The demand for maritime space requires an integrated planning and management approach, which should be based on solid scientific knowledge and reliable mapping of the seabed. One of the widely used seabed mapping methods is underwater imagery. The main advantage of this method is its simplicity, enabling rapid collection of large amounts of data, and, hence, cost-effectiveness. However, only a small part of information available in underwater imagery archives is being extracted due to labor-intensive and time-consuming analysis procedures. Efforts to develop automated techniques for UW image processing are challenged by the specifics of the UW environment, such as heterogeneity of substrate, variation in lighting, color instability, etc. Emerging novel deep learning (DL) methods open opportunities for more effective, accurate and rapid analysis of seabed images than ever before. Our project aims to develop an automated seabed imagery recognition and quantification method based on the DL approach. The project consortium brings together specialists in signal, image and video processing, and marine benthic ecologists with long-term experience in UW research. We plan to develop a user-friendly system, flexible enough to use in a variety of marine environments. To test system’s capabilities, video material collected in the Arctic Ocean, Baltic Sea, Mediterranean Sea and other world regions will be used.
Project funding:
Projects funded by the Research Council of Lithuania (RCL), Projects carried out by researchers’ teams
Project results:
A unique, precisely annotated, database of seabed images, which will be of international value for the development and training of systems for seabed video analysis.
A novel, unsupervised learning-based, method allowing to exploit unlabelled (unannotated) seabed images for training DCNN, dedicated to analysis of seabed videos. This is very important, since annotation of large sets of underwater images is an extremely labor-intensive task.
A novel method and algorithms for diversifying models and combining them into a committee using Conditional Random Fields (CRF) or Dempster-Shafer theory. This is important aiming to achieve high classification accuracy, since high variations in illumination, camera orientation, water limpidity and other sensitive factors are characteristic to seabed images.
Period of project implementation: 2019-06-01 - 2022-06-30
Project coordinator: Kaunas University of Technology
Project partners: Klaipėda University
