Deep Learning for Saffron Adulteration Detection from Stigma Images: A Comparative Study of AlexNet, ResNet, and VGG16

Technical Summary: Deep Learning for Saffron Adulteration Detection from Stigma Images

Problem Statement
Saffron (Crocus sativus L.) is a high-value spice susceptible to economically motivated adulteration (EMA), where fraudulent materials such as dyed corn silk, safflower stamens, and colored plant roots are mixed with authentic stigmas to mimic appearance and increase bulk weight. While conventional laboratory techniques like HPLC, GC–MS, and UV–Vis spectroscopy offer high accuracy, they are destructive, time-consuming, costly, and require specialized infrastructure, limiting their utility for rapid postharvest screening. Existing machine-vision approaches relying on handcrafted features often fail to generalize under variable lighting or natural morphological variability. There is a need for a rapid, non-destructive, and cost-effective automated system to authenticate saffron quality in postharvest pipelines.

Methodology
The study employed a comparative analysis of three Convolutional Neural Network (CNN) architectures—AlexNet, ResNet, and VGG16—using transfer learning to distinguish between authentic and counterfeit saffron stigmas.

• Dataset: A curated dataset of 1,020 expert-labeled RGB images was assembled, comprising 719 authentic samples (covering Sargol, Negin, and Pushal grades) and 301 counterfeit samples (lab-prepared mimics of common adulterants). To prevent data leakage, each physical sample was photographed only once.
• Preprocessing: Images were captured using a Canon 700D DSLR. Preprocessing involved cropping 10% of border margins, applying binary masks to isolate filaments, and homogenizing the background via RGB thresholding (replacing pixels with values >180 with white). Images were resized to architecture-specific inputs (e.g., 227×227 for AlexNet) and normalized.
• Data Augmentation: An online augmentation strategy was applied during training, including random rotations (0–360°) and horizontal/vertical flips to enhance robustness. Validation and test sets were not augmented.
• Model Configuration: The final classification layers of pre-trained models (ImageNet weights) were replaced with a two-unit fully connected layer and softmax activation. Models were fine-tuned end-to-end using Stochastic Gradient Descent (SGD) with momentum (0.9), a mini-batch size of 64, and an initial learning rate of 1e-4 for 50 epochs.
• Implementation: Training was conducted on a CPU-based system (Intel Core i5-6200U, 8GB RAM) using MATLAB R2020a. The dataset was split into 80% training and 20% validation, with a separate test set for final evaluation.
• Evaluation: Performance was measured using Accuracy, Sensitivity (Recall for counterfeit), Specificity (Recall for authentic), Precision (PPV), and Negative Predictive Value (NPV).

Key Results
The study evaluated the models on their ability to detect visual adulteration:

• AlexNet: Achieved the highest overall accuracy of 96.49%. It demonstrated 100% sensitivity (correctly identifying all counterfeit samples) and 93.75% specificity. It showed rapid convergence within 15–20 epochs.
• ResNet-50: Achieved 94.74% accuracy. It exhibited 100% specificity and 100% precision but lower sensitivity (90%), resulting in three false negatives (misclassifying adulterated samples as authentic).
• VGG16: Achieved 87.72% accuracy. While it reached 100% sensitivity, its specificity was significantly lower (79.41%), leading to a higher rate of false positives (misclassifying authentic samples as adulterated). The authors attribute this underperformance to the model's depth and parameter count relative to the dataset size and the computational constraints of CPU-based training, which limited effective convergence.

Significance and Claims
The paper claims that deep learning, specifically CNNs, offers a feasible, rapid, and non-destructive alternative to chemical analysis for saffron authentication. The primary contribution is demonstrating that even relatively compact architectures like AlexNet can effectively capture the salient morphological and chromatic features necessary to distinguish genuine stigmas from visual counterfeits, outperforming deeper models in this specific context due to better optimization dynamics on moderate-sized datasets.

The authors position this work as a foundation for integrating AI-driven screening into postharvest quality control, grading, and export inspections. They emphasize that the high sensitivity of the AlexNet model is particularly valuable for screening applications where missing an adulterated sample (false negative) poses a greater economic risk than a false alarm.

Limitations and Future Directions
The authors explicitly acknowledge several limitations:

1. Imaging Conditions: The study utilized controlled illumination and uniform backgrounds, which may not fully replicate complex real-world postharvest environments with overlapping filaments and variable lighting.
2. Validation: Performance metrics were based on a single random data split without k-fold cross-validation.
3. Hardware Constraints: Training on a CPU, particularly for VGG16, likely hindered optimal convergence.
4. Generalization: The study did not employ cross-validation or test on diverse, large-scale datasets.

Future work suggested by the authors includes validating models on larger and more diverse datasets, employing k-fold cross-validation, utilizing GPU acceleration, integrating explainable AI (XAI) for transparency, and developing lightweight models for on-device inference (e.g., via smartphones) for real-world deployment. The authors also propose exploring hybrid imaging-spectroscopy pipelines for more robust authentication.