Traffic Sign Recognition – CNN vs MobileNetV2

Description

This project implements a Traffic Sign Recognition system using the GTSRB (German Traffic Sign Recognition Benchmark) dataset.
It is part of the Self-Paced Elevvo Machine Learning Internship (Industry-Level Project).

The notebook compares two approaches for classifying traffic signs:

1. Custom Convolutional Neural Network (CNN) – built from scratch.
2. MobileNetV2 (Pretrained Model) – fine-tuned for traffic sign classification.

---

Features

• Image preprocessing and normalization
• Data augmentation for better generalization
• Two model approaches:
  • Custom CNN
  • MobileNetV2 (transfer learning)
• Evaluation using accuracy, precision, recall, and F1-score
• Visual performance comparison between models

---

Project Workflow

1. Dataset Loading and Analysis

   • GTSRB dataset
   • Images resized and normalized

2. Data Preprocessing & Augmentation

   • Applied transformations (rotation, zoom, shifts)

3. Model Development

   • Custom CNN with convolution, pooling, dropout, and dense layers
   • MobileNetV2 pretrained on ImageNet and fine-tuned

4. Training & Validation

   • Trained both models with early stopping and validation tracking

5. Evaluation

   • Compared class-wise precision, recall, and F1-score

⚙️ Installation

Install the dependencies with:

pip install tensorflow keras matplotlib seaborn scikit-learn opencv-python

Dataset: GTSRB - German Traffic Sign Recognition Benchmark

---

🚀 Usage

1. Clone this repository:

   git clone https://github.com/yourusername/traffic-sign-recognition.git
   cd traffic-sign-recognition

2. Open the notebook:

   jupyter notebook traffic_sign_recognition.ipynb

3. Run all cells to train and evaluate both models.

---

🏷️ Tags & Features

• Tags: Deep Learning, CNN, Transfer Learning, MobileNetV2, Image Classification, Traffic Signs, Computer Vision, Internship Project
• File Types: .ipynb (Jupyter Notebook), dataset images (.png)

---

📊 Output Features

The notebook outputs:

• Training vs Validation vs Test Accuracy & Loss plots
• Class-wise Precision, Recall, and F1-score plots

Test Accuracy comparison between 2 models

This shows how the pretrained model generalizes better

Class-wise F1-score Comparison

This plot compares the performance of the CNN and MobileNetV2 across different traffic sign classes.

✅ Conclusion

• The Custom CNN performed well overall but showed weaknesses in some classes, especially rare or complex signs.
• The MobileNetV2 model consistently outperformed the custom CNN with higher accuracy, precision, recall, and F1-scores.
• This shows that transfer learning with pretrained models is more effective for real-world problems like traffic sign recognition.
• In practical applications such as autonomous driving, this higher accuracy can directly improve safety and reliability.