BeejX LeafModel (MobileNetV2)

Edge-AI Solution for Crop Disease Recognition (Offline & Mobile-First).

Quick Start

1. Setup

pip install -r requirements.txt

2. Prepare Data

Put your raw crop images in data/ and run the organization script:

python src/scripts/organize.py

Creates a flattened data_processed/ directory ready for training.

3. Train

python src/train.py

Outputs model.tflite to exports/ folder.

---

Architecture

We use MobileNetV2 (Quantized) for < 2.5MB model size, enabling real-time inference on low-end Android devices in rural areas.

• Read Full Architecture Docs regarding Input Size, Alpha, and Hyperparameters.

Tech Stack

• Core: TensorFlow 2.19.0, Keras
• Pipeline: Custom BeejXDataLoader with Real-time Augmentation (Rotation/Zoom).
• Optimization: Post-training Quantization (Float32 -> Int8).
• Handling Imbalance: Algorithmic Class Weighting (sklearn).

Data Version Control (DVC)

Used data version control --- data_processed folder:

dvc init
dvc add data_processed