YOLOv8l-Voronoi ensemble learning for enhanced detection and segmentation of nanoparticle agglomerates in micro fluidized beds

This study proposes an integrated YOLOv8l-Voronoi ensemble framework for accurate detection and segmentation of nanoparticle agglomerates in micro-fluidized beds. To address challenges posed by high density, small size, and complex morphologies of agglomerates, an adaptive overlapping tiling strategy is introduced, reducing the agglomerate miss rate from 19.4% to 13.1% compared to the baseline YOLOv8l model. A Voronoi polygon partitioning combined with a support vector machine (SVM) is further incorporated to suppress false alarms caused by illumination variations, and a cascaded U-Net model is employed for pixel-level boundary segmentation. Experimental results show that the proposed framework achieves an mAP@0.5 of 90.1% on a self-built dataset, representing an 8.4 percentage point improvement over the baseline, while maintaining a high segmentation accuracy (MIoU>0.87). The method provides a robust and precise solution for online monitoring and quantitative analysis of nanoparticle aggregation behavior.