Multi-class fruit ripeness detection using YOLO and SSD object detection models
Pooja Kamat, Shilpa Gite, Harsh Chandekar, Lisanne Dlima, Biswajeet Pradhan
Abstract
Accurate fruit ripeness detection is critical to reducing post-harvest losses and improving quality control in agricultural systems. This study benchmarks four object detection models—YOLOv5, YOLOv6, YOLOv7, and SSD-MobileNetv1—for multi-class ripeness classification of strawberries and avocados across four stages: unripe, partially ripe, ripe, and rotten. The dataset, captured under natural conditions, has been manually annotated and published for public access. YOLOv6 achieved the highest mean Average Precision (99.5%) and demonstrated a strong balance between accuracy and real-time inference speed (85.2 FPS). All models were evaluated using standard classification metrics and cross-validated through a 5-fold approach to ensure robustness. The results indicate YOLOv6 as the most reliable model for smart fruit sorting and quality monitoring applications. This study offers a reproducible benchmarking pipeline and contributes toward the development of deployable deep learning solutions in precision agriculture.