A cost-aware comparative study of deep learning models for industrial tomato quality grading
Auteur(s): Kabengele K Hervek, Mbayandjambe M Alibor, Nkwimi B Grevi, Tashev Tasho, Kyamakya Kyandoghere, Yang Xiaopeng, Kasereka K SelainNom de la revue/Journal: Scientific African
Mois: septembre
Volume: 33
Numéro: e03454
Année: 2026
pages: 1-16
DOI: 10.1016/j.sciaf.2026.e03454
Résumé
Tomato freshness sorting is a key operation in the food processing industry. Most existing approaches rely on binary sorting, labeling tomatoes as either fresh or rotten. This simplistic approach leads to the discard of large numbers of edible tomatoes. A 2019 study on Barilla Tomato and Basil sauce production in Italy found that 20.2% of tomatoes in the 400 g format were rejected. About 80% of those discarded tomatoes were still edible. They included green, slightly damaged, squeezed, or mislabeled fruit. This paper addresses these inefficiencies using three deep learning models: Xception, YOLOv5, and Swin Transformer. Each model was trained and evaluated on a curated three-class dataset for freshness stage recognition. The three classes are fresh, at risk, and rotten. The novelty of this work lies in the three-stage industrial reformulation of the sorting problem, the curated benchmark design, and the cost-aware comparative analysis. To assess generalizability, all models were additionally evaluated on a second independent dataset, FGrade, through a controlled cross-dataset protocol. On the primary dataset, the Swin Transformer achieved the highest accuracy (99.56%). Xception led on the FGrade benchmark (86.70%). These results show that dataset characteristics influence architecture selection. Across both datasets, all three models reduce estimated economic losses by more than 80% compared to manual sorting. This confirms their industrial viability. The results support the use of automated multi-stage sorting to reduce food waste and improve profitability.