Domain adaptive image semantic segmentation method based on GAN

Abstract

The invention relates to the technical field of image semantic segmentation, in particular to a domain adaptive image semantic segmentation method based on GAN. The method comprises a semantic segmentation network, a generative adversarial module, a spatial receptive field module and an adversarial module group. The target domain is a label-free data set, the source domain is a label data set, andthe task is to obtain a semantic segmentation label graph of the target domain. The inside of a classical image semantic segmentation network can be regarded as an encoder and a decoder, and an inputimage is encoded and decoded to obtain an output image with the same size. Corresponding adversarial training auxiliary modules are added to an encoder and a decoder respectively to reduce the domaindrift problem caused by domain adaptation. According to the invention, the problem of low accuracy of unsupervised image semantic segmentation in the prior art is solved.

Description

technical field [0001] The invention relates to the technical field of image semantic segmentation, in particular to a GAN-based domain-adaptive image semantic segmentation method. Background technique [0002] At present, the deep neural network is very capable of learning a very good visual model under the premise of big data. But collecting label data, even pixel-level label data, is very difficult. According to reports, it takes at least 90 minutes to manually label a picture with pixel-level labels. For unlabeled semantic segmentation of image datasets, that is, unsupervised semantic segmentation, one of the more popular options in recent years is to use virtual data (such as images in game scenes, we can easily obtain pictures and marked pixel-level labels, several orders of magnitude faster than manual labeling) to generate image labels for real-world scenes. However, simply applying the model trained on the virtual data to the actual picture results in a poor pictu...

AI Technical Summary

Problems solved by technology

[0004] Aiming at the deficiencies of the prior art, the present invention discloses a domain-adaptive image semantic segmentation method based on GAN, in order to solve the unsupervised image semantic segmentation in the prior art low accuracy problem

Images