Underwater computer vision

"In recent years, Deep Learning has been applied pervasively in many fields, leading to improvements in several tasks, from object recognition to speech synthesis. Lately, it is also approaching marine exploration, which is more and more relevant not only from a scientific point of view, but also for the industry. Underwater Computer Vision is a particularly important subfield. For example, for the study of marine species, sustainable fishing, monitoring of archaeological sites, and autonomous driving of unmanned underwater vehicles. Nevertheless, it is hindered by the peculiar characteristics of underwater images, since the propagation of light through water degrades them. Another aspect that makes it very challenging is data scarcity: while several datasets are being produced and published for fish classification or detection, for underwater image enhancement, for coral segmentation, and a few other tasks, it is very hard to find labeled images for many other tasks, such as fish weight estimation or smart fish feeding management. On the other hand, in some cases, a paired dataset can be obtained from any set of underwater images with some well-reasoned preprocessing. This is the case for Image Super-Resolution (ISR), the first problem explored in this thesis, as a tool for improving underwater image transmission. The picture is taken on an underwater edge device, which applies compression and downscaling, then the image can be transmitted on dry land or on the cloud, even with constrained bandwidth requirements, and finally brought back to the original quality via super-resolution. A state-of-the-art model for ISR in the wild was studied and adapted to the underwater domain. Then, the focus is shifted to Underwater Object Counting for the design of a smart system for efficient fish feeding: when fish are fed with an amount of food pellets that exceeds their needs, it falls towards the seabed, and can be counted to determine when to stop providing food, thus reducing wastes and sea pollution. Since object counting is a supervised task, a considerable amount of labeled images is generally needed. Moreover, being the application field very specific and under-researched, common approaches for few-shot or zero-shot object counting did not provide satisfactory results. Therefore, a dataset was collected in a fish farm in Roses, in collaboration with the University of Girona, and labeled with a semi-automatic approach. The new dataset finally allowed to train a dedicated model."