Physics driven deep learning based simulation of side scan sonar images

We propose a novel method that leverages the combination of Blender and K-Wave software to generate a realistic acoustic dataset for side-scan sonar images. Side scan sonar (SSS) images have numerous applications in the underwater environment. However, the challenges and high cost of acquiring SSS images in real experiments and the lack of benchmark open-source datasets encourage the researchers to consider the generation of simulated SSS image datasets. There have been several attempts in the literature to combine open-source computer graphics design software such as Blender with generative adversarial networks (GANs) to generate synthetic SSS images. However, the speckles generated in all such approaches have been unrealistic; hence, the DL approaches trained for object detection employing such data are bound to fail. In the proposed method, we first generate acoustic images from Blender. Then, the CycleGAN is applied for style transfer to generate realistic acoustic images based on the speckle generated using k-Wave. The K-wave, Blender, and cycleGAN are used to generate a closed set of SSS images. We then leverage ConSinGAN to generate varying resolution high-quality sonar images from a closed set of SSS images. As per our knowledge, this is the first attempt that uses k-wave in conjunction with Blender and CycleGAN to generate SSS images. The proposed method is cost-effective and efficient alternative for acquiring large SSS datasets.