Plant-Disease-Classification

A ResNet50-DCBAM Framework for Plant Disease Classification on the PlantVillage Dataset

---

Project Overview

This repository provides a reproducible implementation of a deep learning framework for plant disease classification.

The project investigates an architecture based on a pretrained ResNet50 backbone enhanced with DCBAM (Dual Channel & Spatial Attention Module) for image-based plant disease recognition.

The main objective is to improve discriminative feature learning by combining:

• ImageNet pretrained feature extraction
• ResNet50 hierarchical feature representation
• Channel and spatial attention refinement using DCBAM
• Flexible multi-level feature selection
• Efficient classification using global average pooling

The framework is designed for controlled experiments and reproducible research.

Overview of the proposed ResNet50-DCBAM framework for plant disease classification.

---

Key Features

• ResNet50 backbone with pretrained ImageNet weights
• DCBAM attention integration
  • Channel attention
  • Spatial attention
• Flexible feature-level selection (c2, c3, c4, c5)
• Official PlantVillage train/test split support
• Deterministic, stratified train/validation splitting
• Optional dataset integrity validation
• Data augmentation during training
• Automatic checkpoint management
• Validation-based model selection:
  • Best Validation Loss
  • Best Validation Accuracy
• Classification report generation
• Normalized confusion matrix visualization
• Experiment directory organization

---

Project Structure

Plant-Disease-Classification/
│
├── data/
│   └── PlantVillage/
│       ├── raw/
│       └── splits/
│
├── docs/
│   ├── figures/
│   │   └── architecture_overview.png
│   │
│   └── project_structure.md
│
├── experiments/
│   ├── checkpoints/
│   ├── logs/
│   ├── plots/
│   └── results/
│
├── src/
│   ├── config/
│   │   └── config.py
│   │
│   ├── datasets/
│   │   └── plant_dataset.py
│   │
│   ├── models/
│   │   ├── attention/
│   │   │   └── dcbam.py
│   │   │
│   │   ├── backbone/
│   │   │   └── resnet_dcbam.py
│   │   │
│   │   ├── classifier.py
│   │   └── plant_model.py
│   │
│   ├── training/
│   │   ├── train.py
│   │   └── test.py
│   │
│   └── utils/
│       ├── plot_logs.py
│       └── export_requirements.py
│
├── README.md
├── requirements.txt
├── LICENSE
└── .gitignore

---

Dataset

The project uses the official PlantVillage Dataset training and test splits.

To prevent test-set leakage during model selection, the official training split is further divided into training (90%) and validation (10%) subsets using a deterministic, stratified split.

The official test split is kept separate from training, early stopping, and checkpoint selection, and is used for final evaluation.

Split	Images	Purpose
Training	39,236	Model training
Validation	4,360	Early stopping and checkpoint selection
Test	10,709	Final evaluation

Dataset structure:

data/
└── PlantVillage/
    ├── raw/
    │   └── color/
    └── splits/
        ├── color_train.txt
        ├── color_train_internal.txt
        ├── color_val.txt
        └── color_test.txt

The internal training and validation split files are generated automatically using a fixed random seed.

Each split file contains relative image paths, for example:

raw/color/Tomato___Late_blight/image_name.JPG

---

Installation

Clone the repository:

git clone https://github.com/AhmadJeddi/Plant-Disease-Classification.git
cd Plant-Disease-Classification

Install dependencies:

pip install -r requirements.txt

Recommended environment:

• Python >= 3.10
• PyTorch >= 2.5

---

Configuration

All experiments are controlled through:

src/config/config.py

Main configurable parameters:

• Dataset paths
• Image size
• Batch size
• Learning rate
• Training epochs
• Weight decay
• Early stopping patience
• Feature extraction level
• Pretrained backbone option
• Experiment directories

Example:

IMAGE_SIZE = 224

FEATURE_LEVEL = "c5"

PRETRAINED = True

BATCH_SIZE = 32

---

Model Architecture

The proposed model consists of three main components:

Input Image
     |
     v
ResNet50 Backbone
     |
     +--> DCBAM Attention Refinement
     |
     v
Feature Selection (c2/c3/c4/c5)
     |
     v
Global Average Pooling
     |
     v
Linear Classifier
     |
     v
Disease Class Prediction

ResNet50 Backbone

The backbone extracts hierarchical visual representations:

• c2
• c3
• c4
• c5

The selected feature map is passed to the classifier.

DCBAM Attention

DCBAM improves feature representation through:

• Channel attention:
  • Adaptive average pooling
  • Adaptive max pooling
  • Shared MLP transformation
• Spatial attention:
  • Channel-wise average pooling
  • Channel-wise max pooling
  • Convolution-based spatial refinement

---

Training

Run training:

python src/training/train.py

Training includes:

• Deterministic, stratified 90/10 train-validation split
• Optional image integrity validation before training
• Cross Entropy Loss
• Adam optimizer
• Data augmentation
• Validation-based checkpoint selection
• Early stopping based on validation loss
• Automatic checkpoint saving
• Training curve visualization

Generated checkpoints:

experiments/
└── checkpoints/
    ├── best_loss.pth
    └── best_acc.pth

---

Evaluation

Run evaluation:

python src/training/test.py

Evaluation provides:

• Classification report
• Precision
• Recall
• F1-score
• Normalized confusion matrix
• Saved experiment results

Results are stored in:

experiments/
└── results/

Visualizations:

experiments/
└── plots/

---

Experiments

Two checkpoints are saved based on validation performance:

• Best Validation Loss: Selected using the lowest validation loss.
• Best Validation Accuracy: Selected using the highest validation accuracy.

Both checkpoints are evaluated using the same test protocol.

---

Results

The model was evaluated on the official PlantVillage test split, containing 10,709 images across 38 classes, after training with an independent validation subset.

The following results correspond to the checkpoint selected using the lowest validation loss.

Metric	Score
Accuracy	99.65%
Macro Precision	99.37%
Macro Recall	99.58%
Macro F1-score	99.47%
Weighted F1-score	99.66%

These results were obtained using the revised evaluation protocol, where the test split is not used for early stopping or checkpoint selection.

---

Reproducibility

The project provides:

• Fixed random seed
• Centralized experiment configuration
• Deterministic, stratified train/validation splitting
• Separate test set for final evaluation
• Consistent class mapping across dataset splits
• Optional dataset preflight validation
• Saved model checkpoints
• Automated training and evaluation logging

---

Acknowledgements

This project builds upon the ResNet50-DCBAM implementation from:

ResNet-DCBAM

The original implementation was integrated into this framework with improved documentation, project-specific organization, and adaptation for the Plant Disease Classification task.

---

License

This project is licensed under the MIT License.

© Ahmad J. Zahed — Please keep the copyright notice and attribution when using or modifying this code. For full details, see the LICENSE file.