Deep Learning image segmentation for extraction of fish body measurements and prediction of body weight and carcass traits in Nile tilapia

Individual measurement of traits of interest is extremely important in aquaculture, both for production systems and for breeding programs. Most of the current methods are based on manual measurements, which are laborious and stressful to the animals. Therefore, the development of fast, precise and indirect measurement methods for traits such as body weight (BW) and carcass weight (CW) is of great interest. An appealing way to take noninvasive measurements is through computer vision. Hence, the objectives in the current work were to: (1) devise a computer vision system (CVS) for autonomous measurement of Nile tilapia body area (A), length, height, and eccentricity, and (2) develop linear models for prediction of fish BW, CW, and carcass yield (CY). Images from 1653 fish were taken at the same time as their BW and CW were measured. A set of 822 images had pixels labeled into three classes: background, fish fins, and A. This labeled dataset was then used for training of Deep Learning Networks for automatic segmentation of the images into those pixel classes. In a subsequent step, the segmentations obtained from the best network were used for extraction of A, length, height, and eccentricity. These variables were then used as covariates in linear models for prediction of BW, CW, and CY. A network with an input image of 0.2 times the original size and four encoder/decoder layers achieved the best results for intersection over union on the test set of 99, 90 and 64 percent for background, fish body and fin areas, respectively. The overall best predictive model included A and its square as predictor variables and achieved R-2 of 0.96 and 0.95 for fish BW and CW, respectively. Overall, the devised CVS was able to correctly differentiate fish body from background and fins, and the extracted area of the fish body could be successfully used for prediction of body and carcass weights.