Boosting low-resolution algae detection via cross-resolution knowledge distillation and transfer learning

To address the critical challenge of degraded detection performance caused by low-resolution microscopic inputs in field-based harmful algae monitoring, this paper proposes a cross-resolution knowledge distillation framework integrated with transfer learning. The method transfers fine-grained feature representations from a high-resolution YOLO teacher model to a low-resolution YOLO student model, thereby enabling accurate and efficient harmful algae identification at no additional inference cost.

Built upon the EMDS-7 environmental microorganism dataset, we employ a YOLO student network with a low input resolution, initialized with COCO pre-trained weights. A YOLO teacher model, trained at 1280 × 1,280 resolution, provides high-resolution feature supervision. To transfer the fine-grained spatial cues, we design a multi-scale attention-based distillation loss that aligns the spatial attention maps between student and teacher at the P3, P4, and P5 neck levels. The overall training objective is a weighted sum of the original detection loss and the feature distillation loss.

On the test set of EMDS-7, the proposed method achieves a mAP50-95 of 0.78832 and a mAP50 of 0.87897, outperforming the undistilled baseline (0.78434 and 0.87341) as well as other methods such as YOLOv5s, YOLOv8s, YOLOv10s, YOLOv12s and YOLOv26s. Precision and recall reach 0.87542 and 0.84143, respectively. Ablation studies on the distillation weight $\lambda$ show that $\lambda =0.5$ gives the best performance, with lower or higher values leading to degraded results.

The proposed cross-resolution knowledge distillation framework effectively mitigates the fine-grained feature degradation inherent in low-resolution algae imagery, while transfer learning provides the student model with robust visual priors. The resulting lightweight detector, operating at low resolution, is well-suited for deployment on field-portable microscopic platforms. This work offers a viable pathway toward intelligent harmful algae monitoring in resource-constrained field settings.