Domain adaptive unsupervised image segmentation method based on generative adversarial and class feature distribution

Abstract

The invention discloses a domain adaptive unsupervised image segmentation method based on generative adversarial and class feature distribution. The method is characterized in that unsupervised semantic segmentation is carried out on a target domain image by adopting adversarial learning and a self-learning method based on class feature distribution, wherein the invention specifically comprises two networks of a main semantic segmentation network and an auxiliary semantic segmentation network, and two modules of a domain alignment adversarial module and a pseudo-label optimization module. Compared with the prior art, the method has the advantages of high segmentation precision and multi-granularity semantic consistency, unsupervised semantic segmentation of the target domain image is realized by utilizing adversarial learning and self-learning based on class feature distribution, the multi-granularity semantic consistency between the domains is enhanced by aligning the domains on global and class levels, and the problem that a traditional unsupervised segmentation method is insufficient in segmentation precision is solved.

Description

technical field [0001] The invention relates to the technical field of image semantic segmentation, in particular to a domain-adaptive unsupervised image segmentation method based on generative confrontation and class feature distribution. Background technique [0002] In the past decade, especially after the full convolutional neural network was proposed, convolutional neural networks have been commonly used for semantic segmentation tasks. However, supervised semantic segmentation of scenes puts high demands on the data, because manual pixel-by-pixel annotation requires a lot of manpower and time. According to statistics, it takes at least 90 minutes to manually label the pixel-level labels of each real scene map. Therefore, the accuracy improvement of unsupervised image segmentation has become the focus of research in recent years. Recent advances in computer graphics have enabled simulators to generate labeled synthetic images with low effort, which has inspired people...

AI Technical Summary

Problems solved by technology

But the former cannot eliminate the error of class-level segmentation, while the latter is limited to local areas and ignores the overall information

Images