Picture classification method and system based on online domain adaptive deep learning

Abstract

The invention provides a picture classification method and system based on online domain adaptive deep learning, and the method comprises the steps: extracting the features of source domain picture data and target domain picture data through a feature extraction network, and obtaining a source domain feature and a target domain feature; explicitly expressing the category condition distribution probability in the domain based on the distance from the source domain feature and the target domain feature to the clustering center; updating a feature extraction network and a category clustering center by minimizing a relative entropy distance between a source domain data label and an intra-domain category condition distribution probability; obtaining category allocation of a target domain sample by solving a transportation problem, re-updating a feature extraction network and a clustering center by minimizing a relative entropy distance between category allocation and intra-domain category condition distribution probability, inputting a to-be-classified picture into the re-updated feature extraction network, and obtaining picture features of the to-be-classified picture; and calculating the probabilities of the picture features and all category centers in the re-updated category clustering center, and taking the category with the maximum probability as a classification result.

Description

technical field [0001] The present invention relates to the fields of AI application technology and picture classification technology, and in particular to a picture classification method and system based on online domain self-adaptive deep learning. Background technique [0002] Deep convolutional neural network methods are based on the assumption that the training dataset and application scenarios have the same distribution. When there is a domain offset between the actual scene and the training data set, the performance of the model will be greatly reduced. The existing solution is to use the data of the target domain to fine-tune the model, but this method requires a large amount of human-labeled target domain data. Unsupervised domain adaptation (UDA) methods use unlabeled target domain data to address the performance degradation caused by domain shift. Existing clustering-based unsupervised domain adaptation algorithms rely on traditional offline clustering methods s...

AI Technical Summary

Problems solved by technology

[0004] The present invention designs an online clustering algorithm based on transportation problems, and forms a method and system for performing tasks using domain-adaptive deep learning to solve the problem of negative transfer

Images