AI-based non-destructive food screening for detecting calcium carbide (CaC2)-treated bananas
Abstract Artificial fruit ripening using calcium carbide (CaC 2 ) poses a significant food safety concern because prolonged consumption may expose consumers to harmful chemicals. Existing detection methods typically require destructive laboratory analysis, specialized equipment or expert interpretation, limiting their use for rapid on-site screening. Although several AI-based approaches have recently been proposed, they are generally evaluated on individual banana images under controlled conditions, with limited investigation of robustness, computational efficiency and realistic deployment scenarios. This study presents an AI-based framework for the non-destructive detection of calcium carbide-treated bananas using conventional smartphone captured images. Two complementary AI frameworks, Deep Classification with a Frozen Backbone (DCF) and Deep Feature-Based Shallow Classification (DFS), were systematically compared using seven state-of-the-art convolutional and transformer backbones. Unlike previous studies, the proposed framework was evaluated under realistic inspection scenarios containing multiple bananas with different proportions of naturally ripened and calcium carbide-treated fruits, together with comprehensive robustness, explainability and computational efficiency analyses. The DCF-based MobileNet-v2 achieved the highest hold-out classification accuracy of 99.74%, whereas the DFS-based MobileNet-v2 + Logistic Regression (LR) framework provided the best trade-off between predictive performance and computational efficiency. The MobileNet-v2 + LR framework achieved 95.46% hold-out accuracy, 95.40% 5-fold cross-validation accuracy and correctly classified all 20 images (99.9% mean prediction confidence) in realistic multi-banana inspection, while requiring only 0.65 Giga Floating-Point Operations (GFLOPs) and 0.177 seconds per image. Grad-CAM visualizations further confirmed that the model learned meaningful fruit-specific representations. These findings demonstrate that DFS models combined with smartphone imaging provide an efficient and practical solution for non-destructive food adulteration screening.
Authors
- Faisal Tariq (ORCID: https://orcid.org/0000-0002-0580-8805)
- Md. Tarequl Islam (ORCID: https://orcid.org/0000-0001-8076-2379)
- Md Shakhawat Hossain
- Elman Alam (ORCID: https://orcid.org/0009-0005-0534-9361)
- A. S. M. Ashraf Mahmud
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-15
- DOI
- https://doi.org/10.1038/s41598-026-79484-0
- Primary Topic
- Advanced Neural Network Applications
- Type
- article
- Field-Weighted Citation Impact
- 0.00