Comparative Evaluation of YOLO Architectures with Optimized Preprocessing for Multi-Class Microalgae Detection

Accurate detection and counting of microalgae species are essential for biomass monitoring, contamination control, and smart cultivation management in biotechnology applications. However, automated analysis of microscopy images remains challenging because of dense cell distributions, small object sizes, low-contrast boundaries, and inter-class visual similarity. This study presents a two-stage deep learning framework for multi-class microalgae detection using recent YOLO architectures. In the first stage, five preprocessing pipelines were comparatively evaluated using YOLOv8s to determine an effective image enhancement strategy; LAB color space conversion combined with CLAHE and sharpening visibly improved cell boundary definition. In the second stage, the selected preprocessing configuration was fixed and four lightweight detectors (YOLOv8s, YOLOv9s, YOLOv10s, and YOLOv11s) were benchmarked under identical training conditions. YOLOv9s achieved the best overall performance with the highest precision (0.867), F1-score (0.824), and mAP50 (0.889). The findings indicate that detector architecture and preprocessing strategy jointly influence microscopic algae detection performance, and that model recency alone does not guarantee superior results.