Pseudo label loss unsupervised adversarial domain adaptive picture classification method based on Gaussian uniform mixture model

Abstract

The invention discloses a pseudo label loss unsupervised adversarial domain adaptive picture classification method based on a Gaussian uniform mixture model. According to the method, knowledge is migrated to a related target domain in a cross-domain manner by using a large amount of available annotation data of a related source domain through a transfer learning or domain adaptation method to obtain target data with labels; according to the domain adaptation method, Gaussian uniform mixture model detection outliers and a deep neural network are fused for image classification, the Gaussian uniform mixture model is used for modeling a cosine distance from target sample features of each class to a class mean value, and a target sample posterior probability is obtained and serves as an importance degree for estimating a target sample pseudo label; adding auxiliary pseudo tag loss proposed based on a target sample pseudo tag generated in the training process into training of the neural network; meanwhile, conditional entropy loss is minimized, so that learned features are far away from a decision boundary; a large number of experiments prove that the method can improve the picture classification accuracy of the deep network model.

Description

technical field [0001] The invention relates to the technical field of deep learning image classification, and mainly relates to an unsupervised adversarial domain adaptive image classification method based on a Gaussian uniform mixture model. Background technique [0002] Current state-of-the-art achieves impressive results on large-scale datasets for image classification or action recognition. However, acquiring such large annotated datasets is prohibitively expensive and requires knowledge transfer from existing annotated datasets to specific unlabeled data. If labeled and unlabeled data have different characteristics, then they are sampled from two different domains. In particular, data sets collected from the Internet, such as from platforms where videos or images are shared, are very different from the data that applications need to process. Due to domain shifts common in computer vision, the data distribution between the source and target domains can vary greatly. ...

AI Technical Summary

Problems solved by technology

[0005] Although the current domain adaptive methods have shown good performance in practical applications, they cannot effectively use pseudo-labels to improve the accuracy of image classification, and still face huge challenges.

Images