Classification of aquatic creatures is essential for biodiversity research in ecosystems. The minor differences between species and the underwater conditions affecting image quality pose significant challenges for image classification models. This type of classification, where images belong to a single metaclass, is referred to as Fine-Grained Visual Classification (FGVC). A key aspect of FGVC is the ability to distinguish these subtle differences that determine membership in a specific class. Therefore, in this study, an attention-based CNN model with a skip-connection mechanism was proposed. The model was tested on seven publicly available datasets, and for four of them - Croatian Fish Dataset, DeepFish, Large-Scale Fish, and Fish-Pak - it achieved the highest score, respectively 96.88%, 100%, 100%, and 100%. The presented model is significantly lighter compared to existing methods: it has 1.4 million parameters, which, combined with very high accuracy, enables its application in resource-constrained environments.