Students use Deep Vision to automatically recognize fish species
French students at the Institute of Marine Research (IMR) in Bergen have spent their summer using images from Deep Vision to automatically analyze them for different species of fish.
The two French masters students from the University of Nice, Thomas Mahiout and Tiffanie Schreyeck, were employed to look at IMR’s extensive digital data collection. Their task was to suggest a method for automating the analysis of these data, which are still largely done manually by the Institute.
Amongst these data were images collected from the Deep Vision system. The underwater trawl camera collects stereo images of all fish passing through the trawl. Developed in close collaboration between Scantrol Deep Vision and IMR, the system has secured the Institute an extensive collection of images over the past decade.
Machine learning
Over the summer, the students developed and trialled a prototype for automatically recognizing the species of a selection of fish from Deep Vision by use of machine learning. Images from a recent cruise with IMR were used to train the software to identify three different species – herring, mackerel and blue whiting.
Using 150 real images as a starting point, the students edited these to a variety of images to represent the fish from all different angles. In the end, the software was trained for a total of 7000 images containing the three different species.
Improved estimates and sustainability
– The job they have done is very important to us, says project leader at IMR, Ketil Malde:
– Being able to automatically classify fish gives us the opportunity to improve estimates based on acoustic data. With a trawl camera, we no longer need to collect the fish for sampling, and the damage on the fish is reduced.
Innovation Norway supports the development of Deep Vision for use on IMR’s research vessels. The system is planned to be employed as a standard research tool at the Institute by 2017.
PHOTO: French students Thomas Mahiout and Tiffanie Schreyeck have spent their summer developing a method for automatically recognizing fish species from Deep Vision images. (Erlend Astad Lorentzen/Institute of Marine Research)
Related: