🚀 Thrilled to unveil LeafLens: AI-powered disease diagnosis for smarter agriculture! 🌿🔍
After weeks of intense development, model training, and debugging, my team and I are excited to share our capstone project: LeafLens.
Every year, significant crop yields are lost to undiagnosed diseases. We wanted to build a tool that bridges the gap between complex computer vision and accessible technology for farmers and gardeners. LeafLens allows a user to upload a photo of a single leaf and instantly receive a diagnosis across 38 different disease classes for 14 crops.
What makes this project unique isn't just that it works—it's what’s happening under the hood. ⚙️
We didn't want to just rely on a black-box API. We designed a multi-model architecture that lets users choose their "AI Expert" and compare different machine learning paradigms:
🧠 The Specialist (Deep Learning): We fine-tuned a PyTorch EfficientNet-B3 CNN on the PlantVillage dataset. This model achieves state-of-the-art accuracy by analyzing the raw image pixels.
🕸️ The Pattern Seeker (Custom ML): This was the biggest technical challenge. We built a Neural Network entirely from scratch using just NumPy—no high-level Keras/PyTorch abstractions for the training loop!
📊 The Statistician (Probabilistic): A Gaussian Mixture Model (GMM) that diagnoses health based on statistical feature clusters.
The Coolest Technical Part? The Hybrid Pipeline. To make the custom Neural Net and GMM work, we used a fascinating transfer learning approach. We used the pre-trained PyTorch model as a feature extractor to pull out 1,536 complex visual traits from an image, compressed them down using Principal Component Analysis (PCA), and fed those compressed mathematical representations into our custom math models.
We wrapped the entire ensemble in a sleek Streamlit web application for easy deployment.