Related Experiment Videos
OPTIFARM: Benchmarking YOLO Architectures for Location-Robust Potato Quality Detection
Tadej Peršak1, Marko Simonič1, Jernej Hernavs1
1Faculty of Mechanical Engineering, University of Maribor, Smetanova 17, SI-2000 Maribor, Slovenia.
Foods (Basel, Switzerland)
|June 26, 2026
Summary
Automated potato sorting using low-cost RGB cameras and deep learning is feasible. The YOLO26 model achieved the best performance in detecting potato quality across different farm locations, proving system readiness.
Area of Science:
- Agricultural Engineering
- Computer Vision
- Machine Learning
Background:
- Manual potato sorting is labor-intensive, subjective, and limits scalability.
- Automated quality detection is crucial for efficient post-harvest processing.
Purpose of the Study:
- To assess the feasibility of a low-cost RGB optical system for automated potato quality detection.
- To benchmark deep learning object detection models for potato sorting.
Main Methods:
- Constructed a controlled imaging platform with commodity hardware.
- Collected and annotated a dataset of 19,805 potato instances across 1361 images.
- Benchmarked 25 YOLO model configurations (YOLOv8-YOLO26) using cross-location evaluation.
Main Results:
- All models showed high in-distribution performance (F1 ≥ 0.906).
- Cross-location performance varied (F1: 0.792-0.918), with YOLO26 achieving the best results (F1=0.918).
- Feed detection emerged as a generalization bottleneck.
Conclusions:
- Affordable RGB-based potato sorting systems are technically viable.
- Cross-location evaluation is essential for assessing real-world deployment readiness.
- Transferable representations are achievable with standard supervised training.
