Research On Sonar Image Enhancement Method Based On Generative Adversarial Network

In military and civil fields,underwater imaging technology is an important means of underwater exploration.The scope and distance of optical imaging are limited,especially in muddy water,the transmission of optical information is seriously interfered.Acoustic signal has a long distance and good penetration,can resist certain interference-However,the underwater environment is complex and changeable,and affected by a series of factors such as noise,sonar image generally has serious noise and low resolution,which has a great impact on the follow-up research work of sonar image.In order to improve the processing efficiency of sonar image,noise removal,resolution improvement and contrast enhancement can lay a good foundation for the follow-up research.Traditional image enhancement methods have some problems such as uncleanness of noise removal and lack of detail information.Generating antagonistic network can achieve more prominent enhancement effect.The main research contents of this paper are as follows:(1)In sonar image denoising research,the choice of DnCNN model applied to the sonar image denoising,the network model,will generate network as denoising,the residual network structure,network and skip the connection through the neural network to pick up the noise in the image noise at the same time learning image detail characteristics,in view of the sonar image sample to study the noise image with noise image connection and difference,to ensure that the network training speed and restore more sonar image details.SSIM and MSE loss functions are added on the basis of the original loss function,and the sonar image after denoising is made clearer and more real through confrontation training.(2)In the sonar image super-resolution study,on the basis of the original SRGAN network,the paper on the network structure and the loss function improvement and optimization of concrete improvement methods are as follows:will be generated in the network residual replace ordinary convolution layer block structure of empty convolution layer,delete the batch standardization,reduce the consumption of resources and expand the receptive field,thus improve the efficiency of network training.In order to solve the problem of model collapse,a gradient penalty term is added to the discriminant network loss function to increase the stability of training.In this paper,two aspects of sonar image enhancement technology are studied based on the denoising and reconstruction of sonar image in generated adversation network.In sonar image denoising work,the visual effect after noise removal is better than other methods,and higher objective evaluation index is obtained.Through the training and testing of different intensity and type of noise,it is verified that the denoising network model has good generalization.In the sonar image super resolution work,four classical image super resolution algorithms are compared with the improved SRGAN algorithm.The experimental results show that the reconstructed sonar image quality of the improved SRGAN network is better than that of other super resolution methods,which can recover more sonar image texture information and have good visual perception effect.