Dataset 1 : /kaggle/input/corn-or-maize-leaf-disease-dataset/data

🌽 Maize Leaf Disease Classification

---

📁 Dataset Info

• Dataset Path: /kaggle/input/corn-or-maize-leaf-disease-dataset/data
• Detected Classes:
  • 🌱 Blight
  • 🍂 Common Rust
  • 🍁 Gray Leaf Spot
  • ✅ Healthy

🔄 Dataset Split

Split	Percentage
🏋️‍♂️ Training	80%
🧪 Validation	10%
🧾 Test	10%

---

🤖 Project Objective

This program classifies maize leaf images into the above four categories using:

• 📷 CNN with MobileNetV2
• 🧠 CBAM (Convolutional Block Attention Module) for better feature attention
• 💡 Support Vector Machine (SVM) on extracted features to enhance classification performance

---

🧠 CBAM Attention Module

CBAM helps the network focus on important regions of the image.

📌 Structure

• 🔴 Channel Attention
  GlobalAvgPool + GlobalMaxPool → MLP → Add + Sigmoid

• 🔵 Spatial Attention
  Channel-wise mean/max → Concat → Conv2D → Sigmoid

✅ This is implemented in the function cbam_block(input_feature)

---

🧬 Full Architecture

Input → MobileNetV2 → CBAM → GAP → Dense(256) → BN → Dropout(0.5) ↓ Dense(128) → BN → Dropout(0.5) ↓ Dense(4, softmax)

---

🏋️‍♂️ Training Results

Metric	Value
🎯 Accuracy	97.16%
📉 Final Loss	0.4247
⏱️ Epochs	~41

📈 Training Graph: Accuracy and loss plotted across epochs.

---

📊 Evaluation (CNN Output)

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

📊 Confusion Matrix: Displayed using Seaborn heatmap.

---

✅ Final Summary

🔍 Component	🔢 Result/Value
🎯 Deep Model Accuracy	97.16%
🧠 SVM Accuracy	98.11%
🧩 Base Architecture	MobileNetV2 + CBAM
📐 Feature Dimension	128
🏋️ Training Images	3348
🧾 Test Images	423
🏆 Best Classifier	SVM on features
❌ Misclassifications	Minimal (mostly in Gray Leaf Spot)

Dataset 2: /kaggle/input/maizeleaf/MaizeLeafDataset

🌽 Maize Leaf Disease Classification – Advanced (5-Class)

---

1. Dataset and Preprocessing

📂 Dataset Structure

• Source: /kaggle/input/maizeleaf/MaizeLeafDataset
• Classes:
  • 🍂 Common Rust
  • 🍁 Gray Leaf Spot
  • ✅ Healthy
  • 🌫️ Northern Leaf Blight
  • ❌ Not Maize Leaf

📸 Sample Visualization

• 4 random images per class displayed using matplotlib to ensure data integrity.

🔀 Dataset Splitting

Split	Percentage	Destination Folder
🏋️‍♂️ Train	80%	/kaggle/working/split_data/train/
🧪 Validation	10%	/kaggle/working/split_data/val/
🧾 Test	10%	/kaggle/working/split_data/test/

---

2. Image Data Generators (Augmentation)

• Training Generator includes:

  • rescale, rotation_range, horizontal_flip, vertical_flip, brightness_range, zoom_range, fill_mode

• Validation & Test:

  • Only rescale

• Output Shape: (224, 224, 3)

• Labels: One-hot encoded

---

3. 🧠 CBAM Attention Mechanism

The Convolutional Block Attention Module (CBAM) improves the model’s ability to focus on relevant leaf regions.

🔍 Module Structure:

• 🔴 Channel Attention:
  GlobalAvgPool + GlobalMaxPool → MLP → Add → Sigmoid
• 🔵 Spatial Attention:
  Mean + Max → Concatenate → Conv2D → Sigmoid

💡 Integrated via cbam_block(input_feature) into the main model.

---

4. 🏗 Model Architecture

📐 Structure:

Input → MobileNetV2 → CBAM → GAP → Dense(256) → BN → Dropout(0.5) ↓ Dense(128) → BN → Dropout(0.5) ↓ Dense(5, softmax)

• 🧠 Backbone: MobileNetV2 (include_top=False, pretrained on ImageNet)
• 🔧 Trainable Parameters: ~3M
• 🧪 Activation: softmax
• 📉 Loss: CategoricalCrossentropy (with label_smoothing=0.1)
• ⚙️ Optimizer: Adam (lr=0.001)

🧬 Callbacks:

• EarlyStopping(patience=5)
• ReduceLROnPlateau

---

5. 🏋️ Training Pipeline

• 📈 Training completed in ~45 epochs
• ✅ Final Validation Accuracy: ~99.66%
• 💾 Model saved as: final_trained_model.h5

---

6. 📊 Evaluation (CNN)

📉 Test Results:

• Loss: ~0.43
• Accuracy: ~99.00%

📌 Classification Metrics:

Metric	Value
🎯 Precision (weighted)	0.9903
🔁 Recall (weighted)	0.9899
🧠 F1-Score (weighted)	0.9899

🔍 Confusion Matrix:

• Visualized using Seaborn heatmap
• Minimal misclassification across all 5 classes

---

✅ 7. 📦 Feature Extraction (CNN + SVM Hybrid)

• 🧬 Output: 128-dimensional feature vector per image (from CBAM-enhanced MobileNetV2)

📌 Feature Shapes

Set	Shape
🏋️ Train	(7079, 128)
🧾 Test	(891, 128)

---

📈 Performance

🧪 Metric	📊 Value
🎯 Test Accuracy	99.44%
🧠 F1, Precision, Recall	> 0.98 for all classes

🚀 SVM slightly outperforms the CNN softmax classifier!

---

✅ Conclusion & Summary

🧩 Component	💡 Description
🏗 Model Backbone	MobileNetV2 (ImageNet pre-trained) + CBAM
🧠 Classification	Softmax and SVM (on extracted 128-dim features)
🖼️ Input Size	224x224 RGB
🔄 Data Augmentation	✅ Applied to training set
🔎 CBAM	✅ Channel + Spatial attention
🎯 Final Accuracy	~99.4% (combined CNN + SVM performance)
🌽 Use Case	Maize Leaf Disease Classification
🚀 Deployment Ready	✅ Saved model (.h5) and feature extractor exported