• Sort fishes on a boat before they enter the filleting machine at 360 ppm.
• Accept, Reject or Recycle based on the species, the fish conditions and its orientation in the metal pocket

The challenge

• Typical case of non-linearity where it is impossible to model mathematically all the possible variations of the fishes.
• Fish scales are highly reflective
• The size and scales of a same species of fishes can vary depending on the season and the location of the catch.
• The system must be compact, waterproof and require zero maintenance while running 24 hrs/day during sea expeditions
• The interface must be very simple to operate: a “boxing-glove” type keypad which the fishermen can use to setup and teach the camera themselves

• View movie
• Print application note
• 8 boats equipped by Pisces Industries with configurations of 3, 4 and 8 cameras per boat
• The Matrox Iris cameras are loaded with a CogniSight run-time engine, trained to classify herrings and programmed to activate reject brushes according to what they recognize
• Training was done on approximately 3000 images, accumulated during several expeditions throughout the years.
• Final recognition engine of approximately 100 neurons and accuracy of  98%