A Domain Generalization Model Based on Multilevel Domain Adversarial Invariant Representation Learning

Abstract

Due to the influence of various factors such as cross-scene shooting height, shooting time, imaging angle of view, weather, etc., the performance of the defect detection model degrades in cross-scene situations. In this paper, we proposed a domain generalization model based on multilevel domain adversarial invariant representation learning, which included the designed defect space feature and channel feature enhancement modules. Among them, the defect space feature enhancement module uses the shallow feature invariance score to guide the model to extract cross-scene invariant representations, and at the same time, design and use the defect importance information in the shallow features to enhance the representation of defects in the shallow features. The channel feature enhancement module assigns learning weights to different channels of each sample according to the similarity between the sample and the whole, and the importance of different feature channels to defects, so as to better guide the model to extract cross-scene invariant representations. The experimental verification on the photovoltaic cell data set collected in outdoor natural scenes shows that, proposed model has stronger generalization ability than existed domain generalization methods.