Image Segmentation Applied to Multi-species Phenotyping in Fish Farming

Fish farming has been gaining prominence in recent years, almost doubling world production in a ten-year period. Fish farming products are an important source of protein in coastal or insular countries, as they are the most abundant natural resource in these regions and are part of the daily diet of their populations. Recent technological advances allow the evolution of the practice of fish farming through new tools such as Artificial Intelligence and Internet of Things. Among the tasks that stand out is the phenotyping of animals raised in captivity during various stages of growth to assess the interaction of the species with the environment or even the prevalence of certain characteristics after successive reproductive selections. Phenotyping is usually performed manually through a digital image taken by an expert and post-processed in specialized measurement software using an object of known size as a reference in the image. Based on this problem, this work proposes a Computer Vision System to automate the phenotyping of two common species in Brazilian fish farming: Piaractus Mesopotamicus (Pacu) and Colossoma macropomum (Tambaqui). The results indicate a positive correlation between the measurements performed by humans and the proposed Computer Vision system, presenting itself as a viable alternative to accelerate the process of collecting information for reproductive selection in commercial fish farming.