A Lightweight Attention-Enhanced CNN-SVM Hybrid Approach for Maize Leaf Disease Classification

A hybrid deep learning framework integrating MobileNetV2, Convolutional Block Attention Module (CBAM), and Support Vector Machine for robust maize leaf disease classification.

---

Table of Contents

• Overview
• Key Features
• Architecture
• Datasets
• Methodology
• Installation
• Usage
• Results
• Project Structure
• Citation
• License

---

Overview

This repository presents a lightweight yet high-performance hybrid classification framework for automated maize leaf disease detection. The proposed architecture leverages a MobileNetV2 backbone enhanced with Convolutional Block Attention Module (CBAM) for discriminative feature extraction, followed by a Support Vector Machine (SVM) classifier operating on the learned feature representations. The system is designed to operate efficiently on resource-constrained environments while achieving state-of-the-art classification accuracy across multiple disease categories.

---

Key Features

Feature	Description
Lightweight Backbone	MobileNetV2 with ImageNet pre-trained weights for efficient feature extraction
Attention Mechanism	CBAM (Channel + Spatial Attention) for enhanced focus on disease-specific regions
Hybrid Classification	Dual-stage pipeline: CNN feature extraction + SVM classification
Multi-Dataset Support	Compatible with 4-class and 5-class maize leaf disease datasets
Data Augmentation	Comprehensive augmentation pipeline for improved generalization
Model Persistence	Trained models and feature extractors exported in .h5 format

---

Architecture

Core Pipeline

Input (224 x 224 x 3)
    |
    v
[ MobileNetV2 Backbone ]  (ImageNet Pre-trained, include_top=False)
    |
    v
[ CBAM Attention Block ]  (Channel Attention + Spatial Attention)
    |
    v
[ Global Average Pooling ]
    |
    v
[ Dense(256) -> BatchNorm -> Dropout(0.5) ]
    |
    v
[ Dense(128) -> BatchNorm -> Dropout(0.5) ]  <-- Feature Extraction Layer
    |
    v
[ Dense(4/5, Softmax) ]  <-- CNN Classifier

CBAM Attention Module

The Convolutional Block Attention Module is integrated sequentially after the MobileNetV2 backbone to refine feature maps through:

1. Channel Attention

   • Global Average Pooling + Global Max Pooling
   • Shared MLP with reduction ratio
   • Element-wise addition and Sigmoid activation

2. Spatial Attention

   • Channel-wise mean and max pooling
   • Concatenation along channel dimension
   • Convolutional layer (7x7) and Sigmoid activation

---

Datasets

Dataset 1: Corn/Maize Leaf Disease Dataset (4-Class)

Attribute	Value
Source	/kaggle/input/corn-or-maize-leaf-disease-dataset/data
Classes	Blight, Common Rust, Gray Leaf Spot, Healthy
Training	80% (3,348 images)
Validation	10%
Test	10% (423 images)

Dataset 2: MaizeLeafDataset (5-Class)

Attribute	Value
Source	/kaggle/input/maizeleaf/MaizeLeafDataset
Classes	Common Rust, Gray Leaf Spot, Healthy, Northern Leaf Blight, Not Maize Leaf
Training	80% (7,079 images)
Validation	10%
Test	10% (891 images)

Preprocessing Configuration

IMAGE_SIZE = (224, 224)
BATCH_SIZE = 32
COLOR_MODE = 'rgb'

# Training Augmentation
rescale=1./255,
rotation_range=20,
horizontal_flip=True,
vertical_flip=True,
brightness_range=[0.8, 1.2],
zoom_range=0.2,
fill_mode='nearest'

# Validation / Test
rescale=1./255

---

Methodology

Training Configuration

Parameter	Value
Optimizer	Adam (learning_rate=0.001)
Loss Function	Categorical Crossentropy (label_smoothing=0.1)
Activation	Softmax
Trainable Parameters	~3 Million
Callbacks	EarlyStopping (patience=5), ReduceLROnPlateau

Feature Extraction for SVM

The 128-dimensional feature vectors are extracted from the penultimate dense layer of the trained CNN. These embeddings serve as input to an SVM classifier with the following characteristics:

• Feature Dimension: 128
• Feature Shape (Train): (7,079, 128)
• Feature Shape (Test): (891, 128)

---

Installation

Prerequisites

• Python >= 3.8
• TensorFlow >= 2.8
• Keras >= 2.8
• scikit-learn
• NumPy, Pandas, Matplotlib, Seaborn

Setup

# Clone the repository
git clone https://github.com/Soyebsoyeb/Maize_Disease_Classification.git
cd Maize_Disease_Classification

# Create virtual environment
python -m venv venv
source venv/bin/activate  # On Windows: venv\Scripts\activate

# Install dependencies
pip install -r requirements.txt

Dependencies

tensorflow>=2.8.0
keras>=2.8.0
scikit-learn>=1.0.0
numpy>=1.21.0
pandas>=1.3.0
matplotlib>=3.4.0
seaborn>=0.11.0
pillow>=8.3.0

---

Usage

Training the Model

from model import build_model
from data_loader import load_data

# Load dataset
train_gen, val_gen, test_gen = load_data(dataset_path="/path/to/dataset")

# Build model with CBAM attention
model = build_model(
    input_shape=(224, 224, 3),
    num_classes=5,
    use_cbam=True,
    backbone="mobilenetv2"
)

# Train
history = model.fit(
    train_gen,
    validation_data=val_gen,
    epochs=50,
    callbacks=[early_stopping, reduce_lr]
)

# Save model
model.save("final_trained_model.h5")

Feature Extraction and SVM Classification

from feature_extractor import extract_features
from sklearn.svm import SVC

# Extract 128-dim features
X_train_features = extract_features(model, train_gen, layer_name="dense_128")
X_test_features = extract_features(model, test_gen, layer_name="dense_128")

# Train SVM classifier
svm = SVC(kernel='rbf', C=1.0, gamma='scale')
svm.fit(X_train_features, y_train)

# Evaluate
accuracy = svm.score(X_test_features, y_test)
print(f"SVM Test Accuracy: {accuracy * 100:.2f}%")

Inference on New Images

from inference import predict_disease

prediction = predict_disease(
    image_path="path/to/leaf_image.jpg",
    model_path="final_trained_model.h5",
    class_names=["Blight", "Common_Rust", "Gray_Leaf_Spot", "Healthy"]
)
print(f"Predicted Class: {prediction}")

---

Results

Dataset 1: 4-Class Classification

CNN Performance

Metric	Value
Test Accuracy	97.16%
Final Loss	0.4247
Epochs	~41

CNN Classification Report

Class	Precision	Recall	F1-Score	Support
Blight	0.96	0.93	0.95	116
Common Rust	1.00	1.00	1.00	132
Gray Leaf Spot	0.87	0.93	0.90	58
Healthy	1.00	1.00	1.00	117
Weighted Avg	0.97	0.97	0.97	423

Hybrid CNN-SVM Performance

Component	Result
SVM Accuracy	98.11%
Feature Dimension	128
Best Classifier	SVM on extracted features

---

Dataset 2: 5-Class Classification

CNN Performance

Metric	Value
Validation Accuracy	99.66%
Test Accuracy	99.00%
Test Loss	0.43
Epochs	~45

CNN Classification Report

Metric	Value
Weighted Precision	0.9903
Weighted Recall	0.9899
Weighted F1-Score	0.9899

Hybrid CNN-SVM Performance

Metric	Value
Test Accuracy	99.44%
F1 / Precision / Recall	> 0.98 (all classes)

Comparative Summary

Component	Specification
Model Backbone	MobileNetV2 (ImageNet pre-trained) + CBAM
Classification	Softmax and SVM (128-dim features)
Input Size	224 x 224 RGB
Data Augmentation	Applied to training set
CBAM Integration	Channel + Spatial attention
Final Accuracy	~99.4% (combined CNN + SVM)
Deployment Ready	Saved model (.h5) and feature extractor exported

---

Project Structure

Maize_Disease_Classification/
|
|-- data/
|   |-- corn-or-maize-leaf-disease-dataset/    # Dataset 1 (4-class)
|   |-- maizeleaf/MaizeLeafDataset/             # Dataset 2 (5-class)
|
|-- models/
|   |-- cbam.py                                 # CBAM attention module
|   |-- mobilenetv2_cbam.py                     # Full architecture
|   |-- final_trained_model.h5                  # Saved trained model
|
|-- utils/
|   |-- data_loader.py                          # Data loading and augmentation
|   |-- feature_extractor.py                    # Feature extraction pipeline
|   |-- visualizations.py                       # Confusion matrix and plots
|
|-- notebooks/
|   |-- training_pipeline.ipynb                 # End-to-end training notebook
|   |-- svm_evaluation.ipynb                    # SVM classification notebook
|
|-- results/
|   |-- confusion_matrix.png
|   |-- training_curves.png
|   |-- classification_report.txt
|
|-- README.md
|-- requirements.txt
|-- LICENSE

---

Citation

If you use this work in your research, please cite:

@article{maize_disease_classification_2024,
  title={A Lightweight Attention-Enhanced CNN-SVM Hybrid Approach for Maize Leaf Disease Classification},
  author={Soyeb},
  journal={GitHub Repository},
  year={2024},
  url={https://github.com/Soyebsoyeb/Maize_Disease_Classification}
}

---

License

This project is licensed under the MIT License - see the LICENSE file for details.

---

Developed with precision for agricultural AI applications.

Report Bug · Request Feature