Semi-supervised Semantic Segmentation Based on Confrontation Network
Rui Di, Dan Huang · 2021 International Conference on Electronic Information Engineering and Computer Science (EIECS) · 2021
We propose a semi-supervised semantic segmentation method using adversarial networks. Different from most existing discriminators, we use spatial resolution to distinguish the predicted probability map and the real ground segmentation distribution. The identifier can improve the accuracy of semantic segmentation by coupling the antagonism loss of the model and the standard cross entropy loss. In addition, full convolution discriminator is used to achieve semi-supervised learning, which can provide additional monitoring signals. Compared with the existing methods using weak labeled images, our method enhances the segmentation model. The experimental results on VOC2012 data set prove the effectiveness of the algorithm.