Adaptive Multi-Scale Fusion Enhanced RT-DETR for Efficient Cyanobacteria Detection in Microscopic Images

Accurate and efficient detection of cyanobacteria in microscopic images is important for automated water-quality monitoring, but remains challenging because of complex aquatic backgrounds, large scale variation, and uneven sample quality. This study proposes an adaptive multi-scale fusion enhanced RT-DETR framework for cyanobacteria detection. The baseline RT-DETR-R18 is improved by incorporating the SeFaster module for efficient feature extraction, the high-level screening-feature fusion pyramid network for semantic-guided multi-scale fusion, and the Wise-IoU loss for more stable localization learning under mixed-quality samples. Experiments on the reorganized EMDS-7 dataset show that the proposed method achieved 79.05% mAP@0.5, 66.03% mAP@0.5:0.95, 16.31 M parameters, 54.6 G FLOPs, and 70.85 FPS. The proposed model also obtained the highest mAP@0.5 across the seven cyanobacteria categories. Moreover, cross-dataset evaluations further suggest the stability and transferability of the model. These results indicate that the proposed framework demonstrates potential for effective cyanobacteria detection in microscopic images with a good balance between detection accuracy and computational efficiency.