Research Of Transfer Learning In Image Classification

With the rapid development of deep networks and the expansion of various kinds of data,people are more and more inclined to solve a variety of the visual tasks in real life by largescale datasets and deep models,such as the image classification problem.While in many real applications,there are a few or no labeled data,or even some categories are not available.The cost of data annotation is too expensive,so it's difficult to train a valid classifier.However,there are enough labeled data in other similar but different areas.Then how to use the data from these similar areas to assist the target task is an important problem in transfer learning.This thesis mainly studies three transfer learning problems in image classification scenery.They are single-source transfer learning,multi-source transfer learning and zeroshot classification.The main contents are as follows.Firstly,this thesis studies the unsupervised single source domain adaptation problem.The method of adversarial domain adaptation based on the joint discriminative representations was proposed.This method uses the adversarial learning between the deep network and the domain discriminator to learn the transferrable features.And we expands the input of domain discriminator so that it takes into account both the feature space and the label space which with the discriminative information.The pseudo-label of target data is gradually corrected by iteration.At the same time,we train the classifier under the supervision of the source domain label information and constrain the classifier,so as to have a better performance in target data.Experiments on public datasets show that this method can classify the target domain data well without label information.Secondly,this thesis studies the problem of unsupervised multi-source transfer learning.For the case that the multiple source domains are similar but different from each other and they all different from target domain,a multi-source transfer learning method based on adversarial and weighted strategy was proposed.The feature extraction process is divided into two parts.The domain sharing and domain-specific network is used in turn to the extract the image features from the source domain and target domain.The adversarial learning between the domain discriminator and the feature extraction network in each source-target combination ensures that the extracted feature invariant and promote knowledge transfer.Then we align the condition distribution in the feature space,further strengthen the discrimination of the feature.One classifier is trained on each source domain,and the classification results of the target domain data are combined by the predictable values of multiple source domain classifiers.Experiments on multiple datasets show that the method can combine the advantages of all source domains and improve the classification performance of the target domain data.Finally,this thesis studies the special situation of transfer learning,zero-shot learning.From the perspective of data generation,a zero-shot learning method based on generation model is given.With semantic feature as conditional information,a semantic condition generation model is established,and the features are distinguished by the classification and the category feature alignment.We make the generated features can fully represent semantic information through a map from visual feature to semantic representation.We use the well-trained generate model to generate the visual representations of unseen classes,thus turning the problem into a supervised classification problem.Then a secondary classification model was given to predict the test samples during the test phase.Experimental results on multiple zeroshot datasets show that the method based on this improved generation model can improve the classification performance of the test set,especially the classification accuracy of unseen classes.