Domain adaptive learning method

Abstract

The present invention belongs to the computer vision field and relates to a domain adaptive method. The objective of the invention is to solve the problem of domain migration caused by the difference of the distribution of training categories and test categories in object classification. According to the method of the present invention, an expression described in the descriptions of the invention is adopted to indicate a training data set containing n samples; x in the expression, which belongs to R, indicates a vision space; ZS in the expression, which belongs to a (Z1,... ZN) set, contains N categories; an expression U={Xi}<m>i=1 is adopted to express m test samples from M categories, and training categories and test categories do not intersect with each other; each category is characterized by one vector embedded in a category semantic space y which belongs to R<q>, wherein R is a semantic space or a text description space; p and q represent the dimensionalities of the vision space and the category semantic space; according to the basic principle of a cross-modal embedding method, the training data S are adopted to learn a transfer matrix W<*> which belongs to R<p*q>, so that vision samples can be transferred from the vision space X to the category semantic space y; in a test stage, the transfer matrix W<*> in the learning stage is utilized to map test samples x into the category semantic space. The method of the invention is mainly applied to computer vision conditions.

Description

technical field [0001] The present invention relates to a domain adaptive technology oriented to the field of computer vision, specifically, to a domain adaptive learning method. Background technique [0002] In the traditional object classification technology, the category of the test sample must be included in the category that appeared in the training stage. In order to obtain better classification performance, each category often requires a large number of training samples. In reality, samples of some categories are difficult to obtain. For scenarios where there are few or missing training samples in some categories during the training phase, the training model cannot obtain the data distribution of the test category. Using the model learned from the training category to predict the samples of the test category is prone to the problem of domain shift. Take zero-shot classification as an example for introduction. [0003] Zero-shot classification is a scenario where th...

AI Technical Summary

Problems solved by technology

For scenarios where the test category has fewer or missing samples during the training phase, the problem of domain shift is prone to occur.

Images