Cultivar Shift in YOLO-Based Potato Quality Grading
An optical potato grading line may process several cultivars. We examined how much accuracy a detector trained on one cultivar lost on another, whether a jointly trained model matched cultivar-specific models and whether compact models matched larger ones. Twenty-five You Only Look Once (YOLO) variants (five families, five sizes each) were trained on Belmonda, Prada or both (75 runs, one seed) to detect tubers and grade them as edible, feed or rotten. Models were evaluated on held-out images from the same acquisition sessions (95 per cultivar). Relative to target-domain specialists, the mean shortfall in mean average precision (mAP50–95) was 0.336 from Belmonda to Prada and 0.162 in the reverse direction; it was larger from Belmonda to Prada for all 25 variants. Grade confusions, mostly involving the feed class, accounted for 96–97% of false negatives, whereas missed tubers were rare. Joint models differed from the specialists by −0.014 to +0.017 mAP50–95 on each cultivar; small joint models met a 0.01 margin against larger same-family specialists in 17 of 30 comparisons. For these two cultivar domains, a jointly trained model performed similarly to same-architecture specialists; as cultivar was confounded with lot and season, validation on independent lots is required.
Authors
- Tadej Peršak
- Simon Klančnik (ORCID: https://orcid.org/0000-0001-5922-3792)
- Jernej Hernavs (ORCID: https://orcid.org/0009-0001-1518-803X)
- Marko Simonič (ORCID: https://orcid.org/0009-0005-7145-6069)
Institutions
- University of Maribor (SI)
Publication Details
- Journal
- Foods
- Published
- 2026-11-18
- DOI
- https://doi.org/10.3390/foods15224138
- Primary Topic
- Smart Agriculture and AI
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- Javna Agencija za Raziskovalno Dejavnost RS