<![CDATA[Accurate fruit condition classification is essential for automated food safety assessment, particularly due to health risks associated with chemical contaminants such as formalin. However, reliable generalization in automated inspection systems remains challenging because limited visual variation in image datasets often leads to overfitting in deep learning models. To address this challenge, this study proposes an EfficientNetB3-based framework that integrates geometric data augmentation with a structured two-stage fine-tuning strategy to improve robustness and training stability. The proposed model achieved 99% test accuracy with consistent cross-dataset performance. The framework also demonstrated stable optimization behavior across training stages, indicating improved generalization capability. From a practical perspective, the proposed approach may support scalable food quality monitoring and automated sorting in agricultural supply chains, as well as preliminary food safety screening in large-scale inspection processes.]]>
Copyrights © 2026
