Crop Disease Classification for Edge Devices

This project demonstrates how to build and optimize a deep learning model for classifying crop diseases on resource-constrained edge devices. It utilizes a lightweight MobileNetV2 architecture and explores different quantization techniques to reduce model size and improve inference speed while maintaining high accuracy.

📌 Key Features

• Lightweight Architecture: Uses the MobileNetV2 model pre-trained on ImageNet for efficient feature extraction.
• Model Compression: Implements two post-training quantization techniques:
  1. Post-Training Quantization (PTQ): Calibrates the model using a representative dataset to convert weights and activations to 8-bit integers.
  2. Dynamic Range Quantization (DRQ): A simpler technique that quantizes weights to 8-bit integers while dynamically quantizing activations at inference time.
• Comprehensive Evaluation: Compares the full-precision model against the quantized models based on accuracy, F1-score, and file size.
• Data-Specific: Trained and evaluated on the Rice Leaf Disease Dataset.

📦 Prerequisites

Before running the code, ensure you have the following libraries installed:

• tensorflow
• tensorflow-model-optimization
• numpy
• scikit-learn
• matplotlib

🚀 Installation

You can install all the required packages using pip:

pip install tensorflow==2.15.0 tensorflow-model-optimization numpy scikit-learn matplotlib

📁 File Structure

The project assumes the following directory structure. Place the dataset in the root folder as shown.

.
├── rice_leaf_disease_dataset/
│   ├── train/
│   │   ├── Bacterial Leaf Blight/
│   │   ├── Brown Spot/
│   │   └── ...
│   └── validation/
│       ├── Bacterial Leaf Blight/
│       └── ...
├── leaf_disease_classification.ipynb              # The main script to train and quantize models
└── README.md

💻 Usage

1. Download the Dataset: Obtain the Rice Leaf Disease Dataset and place it in the project's root directory.
2. Update Path: In the main.py script, update the data_dir variable to the correct path of your dataset.
3. Run the Script: Execute the main script from your terminal.

python main.py

The script will perform the following steps:

1. Train the full-precision MobileNetV2 model.
2. Save the full-precision model as rice_leaf_fp32.h5.
3. Perform PTQ and save the quantized model as rice_leaf_quantized.tflite.
4. Perform DRQ and save the quantized model as rice_leaf_drq.tflite.
5. Print evaluation metrics and file sizes for all three models.

📊 Results

The following table summarizes the key metrics, showcasing the trade-off between model accuracy and size.

Model	Accuracy	Size (MB)	F1-Score (Macro)
Full-Precision	~94%	~9 MB	0.9374
PTQ (Quantized)	83%	~2.5 MB	0.8245
DRQ (Quantized)	92.23%	2.59 MB	0.9212

Analysis: The DRQ model provides a superior balance of size reduction and accuracy retention, making it the most suitable for edge deployment in this scenario.

🙏 Acknowledgements

• Dataset: Rice Leaf Disease Dataset hosted on Kaggle.