Assessing freshwater algae is pivotal for understanding aquatic ecosystems, yet data imbalances and scarcities in algal genera complicate the efficacy of detection models. This study introduces an innovative method to tackle this issue by leveraging the capabilities of StyleGAN2-ADA for data augmentation, thereby enhancing the detection performance for algal images. By generating artificial algal instances using StyleGAN2-ADA, this approach provides a robust solution to the critical issues of data scarcity and imbalance prevalent in freshwater algal datasets. The method was applied to a dataset of 645 images of various freshwater algae genera. A Cascade Mask R-CNN model, trained on a combined dataset of real and GAN-generated multi-genera algal images using the Swin-Tiny backbone model,...[ Read more ]

Assessing freshwater algae is pivotal for understanding aquatic ecosystems, yet data imbalances and scarcities in algal genera complicate the efficacy of detection models. This study introduces an innovative method to tackle this issue by leveraging the capabilities of StyleGAN2-ADA for data augmentation, thereby enhancing the detection performance for algal images. By generating artificial algal instances using StyleGAN2-ADA, this approach provides a robust solution to the critical issues of data scarcity and imbalance prevalent in freshwater algal datasets. The method was applied to a dataset of 645 images of various freshwater algae genera. A Cascade Mask R-CNN model, trained on a combined dataset of real and GAN-generated multi-genera algal images using the Swin-Tiny backbone model, showed remarkable improvements in detection accuracy. The mAP scores for bounding box detection demonstrated a substantial increase, with a notable enhancement in detecting rare algal genera. The strategy of adding 50% more artificial data to the training set proved beneficial, pushing the model performance further without necessitating excessive usage of artificial data. Moreover, this GAN-based data augmentation technique was successfully applied to both Swin-Tiny and ResNet-50 backbone models, indicating its versatility and wide applicability across various machine learning models. The substantial increase in mAP scores across these models validates the robustness of GAN-based data augmentation, regardless of their original performance level. The study took a step forward by deploying the freshwater algae detection model as an online service. By converting the model to the ONNX format and hosting it on the HuggingFace Space cloud platform, real-time and precise algae detection was made accessible to a broader range of users, greatly enhancing efficiency and flexibility. The model maintains high performance while significantly reducing the time required for microscope algae detection. In conclusion, this study presents a promising data augmentation solution for algae detection and monitoring, offering a noteworthy contribution to aquatic ecology research and applications challenged by imbalanced and scarce data.