Building Classification In Remote Sensing Image Based On Generative Adversarial Network And Self-attention

Extraction of buildings from high-resolution optical images is a fundamental application in remote sensing,and the rich texture features brought by its high resolution provide a basis for distinguishing buildings from other feature classes.However,the cost of acquiring highresolution optical images remains high,and the labor cost of manually extracting building patches is also increasing,so an efficient and low-cost remote sensing building extraction method is particularly important nowadays.In the semantic segmentation,models also need a large number of data to support training.If the dataset is small,the model is prone to overfit or be poor performance.In the generation of optical images,if the number of samples is very sparse,the images generated by the model would not be able to achieve the expectation.How to use machine learning to generate sufficiently realistic virtual images and how to suppress overfitting while training with extremely sparse samples are the focus of this paper.For the above two problems,two different networks are proposed,the main contents are as follows:(1)Style GAN2,which is used to generate images of buildings in remote sensing images to augment the dataset for semantic segmentation model.In the production of the dataset,due to the sparse data,the following points are required to be paid attention: the target building in the dataset needs to be in the center of the image or the same position of the image as much as possible;the features around the target building are similar;the number of buildings in an image is only one,and the corner points of each building are included;the orientation of the building should be normalized,and there should be no obstructions above the buildings.In the training,an efficient self-attention module,a top-k training mechanism and a differentiable data augmentation method are added to the network to achieve better results.Finally,after using the three improved methods simultaneously,the FID value of the generated images reaches 31.829 and generates sufficiently realistic virtual remote sensing building images.(2)Swin Transformer,an image semantic segmentation network,can extract the building regions in remote sensing images.A large number of self-attentive modules and a light network structure are used in the network,and with the help of the virtual building image samples generated by the generative model,a better extraction effect is achieved with a faster training speed.Finally,the m Io U value of extracted building images reaches up to 88.39%,and the m F1 value and overall accuracy reach up to 90.35% and 94.59%,which effectively solves the model training problem when the training samples are sparse and further reduces the difficulty of training a qualified building extraction model for remote sensing images.