A GAN-based Super Resolution Model for Efficient Image Enhancement in Underwater Sonar Images

Acoustic imaging systems dominate in underwater imaging due to their unique ability to illuminate objects on the seabed, even in dark or turbid water conditions. These systems mounted on an autonomous underwater vehicle (AUV) are being used for a variety of civilian and military applications. Mine detection and classification is a predominant application. The raw images captured using these systems are usually noisy and poor in their resolution. Consequently, methods to enhance sonar images are necessary to aid further processing and classification of these acquired scenes. Inspired by the developments in the field of deep learning in different areas of computer vision, this study explores efficient deep neural networks for acoustic image super resolution. The study is performed on a custom-made sonar image dataset to handle the deficiency of public datasets in the domain. We employ a Generative Adversarial Network (GAN) deep learning model i.e. pre-trained ESRGAN and make use of transfer learning to achieve our goal with limited data samples. We use the model published by the original authors, Xintao Wang et al and experiment with our proposed method in three ways. a) Direct use of pre-trained model b) Fine-tuning the model with VGG-19 feature extractors at the discriminator and c) Finetuning the model with ResNet-34 feature extractors at the discriminator. The super resolved images are validated through image quality assessment metrics like PSNR, SSIM, and Perceptual index.