Attention Dual Adversarial Remote Sensing Image Semantic Segmentation

Existing semi-supervised remote sensing image semantic segmentation methods neglect to improve the stability of the adversarial network, so that the adversarial network cannot be effectively used to assist segmentation network training, which limits the further improvement of semantic segmentation accuracy. To this end, this paper first introduces a dual confrontation network, and plays a three-way game with the generator to make the network converge as soon as possible, and combines the vertical and cross attention network to propose an attention dual confrontational semantic segmentation model for remote sensing images. The model can not only use dual confrontation training to improve the stability of the network, but also use the global context relationship of pixels to predict by introducing an attention mechanism, thereby improving the accuracy of remote sensing image semantic segmentation. The experimental results on the public remote sensing data set show that the MIOU of this method on US2D dataset reached 69.65%, which are higher than the existing fully-supervised and semi-supervised methods. The effectiveness of the method proposed in this paper is verified.