Semi-supervised Remote Sensing Image Semantic Segmentation Method Via Deep Generative Adversarial Networks

As a basic technology for image understanding,image semantic segmentation is a field that has been studing in the field of artificial intelligence.Image semantic segmentation performs feature extraction for all pixels in the image and then classifies them into different groups according to semantic information.Semantic segmentation of remote sensing images is an application of semantic segmentation to remote sensing images.It can obtain the category information of surface objects while obtaining its spatial information.It is the basis for applications such as environmental change monitoring and has high research value.At present,deep convolution models have good applications in semantic segmentation of remote sensing images,but the training of these models requires many labeled data.In some cases,remote sensing images are difficult to obtain and label,which requires the use of as little labeling data as possible to obtain satisfactory results.Based on the deep convolution model,this thesis uses a generative confrontation network and attention mechanism to propose a semisupervised remote sensing image semantic segmentation method.The innovations are as follows:(1)Aiming at the problem of less labeled data in remote sensing image datasets and poor segmentation effect of existing methods.The mainstream deep learning model is adopted as the generator,the full convolution discriminator is introduced,and the labeled data and unlabeled data are used for training.In the first stage,the use of labeled data for training allows the discriminator to learn the spatial distribution close to the labeled image.In the second stage,the entire network is trained using unlabeled data,and the discriminator assists the generator training according to the spatial distribution learned in the previous stage.(2)The current semi-supervised remote sensing image semantic segmentation method does not pay attention to the problem of long-distance correlation.By introducing crisscross attention,a multi-scale attention module is proposed to further improve the feature extraction capability of the generator,thereby capturing useful context information,further improving the accuracy of target edge segmentation,and helping to solve the problem of visual understanding.(3)Aiming at the problems of instability and difficulty in convergence of current generative adversarial network training.The discriminator is further improved,the stability of the network is improved through the dual discriminator,and the auxiliary training effect of the discriminator on the generator is further improved.One of the discriminators rewards the data generated by the generator,and the other discriminator gives negative feedback to the data generated by the generator,and uses the complementary statistical characteristics of the dual confrontation network to train the entire network,thereby improving the generated samples the quality of.The three-party game contributes to the stability of the network and the improvement of the overall accuracy of the network.Through experiments on public datasets,the multi-scale generative confrontation method can effectively improve the accuracy of semantic segmentation of remote sensing images with a small amount of labeled data.on this basis,multi-scale attention introduces an attention module to focus on contextual information,and further Improve the accuracy of target edge segmentation.Finally,in view of the difficulty of convergence of current semi-supervised methods,this thesis proposes an attention dual confrontation method.In the commonly used evaluation index MIOU,the performance of the method proposed in this thesis exceeds the current semantic segmentation model of remote sensing images,which proves the effectiveness of the method proposed in this thesis.This thesis has 43 figures,15 tables,and 72 references.