Multi-label active learning method based on condition-dependent label set

Abstract

The invention discloses a multi-label active learning method based on a condition-dependent label set. First, a condition-dependent label set of labels is mined under the condition of weak labeling. For each time of iteration, the information entropy and the relative entropy of each sample-label pair in a current unlabeled sample pool are calculated on the basis of the condition-dependent label set. Then, the information entropy and the relative entropy are integrated to get the amount of information of each sample-label pair. Finally, the sample-label pair with the largest amount of information is selected as an input object for training an active learning model in the current iteration. Experiments show that the active learning method has good generalization performance compared with other methods for mining the relationship between labels under the condition of weak labeling.

Description

technical field [0001] The invention relates to the technical field of multi-label active learning, in particular to a multi-label active learning method based on a condition-dependent label set. Background technique [0002] With the rapid development of computers, the research on multi-label image classification has become a research hotspot in related fields, and has received more and more attention in academia and business circles. The role of the multi-label active learning method is to deal with the classification problem of multi-label images, that is, to obtain a classifier through learning, which can assign multiple relevant labels to the image according to the content of the image. [0003] At present, the mainstream of research in the field of multi-label active learning is the research on multi-label active learning methods based on "sample-label pairs". Because compared with the multi-label active learning method based on "sample", the multi-label active learni...

AI Technical Summary

Problems solved by technology

[0006] The purpose of the present invention is to provide a multi-label active learning method based on a condition-dependent label set, which is used to solve the sample "weak label" problem encountered in the multi-label active learning method based on "sample-label pairs", or in Mining and exploiting label relationships under "weak labeling" conditions for active learning

Images