N.I.A.E., AG ENG 84 and the Journal

Abstract

Fish mass estimation is extremely important for farmers to get fish biomass information, which could be useful to optimize daily feeding and control stocking densities and ultimately determine optimal harvest time. However, fish tail fin mass does not contribute much to total body mass. Additionally, the tail fin of free-swimming fish is deformed or bent for most of the time, resulting in feature measurement errors and further affecting mass prediction accuracy by computer vision. To solve this problem, a novel non-supervised method for fish tail fin removal was proposed to further develop mass prediction models based on ventral geometrical features without tail fin. Firstly, fish tail fin was fully automatically removed using the Cartesian coordinate system and image processing. Secondly, the different features were respectively extracted from fish image with and without tail fin. Finally, the correlational relationship between fish mass and features was estimated by the Partial Least Square (PLS). In this paper, tail fins were completely automatically removed and mass estimation model based on area and area square has been the best tested on the test dataset with a high coefficient of determination (R2) of 0.991, the root mean square error (RMSE) of 7.10 g, the mean absolute error (MAE) of 5.36 g and the maximum relative error (MaxRE) of 8.46%. These findings indicated that mass prediction model without fish tail fin can more accurately estimate fish mass than the model with tail fin, which might be extended to estimate biomass of free-swimming fish underwater in aquaculture.