Research On The Style Based Countermeasure Network For Sea Surface Remote Sensing Target Image Generation

In recent years,deep learning-based remote sensing image object detection has been widely used in many scenarios and fields.In this thesis,taking marine remote sensing targets as an example,due to the high cost and unstable imaging quality of satellite remote sensing image data acquisition,there are very few training samples available for deep learning-based object detection.Therefore,to obtain a large quantity of high-quality,controllable style marine remote sensing target images more easily,this thesis proposes a Style-based marine remote sensing target image generation adversarial network.This method utilizes a generative adversarial network(GAN)and combines Style control mechanism to provide image samples for marine remote sensing target detection training.The main research contents are as follows:1.The progressive growing of GANs(PGGAN)model was studied and improved.Since the marine remote sensing target image generation network studied in this thesis is trained in a progressive manner,to improve the quality of the generated remote sensing images,the PGGAN network was studied and improved.Firstly,aiming at the problems of sample detail distortion and pattern collapse in this method when generating marine remote sensing target images,the loss function of the discriminator was redesigned,and a 1-Lipschitz condition-adaptive penalty loss function was proposed to improve the measurement accuracy of the distance distribution between the generated samples and the real samples by the discriminator.Secondly,to avoid problems such as gradient disappearance and overfitting,this thesis introduces the residual network "Resnet" structure into the generator and discriminator,allowing certain layers of the network to skip the connection of the next layer neurons,weakening the strong connection between each layer,and making the image features better transmitted between layers.2.A Style-based marine remote sensing target image generation adversarial network was designed.The main work is reflected in:(1)introducing an adaptive penalty loss function into the discriminator,which makes the generated image details more realistic;(2)to solve the problem that the ship targets in the generated results cannot be well distinguished from the coast and port,this thesis introduces self-attention mechanism into the network,which allows the network to ignore non-target features and learn the features of specific targets in the sample;(3)introducing adaptive data augmentation technology into the discriminator,which reduces the dependence of the network on the number of samples;(4)adding Style control mechanism to the generator to achieve style control of the target image and enhance the richness of the generated sample style.In summary,the Style-based marine remote sensing target image generation adversarial network proposed in this thesis is an effective data augmentation method.Firstly,the loss function was redesigned in the improved PGGAN model,and the generator and discriminator network structures were improved.In addition,this thesis proposes a Style-based marine remote sensing target image generation adversarial network,which introduces self-attention mechanism and adaptive data augmentation mechanism based on PGGAN,and finally adds style control synthesis mechanism.The experimental results show that this method is superior to existing algorithms in generating marine remote sensing target images and can be used to improve the training effect of marine remote sensing target detection.