Super-Resolution of Underwater Sonar Image Based on Generative Adversarial Network

As an indispensable sensor for obtaining ocean resources, sonar can provide rich underwater observation information. However, due to the limitations of sonar equipment, the resolution of underwater sonar images collected is always low, resulting in some inexplicit underwater targets. To obtain high-resolution underwater information, image super-resolution reconstruction, as an efficient method, can significantly reduce hardware requirements. Super-resolution refers to the technology of restoring image details based on known image information. To achieve high-quality super-resolution for underwater sonar images, we propose a ConvGAN image super-resolution reconstruction model. We design the generator model with the improved ConvNeXt Block which enables the generator to generate rich details. At the same time, we use a VGG-Style discriminator and make its output a patch score map, which is used to guide the generator to recover complex texture details for sonar images. In addition, we compare with other state-of-the-art super-resolution networks on two different datasets. Quantitative and visual comparisons show the effectiveness of our proposed model.