Rapid Starch Particle Sizing by YOLOv8n

Accurate measurement of starch particle size is essential for quality control in the food industry. This study developed a rapid image-based detection and analysis method for starch particles based on an improved YOLOv8n network. The model incorporates a P2 small-object detection layer, a simple attention module (SimAM), and the Wise Intersection over Union v3 (WIoU3) loss function to balance segmentation accuracy and computational efficiency. Experimental results showed that the improved model reached an mAP@0.85 of 93.31%, which is 3.50 percentage points higher than that of the baseline, while reducing the number of parameters by about 32.8% and achieving an inference time of 2.4 ms per image on an RTX 3080 GPU. Compared with several lightweight models, the improved model showed higher boundary-sensitive segmentation accuracy while maintaining relatively low model complexity. Comparison with the laser particle size analyzer (LPSA) showed a relative error of 1.85% in the median diameter (D50) and similar particle size distribution trends under the tested conditions. Moreover, the method enabled quantitative analysis of particle circularity, showing a consistent peak near 0.85 across different image-set sizes. Overall, the proposed approach provides a rapid image-based method for starch particle characterization and shows potential for particle-size and morphology analysis in food-related applications.