GAN-based Sea Surface Image Generation And Effect Evaluation

In recent years,with the continuous development and progress of satellite remote sensing technology and deep learning technology,optical remote sensing image target detection based on deep learning has been widely used in many fields.When it is necessary to search and identify sea targets,the acquisition cost of satellite remote sensing images is so high that the samples of remote sensing sea target images are scarce,which is insufficient to provide sufficient sample images for learning and training.Therefore,in order to obtain more sea surface target image samples,this paper focuses on the generation method of remote sensing sea surface target images based on the generative confrontation network(GAN for short)(sea surface targets take ships as an example).The main research contents are as follows:1.The deep convolution generated adversarial network is studied and improved(Referred to as DCGAN).Aiming at the problems of blurred image samples and mode collapse,the original deep convolution generating adversarial network based on the Wasserstein distance in WGAN has been improved in terms of loss function.And by introducing the residual network Resnet,the generator structure is improved,and the image quality of the remote sensing sea surface target is improved.2.Under the condition of few samples,a generative adversarial network based on conditional(Referred to as conditional GAN)background is proposed.Since the deep convolutional generative adversarial network relies on a large number of training samples and too few remote sensing sea target image samples,it is difficult to generate high-quality samples required for target detection.The conditional generative adversarial network proposed in this paper can generate high-resolution,high-quality remote sensing ship images with a small number of samples.The main work is reflected in:(1)The attentional mechanism of machine translation and computer vision is introduced into the adversarial antagonistic network to learn the characteristics of specific targets in the samples,ignore the characteristics of non-targets,and reduce the dependence on the number of samples.(2)The class of generated ships is controlled by a color value of conditional mask.(3)In order to restore the sea surface background as much as possible when generating ships,the u-net network is used as the generator.(4)L1regularization loss is added to the loss terms of generative adversarial network to make the details of the generated image more realistic.(5)The subjective and objective evaluation methods are used to verify the image quality of the conditional generative adversarial network proposed in this paper.In summary,by improving DCGAN,this paper makes the generated remote sensing sea surface target sample image clearer and reduces the occurrence of mode collapse.The proposed generative adversarial network based on conditional background can be used to generate a large number of high-quality remote sensing sea surface target images,which solves the problem that it is difficult to support sample training with too few remote sensing image samples.