Tomato Leaf Disease Classification (Transfer Learning + Grad-CAM)

This repository contains the training and inference code for the research:

"Comparative Study of Lightweight and Deep CNN Architectures for Tomato Leaf Disease Classification Using Transfer Learning and Grad-CAM Visualization"

Overview

This project evaluates multiple Convolutional Neural Network (CNN) architectures for automated tomato leaf disease classification using a transfer learning approach and Grad-CAM for model interpretability.

Architectures evaluated:

• ResNet50 (deep CNN)
• MobileNetV2 (lightweight CNN)
• EfficientNetB0 (balanced architecture)

The study analyzes the trade-offs between classification performance, computational efficiency, and interpretability.

The dataset used is the PlantVillage tomato leaf subset, consisting of 18,160 RGB images across 10 classes (nine diseases + one healthy class).

---

Repository Structure

├── Tugas2_Visi_Komputer_Kelompok_10.ipynb            # Training notebook
├── Tugas2_Visi_Komputer_inferensi_Kelompok_10.ipynb  # Inference notebook
└── README.md

---

Training Scenario

All models were trained under identical settings to ensure a fair comparison.

Data Split

• Training: 80%
• Validation: 10%
• Testing: 10%

Preprocessing

• Image resizing to 224 × 224
• Architecture-specific preprocessing aligned with ImageNet
• Data augmentation: horizontal flip, small rotations, and slight zoom

Training Configuration

• Optimizer: Adam
• Batch size: 32
• Epochs: 30
• Loss function: Categorical Cross-Entropy
• Early stopping to reduce overfitting

Transfer Learning Schemes

1. Freeze-all layers (train classifier only)
2. Fine-tuning with 20% unfrozen layers
3. Fine-tuning with 40% unfrozen layers

Grad-CAM was applied to the best-performing model to verify that predictions focus on disease-relevant regions.

---

Results

• EfficientNetB0 with 40% fine-tuning achieved the highest classification performance across evaluation metrics.
• MobileNetV2 delivered the fastest inference time, making it suitable for mobile or edge deployment.
• ResNet50 provided stable performance but required higher computational cost.

Grad-CAM visualizations showed that the model consistently focuses on biologically meaningful regions such as lesions, discoloration patterns, and infected areas.

---

Conclusion

This study demonstrates that model selection should depend on deployment needs:

• EfficientNetB0 is preferred when accuracy and interpretability are the primary goals.
• MobileNetV2 is more suitable for real-time applications requiring low latency.

Although EfficientNetB0 is parameter-efficient, it exhibited longer inference times, highlighting that parameter efficiency does not always translate to runtime efficiency.

Future work may explore model optimization techniques, transformer-based architectures, and evaluation on real-world datasets with more complex backgrounds.