Unsupervised two-stage field adaptive method

Abstract

The invention provides an unsupervised two-stage field adaptive method. According to the method, the accuracy of the pseudo-label is enhanced by using the continuously updated target domain projection subspace, the precise pseudo-label is fed back to the target domain projection subspace, and the two stages are alternately updated, so that a discrimination subspace with optimal classification performance can be obtained. Specifically, in the first stage, two projection subspaces are used for mapping data of a source domain and a target domain into corresponding low-dimensional subspaces, and meanwhile, a feature alignment loss function based on MMD is used for clustering samples of the same category from a cross-domain category space, so that conditional distribution differences are further reduced; and a graph regular term is constructed in a target domain projection subspace according to semantic information (the source domain is label information and the target domain is pseudo label information) of all samples and distance information between two sample points, and adjacent information of original data is kept. And in the second stage, training an SVM classifier by using the subspace projection obtained in the first stage, calculating a pseudo label of a target domain sample, and feeding back the precise pseudo label to the first stage. The method has the advantages that geometric structure information and prior label information in the knowledge migration process are considered, and the accuracy of image classification can be effectively improved.

Description

technical field [0001] An Unsupervised Two-Stage Domain Adaptation Method. Background technique [0002] Image classification based on traditional machine learning assumes that the test samples and training samples need to satisfy independent and identical distribution, and at the same time, a large number of labeled samples with the same distribution as the test data are required to ensure the generalization performance of the model. These two conditions are difficult to achieve in real applications. Domain adaptation can break through the implicit assumptions of traditional machine learning. It is expected to use labeled source domain data and unlabeled target domain data to build a cross-domain learning model, which can solve the problem of scarcity of training sample labels and the difficulty of truly satisfying independent and identical distribution in machine learning. A question of conditions. Contents of the invention [0003] The present invention aims at the pro...

AI Technical Summary

Problems solved by technology

[0003] The present invention aims at the problem of insufficient reduction of inter-domain distribution differences caused by learning a projection subspace in the current domain adaptive method based on low-rank subspace learning, and proposes a two-stage pseudo-label accurate domain adaptive method, using categories that are meaningful for classification Prior information and local structure information, explore the inherent laws hidden in the bottom layer of the data, and improve the robustness, generalization and efficiency of the cross-domain image classification model

Images