Plant Disease Detection using CNN (PlantVillage)

Crop diseases destroy a significant share of yields every season, and early visual diagnosis is still the main line of defense for most farmers. Plant pathologists cannot be everywhere, so an automated leaf-image classifier that runs on a phone photo is genuinely useful — but most student builds of this idea train on tiny, clean datasets and collapse on real field photos. This project does it properly: a ResNet-50 backbone fine-tuned on the full PlantVillage corpus (54,305 images across 38 crop-disease classes), with augmentation that simulates field conditions, a held-out test evaluation with per-class metrics, and a working web demo that classifies uploaded leaf photos and explains the result. Every design choice — backbone, input size, augmentation, class handling — is documented so the viva discussion stays on solid ground.

1. The 54,305 PlantVillage images are loaded and split with stratification into train, validation and test sets.
2. Images are resized to 256×256 and normalized with ImageNet statistics; augmentation (rotation ±20°, flips, brightness, blur) is applied on the training split.
3. A ResNet-50 backbone pretrained on ImageNet is loaded; its 1000-class head is replaced with a 38-class classifier and dropout.
4. The head is trained first with the backbone frozen, then the top blocks are unfrozen for fine-tuning with a lower learning rate.
5. Training runs up to 40 epochs with early stopping on validation accuracy; the best checkpoint is kept.
6. The held-out test set is evaluated once: overall accuracy, per-class precision/recall/F1 and the confusion matrix are exported for the report.
7. In the web demo, an uploaded leaf photo goes through the identical preprocessing pipeline and the saved model returns the prediction with confidence scores and treatment guidance.

Tech stack:

• Python 3, TensorFlow/Keras
• ResNet-50 (transfer learning)
• PlantVillage dataset
• NumPy, Matplotlib, scikit-learn (metrics)
• Jupyter Notebook (training & evaluation)
• HTML/CSS/JavaScript (diagnosis demo)

Dataset & model details

• Dataset: PlantVillage — 54,305 labeled leaf images, 38 classes (14 crop species × healthy/diseased states), released by Hughes & Salathé (2015). Open-access via the PlantVillage project.
• Task: 38-class image classification; input = 256×256×3 leaf photo, output = probability distribution over crop-disease classes.
• Model: ResNet-50 backbone (ImageNet-pretrained) + custom head: global average pooling → Dense(512, ReLU) → Dropout(0.5) → Dense(38, softmax). Approximately 24M parameters, backbone partially frozen.
• Metrics: Test top-1 accuracy ~94% (design target for the built-to-order training run), per-class precision/recall/F1, confusion matrix. No accuracy is claimed as measured until the training run is executed for the order.

• [ResNet-50 transfer learning] ImageNet-pretrained backbone fine-tuned on PlantVillage, with the classifier head replaced for 38 crop-disease classes — full rationale documented.
• [Leaf-diagnosis web demo] Upload any leaf photo; the app preprocesses it (256×256, normalized) and returns the predicted disease with per-class confidence bars.
• [Treatment guidance] Each predicted class maps to a cause and recommended action (cultural control, fungicide class), so the demo output is agronomically useful.
• [Field-condition augmentation] Rotation, flips, brightness jitter and blur simulate phone-camera field photos during training.
• [Full training notebook] Data loading, stratified splits, augmentation, model definition, training loop and evaluation in one reproducible notebook.
• [Per-class evaluation] Precision, recall and F1 for all 38 classes plus a confusion matrix, highlighting the genuinely confusable disease pairs.
• [Training curves] Accuracy and loss plots with interpretation notes for the report and viva.

• Complete training & evaluation Jupyter notebook
• Trained ResNet-50 model file (.h5) with preprocessing code
• Leaf-diagnosis web demo with treatment guidance
• Per-class metrics, confusion matrix and training-curve plots
• Project report PDF (background, CNN theory, methodology, results)
• PPT presentation for final review
• Viva Q&A preparation document (transfer learning, augmentation, overfitting, metrics)