transferred to the unlabeled target domain.Based on this idea,there are many difficult problems of existing unsupervised domain adaptation methods that need to be further considered.First,class-imbalance is a very common and practical problem.The existing methods explore data augmentation and pseudolabel generation methods based on the original images,but they do not generate new usable images.Especially in medical images,the image synthesis methods are prone to be ineffective due to non-conformity and low credibility.Second,with the significant increase in computing power,to improve the precision of image classification,the existing methods either build more and more complex models or stack more and more modules.However,these methods ignore the information interaction and knowledge transfer between different modules in the model.Third,when the source data cannot be directly accessed due to privacy protection and data scale,the single knowledge transfer approach cannot fully exploit the inter-domain information,resulting in the lack of attention to the feature information of the original images and insufficient training of the model.Forth,the common label set in universal domain adaptation is critical.It can be used to obtain the private label sets,weaken negative transfer and classify images.However,the existing universal domain adaptation methods do not predict and delve into the common label set.To deal with the above problems,the research works of this dissertation are as follows:(1)Unsupervised domain adaptation method based on deep network fine-tuning and self-learning.In order to solve the problems of multimodality heterogeneity,insufficient labels and class-imbalance of images,a model for cross-domain classification is proposed,which is trained by a two-stage progressive fine-tuning approach.First,based on transfer learning,general knowledge is learned directly from the model that pre-trained on a large dataset,and it can classify source images after fine-tuning.Then,to apply the fine-tuned model to the target domain,the self-learning strategy is used to generate pseudo labels and the model is fine-tuned again with the proposed adaptive layer,so that the model can effectively classify the target images.(2)Unsupervised domain adaptation method based on coherent training and cooperative learning.The above proposed method tends to cause insufficient information interaction between modules during multi-stage training.Moreover,the correctness of the constructed pseudo labels depends on the ability of the classification model.Aiming at these limitations,a coherent cooperative training framework is designed instead of a phased training model.The intermediate images of the generator are innovatively used for data augmentation,so as to obtain the more accurate pseudo labels to construct a balanced training dataset.The proposed framework consists of two classifiers for source images and target images respectively,as well as a generator for image style translation and data augmentation.In the coherent training process,the parameters of all modules are successively updated,and the data is transfered between different modules to realize knowledge transfer and collaborative training.The final classification predictions are the voting results of two classifiers.(3)Domain consistent knowledge transfer for source-free domain adaptation method.The above two proposed methods require access to the source data during training.For the source-free domain adaptation without access to the source data,a domain consistent knowledge transfer method is proposed.To avoid insufficient training and over-fitting of the target model,the progressive two-stage training model is adopted,and a variety of knowledge transfer approaches are combined to ensure consistent knowledge transfer between different domains.Firstly,considering that better feature alignment is beneficial to train the generator for image style translation and train the target model for target image classification,a feature distribution consistency strategy is designed to align the feature distributions of images.Then,the pseudo-labels generated by the classification consistency strategy are used to fine-tune the target model and enhance the classification performance of the target model.(4)Prediction of common labels for universal domain adaptation method.The previous proposed methods are all based on the closed domain adaptation setting where the label sets of the source domain and the target domain are the same.For the universal domain adaptation where the source domain and the target domain contain common labels and private labels,a universal domain adaptation method based on the prediction of common labels is proposed.Compared with the existing methods,it pays more attention to the prediction and utilization of common labels.Firstly,to predict the common labels,a method named category separation via clustering is designed,and a new evaluation metric named category separation accuracy is also proposed to measure the accuracy of the predicted common labels.Then,to weaken the negative transfer,the source data is selected based on the predicted common labels to fine-tune the model to better align the image feature distributions of different domains.Finally,the target images are classified by clustering results and predicted common labels.Based on the different distributions of image features in unsupervised classification tasks,the above four unsupervised domain adaptation methods are studied in this dissertation.Starting from the common class imbalance problem,several problems of the existing unsupervised domain adaptation methods are gradually solved.For example,insufficient information interaction between different modules,the unavailability of source data,the prediction and utilization of common labels when the label sets of different domains are inconsistent,etc.In this dissertation,the proposed methods are extensively tested on several public datasets.The experimental results show that the proposed methods can not only solve some problems in the existing unsupervised domain adaptation methods,but also fully improve the robustness,generalization and self-adaptability of the models,and stabilize the performance of image classification fruitfully.