Research On Unsupervised Domain Adaptive Method Based On Image Recognition

Deep neural networks have been successfully applied to many machine learning tasks and have shown amazing performance improvements.However,Traditional machine learning algorithm assumes that training data and test data follow the same distribution,which is often not true in real-world applications.If the distribution of training data and test data is greatly different,the performance of the classifiers trained by traditional machine learning algorithms will be greatly reduced.Domain adaptive learning is about how to adapt a classifier from a source domain(training data)to a target domain(test data).It can effectively extract the domain invariant features of the two fields,and improve the final classification accuracy.Therefore,this paper focuses on the domain adaptation problem in image recognition,and proposes a better improved algorithm to solve the domain adaptation problem based on the shortcomings of the existing domain adaptive methods.The main research contents of this paper are as follows:(1)This paper first expounds the research background and development status of domain adaptive learning,and makes a detailed comparison and analysis of domain adaptive methods.By analyzing the domain adaptive methods of different ideas,the advantages and disadvantages of their existence are compared,which leads to the research direction of the domain adaptive problem.(2)A multi-weighted subspace alignment method is proposed for the characteristics of source domain samples and target domain samples.The weighted subspace is generated by re-weighting the sample data;then the weighted subspace fitting is performed by the PCA method and the greedy strategy proposed in this paper;finally,the source subspace and target subspace are aligned by using the differential geometry tool of the Grassmann manifold.And a nearest neighbor classifier is learned to obtain the data representation of the source domain and target domain.(3)The most compelling aspect of deep networks may be that they learn a feature representation that is especially well suited for the particular prediction task entirely from the data itself.For this aspect,a neural network structure based on multi-layer correction is constructed.The correction is used to correct the internal representation of the target data so that they simulate the source data.For the residual layer of the target data,we make the data classifier of the source domain adapt to the target domain by additive correction,and use the additive superposition to perfectly align the data representation of the source domain and the target domain.Then,if the prior distribution of the class is not considered,it is easy to ignore the weight deviation of the class,which leads to the decrease of domain adaptation performance.Therefore,we introduce the auxiliary weight of a specific class to override the set source sample.In this way,the re-weighted source data shares the same category weight as the target data.On this basis,we use multiple weights MMD to modify the fully connected layer and increase the representation ability of the network.Finally,the domain invariant features obtained by the learning are extracted and classified to obtain the final recognition effect of the target image.