MicroalgaeNet: Enhancing recognition of long-tailed marine microalgae images through multi-expert networks and feature compression

The recognition of marine microalgae is crucial for assessing the ecological status of natural water bodies. This study presents a novel approach to address the challenges of inefficiency, limited Precision, and long-tailed data distributions inherent in manual microscopic examination. Unlike previous works focusing mainly on data augmentation or re-weighting, we introduce for the first time a ResNeXt-50-based multi-expert network, coupled with an exponential-function-based feature compression mechanism, to improve the recognition of marine microalgae images. This innovative approach specifically mitigates the impact of class imbalance on classification performance. The method is evaluated on the WHIO-Plankton dataset, which comprises images of 23 marine microalgae species. Our proposed method sets a new state-of-the-art (SOTA) benchmark, achieving a leading average precision of 88% and an average recall of 86.62%. The model significantly outperforms the baseline ($p<0.05$), demonstrating its enhanced ability to identify tail categories. Furthermore, with an inference latency of 8.556 ms, our model demonstrates strong feasibility for real-world deployment. These results indicate that the proposed approach can effectively enhance the recognition performance of marine microalgae, offering valuable support for marine microbial research and life science applications.