Research On Remote Sensing Sample Expansion Technology Based On Generative Adversarial Network

Since entering the information age,related technologies based on deep learning algorithms have developed rapidly,and have achieved fruitful results in various fields of artificial intelligence.At the same time,with the rapid development of remote sensing technology,people have been able to obtain remote sensing images with different spatiotemporal and spectral resolutions on different remote sensing platforms.The rich feature information of remote sensing images promotes the continuous development of remote sensing image object detection technology.However,in specific application scenarios,it is difficult to obtain a sufficient number of public data sets,and the remote sensing image object detection algorithm is difficult to train to achieve the expected effect,which further limits the application of deep learning object detection algorithms in the field of remote sensing images.Therefore,it is imperative to study the sample expansion technology of remote sensing images.Based on the generative adversarial network,this paper studies the remote sensing image sample expansion technology,and mainly does the following work:This paper first uses VAE and WGAN models to generate remote sensing images,and compares and analyzes the generated results of the two models,analyzes the shortcomings of the images generated by VAE and WGAN,and then improves the model according to the existing problems.This paper conducts research on the basis of a small sample data set.In order to prevent the discriminator of the model from overfitting,a discriminator data enhancement strategy is introduced to train the discriminator.Data augmentation.This paper proposes a WGAN model based on variational autoencoder,which combines variational autoencoder and WGAN network,so that the generated image is closer to the original sample distribution,the image features are clearer,and the samples are more diverse.In order to solve the problem of low resolution of generated images,a residual block structure is introduced on the basis of the above improved model,which deepens the depth of the network and improves the resolution of the generated remote sensing images.The generated remote sensing images are evaluated by IS,FID and Tenengrad gradient function indicators.The experimental results show that the images generated by the improved model are better.In order to verify the effectiveness of the generated remote sensing images for sample expansion,this paper designs a comparative experiment,using the object detection model Faster R-CNN to compare the effects of different proportions of sample expansion on the precision and recall rate of object detection.The experimental results show that the expanded sample set can effectively improve the performance of object detection.