Generative Adversarial Networks Based Sonar Image Synthesis And Blind Denoising Methods

Algorithm research and analysis needs supports of collected data from real world,however,for some special data which is not open source or expensive to collect,such as sonar image data makes troubles to the process of experiment.Using image synthesis technology could help to partly solve this problem.From the perspective of deep learning,based on the great simulation power of Generative Adversarial Networks,this paper systematically studies the sonar image synthesis and the blind denoising with Generative Adversarial Networks.To select a more stable baseline,this paper uses salmon fish sonar images as target,tests those synthesized images of several networks with Frechet Inception Distance and finally chose WGAN-GP as the baseline.Meanwhile,this paper divides the generate images of WGAN-GP to high quality and low quality according to human vision,then trains a classifier with these divided data.The classifier has a good performance on test data and real samples shows that the classifier could optimize the generate images' quality.In addition,learns from the idea of modeling the noise with Generative Adversarial Networks to achieve optical image blind denoising,this paper achieves blind denoising of sonar images with good objective and subjective results without any sonar parameters.Then,this paper tests the method on data that collected from same location but with different quality through a same network.Meanwhile,according to the features of sonar images,this paper proposes a global pixels' gray value means based restriction for noise extraction,and validated this proposed restriction with experiments and analysis.