Weakly supervised learning and efficient deep models deployed on an ROV for the detection and the classification of underwater objects – ROV-Chasseur

With 18,000 km long coastlines and more than 11 millions km^2 of the exclusive economic zone (EEZ), the French maritime environment has a significant impact not only on the social and economic side but also on the side of defense. The second world EEZ cannot be ignored, but our means of patrol and surveillance remain limited. As part of this project, we are interested in developing an intelligent ROV-machine to efficiently detect and classify specific underwater objects. He is particularly interested in two concrete applications in defense and in civilian clothes: the detection / classification of underwater mines (and objects qualified as ``potentially dangerous'') and the detection/classification of fish and animal species. These make it possible to contribute to economic development and surveillance, protection and defense by concentrating on certain activities linked to the sea. First of all, this project makes it possible to protect and evaluate fishery resources by relying on the detection and classification of fish to decide the optimal and legal areas to catch fish according to their frequency, size and species, etc. It also contributes to the protection and management of natural marine and coastal environments (monitoring of corals and protected species). In addition, this project promotes the guarantee of maritime navigation routes, as well as the protection of naval bases, ports, ships, against submarine mines which have a significant impact in the defense context. Its second application concerning the detection and classification of underwater mines improves the underwater demining process which is traditionally very expensive and time consuming.

On the scientific side, this project is interested in technological challenges in the development and deployment of deep models on the platform of an ROV working in an underwater environment. The objective of this project is to bring innovations in weakly supervised learning and in the design of effective deep models. These make it possible to overcome challenges in designing an intelligent ROV for reconnaissance tasks in an underwater environment. First, the weakly or auto-supervised learning approach is addressed to deal with the lack of underwater annotated data. Second, the design of efficient deep models is carried out by relying on the compression of deep neural networks allowing them to be deployed to the platform of an ROV having limited computation and memory resources. New databases for the detection and classification of specific underwater objects (mines, fish) are also being built within the framework of this project to facilitate the learning of deep models as well as the evaluation of various existing methods.