🌉 Bridge Defect Image Detection and Classification

A CNN-based image classification project that automatically detects and classifies bridge surface defects using deep learning.

---

📋 Project Overview

This project implements a deep learning pipeline to classify bridge surface defects into 6 categories using Convolutional Neural Networks (CNN). Two models are trained and compared:

• Custom CNN — built from scratch with 4 convolutional blocks
• MobileNetV2 — transfer learning with ImageNet pretrained weights

Model	Validation Accuracy	Parameters	Pretrained
Custom CNN	49.90%	720,806	No
MobileNetV2	66.60%	2,588,486	Yes (ImageNet)

---

🗂️ Dataset

Dataset: MultiClassifier Bridge Defect Dataset
Total Images: 2,411
Classes: 6
Image Size: 224 × 224 pixels (after preprocessing)
Split: 80% Train (1,932) / 20% Validation (479)

Class	Images	Percentage
Cracks	789	32.7%
No Defect	452	18.7%
Spalling	427	17.7%
Efflorescence	311	12.9%
General	264	10.9%
Scaling	168	7.0%

---

📁 Repository Structure

bridge-defect-classification/
│
├── main_notebook.ipynb          # Complete Jupyter notebook (all 15 cells)
│
├── multi_classifier_data/       # Dataset folder
│   └── MultiClassifier/
│       ├── cracks/
│       ├── efflorescence/
│       ├── general/
│       ├── no defect/
│       ├── scaling/
│       └── spalling/
│
├── outputs/
│   ├── figures/                 # All saved output figures
│   │   ├── class_distribution.png
│   │   ├── sample_images.png
│   │   ├── cnn_training_curves.png
│   │   ├── cnn_confusion_matrix.png
│   │   ├── mobilenet_training_curves.png
│   │   ├── mobilenet_confusion_matrix.png
│   │   ├── model_comparison.png
│   │   └── gradcam.png
│   └── models/                  # Saved trained models
│       ├── best_cnn.keras
│       └── best_mobilenet.keras
│
└── README.md

---

⚙️ Setup Instructions

1. Clone the Repository

git clone https://github.com/RohanRaj369/bridge-defect-classification.git
cd bridge-defect-classification

2. Create a Virtual Environment

python3.11 -m venv venv
source venv/bin/activate        # Mac/Linux
# venv\Scripts\activate         # Windows

3. Install Dependencies

pip install tensorflow scikit-learn matplotlib seaborn numpy pillow opencv-python pandas ipykernel jupyter

4. Add the Dataset

Place the dataset inside the project folder so the structure matches:

bridge-defect-classification/
└── multi_classifier_data/
    └── MultiClassifier/
        ├── cracks/
        ├── efflorescence/
        ...

5. Run the Notebook

jupyter notebook main_notebook.ipynb

Or open in VS Code and run cells one by one with Shift+Enter.

---

🧠 Model Architecture

Custom CNN

Input (224×224×3)
→ Block 1: Conv2D(32) + BN + Conv2D(32) + MaxPool + Dropout
→ Block 2: Conv2D(64) + BN + Conv2D(64) + MaxPool + Dropout
→ Block 3: Conv2D(128) + BN + Conv2D(128) + MaxPool + Dropout
→ Block 4: Conv2D(256) + BN + MaxPool + Dropout
→ GlobalAveragePooling2D
→ Dense(512) + BN + Dropout(0.5)
→ Dense(6, softmax)

MobileNetV2 (Transfer Learning)

Input (224×224×3)
→ MobileNetV2 base (frozen, pretrained on ImageNet)
→ GlobalAveragePooling2D
→ Dense(256) + BatchNorm + Dropout(0.4)
→ Dense(6, softmax)

---

📊 Results

MobileNetV2 Classification Report

Class	Precision	Recall	F1-Score
Cracks	0.76	0.75	0.75
Efflorescence	0.53	0.68	0.60
General	0.42	0.31	0.36
No Defect	0.88	0.70	0.78
Scaling	0.37	0.52	0.43
Spalling	0.62	0.66	0.64
Weighted Avg	0.66	0.65	0.65

Key Findings

• MobileNetV2 outperformed Custom CNN by +16.7% accuracy
• MobileNetV2 converged in just 8 epochs vs 30 for Custom CNN
• no defect and cracks were the easiest classes to classify
• general was the hardest class due to visual ambiguity
• Grad-CAM confirmed the model focuses on crack textures and surface patterns

---

🖼️ Output Figures

Figure	Description
class_distribution.png	Bar chart of images per class
sample_images.png	Sample grid from all 6 classes
cnn_training_curves.png	CNN accuracy & loss over epochs
cnn_confusion_matrix.png	CNN confusion matrix
mobilenet_training_curves.png	MobileNetV2 accuracy & loss over epochs
mobilenet_confusion_matrix.png	MobileNetV2 confusion matrix
model_comparison.png	Side-by-side comparison table
gradcam.png	Grad-CAM heatmap visualization

---

🛠️ Technologies Used

• Python 3.11
• TensorFlow / Keras 2.21.0
• scikit-learn — classification report, confusion matrix
• Matplotlib / Seaborn — visualizations
• OpenCV — Grad-CAM heatmap overlay
• Pillow — image loading
• NumPy / Pandas — data handling

---

📌 Notebook Structure

Cell	Description
1	Import libraries
2	Configuration and dataset path
3	Dataset exploration and class distribution
4	Sample image visualization
5	Preprocessing and data augmentation
6	Custom CNN model architecture
7	Train Custom CNN
8	Plot CNN training curves
9	Evaluate CNN — confusion matrix and report
10	MobileNetV2 transfer learning model
11	Train MobileNetV2
12	Evaluate MobileNetV2
13	Model comparison table
14	Grad-CAM visualization
15	Final summary

---

👤 Author

Rohan Raj
Machine Learning Course — Phase 2 Project
June 2026

---

📄 License

This project is submitted as part of an academic course assignment. Dataset credits go to the original dataset creators.