Salmon fish (Salmonidae specie) is a critical component of the trophic dynamics within Scandinavian aquatic ecosystems, especially in Norway. This region has witnessed the proliferation of aquaculture facilities along its coastline, reflecting the industry’s significance. These aquaculture enterprises meticulously record mortality rates and the mass of individual salmonids, either expressly or on average. Manual execution of these tasks, however, is labor-intensive and prone to inaccuracies. Various scholars have advocated for the utilization of three-dimensional (3-D) imaging technologies for mass estimation, yet the economic implications render this approach less feasible. Concurrently, attempts to employ two-dimensional (2-D) imagery for this purpose have been made, albeit with limited success in achieving precise mass estimations.
This investigation proposes a methodology for estimating the mass of salmon utilizing two-dimensional (2-D) imagery. Initially, a dataset comprising over 800 images of salmon, with known lengths and masses ranging from 1 kg to 6.5 kg, was compiled as Figure 1. The average mass of the Salmonidae samples was recorded at 4.28 kg. A Mask-Region based Convolutional Neural Network (CNN) algorithm was employed to develop a segmentation model, trained on 250 annotated images of salmon. Notably, during annotation, fin regions were excluded from segmentation due to potential damage or absence, ensuring model robustness. Post-segmentation, the model was evaluated using the remaining images (Figure 2), where the pixel area and length of each fish were determined. Additionally, the ratio of area to length was calculated.
 Subsequently, a simplistic neural network architecture, consisting of two layers, was developed. This network was trained and validated using features such as pixel length, area, and the previously calculated area-to-length ratio, aiming to predict the actual mass of the fish. The predictive model was tested on data from 762 salmon, achieving an average accuracy exceeding 98.2%. The model exhibited a mean error of 77.8 grams, with the minimum and maximum errors being 23 and 165 grams, respectively. Figure 3 illustrates the correlation between the actual and predicted masses of the salmon.