Underwater Acoustic Image Enhancement by Using Fast Super-Resolution with Generative Adversarial Networks

Acoustic sensors play a fundamental role in underwater applications. They are used to perform a wide variety of tasks: from the perception of the surrounding environment to the support of inertial sensors in navigation strategies. The quality of the acquired images deeply affects the obtained results and, consequently, image enhancement approaches need to be developed and tested. Super-Resolution (SR) techniques are employed to reconstruct one high-resolution image by composing a sequence of low-resolution ones. By applying these strategies, the information content of an image can be considerably increased, but the required computational time is incompatible for real-time employment. Due to this limitation, an SR Generative Adversarial Network (SRGAN) approach has been developed in the presented work, where the SR images are used during the training phase of the GAN framework. The proposed approach, which has been developed for images provided by a Forward-Looking Sonar (FLS), can guarantee a solid trade-off between the quality of the generated high-resolution image and the run-time execution.