A Model for Large-Scale Fish Dataset Classification: SIM-CONV

Image based aquatic organism recognition technology is one of the key technologies in the current field of aquatic ecology. By accurately identifying aquatic organisms, the structure and function of aquatic ecosystems can be better understood, providing scientific basis for the protection and management of aquatic organisms. However, current aquatic organism recognition methods have problems such as limited coverage of species and single classification methods. Based on this, this paper proposes a model SIM-CONV for fish classification in large-scale fish datasets. Due to the complexity of aquatic species, in order to provide more scientific classification results, this paper uses aquatic classification method to classify data at the phylum, genus, and species level, and classifies fish at the order and family classification levels. Common and rare samples are established for each classification level. The large-scale fish dataset FishNet is used as the experimental target, and the SIM-CONV model is obtained by introducing SIMAM attention mechanism to improve the architecture of ConvNext network and class balance training technology, which improves the accuracy of fish classification. The SIM-CONV model improved the classification accuracy of majority and rare classes at the order and family levels on the FishNet dataset by 0.83%, 2.27%, 6.45%, and 3.89%, respectively.