Revolutionizing Subaquatic Perception

Abstract

Studying the underwater world was considered interesting and challenging. The light was bent, colours were lost in depth, and the visibility was obscured with sediments, and it became difficult to study marine life and structures. This image has changed with the recent advancement in artificial intelligence (AI) and computer vision. Smarter image processing, deep learning, and sensor fusion were employed and the underwater environments could now be perceived with much better clarity. These inventions gave rise to new opportunities: shipwrecks could be discovered, coral reefs could be tracked, the independence of underwater vehicles was enhanced, and conservation could be facilitated. But there were still difficulties. Based on the concept of the mantis shrimp, polarization-based systems were designed to improve clarity, and self-supervised learning was developed to allow AI to learn using raw images and not huge data volumes. The low-weight models were created in such a way that it was possible to perform real-time processing on small robots. Moreover, generative adversarial network (GAN)-based technology was also presented to reconstruct missing colours and details in an impressive real-life manner. A combination of these developments led scientists to study the oceans in a more promising way that had not been imagined. Artificial environments that imitated underwater conditions were used as a source of large amounts of training data, and Vision Transformers were used to push the boundaries of underwater object recognition and scene interpretation. This chapter gave an explanation of these transformative technologies in detail, along with how they worked, their position at that time, and their future direction. The most prominent methods were contrasted, the current drawbacks have been pointed out, and the way forward has been described to transform underwater computer vision into something smarter, more realistic, and more applicable to physical scenarios.