Computer Vision Real-Time Impurity Detection in Sour Cassava Starch Production

This paper presents a computer vision-based framework for real-time impurity detection in cassava conveyor belts used in starch production. A YOLOv11m model was trained using transfer learning on a dataset collected from artisanal processing units, incorporating data augmentation to enhance generalization. The model achieved 80% precision and over 80% recall, demonstrating reliable detection performance despite class imbalance and visual similarity between cassava and impurities. Real-time validation using a mobile-based workflow confirmed robustness under varying conditions. The proposed approach shows potential for improving process efficiency, product quality, and automation in small- and medium-scale cassava starch industries.