Potato Disease Classification

An end-to-end computer vision system that classifies potato leaves as Healthy, Early Blight, or Late Blight from a single photo. Trained on the PlantVillage dataset with a convolutional neural network, packaged behind a FastAPI inference endpoint, and served through a lightweight Streamlit interface intended for farmers and field workers.

Late diagnosis of potato foliar diseases is a major yield-loss driver for smallholder farmers. Existing tooling is either lab-bound or requires strong connectivity. The goal was to ship a fast, offline-friendly classifier that runs on a phone photo and returns actionable output in under a second.

• 01Data pipeline: PlantVillage images resized to 256×256, augmented with random flips, rotations, and zoom to reduce overfitting on background artifacts.
• 02Model: a compact CNN (3 conv-pool blocks → dense head) trained with sparse categorical cross-entropy and Adam. Class weights tuned to handle mild dataset imbalance.
• 03Serving: model exported to TensorFlow SavedModel format and loaded once per FastAPI worker; predictions returned as JSON with class probabilities.
• 04UI: Streamlit front-end for drag-and-drop images that calls the FastAPI endpoint over HTTP.

• Python
• TensorFlow
• CNN
• FastAPI
• Streamlit

• Real-time inference under 300ms per image on CPU
• Confidence score returned alongside the predicted class
• Image preprocessing handled server-side so mobile clients stay thin
• Swagger docs auto-generated from FastAPI schemas for easy integration

• Class imbalance between healthy and late-blight samples was hurting recall — solved by class weighting and targeted augmentation.
• Streamlit and FastAPI had to run as separate processes; introduced a small shared config module to avoid drift between the two.
• Cold-start latency on first inference — solved by warming the model with a dummy tensor at startup.

• 95%+ accuracy on held-out test set across all three classes
• Sub-300ms average inference on commodity CPU
• Reusable inference module packaged for other crop-disease datasets

• A small, well-augmented CNN beats a large under-regularized one on niche image datasets.
• Serving and modelling are two very different disciplines — keeping them in separate modules paid off during iteration.
• Streamlit is great for demos but the real product surface belongs behind a proper HTTP API.