FreISH: A Database of Japanese Freshwater Fish in the Wild and Baseline Recognition Evaluations

Fish identification is more challenging than that of other animals due to mismatches between their names and appearances. Building a fish database is crucial for fish research and the creation of digital fish-related applications. Though some fish databases exist, they often operate under specific settings like fixed lighting, simple backgrounds, or solo fish imagery. In our research, we developed a database (FreISH) capturing Japanese freshwater fish in their natural settings. This preliminary version features 1,200 images of twelve native Japanese freshwater species, each represented by 100 images displaying different postures. We gauged the species recognition rate using the VGG-19-BN deep learning model on this dataset. Additionally, we contrasted the efficiency of the "fine-tuning" and "transfer learning" techniques to find the optimal approach.