Automated Grape Classification Using a YOLOv11n-Based Computer Vision System

Abstract

This study presents the implementation and evaluation of a computer vision system based on the YOLOv11n model for the automated classification of grapes intended for wine production at a Peruvian agroindustrial company. The proposed system replaces manual visual inspection, whose average precision is approximately 76%, with a lightweight real-time detection model trained on 5,137 augmented images annotated using Roboflow. The model achieved a mean precision of 89.1% (95% CI: 87.4–90.8%), a recall of 88.3% (95% CI: 86.1–90.4%), and an mAP@50 of 92.5% (95% CI: 91.2–93.8%) for the Premium grape category, representing a 13% improvement in precision compared to traditional manual inspection. Lower performance was observed for the Bulk category (76.1% precision and 61.4% recall; 95% CI: 58.0–64.8%), mainly due to morphological variability among the grape clusters. The results confirm the technical feasibility and scalability of YOLOv11n as an effective tool for optimizing quality control processes in viticulture, demonstrating its capability for real-time inference under uncontrolled environmental conditions.