Small-target and diversity oriented underwater sonar image augmentation

Abstract Underwater sonar images are crucial in areas like oceanographic research for mapping the seabed and detecting resources, and in marine biology for understanding habitats. They are also important for naval and military uses such as navigation and surveillance. However, due to equipment and environmental limitations, the number of image samples is restricted, impeding further data-driven AI research. Although some works have explored data augmentation of underwater sonar images, they still face the following two problems: 1) inability to generate small-target images; 2) limited diversity of generated images. Toward this end, in this paper we propose a small-target and diversity oriented underwater sonar image augmentation method. Specifically, for small-target images, we propose to first detect and extract the target objects in the seabed sonar images, then perform scale scaling, and fuse them onto the background image using the Poisson fusion algorithm; for diverse images, we ingeniously combine mainstream image generation methods, including GAN, VAE, and Diffusion Models, using the diversity of the generative models to ensure the diversity of the generated images. Meanwhile, we design a Mixture-of-Experts (MoE) enhanced discriminator in GAN to screen the images generated by the three generative models to ensure the quality of the final augmented images. Experimental results prove that our method can effectively increase the proportion of small-target images and ensure the diversity of the augmented images, which further boost related researches based on underwater sonar images.