Automated Marking of Underwater Animals Using a Cascade of Neural Networks

In this work, a multifactorial problem of analyzing the seabed state of plants and animals using photo and video materials is considered. Marine research to monitor benthic communities and automatic mapping of underwater landscapes make it possible to qualitatively assess the state of biomes. The task includes several components: preparation of a methodology for data analysis, their aggregation, analysis, presentation of results. In this work, we focused on methods for automating detection and data presentation. For deep-sea research, which involves the detection, counting and segmentation of plants and animals, it is difficult to use traditional computer vision techniques. Thanks to modern automated monitoring technologies, the speed and quality of research can be increased several times while reducing the required human resources using machine learning and interactive visualization methods. The proposed approach significantly improves the quality of the segmentation of objects underwater. The algorithm includes three main stages: correction of image distortions underwater, image segmentation, selection of individual objects. Combining neural networks that successfully solve each of the tasks separately into a cascade of neural networks is the optimal method for solving the problem of segmentation of aquaculture and animals. Using the results obtained, it is possible to facilitate the control of the ecological state in the world, to automate the task of monitoring underwater populations.