Research On Acoustic Image Super-Resolution Based On Generative Adversarial Network

In the field of underwater wireless communication,the acoustic signals have the advantages of little attenuation,long propagation distance and high fidelity comparing with electromagnetic wave signals.Therefore,the acoustic wave has been widely used for submarine topography,underwater search-and-rescue and ship navigation.In this paper,the acoustic signals mainly involved in side-scanning sonar images.The contrast of the sonar image is very poor due to the complexity of the underwater environment and the characteristic of sonar imaging.The low resolution and the blurred details make it difficult to analysize and recognize underwater targets in the sonar images.Improving the quality of sonar images,such as improving resolution,enhancing contrast,and removing noise,can lay the foundation for subsequent work such as image analysis and recognition.It is also a necessary condition for improving the efficiency of sonar image processing.Therefore,for the low-resolution sonar images,the super-resolution technology can improve the visual effect of images and reduce the difficulty of sonar image processing.The super-resolution technology refers to the reconstruction of one or more low-resolution images into the high-resolution images through some techniques.Comparing with the traditional super-resolution algorithm,the deep learning algorithm has achieved better image reconstruction effects.The paper researches the Generative Adversarial Network,which is the state-of-the-art super-resolution method in the field of deep learning.Based on the Conditional Generative Adversarial Network(CGAN),the paper designs a Super-Resolution Conditional Generative Adversarial Network(SR-CGAN),which is used as the output of the conditional control network directed against the characteristics of low brightness and poor contrast of sonar images,so as to get the sonar images with enhanced brightness and contrast.The paper also proposes a Deeper Super-Resolution Generative Adversarial Network with Gradient Penalty(DGP-SRGAN),which is employed for super-resolution experiments of sonar images.Experiments show the proposed method has higher Mean Opinion Score(MOS)than traditional methods.Compared to the SRGAN algorithm,it has more realistic details and textures.We make following optimizations based on SRGAN: for the purpose of weakening the checkerboard effect,the generator networkfinally uses the proposed residual nearest neighbor resize convolution to perform up sampling of four times amplification to weaken the checkboard artifacts;the number of residual blocks in the generator network is increased to deepen the network level and improve the network's ability of feature extraction and expression.The number of residual blocks has increased from16 to 32,and the network depth has doubled.The discriminator network loss function increases a gradient penalty term for faster convergence.In the paper,the experimental results are verified on the side-scanning sonar image dataset.Six classical image super-resolution algorithms are compared with the proposed SR-CGAN algorithm and the improved DGP-SRGAN algorithm.The experimental results show that the SR-CGAN network designed in this paper can control the output of super-resolution images well according to the input conditions.Compared with other classical algorithms,DGP-SRGAN algorithm plays a great role in reconstructing the high-frequency details and rich texture of sonar images,and the super-resolution reconstruction quality of sonar images has been improved.