A Deep Learning based Submerged Body Classification Using Underwater Imaging Sonar

Recognizing a submerged body in turbid water is extremely challenging despite the clear necessity from a diver or a submersible. Above all, the water turbidity and limited light condition prohibit clean image quality. Even with a visible image, the conventional feature-based approaches would be limited due to the diverse form and intensive noise level in the target object image. To tackle this issue, we propose an automatic submerged body classification using the multibeam sonar widely applicable in underwater. To learn the sonar images, we adopted Convolutional Neural Network (CNN)-based models, AlexNet and GoogLeNet. Experimental validation was performed using two sets of underwater sonar image data from a submerged body of a dummy. Our training and testing include different level of water clarity and noise; clean water in an indoor pool and a turbid water in the ocean.