Research On Sonar Image Denoising Method Based On GPU

With the increasing attention to marine resources and development,imaging sonar developed and the amount of data included increased dramatically.At the same time,due to the complexity of underwater environment,sonar images are seriously affected by many kinds of complex and serious noises.Excellent sonar image denoising algorithm is usually complex and time-consuming,which is difficult to meet the real-time requirements of applications.The GPU has a large number of computing units,and the performance in large-scale data computing,processing speed and other aspects greatly exceeds the CPU,and shows great advantages and potential.Using GPU to conduct sonar image denoising parallel acceleration research has good feasibility and engineering application value,which can greatly improve the operation efficiency and ensure the real-time performance.Based on the characteristics of slow water environment changes and relatively fixed noise sources,this paper chooses two waters of East Lake of University of Electronic Science and Technology of China and Heilongtan Reservoir of Meishan,Sichuan as test sites.According to the principle of marine environmental noise measurement,a test platform was built to carry out the acquisition test of the noise characteristics of the water area,and different Gaussian noise models were established based on the collected data.Then in COMSOL Multiphysics,the model is used as the background sound field input,and the echo data is obtained by simulation modeling.On this basis,the synthetic aperture imaging method is used to simulate the echo imaging,so as to obtain two echo image under the influence of the water area noise,which appears as obvious speckle noise and seriously interferes with the image quality.Then,the sonar image denoising algorithm is analyzed in detail,and three image denoising methods including guided filtering,K-SVD dictionary learning and non-local mean are selected as the research objects,and then the sonar image denoising effect and performance on the selected water area is studied.After that,carry out simulation experiments,combined with PSNR,FSIM and other evaluation indicators and the real-time targets of this paper to comprehensively evaluate the performance of the algorithm.Finally,it is aimed at different situations where the guided filtering operation speed is fast but the effect is poor,the K-SVD speed is slow but the effect is better,and the non-local mean efficiency is extremely poor but the effect is excellent.From the perspective of real-time performance in practical applications,the NVIDIA Ge Force GTX 1080 Ti is used as the platform and the parallel acceleration of the sonar image denoising algorithm is realized based on CUDA.By decomposing the computationally intensive and time-consuming basic operation modules in the algorithm,the two basic operations of matrix multiplication and matrix inversion are optimized in parallel,and then the parallel analysis and simulation of the three sonar image denoising algorithms are realized.Through the simulation experiment,it is found that the effect of several algorithms on CPU and GPU is the same,but there are differences only in operation efficiency.The acceleration effect of matrix multiplication and matrix inversion is related to the data scale.When the data scale is small,the acceleration effect is not obvious.In the scale of 2048*2048,the acceleration ratio can reach more than 300 or even more than 400 times.For the sonar image of 210*210,the guided filtering acceleration is relatively small,but the running time is still the best,and the K-SVD dictionary learning acceleration ratio is more than 10 times,which improves the running efficiency.In addition,the non-local mean acceleration ratio is more than 15 times,which greatly improves the running efficiency.