A Transfer Classification Learning Method Preserving the Sparse Structure of Image Classification

Abstract

The invention discloses a migration classification learning method that maintains the sparse structure of image classification, finds two source domains and target domains that are similar in distribution but different, the source domain contains label data, and first uses a supervised classification method to train a classifier on the source domain , and use the classifier to predict the pseudo-label of the target domain data; secondly, use the maximum mean difference to construct the marginal distribution and the conditional distribution item of the source domain and the target domain data respectively, and combine the two to form a joint distribution item; then, use the effective projection The sparse learning toolkit constructs the sparse representation matrix S on all data to construct the sparse structure preserving term; then, constructs the structural risk minimization term using the structural risk minimization principle; finally, constructs the structural risk minimization term, the joint distribution term and the sparse The structure-preserving items are combined to construct a unified transfer classification learning framework, and the classification function representation theorem including the kernel function is used to solve the framework to obtain a classifier that can finally be used to predict the target domain category.

Description

technical field [0001] The invention relates to the technical field of transfer learning classification in machine learning, in particular to a transfer classification learning method for maintaining the sparse structure of image classification. Background technique [0002] Traditional machine learning has two basic assumptions, one is that the test samples and training samples should satisfy independent and identical distribution, and the other is that there must be enough available training samples. However, these two conditions are often not easy to meet, so transfer learning came out as the times require. Transfer learning is a method of using known content knowledge to solve related but different fields. Use the knowledge already known in a field to solve the problem that there are only a few labeled samples or even no labeled samples in the learning target field. Often, the smaller the difference between the two fields and the more factors they share, the easier and ...

AI Technical Summary

Problems solved by technology

However, since manifold regularization is used to describe local structural information, this regularization framework is not very satisfactory for image classification.

Images