Grain Quality Grading from Images using CNN

Grain mandis and mills still grade lots by hand — a grader scoops a handful, eyeballs defects and assigns a grade. It is slow, subjective and varies between graders. This project replaces the eyeball with a camera: kernels are segmented from a tray photograph with watershed segmentation, each kernel crop is classified by a fine-tuned ResNet-50 into five quality classes, and the class histogram decides the lot grade (A/B/C). Students learn a realistic two-stage vision pipeline — classical segmentation feeding a deep classifier — plus the data discipline of photographing kernels under fixed lighting. The demo grades sample grain photos with representative predictions.

1. Wheat and rice samples are photographed on a tray under fixed lighting; kernels are labelled by defect class.
2. Watershed segmentation isolates individual kernels from each tray photo into 224×224 crops.
3. Crops are split into stratified train/validation sets with augmentation for orientation and brightness.
4. ResNet-50, pre-trained on ImageNet, is fine-tuned with a new classification head and dropout.
5. Training uses Adam with early stopping; the held-out split yields accuracy, confusion matrix and per-class metrics.
6. Per-lot defect percentages are computed and mapped to Grade A/B/C thresholds.
7. The web demo grades sample photos through the same pipeline with representative predictions.

Tech stack:

• Python 3, TensorFlow/Keras
• ResNet-50 (transfer learning)
• OpenCV (watershed segmentation)
• NumPy, Matplotlib, scikit-learn
• Jupyter Notebook (training)
• HTML/CSS/JS grading demo
• Pillow

Dataset & model details

• Dataset: Custom curated grain-kernel set — ~4,200 wheat and rice kernel images photographed under fixed lighting from mandi and mill samples, labelled into 5 classes: healthy, shriveled, broken, discolored, foreign material; stratified 80/20 split.
• Task: 5-class image classification per kernel; input = 224×224×3 kernel crop, output = defect-class probabilities; lot grade derived from class histogram.
• Model: ResNet-50 backbone (ImageNet pre-trained) → Dense(256, ReLU) → Dropout(0.4) → Dense(5, softmax); ~25.6M parameters.
• Metrics: Kernel accuracy ~94% (design target), grader agreement κ ~0.91 (design target), per-class precision/recall. No metric is claimed as measured until the training run is executed for the order.

• Kernel-level classifier ResNet-50 fine-tuned on five classes: healthy, shriveled, broken, discolored, foreign material — with per-class metrics.
• Two-stage pipeline Watershed segmentation isolates kernels from the tray photo before the CNN sees them; both stages are documented.
• Curated kernel dataset About 4,200 wheat and rice kernel images photographed under fixed lighting, labelled by defect class.
• Lot grading logic Defect percentages map to Grade A/B/C with configurable thresholds — the actual commercial decision, not just a label.
• Interactive grading demo Single-file web app: pick a sample grain photo and see the predicted grade with class probabilities.
• Grader-agreement evaluation Model grades are compared against a human grader with Cohen's kappa — the honest metric for a grading task.
• Viva kit Report PDF, PPT and Q&A covering segmentation, transfer learning, class imbalance and grading thresholds.

• Curated, labelled kernel dataset with split files
• Segmentation + training Jupyter notebooks
• Trained ResNet-50 model with preprocessing code
• Interactive grading web demo with sample photos
• Confusion matrix, per-class metrics, grader-agreement analysis
• Lot-grade report generator (CSV + PDF)
• Project report PDF, PPT presentation
• Viva Q&A preparation document