Medicinal Plant Identification from Leaves using CNN

Ayurvedic practitioners, herbal product makers and students all face the same problem: many medicinal plants look alike to non-experts, and misidentification has real consequences. This project trains a MobileNetV3-Large classifier on about 2,800 leaf photographs across six medicinal species, with green-hue segmentation isolating the leaf so the model learns leaf morphology instead of backgrounds. MobileNetV3 is chosen deliberately so the model runs offline on a phone in the field. Students learn leaf-image preprocessing, augmentation that respects leaf shape, and mobile deployment via TFLite. The demo identifies sample leaf photos with representative predictions and usage notes — with an explicit educational-use disclaimer.

1. Leaf photographs are collected from herbal gardens and nurseries and labelled by species.
2. Green-hue segmentation masks isolate the leaf from the background in each photo.
3. Masked leaves are resized to 224×224, normalized and split into stratified train/validation sets.
4. MobileNetV3-Large, pre-trained on ImageNet, is fine-tuned with a new classification head.
5. Augmentation uses rotation, flip and perspective warp to simulate varied photo angles.
6. Training uses AdamW with early stopping; the held-out split gives accuracy and per-species metrics.
7. The model is exported to TFLite; the web demo identifies sample leaf photos with representative predictions.

Tech stack:

• Python 3, TensorFlow/Keras
• MobileNetV3-Large
• OpenCV (leaf segmentation)
• NumPy, Matplotlib, scikit-learn
• Jupyter Notebook (training)
• HTML/CSS/JS identification demo
• TFLite export

Dataset & model details

• Dataset: Custom curated medicinal-leaf set — ~2,800 leaf photographs from herbal gardens and nurseries in the Pune region, labelled into 6 species: Tulsi, Neem, Aloe vera, Moringa, Ashwagandha, Giloy; stratified 80/20 split.
• Task: 6-class image classification; input = 224×224×3 segmented leaf, output = species probabilities.
• Model: MobileNetV3-Large backbone (ImageNet pre-trained) → GlobalAveragePooling → Dropout(0.25) → Dense(6, softmax); ~5.4M parameters; green-hue segmentation upstream.
• Metrics: Species accuracy ~94% (design target), top-2 accuracy ~98% (design target), per-species precision/recall. No accuracy is claimed as measured until the training run is executed for the order.

• Six-species leaf classifier Tulsi, Neem, Aloe vera, Moringa, Ashwagandha, Giloy — common Indian medicinal plants, with per-species metrics.
• Leaf segmentation Green-hue masking isolates the leaf from the background before classification, removing background bias.
• Curated leaf dataset About 2,800 leaf photographs from herbal gardens and nurseries, labelled by species.
• Phone-ready model MobileNetV3-Large exported to TFLite (~3.4 MB) for fully offline field identification.
• Traditional-use notes Each prediction shows documented traditional uses — educational reference, clearly disclaimed.
• Interactive identification demo Single-file web app: pick a sample leaf photo and see the predicted species with probability bars.
• Viva kit Report PDF, PPT and Q&A covering leaf preprocessing, MobileNetV3, augmentation and TFLite deployment.

• Curated, labelled medicinal-leaf dataset with split files
• Segmentation + training Jupyter notebooks
• Trained MobileNetV3 model + TFLite export
• Interactive identification web demo with sample photos
• Confusion matrix and per-species metrics
• Traditional-use reference notes
• Project report PDF, PPT presentation
• Viva Q&A preparation document

• Educational and prototype use only — never use predictions for actual medicinal consumption decisions.
• Some species have lookalike non-medicinal relatives; the report documents the confusion pairs.
• 94% is a design target, stated honestly — the report documents the actual achieved figure.
• Damaged, dried or partial leaves degrade accuracy; fresh whole leaves are recommended.
• Only the six covered species are supported; new species need new labelled data.