Domain generalization image recognition method based on causal decoupling generation model

Abstract

The invention discloses a domain generalization image recognition method based on a causal decoupling generation model. The purpose of domain generalization is to learn domain invariant representation from a plurality of source domain data to be well generalized to an invisible target domain, but the biggest challenge of learning domain invariant features is to decompose semantic information and domain information from an entangled feature space. Considering that causal features have the cross-domain invariant characteristic, the invention provides a causal decoupling characterization model. Firstly, a cross-domain stable causal structure model is introduced as prior; then, a generation model based on a causal structure is constructed, and modeling is carried out on task-related and domain-related features respectively; particularly, the bidirectional causal dependence between the two hidden features is relieved through an intervention means, so that the influence of the domain related features on the prediction task is effectively eliminated. Results prove that the method provided by the invention can effectively decouple task-related and domain-related features, and exceeds most methods for solving domain generalization.

Description

technical field [0001] The invention belongs to the technical field of causal representation learning and decoupling method fusion processing domain generalization, and in particular relates to a domain generalization image recognition method based on a causal decoupling generation model. Background technique [0002] The development of deep neural networks has led to great success in computer vision, especially when the training data and test data follow the same distribution. But learning how to generalize deep neural networks to data outside the training distribution remains a fundamental but challenging problem in machine learning. The goal of domain generalization is that a model trained on multiple source domains generalizes well to unseen target domains. Learning a domain-invariant representation is proposed as a key technique for solving domain generalization problems. But the above-mentioned methods have a drawback, that is, it becomes difficult to learn a domain-...

AI Technical Summary

Problems solved by technology

For example, a camel is standing in the desert, assuming that there is a neural network that can get a high accuracy rate for camels in the desert (maybe focusing on the feature of desert), but this model may not be able to recognize camels standing on green grass or cow standing in desert

Images