Research On Image Domain Adaptation Based On Deep Learning

In recent years,deep learning technology has played an increasingly important role in each line,but the key to good performance of deep learning algorithms is based on the need to consume a lot of resources to train deep learning models.This process not only It is time-consuming and labor-intensive,and there will be repeated waste in many fields,that is to say,a deep learning model performs well in training and testing using the data in the current scenario,but it performs poorly in a related application scenario,resulting in having to Re-expending the expensive data to re-label the data,and then waste time retraining the model.Therefore,how to use the rich data in related fields to transfer and reuse knowledge on tasks on the current data set has become a hot research topic today.One of the directions is Domain adaptation problem.This paper proposes corresponding domain adaptive algorithms from the perspectives of adversarial learning ideas and semisupervised learning for such problems,as follows:(1)Starting from the search of the invariant feature space of the source and target domains,this paper proposes a A domain adaptive algorithm combining edge distribution alignment and conditional distribution alignment.Aiming at the problems in the current adversarial-based domain adaptive research,it is proposed that the edge distribution of data features should be aligned first,and then the conditional distribution of data features should be aligned.Through experimental results,it can be shown that the algorithm proposed in this paper has an accuracy rate of 96.2 % and 97.1% And 90.4 % on three tasks:MNIST ? USPS,USPS ? MNIST and SVHN ? MNIST on the disgits dataset;on the Office31 dataset,the accuracy of the three tasks A ? W,D ? W,and W ? D is 91.4 %,98.2 %,and 99.8 % Compared with a single type of algorithm,these results are greatly improved.(2)A semi-supervised domain adaptive algorithm model is proposed.Aiming at the problems in the current adversarial-based domain adaptive algorithms,this article explores the use of semi-supervised learning techniques to avoid the use of adversarial ideas for domain adaptation.algorithm.Based on the idea of consistent regularity,a method of obtaining pseudo-labels was designed,which combined the loss of clustering and the loss of alignment according to the clustering structure to directly align the data features of the source and target domains.It can be seen from the experimental results that the accuracy of the algorithm proposed in this paper is 93.2%?98.0% and 99.9% on the three tasks A? W,D ? W,and W ? D on the office31 dataset;the average accuracy of the six tasks on the ImageCLEF-DA dataset reached 87.3 %,and these results exceeded the similar MSTN algorithm.(3)This article expresses the research outlook of adaptive algorithms,and points out that in future research,we should focus on how to solve the problems of domain adaptive algorithms based on adversarial ideas,and we should focus more on Applications in other fields.