Research On Unsupervised And Semi-supervised Domain Adaptation Methods Based On Drop-Out Regularization

Domain adaptation learning is a hot topic in the field of machine learning and pattern recognition.With the continuous progress of deep learning research,the era of manual feature selection is coming to an end.Deep learning can automatically extract representative features from the original data in the best way and can be applied to different domain adaptation tasks more effectively.The existing domain adaptation method based on deep learning aims to reduce the difference of data probability distribution between different domains by mapping the instances of the label-rich source domain and the unlabeled target domain into the same feature space.However,these methods do not take into account the data in the class boundary in the target domain,which leads to fuzzy target characteristics in the class boundary,which hindering the improvement of classification accuracy.In addition,most of the time,these methods align the global image of the source domain and the target domain,failing to realize that not all regions have a positive effect on the domain adaptation task in the process of image domain adaptation,and forced alignment of invalid regions such as background and noise will lead to negative migration.Therefore,"How can the model efficiently distinguish target domain data at the classification boundary?" "How to suppress the interference of invalid regions in the image in the process of domain adaptation and highlight the effective regions?" Become a major concern right now.Inspired by adversarial training and attention mechanism,in view of the above problems,this paper respectively from the perspective of a half and no supervision,puts forward using the Dropout regularization method to produce a can detect in the recognition of the decision boundary of the target domain data,encourage generators for the target domain generation is more distinct characteristic of the target domain.In addition,in order to suppress the possible negative effects of invalid regions of images in the process of unsupervised domain adaptation,an end-to-end trainable attention model was used to replace the original generator and extract the main migration features.This paper has completed the following work:(1)It changes the previous way that domain adaptation learning needs to align the global image of the source domain and the target domain,and uses the attention mechanism in the convolution network,so that the attention network can extract the attention features that areeffective for the domain adaptation task,suppress the interference of invalid information,and enhance the generalization of the domain adaptation model.(2)Different from the previous domain adaptation learning,the label-rich source domain data and the unlabeled target domain data needs to be trained separately in different ways,this paper changes the domain adaptation learning the inherent way of training,adding a small amount of labeled data in the source domain of the target domain data,to form a new "source domain" dataset,the network on the training of source domain data at the same time know the probability distribution of the target domain,improve the model generalization of the target domain.