It may make you smile, but the video recognition of a fish by computer is not, however, obvious. On an underwater image, the luminosity varies, the water can be cloudy, and the fish hide behind algae or rocks…
Scientists have been working on the subject for 10 years, and work published this month by researchers from the National Center for Scientific Research (CNRS) of the National School of Engineers of Brest, shows that by training artificial intelligence systems to recognize fish both on their physical characteristics (shape and color) and on their movement characteristics in the water. It is indeed possible to teach a computer to identify them and classify them by species or family of species. These researchers worked on the basis of 700,000 underwater videos showing 3,000 species of fish. They trained artificial neurons to effectively identify around 20 of them.
The objective is to develop a video counting system for species to more effectively monitor the state of populations, for example in protected areas, or, on the contrary, in areas threatened by human activities. Moreover, at this time, theFrench Research Institute for the Exploitation of the Sea (Ifremer) is testing an intelligent fishing net which also uses this automatic recognition of fish species. This clever net catches only the fish you are sure to keep.
It makes it possible to limit bycatch because it is equipped with an underwater camera, explains Abdelbadie Belmouhcine, researcher at Ifremer in Lorient. This is coupled with an image analysis system which has been trained to recognize around ten species: in particular mackerel, sardines, anchovies and horse mackerel.
When the trawl crosses a shoal of fish that is not the species sought by fishermen, a net opening system is triggered so that these undesirable fish are not captured. This makes it possible to sort underwater, and not on the deck of the boat: according to figures from the World Food Organization, 10% of the fish caught are then thrown back into the sea, not always alive. Ifremer hopes to market a first version of this smart fishing net in 2025.