Research Semantic Segmentation Of Remote Sensing Images Based On Attention Mechanism Of Generative Adversarial Networks

With the technological development and application of satellites and UAVs,the clarity of remote sensing images is getting higher and higher,and the photographic images of terrain and landscape are becoming clearer and clearer.However,the classification of scenes in remote sensing images by manual determination is timeconsuming and laborious.In recent years,classification algorithms of deep learning have been applied in the field of image segmentation,and have obtained good classification accuracy.However,satisfactory classification results cannot be achieved for remote sensing images involving complex spatial information.Due to the huge information redundancy within the remote sensing images,the spectra are not related to each other,and the terrain classification also has the cases of same object and different spectrum and same spectrum and different object.In this paper,we propose a new method to process remote sensing images,which utilizes attention mechanism and generative adversarial network to achieve semantic segmentation.Two methods,channel attention and spatial attention,are also introduced to enhance the target features while avoiding redundant information and improving the operation speed and recognition efficiency of the network.However,the learning ability is not strong.By understanding the training principles of generative adversarial networks,the generation and discriminative approach of generative adversarial networks is incorporated into the model of semantic segmentation,and the semantic segmentation model with attention mechanism is treated as a generator and a convolutional network with two image inputs is designed as a discriminator.The discriminator performance is trained,and the discriminator modifies the weight parameters of the semantic segmentation network to further enhance the feature extraction ability of the model for remote sensing images,thus improving the topographic segmentation accuracy of remote sensing images.By training and testing the model on the publicly available remote sensing image dataset WHDLD,the semantic segmentation networks Unet,Segnet of typical deep learning are fused with attention mechanism,and the accuracy is improved by 0.19%and 0.51% each,which improves the learning efficiency of the model.Based on the generative adversarial network,a new semantic decomposition-based network was designed.The model was made to undergo secondary training,and the accuracy was further improved after the Unet network model with better feature recognition efficiency of remote sensing images was added to the new training mechanism.Through the validation and testing of the dataset,it can be concluded that the semantic segmentation network incorporating the attention mechanism of the generative adversarial network can be very helpful for scene classification of remote sensing images.This research result can provide technical support for remote sensing data to better serve the fields of crop detection,land resource monitoring,and natural disaster monitoring in rural China.