Cycle-GAN-based synthetic sonar image generation for improved underwater classification

One of the main challenges in underwater automatic target recognition is in the data scarcity of underwater sonar imagery. This challenge arises especially in data-driven approaches because of the limited training dataset and unknown environmental conditions before the mission. Transfer learning and synthetic data generation have been suggested as effective methods to overcome this challenge. However, the efficiency and effectiveness of synthetic data generation methods have been limited due to the difficulty from implementing complex acoustic imaging processes and data-driven model's poor performance under domain shifts. In this paper, we present a novel approach to address this challenge by utilizing cycle-Generative Adversarial Networks (GAN) to generate synthetic sonar images to enhance the effectiveness of the training data set. Our method simplifies the process of synthetic data generation by leveraging cycle-GAN, which is a deep Convolutional Neural Network (CNN) for image-to-image translation using unpaired dataset. The cycle-GAN based generation model transfers camera images of ship and plane into realistic synthetic sonar images. Then, these generated synthetic images are used to augment the training data set for the classification model. In this work, the effectiveness of this approach is demonstrated through a series of experiments, showing improvements in classification accuracy. One advantage of the proposed approach is in the simplification of the synthetic data generation process while improving classification accuracy. Another advantage is that the ship and plane sonar image generation model is trained solely on seabed sonar images, which are relatively easy to obtain. This approach has the potential to greatly benefit the field of underwater sonar image classification by providing a more efficient solution for addressing data scarcity.