Research On UUV Side Scan Sonar Image Enhancement Method Based On Generative Adversarial Networks

High-quality sonar images are crucial for identifying targets in the ocean,so how to obtain high-quality sonar images has become a current research hotspot.The side-scan sonar image obtained during the movement of an Unmanned Underwater Vehicle(UUV)is prone to motion blur,and it is difficult to accurately obtain the position and specific shape of small targets through the side-scan sonar image.Although high-quality sonar images can be obtained through many traditional methods,there are problems such as large amount of calculation and narrow scope.The Generative Adversarial Networks(GAN)are simple to calculate,and the effect on improving image quality is better than most traditional methods.This paper will use the advantages of generative adversarial networks to propose two processing methods for removing motion blur and improving resolution of side-scan sonar images.This article conducts in-depth research from three aspects.First,according to the characteristics of the side-scan sonar data obtained by UUV,the generation process of the sonar image motion blur dataset and super-resolution dataset is designed.The first step is to convert the side scan sonar text data in JSF format to the side scan sonar image in BMP format by using Visual C ++ 6.0.In second step,according to the construction characteristics of side-scan sonar image and its noise model,high-quality side-scan sonar image in PNG format is obtained by adding noise and adjusting color.Obtain a high-quality side-scan sonar image in PNG format by other operations.The third step is to design a motion random generation method to generate a motion blur data set based on the inconsistency of the blurred point positions of different motion blurred images;In the fourth step,according to the features of pixel collapse of low-resolution images,a method of double thrice-down sampling and local pixelation fusion is designed to generate low-resolution data sets.Then,for the problem of motion blur in sonar images,an improved lightweight pyramid multi-scale feature fusion generative adversarial networks is proposed.In this paper,Group Normalization(GN)and Parametric Rectified Linear Unit(PRe LU)with parameters are introduced to improve the robustness of the network.In order to accelerate the network training speed,this paper adopts the inverted residual block as the network's down sampling method.Based on the design idea of ??the pyramid multi-scale feature fusion module,this article introduces the residual module to design the entire network framework to improve the network edge refinement ability.In this paper,content loss and Wasserstein distance loss areintroduced into the loss function to improve the stability of training.In this paper,the correctness of the designed motion fuzzy generative adversarial networks is verified by different comparative experiments.Finally,to solve the problem of low resolution of the sonar image,a dual discriminator generative adversarial networks with multi-scale sub-pixel upsampling is proposed.This paper introduces a sub-pixel convolution module,combined with a multi-scale convolution fusion method,and proposes a multi-scale sub-pixel convolution module to perform upsampling operations to achieve the purpose of efficiently restoring high resolution pixel information.It is proposed to use double discriminator to train the network,so that the network can have better generalization ability in the case of fewer samples.In this paper,the correctness of the designed super-resolution generative adversarial networks is verified by different comparative experiments.