Image semantic segmentation method

Abstract

The invention discloses an image semantic segmentation method. According to the method, the problem that semantic segmentation of a synthetic data set image excessively depends on the performance of an image translation model is solved by reversely improving the image translation model through a first image segmentation model; the output data of an optimized image translation model (a first imagetranslation model) and the output data of a first image segmentation model are used to translate an image. Therefore, the first image segmentation model can be optimized again through supervised learning, an image semantic segmentation model (a second image segmentation model) with stronger domain invariant feature distinguishing ability is obtained, semantic segmentation is performed on the target image by adopting the image semantic segmentation model (the second image segmentation model), and image semantic segmentation can be completed. The method solves the problems that in the prior art,the performance of the image translation model greatly influences the segmentation result and cannot be corrected.

Description

technical field [0001] The invention relates to the field of image processing, in particular to an image semantic segmentation method. Background technique [0002] The pixel-level semantic segmentation map is to mark the pixels belonging to different categories on the picture with different labels, and it has the function of eyes in automatic driving. The vehicle takes a picture of the front, obtains the segmentation map through the segmentation model, and obtains the category and location information contained in the road. This information is fed back to the automatic driving system to determine whether to move forward, stop, turn, or other operations. [0003] The main difficulty faced by image segmentation techniques on synthetic datasets is that when artificial labels from the target domain are not used at all, the segmentation model trained from synthetic datasets will experience domain shift in application. Therefore, the main breakthrough point of the existing synth...

AI Technical Summary

Problems solved by technology

[0006] Aiming at the above deficiencies in the prior art, an image semantic segmentation method provided by the present invention solves the problem in the prior art that the performance of the image translation model has too much influence on the segmentation result and cannot be corrected, and provides a method with stronger Image Semantic Segmentation Based on Domain-Invariant Feature Discrimination

Images