Tea Leaf Quality Classification using CNN

Tea quality starts at the pluck: two leaves and a bud commands a premium price, while coarse or damaged leaves drag a whole batch down. In practice, grading plucked leaf is done by experienced buyers at the collection center, eyeballing each basket in seconds — fast but subjective, and impossible to keep consistent across buying days and seasons. Leaf images, on the other hand, carry exactly the visual cues graders use: color, size, uniformity, damage marks. This project builds a CNN classifier that learns those cues: a pretrained ResNet or EfficientNet backbone is fine-tuned on public tea-leaf image collections, so the model starts from strong general vision features and only needs to learn leaf-specific grading distinctions. Grad-CAM overlays make each prediction inspectable, showing which leaf regions pushed the model toward its grade — the visual explanation that turns a black-box classifier into something you can defend in a viva.

1. Plucked-leaf images are loaded, resized to the classifier's input size and normalized with OpenCV.
2. Training applies augmentation — rotation, flips, color jitter — so the model tolerates real collection-center photo variation.
3. A pretrained ResNet or EfficientNet backbone is fine-tuned on the labeled tea-leaf collections, with the classifier head retrained for the grade classes.
4. At inference, the model outputs a quality grade with a probability score for each uploaded leaf photo.
5. Grad-CAM overlays visualize the image regions that drove the prediction, giving an inspectable explanation per grade.
6. The evaluation notebook computes accuracy, F1 and the confusion matrix on the validation split and runs a misclassified-samples review.

Tech stack:

• Python 3.10, PyTorch (CNN training and inference)
• torchvision (pretrained ResNet and EfficientNet backbones)
• Public tea-leaf image collections (ImageFolder-style datasets)
• OpenCV (preprocessing, Grad-CAM overlays)
• Jupyter notebook (evaluation with accuracy and F1)
• NumPy, scikit-learn (metrics and confusion matrix)
• Flask demo app with leaf-image upload interface

• ResNet/EfficientNet CNN pretrained on ImageNet, fine-tuned on public tea-leaf image collections
• Quality-grade classification of plucked-leaf photographs
• Grad-CAM overlays explaining which leaf regions drove each grade
• Data augmentation pipeline (rotation, flip, color jitter) for leaf-image robustness
• Evaluation notebook computing accuracy, F1 and the confusion matrix on the validation split
• Class-wise performance breakdown showing which grades confuse with each other
• Demo app for uploading leaf photos and viewing grades with explanations
• Configurable grade classes matching the buyer's labeling scheme

• Complete source code (dataset loading, augmentation, fine-tuning, inference, demo app)
• Jupyter training and evaluation notebook (buyer-run: train, evaluate, review misclassifications)
• Project report PDF (background, transfer learning design, methodology, evaluation, error analysis)
• PPT presentation for final review
• Viva Q&A preparation document (transfer learning, Grad-CAM, confusion matrix, augmentation)
• Setup guide (environment, dataset download, training, grading your own leaf photos)

• The model can only grade leaves like the ones in its training collections: different tea varieties, growing regions or imaging conditions classify worse, and the report documents this coverage gap openly.
• Plucked-leaf appearance overlaps across adjacent grades, so the confusion matrix will show grade-pair confusions — the notebook reviews these misclassified samples explicitly.
• The system grades visual appearance only; liquor taste, aroma and chemical quality need human cupping and lab analysis, which no image classifier can replace.