Application Of Unsupervised Domain Adaptation In Image Recognition Technology

Pattern recognition algorithms use a large number of labeled sample data as the training set to train the model and then realize the classification of target samples.The two important prerequisites for these algorithms to be effective are that there is a large number of labeled data and these data have the same feature distribution.However,in most practical application scenarios,data is usually missing labels,and labeling a large amount of data is a very time-consuming and laborious matter.In the problem of image recognition,there always are different degrees of distribution discrepancy between a large amount of annotated data that is easily obtained by people and the actual target data which needs to be classified.In order to realize the effective classification of the unlabeled target domain sample by the labeled source domain sample,the transfer learning/domain adaptation algorithm has been gradually explored by the majority of researchers.Based on the traditional machine learning algorithm and the widely popular prototype model of the deep neural network,the author innovates the algorithm and applies the domain adaption algorithm to the image recognition technology.The main work and contributions of this thesis are as follows:(1)A label disentangled analysis domain adaptation algorithm for image cross-domain recognition is proposed.In order to solve the problem of insufficient feature changes and the problem of low model generalization due to the lack of learning the essential features of images,the author proposes an unsupervised domain adaption algorithm based on label disentangle analysis in this thesis.The algorithm increases the degree of freedom of feature transformation by appropriately disentangle the link between the label and the feature and obtains more essential feature information of the image.The algorithm achieves the distribution alignment at the feature level and the label level respectively.At the same time,cross-coupling between the reconstructed feature and the sample feature is introduced to reconstruct the relationship between the feature and the label.Thus,the reconstructed relationship can adapt to the two domains simultaneously and then ensure the class discriminability of the feature.Furthermore,in this thesis,the effectiveness of the proposed algorithm is verified through domain adaptation theory analysis and experiments on multiple image data sets.(2)A deep multi-representation adversarial learning domain adaptation algorithm for image cross-domain recognition is proposed.With the rapid development of deep learning,good domain invariant features can be extracted by fine-tuning the neural network to minimize the distribution discrepancy between the two domains,and corresponding class discriminant constraints can be designed to maintain the discriminability of the features.However,the transferability and discriminability of features are often out of sync,and even there are some contradictions between them.In model design,these contradictions tend to be ignored and then lead to a reduction in the recognition accuracy.Therefore,based on the current popular deep learning,the author further proposes a deep multi-representation adversarial learning domain adaptation algorithm.By learning domain-invariant feature,domain-specific feature,class-invariant feature,and class-specific feature,we can pay full attention to different feature information.Furthermore,we obtain the final consistent and good feature representation in domain invariance and class discriminability through the adversarial learning among the four feature representations.Such a feature representation can enable the two domains to be aligned while maintaining good discriminability of the data.Extensive experiments show the effectiveness of the proposed algorithm.(3)A dynamic weighted learning domain adaptation algorithm for image cross-domain recognition is proposed.Domain transferability learning and class discriminability learning are very important to solve domain adaptation problems.In this thesis,the author innovatively focuses on the possible adverse effects of these two kinds of learning on each other in the learning process and then proposes a dynamic weighted learning algorithm to balance their learning weights.By dynamically weighting domain transferability learning and class discriminability learning in the process of model training,the discrimination missing caused by excessive domain transfer learning and the insufficiency of domain alignment caused by excessive class discriminability learning are avoided.The feasibility of the proposed algorithm is analyzed based on the domain adaptation theory,and the effectiveness of the proposed algorithm is verified on multiple image data sets.