Research And Applications Of Image Separation Based On Generative Adversarial Networks

Generative adversarial networks(GAN)have great potential in learning highdimensional and complex real data distribution,so they have been widely paid attention to in various fields of machine learning.Because it does not rely on any assumptions about distribution,it is easy to generate a sample with real properties directly from a potential space.This powerful performance makes generative adversarial network widely used in image synthesis,image attribute editing,image translation and many other frontier academic research fields.Generating models can be used as a special solution to solve various problems in the field of machine learning technology,such as the lack of labeled data,and some tasks in other fields of technology can be handled by providing a small amount of known information.At present,generative adversarial networks mainly deal with the related problems in the image domain by the way of generating ideas and discriminating ideas competing with each other.This paper takes the development of generative adversation network as the starting point,takes advantage of the variability and compatibility of antagonistic network structure,innovates generative adversation network structure and applies it to the field of blind image separation,aiming to solve the problem of separation with generation idea,which is a useful exploration in the field of generative adversation network development and blind source separation.Therefore,this paper mainly studies the application of generative adversation network in single-channel blind image separation,analyzes the cloud covering problem of high-dimensional remote sensing images from the perspective of source mixing,and migrates generative adversation network to cascade generative adversation network for cloud removal,etc.The main research work is summarized as follows:(1)For extreme underdetermined conditions and the traditional single channel blind source separation methods is difficult to overcome the independent source signals,nongaussian distribution and so on many constraints,and lack of prior knowledge,in the generated against model under the network architecture,this paper proposes a single channel blind image separation method based on attention mechanism,will be ordered based on manifold visual attention mechanism embedded in the separation of network,the key information to enhance the target,to separate the mixed images generated ideas iteration.Experimental results show that the generated admissive network with fused attention mechanism can only learn from single source data without multiple prior constraints,and achieves higher separation accuracy than the classical blind source separation method.Compared with the neural network separation method with known distribution,the proposed method is more effective in separating single channel mixed images.On the basis of this work,the network structure and attention mechanism are improved and another separable network model is proposed.In this model,self-attention mechanism is introduced to solve the problem of image detail blurring and to preserve image details in the process of image separation.Compared with the existing single-channel blind source separation algorithm based on generated adversation network,the new algorithm has more detailed information,and more source signals are separated in the mixed image,and has better separation performance than the traditional blind source separation algorithm.(2)As the image training pair defogging algorithm is difficult to deal with the problems of insufficient training sample pairs and model generalization in remote sensing images,the author believes that the fogged image can be regarded as a mixed source image,and the defogging process can be regarded as a process of extracting ground object source.Based on the above research ideas,a method of defogging remote sensing images based on cascading generated Adversarial Networks was proposed.Unpaired remote sensing images and foggy remote sensing images were used to guide Pagan(Pix Attention Generative Adversarial Networks)for correct defogging.In order to minimize the difference between the remote sensing image with fog generated by the model and the remote sensing image after fog removal,a self-attention mechanism is added to Pagan.The generator generates high resolution detail features,and the discriminant checks the difference between the advanced features in the remote sensing image.Compared with the classical method and neural network defogging method,the cascading generation admittedly network method does not need to train the network repeatedly with a large number of paired data.The experimental results show that the method can effectively remove fog and thin clouds,and is superior to the contrast method in visual effect and quantitative index.