Abstract
Traditional fish classification systems suffer from limited training data and imbalanced datasets, particularly for rare or morphologically complex species. This paper presents a novel Generative Adversarial Network architecture that integrates adaptive identity blocks to preserve critical species-specific features during generation, coupled with species-specific loss functions designed around distinctive characteristics of marine species. Our method introduces adaptive identity blocks that learn to maintain species-invariant features while allowing controlled morphological variations for data augmentation. The species-specific loss function incorporates morphological constraints and taxonomic relationships to ensure generated samples maintain biological plausibility while enhancing dataset diversity. Experimental evaluation on a comprehensive fish dataset containing nine species demonstrated significant performance improvements. Our proposed method achieved 95.1% ± 1.0% classification accuracy, representing a 9.7% improvement over baseline methods and 6.7% improvement over traditional augmentation approaches. While demonstrated on a dataset of 9000 images across nine fish species, these results provide a solid foundation that warrants validation on larger, more taxonomically diverse datasets to establish broader generalizability. Segmentation performance achieved 89.6% ± 1.3% mean Intersection over Union, representing a 12.3% improvement over baseline methods. Critically, our approach showed substantial improvements for morphologically complex species, with expert evaluation by marine biology specialists confirming 88.7% ± 2.0% overall quality and achieving 87.4% ± 1.6% biological validation score. Statistical significance testing confirmed all improvements at p < 0.001 with large effect sizes, and cross-validation demonstrated exceptional consistency across folds. The results validate the effectiveness of our biologically-informed approach for generating high-quality synthetic fish data that significantly improves classification and segmentation performance while maintaining biological authenticity.