Unsupervised domain adaptive semantic segmentation method based on category homogeneity guidance

Abstract

The invention discloses an unsupervised domain adaptive semantic segmentation method based on category homogeneity guidance, and the connotation of category homogeneity comprises similar feature aggregation and heterogeneous feature pushing, i.e., in a domain adaptive process, not only is the feature of the same category target between different domains ensured to be close to the same clustering center as much as possible, but also the feature of the same category target between different domains is ensured to be close to the same clustering center as much as possible; and meanwhile, the feature distribution difference between different types of targets is as large as possible. The domain adaptation effect from coarse to fine and from shallow to deep is realized by constructing a hierarchical domain adaptation strategy of image level-feature level-category level-instance level from similar feature aggregation and heterogeneous feature extension. According to the model constructed by the method, the difference between the source domain and the target domain is comprehensively considered, and the leading performance is realized on the unsupervised domain adaptive semantic segmentation task of the general streetscape semantic data set.

Description

technical field [0001] The invention relates to the technical field of computer vision and pattern recognition, and more specifically relates to an unsupervised domain-adaptive semantic segmentation method based on category similarity and difference guidance. Background technique [0002] Semantic segmentation is to assign a semantic label to each pixel in the image. At present, the semantic segmentation method based on deep learning requires large-scale manual fine-grained annotation, and the time and labor costs of fine-grained annotation are extremely high. Therefore, the existing labeled source domain data is used to train the model, and the unlabeled target domain Effective reasoning, and then unsupervised semantic segmentation on the target domain, this method is called unsupervised domain adaptive semantic segmentation, which has theoretical research value and practical application value. [0003] The core of the unsupervised domain adaptation semantic segmentation t...

AI Technical Summary

Problems solved by technology

However, the current GAN-based methods have the following deficiencies: 1) The GAN-based global adversarial learning method extracts the features of the source domain and target domain images through the feature extraction network, and trains the generator and the discriminator at the same time; but when the features obtained by the generator can be different When correctly identified by the discriminator, the generalization of the trained semantic segmentation model on the target domain is still poor

This is because domain adaptation strategies based on global feature adversarial learning ignore deep intra-class and inter-class differences; 2) Although category-level domain adaptation strategies and instance-level domain adaptation strategies have been proposed, the “pixel confusion” The problem has not been well resolved

Images