Tomato maturity assessment via region-consistency constrained learning
Abstract Accurate tomato maturity assessment based on visual appearance is essential for quality control in agricultural production, yet remains challenging under complex visual conditions due to background clutter and ambiguous maturity cues. Many existing methods can suppress background interference to some extent, but lack explicit mechanisms to ensure that classification decisions are formed within maturity-relevant fruit regions. To address this limitation, this study proposes a region-consistency constrained learning framework for tomato maturity assessment. The framework integrates decision-relevant regions characterized by Class Activation Maps (CAM) with saliency-based spatial priors and enforces consistency between them during training, thereby regulating where classification decisions are formed without modifying the input or intermediate feature representations. Experiments conducted on a real-world tomato dataset demonstrate that the proposed framework consistently improves classification performance across multiple backbone architectures. Statistically significant gains are achieved without introducing additional computational cost during inference, indicating that the proposed framework provides an effective way to enhance the reliability of visual maturity assessment under complex agricultural conditions.
Authors
- Jianxun Lou (ORCID: https://orcid.org/0000-0002-2982-595X)
- Siqi Zhou
- Ran Dong
- Yutong Zhang
- Feng Liang
Institutions
- Electric Power University (VN)
- Northeast Electric Power University (CN)
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-14
- DOI
- https://doi.org/10.1038/s41598-026-76301-6
- Primary Topic
- Smart Agriculture and AI
- Type
- article
- Field-Weighted Citation Impact
- 0.00