Abstract
The component parts of an image analysis based system for estimating the mass of swimming fish have been developed and tested under a limited range of conditions. They comprise automatic techniques for identifying video frames, which contain good images of fish, for identifying the outline of the fish in 3D space, and for determining mass from linear measurements taken from this outline. The major challenges in developing these techniques have been the variable and complex nature of images collected in sea cages, and in identifying morphological models of salmon which are invariant with strain, maturity and culture conditions. Tests show that salmon mass can be estimated from linear dimensions with a mean error of less than 0.5% and that linear dimensions can be extracted automatically from stereo images with a mean error of less than 10%. Preliminary tests of the complete system used 60 images of 17 fish in the weight range 0.7 to 5.7 kg. The mean mass measurement error was 18% with a standard deviation of only 9%. Development of the algorithms is continuing.