Abstract
To support the sustainable use of marine resources, regulations have been proposed to reduce fish discards focusing on the registration of all listed species. To comply with such regulations, computer vision methods have been developed. Nevertheless, current approaches are constrained by their closed-set nature, where they are designed only to recognize fish species that were present during training. In the real world, however, samples of unknown fish species may appear in different fishing regions or seasons, requiring fish classification to be treated as an open-set problem. This work focuses on the assessment of open-set recognition to automate the registration process of fish. The state-of-the-art Multiple Gaussian Prototype Learning (MGPL) was compared with the simple yet powerful Open-Set Nearest Neighbor (OSNN) and the Probability of Inclusion Support Vector Machine (PISVM). For the experiments, the Fish Detection and Weight Estimation dataset, containing images of 2216 fish instances from nine species, was used. Experimental results demonstrated that OSNN and PISVM outperformed MGPL in both recognizing known and unknown species. OSNN achieved the best results when classifying samples as either one of the known species or as an unknown species with an F1-macro of 0.79±0.05 and an AUROC score of 0.92±0.01 surpassing PISVM by 0.05 and 0.03, respectively.