Implementation method of unsupervised domain adaptive model based on label correction

Abstract

The invention discloses an implementation method of an unsupervised domain adaptive model based on label correction. The method belongs to the field of computer image classification, and comprises the following specific steps: 1, constructing an unsupervised domain adaptive model based on an adversarial domain adaptive idea; 2, pre-training the constructed unsupervised field adaptive model; 3, after the pre-training is completed, initializing a label corrector in the constructed unsupervised field adaptive model; 4, adding an inverse relative entropy training label corrector and adding entropy loss assistance clustering. The invention provides an implementation method of an unsupervised field adaptive model based on label correction, which improves the prediction precision on the basis of almost not additionally increasing computing resources and training time, so that the method can be conveniently applied to adaptive models in other fields, wherein the training precision and efficiency are improved by adding the label correction module, and the method has a wide application prospect.

Description

technical field [0001] The invention belongs to the field of computer image classification, and relates to a method for realizing a domain adaptive model, in particular to a method for realizing an unsupervised domain adaptive model based on label correction. Background technique [0002] With the continuous development of information technology, various data are widely collected and stored, but these data lack labels. Transfer learning uses existing knowledge to solve problems in different but related fields. The key to realizing transfer learning is to make full use of existing knowledge. Existing knowledge and connections between mining domains. [0003] Domain adaptation is a sub-direction of transfer learning, which tends to solve inter-domain transfer tasks where the feature space and category space are consistent but the feature distribution is inconsistent. Generally speaking, we call the labeled data set as the source domain (including data and labels), and the dat...

AI Technical Summary

Problems solved by technology

From the initial global feature alignment to the multi-dimensional alignment of the intrinsic structure of the mining data set and assisted by various sample selection strategies, some of these methods may increase the pseudo-label of the target domain sample (the mod...

Images