Enhancing Side-Scan Sonar Imaging: A Comparative Study of Deep Learning Super-Resolution Techniques

In naval operations, detecting nearby underwater objects with submarines poses a significant risk due to factors like magnetic anomaly detectors and lurking submarines. Furthermore, side-scan sonar imaging loses resolution and accuracy at extended distances. This paper presents a deep learning approach to enhance the resolution of side-scan sonar images. We applied state-of-the-art super-resolution techniques including RCAN, ESRGAN, and SwinIR, each based on distinct principles like GAN, residual channel attention, and Swin transformer, to the sonar data. Evaluations using PSNR and SSIM metrics revealed RCAN's superiority in PSNR, ESRGAN's dominance in SSIM, and SwinIR's ability to retain sharp edges and intricate details. Upon assessing model parameters and computational speed, we found that ESRGAN and SwinIR strike a balance between performance and efficiency, making them suitable for real-world applications.