Deep learning for underwater scene classification
Marine data exhibit significant within-class variations, complex between-class boundaries and inconsistent image clarity. This makes underwater scene classification a challenging task. In my research, I investigate the applications of generic Convolutional Neural Networks (CNN) based feature representations for diverse marine image classification tasks e.g., automatic recognition of coral reefs and localization of lobsters in complex underwater scenes. I will devise techniques for the automatic localization of lobster species in complex underwater scenes. Coral reefs are vital to marine ecosystems and lobster is one of the most important commercial marine species, and hence, their automatic classification and detection will lay the groundwork for a better understanding of their growth and habitat behaviours. Sea-grass and kelps provide shelter to various other marine species and assist their growth. I will also propose to investigate the coverage area and density analysis of sea-grasses and kelps using machine learning techniques. Finally, I propose to combine these results and formulate a relationship between species and their habitats and provide experimental evidence to the pre-existing hypotheses by marine scientists.
Deep sea exploration and imaging have given us a great opportunity to look into the vast and complex marine ecosystems. Data acquisition from the sea beds is vital for the scientific understanding of these intricate ecosystems. Widespread monitoring of large areas, remote sensing and tracking of marine species and their habitats, and the automatic annotation of the collected marine data are important research issues that need to be explored in details.
With the increase in global warming, urbanization, human population, large use of sea for shipping, exploration for oil and gas, recreational uses such as boating and industrial trade and activities, there has been a huge impact on the sea, both positive and negative. In order to minimize the negative impact of these activities on the sea, the marine ecosystems need to be monitored regularly. That is where underwater optimal imaging comes to rescue. With the development of optical imaging, standard protocols can be developed for analysing and curtailing the negative impacts for seawater environmental sustainability. Additionally, an exponential increase in the use of digital cameras and videos, intimates the need for automated analysis of such data as well.
In general, it is quite difficult to automate the human tasks in marine species survey and investigation. In this context, a lot of the work has been done in the real-world, above-sea environment, but very little research has been dedicated to underwater world. Therefore, the proposed research focuses on finding a feasible solution that can be considered as state-of-art in the marine domain.