CPU-Constrained Evaluation of Modern Hybrid Transformers and CNN Architectures for Tomato Leaf Disease Classification Using the PlantVillage Datase 🌱

This project evaluates modern CNNs and hybrid Transformer architectures — ResNet-50, ConvNeXt-Tiny, and FastViT-T8 — for tomato leaf disease classification on the PlantVillage dataset.

The focus is on CPU-constrained environments to make plant disease detection feasible in resource-limited regions where GPU access is rare.

---

📌 Features

• Dataset: Tomato subset of PlantVillage (16,012 images, 10 classes).

• Architectures: ResNet-50, ConvNeXt-Tiny, FastViT-T8 (via timm).

• Training:

  • Two-phase transfer learning (frozen backbone → full fine-tuning)
  • Mixed-precision support (AMP)
  • Early stopping & checkpointing

• Evaluation: Accuracy, precision, recall, F1-score, confusion matrix, inference latency.

• Deployment: Designed for CPU-only training & inference.

---

🚀 Results (Validation)

Model	Params (M)	Accuracy	Weighted F1	Latency (s/img)
ResNet-50	25.6	99.50%	0.995	0.0352
ConvNeXt-Tiny	29.0	99.88%	0.9988	0.0508
FastViT-T8	27.4	99.66%	0.997	0.0219

• ConvNeXt-Tiny → Highest accuracy
• FastViT-T8 → Fastest inference
• ResNet-50 → Strong classical baseline

---

📂 Repository Structure

efficient-leaf-disease/
│── datasets/               # PlantVillage (tomato subset)
│── src/                    # Training & evaluation scripts
│   └── train.py            # Main training loop (works for all models)
│── trained_models/         # Saved checkpoints (.pth)
│── evaluation_results/     # Metrics JSON + confusion matrices
│── logs/                   # Training logs (CSV)
│── README.md

---

⚙️ Installation

git clone https://github.com/your-username/efficient-leaf-disease.git
cd efficient-leaf-disease
pip install -r requirements.txt

requirements.txt

torch
torchvision
timm
scikit-learn
matplotlib
seaborn
tqdm
numpy

---

🖥️ Usage

Training a model

python src/train.py --model resnet50
python src/train.py --model convnext_tiny
python src/train.py --model fastvit_t8

Evaluating best checkpoint

After training, evaluation metrics and confusion matrix will be saved in evaluation_results/.

Example:

cat evaluation_results/resnet50_metrics.json

---

📊 Outputs

• evaluation_results/<model>_metrics.json → Accuracy, F1, latency, per-class report
• evaluation_results/<model>_metrics.png → Confusion matrix plot
• trained_models/ → All checkpoints (last.pth, best.pth, per-epoch .pth)
• logs/<model>.csv → Training logs (loss & accuracy per epoch)

---

🌱 Applications

• Edge & mobile AI for agriculture
• Real-time tomato disease detection
• Low-resource decision support for farmers

---

📖 Citation

If you use this work, please cite: Rahman, O., et al. (2025) CPU-Constrained Evaluation of Modern Hybrid Transformers and CNN Architectures for Tomato Leaf Disease Classification Using the PlantVillage Dataset.