High generalization method and system for semantic segmentation of cross-domain road scenes

Abstract

The invention discloses a highly generalized cross-domain road scene semantic segmentation method, including: generating a virtual image and corresponding labels through a game engine; using the virtual image to generate a global / local texture migration image; migrating the virtual image and global / local texture The image is sent to the neural network for training; the global / local texture transfer image trained by the neural network is subjected to consistency constraints; the virtual image trained by the neural network, the local texture transfer image and the local texture transfer image after the consistency constraint are respectively compared with the label Calculate the loss value, and train the semantic segmentation model according to the loss value; use the trained semantic segmentation model to perform semantic segmentation. The present invention implements data enhancement through the global texture migration and local texture migration of the virtual image, attacks the neural network, and forces the model to learn cross-domain invariant shape information; and the method only performs network training in the source domain, realizing reliable cross-domain Segmentation effect, while having a strong generalization performance.

Description

technical field [0001] The invention belongs to the technical field of computer vision, and in particular relates to a method and system for semantic segmentation of cross-domain road scenes with high generalization. Background technique [0002] Image semantic segmentation means that the computer realizes the understanding of the image according to the content of the image, and then performs visual segmentation. In recent years, with the continuous development of artificial intelligence, semantic segmentation technology based on deep learning has been increasingly applied to various aspects such as industrial production, social security, and transportation. The direction is also an inevitable trend of development. Semantic segmentation is the core algorithm technology of unmanned driving. After the image is detected by the on-board camera or lidar, it is input into the neural network, and the background computer can automatically segment and classify the image to avoid obs...

AI Technical Summary

Problems solved by technology

[0004] Aiming at the problems of high dependence on target domain data and poor generalization performance in the prior art, the present invention provides a cross-domain road scene semantic segmentation method that does not depend on target domain data and has considerable generalization performance. The game engine generates virtual data and performs image processing on the virtual data set, attacks the neural network, improves the robustness of the network to texture changes in cross-domain segmentation, and promotes the learning of the shape of objects in the image by the network, thereby enhancing the multi-domain generality of the model. optimized performance

Images