Research Of Cross-domain Image Recognition Based On Transfer Learning

The emergence of large-scale accurately labeled datasets is one of the keys to the great success of today’s image recognition field.However,data labeling is a very time-consuming and expensive task.Transfer learning has received great attention as an effective solution to this problem because it can transfer knowledge from a well-labeled relevant dataset(source data)to the current task(target data),thereby effectively alleviating the scarcity of annotation information.This doctoral thesis studies the transfer learning problem in the field of image recognition,and the main contributions are listed as follows:1.Firstly,for the problem of unsupervised transfer learning,a transfer learning method based on category mean metric is proposed.The previous transfer learning methods ignored the optimization of the classification boundary during adaptation and did not explicitly optimize the distance between the same category and different categories of data,hence the discriminative feature expression could not be extracted.To remedy this problem,this doctoral thesis proposes a transfer method based on category mean metric,which shortens the distance between category mean in different domains but the same category,and increases the distance between different category mean in different domains.The proposed method effectively overcomes the shortcomings of existing methods.2.Secondly,for the heterogeneous transfer learning problem,a transfer learning method based on label and structure consistency is proposed.Under the condition that the feature space is completely heterogeneous,it is a very challenging task to find transferable samples and features of data in different domains.Therefore,the method proposed in this doctoral thesis uses the label information to constrain the feature space and explores the structure of the target data through the label propagation algorithm to update of the training labels.Comparing with existing state-of-the-art methods,the proposed method achieves excellent performance in the scenario of cross-modal,cross-feature and so on.3.Thirdly,for partial transfer learning problems,a transfer learning method based on domain compensation mechanism is proposed.This doctoral thesis focuses on analyzing the internal mechanism that the existing transfer learning methods cannot be directly applied to partial transfer learning problems in principle and proposes to compensate the category space of the target data and ensuring the rationality of data distribution adaptation in different domains.Secondly,it is noticed that the data scales of the same category of the source data and the target data are not the same,and their category biases are not always same,hence a compensation scheme of category space in this paper is proposed.Categories that appear more frequently in the target data but with a small number in the source data can be appropriately weighted.4.Fourthly,for the domain adaptative object detection problem,a transfer learning method with discriminative distribution alignment is proposed.Most of the existing transfer learning methods regard the image as the minimum basic unit of distribution adaptation and ignore the problem of different transferability in different regions of the image.In this paper,we design a sub-network that can focus on more transferable image regions in images from different domains and use the aggregation of scores from multiple regions of interest to calculate the foreground scores of different regions of the image.The training of joint classification loss and distribution adaptation enables data in different domains to obtain discriminative information from the classifier during distribution adaptation,thereby better adapting the distribution of data in different domains.5.Fifthly,for the black-box transfer learning problem,a transfer learning method based on class-conditional contrastive learning is proposed.In this paper,a transductive learning method based on the CLIP model is proposed,which can achieve high classification performance under high noise background through class-conditional contrastive learning.The method proposed in this paper has achieved state-of-the-art performance in many scene classification tasks,including remote sensing image classification,fine-grained image classification and other tasks.In scenarios where the target data is completely lacking in annotations,the proposed model shows great advantages in both accuracy and capacity.