Deep Learning for Enhanced Marine Vision: Object Detection in Underwater Environments

Abstract

This study leverages the Semantic Segmentation of Underwater Imagery (SUIM) dataset, encompassing over 1,500 meticulously annotated images that delineate eight distinct object categories. These categories encompass a diverse array, ranging from vertebrate fish and invertebrate reefs to aquatic vegetation, wreckage, human divers, robots, and the seafloor. The use of this dataset involves a methodical synthesis of data through extensive oceanic expeditions and collaborative experiments, featuring both human participants and robots. The research extends its scope to evaluating cutting-edge semantic segmentation techniques, employing established metrics to gauge their performance comprehensively. Additionally, we introduce a fully convolutional encoder-decoder model designed with a dual purpose: delivering competitive performance and computational efficiency. Notably, this model boasts a remarkable accuracy of 88%, underscoring its proficiency in underwater image segmentation. Furthermore, this model's integration within the autonomy pipeline of visually-guided underwater robots presents its tangible applicability. Its rapid end-to-end inference capability addresses the exigencies of real-time decision-making, vital for autonomous systems. This study elucidates the model's practical benefits across diverse applications like visual serving, saliency prediction, and intricate scene comprehension. Crucially, the utilization of the Enhanced Super-Resolution Generative Adversarial Network (ESRGAN) elevates image quality, enriching the foundation upon which our model's success rests. This research establishes a solid groundwork for future exploration in underwater robot vision by presenting the model and the benchmark dataset.