Grain Quality Grading using Computer Vision

Grain traders, millers and farmer-producer groups still grade grain mostly by eye: an inspector scoops a sample, looks at kernel size, color uniformity and visible defects, and assigns a grade from experience. That is slow, subjective, and two inspectors can grade the same lot differently. Instrumental lab analysis exists but is out of reach for small operations. Computer vision fills this middle ground: a camera, controlled lighting and a classifier that measures exactly the features an inspector looks at — color, size, shape and texture — and applies the same decision every time. This project builds that system as a classical machine-learning pipeline: no deep-learning black box, just interpretable handcrafted features and an SVM or Random Forest classifier trained on labeled reference samples you photograph and label yourself. Because the features are inspectable, the report and viva discussion can explain why a sample got its grade — kernel-size distribution, discoloration share, broken-kernel rate — which is exactly the kind of explainability deep models cannot offer.

1. A grain sample is photographed on a matte uniform background with even diffuse lighting at a fixed camera height.
2. OpenCV converts the image to HSV and segments kernels from the background via thresholding plus morphological cleanup.
3. Contour detection extracts every kernel and computes size and shape features — area, perimeter, aspect ratio, circularity, equivalent diameter.
4. HSV color histograms and GLCM texture features (contrast, homogeneity, energy) are computed per kernel.
5. Features are normalized with StandardScaler and fed to the trained SVM or Random Forest classifier for per-kernel predictions.
6. Per-kernel predictions are aggregated by majority vote into the sample grade; batch mode writes per-sample grades and feature summaries to a report.

Tech stack:

• Python 3.10, scikit-learn (SVM, Random Forest, StandardScaler)
• OpenCV (segmentation, contour extraction, color histograms)
• scikit-image (GLCM texture features)
• NumPy, pandas (feature tables and grade reports)
• Jupyter notebook (buyer-run training and evaluation)
• Matplotlib (feature distributions, confusion matrix plots)
• Flask demo app with sample-image upload interface

• Kernel segmentation from sample photos via HSV thresholding and morphological cleanup
• Per-kernel size and shape features: area, perimeter, aspect ratio, circularity, equivalent diameter
• HSV color histogram features capturing discoloration and color uniformity
• GLCM texture features (contrast, homogeneity, energy) for surface and damage cues
• SVM (RBF kernel) and Random Forest classifiers in scikit-learn with StandardScaler normalization
• Sample-level grading by majority vote over per-kernel predictions, with configurable grade classes
• Batch mode that grades many samples and writes a per-sample grade report
• Evaluation notebook computing accuracy, confusion matrix and per-class F1 on your validation samples

• Complete source code (segmentation, feature extraction, training, grading, reporting, demo app)
• Jupyter training and evaluation notebook (buyer-run procedure: label reference samples, train, evaluate)
• Project report PDF (background, feature design, methodology, evaluation, error analysis)
• PPT presentation for final review
• Viva Q&A preparation document (GLCM, SVM, color spaces, segmentation, grading logic)
• Setup guide (environment, imaging setup, labeling reference samples, grading your own samples)

• The classifier is only as good as the labeled reference samples you provide: it cannot recognize defects or varieties absent from the training labels, and the report documents this openly.
• Grading depends on imaging discipline — inconsistent lighting, background or camera height shifts feature distributions, so the setup guide fixes the imaging protocol.
• The system measures appearance features only; moisture content, internal pests and chemical quality are outside a camera's view and need laboratory tests.