Deep Domain Adaptation Methods For Visual Classification

The maturity of deep learning technology has brought significant effects to various artificial intelligence tasks,and it has a wide range of application prospects.In recent years,various deep learning models have emerged,which has promoted the further development of computer vision,natural language processing,and recommendation systems,especially in the fields of face recognition and recommendation search.However,the effectiveness of deep learning depends on the support of hardware technology and a large amount of annotated data.If the model is applied to a completely new field,it will inevitably consume a lot of manpower to label the original data.Another inevitable problem is that in the real scene,the data distribution of the training set and the test set are inconsistent due to the influence of the external environment,which will cause the model perform perfectly on the train set but perform poorly on the test set.In computer vision,such a problem is called "domain shift".Domain adaptation technology can effectively solve the above two problems.The purpose of domain adaptation research is to adapt the features of the source domain samples to the target domain,and finally achieve better generalization performance on the target domain.Due to the effectiveness of domain adaptation technology and its broad application prospects,it has become an important direction in the field of artificial intelligence.In this paper,two models for unsupervised domain adaptation are proposed to address the shortcomings of existing domain adaptation methods.One is that the traditional deep domain adaptation method only considers aligning the data distribution of the source domain and the target domain in a high-dimensional space,ignoring the change of the data distribution of the target domain.If the data distribution of the target domain changes too much,the information of the target domain will be lost.This will impair the classification performance of the algorithm on the target domain.In this paper,the concept of local information preservation is introduced into the deep domain adaptation network,which reduces the problem of loss of target domain data information caused by feature transformation.Based on this,a deep adversarial reconstruction classification network model for unsupervised domain adaptation is studied.Another is that most of the existing domain adaptation methods are based on a single feature structure to align the distribution between different data domains,so the transferred content only contains part of the effective information,such as only low-level geometric features or only high-level semantic features,a single-feature-based domain adaptation method may produce poor transfer results,and based on this,a multi-scale domain adversarial network model for unsupervised domain adaptation is studied.We conduct experiments on the digit datasets and classical transfer learning datasets,and the corresponding experimental results indicate that our model has superior performance in most cases compared to existing domain adaptation algorithms.