Zhang, D., & Piacentino, M. (2017, August). Advances in Automated Stock Assessment Based on Computer Vision Deep Learning Technologies. In 147th Annual Meeting of the American Fisheries Society. AFS.
Abstract
Today, there exists very large collections of underwater video collected for use to perform stock assessment, but extensive manual efforts are required to perform even limited fish classification and counting for achieving the required assessment data results. We present a rapid fish assessment method leveraging computer vision deep learning technologies to provide both (1) rapid fish annotation and (2) fish classification with fish counting.
During the development of our stock assessment tool we learned image based annotation data from previously collected stock assessment video is often very limited, therefore a tool that first provides annotation methods to train computer vision classification algorithms is greatly needed. With our newly developed semi-automated, analyst-guided annotation tool, rapid annotation methods now enable rapid stock assessment where today very limited annotations exist.
The stock assessment tool, trained from guided annotation data supports classifying, counting and logging each detected fish/object metadata. The tool has been integrated into the NOAA VIAME (Video and Image Analytics in a Marine Environment) open-source video processing framework for research analysts. The tool, with its analyst guided annotation capabilities, can support stock assessment needs for all fish types, environments, and underwater conditions allowing researchers to process new and existing videos.