HySwinFormer: A hybrid deep learning architecture for fine-grained classification of marine microalgae

Accurate identification of microalgae is vital for marine ecological monitoring, algal bloom early warning, and environmental management. However, existing methods often struggle with misclassification due to the morphological similarity and biological complexity of microalgae. To address this, we propose HySwinFormer, a dual-branch deep learning architecture designed for fine-grained classification of microalgal images. The architecture includes a Multi-Scale Modeling Branch (MSMB) for capturing global morphological features and a Detail Enhancement Branch (DEB) for extracting local texture details. These features are dynamically fused, and an enhanced Efficient Multi-Scale Attention (EMA) module is introduced to improve feature representation and discrimination. We built a dataset of 20 Rhizosolenia species, a genus known for subtle interspecies differences, to evaluate the model. Experimental results show that HySwinFormer outperforms existing models in accuracy, F1-score, and Matthews correlation coefficient (MCC), and generalizes well on two public benchmarks: Plankton Set-1.0 and WHOI-Plankton. Finally, we successfully deployed the model on an NVIDIA Jetson edge computing platform, demonstrating its capability for real-time analysis at approximately 25 frames per second (FPS) and thus validating its practical feasibility in resource-constrained, real-world scenarios. In summary, HySwinFormer provides an empirically validated, efficient, and reliable technical solution for building intelligent marine ecological monitoring systems.