An end-to-end deep learning pipeline for classifying traffic signs using the German Traffic Sign Recognition Benchmark (GTSRB) dataset. This project implements and compares MobileNetV2 (for efficiency) and ResNet50V2 (for performance) using TF/Keras.
The goal of this project is to develop a robust traffic sign classifier that balances high accuracy with low inference latency. The system is designed to be deployed as a web application using Gradio.
-
Transfer Learning: Utilizes pre-trained architectures (MobileNetV2, ResNet50V2).
-
Data Pipeline: Automated downloading, extraction and preprocessing of GTSRB dataset.
-
Evaluation: Confusion Matrices, F1-Scores, and Inference Time analysis.
-
Deployment: Interactive web interface for real-time predictions.
-
Clone the repository:
git clone https://github.com/ZohaibHassan16/GTSRB-ML.git cd GTSRB-ML -
Install Dependencies:
It is recommended to use virtual environment.
pip install -r requirements.txt
This project uses the GTSRB dataset from kaggle. You can automatically download it using the included script.
python download_data.py
Option 1: Run the Full Pipeline
The gtsrb.py script handles data loading, training, evaluation and app launching. Just don't forget to update paths.
python gtsrb.py
By default, this will:
-
Train MobileNetV2 and ResNet50V2
-
Fine-tune the models.
-
Evaluate on the test set.
-
Generate comparison plots.
-
Prompt to launch the Gradio app.
Option 2: Jupyter Notebook
Open the notebook in Coloab and rull all the cells, recommended.
We experimented with two architectures to analyze the trade-off between model size/speed and accuracy.
While ResNet50 offers marginally higher accuracy, MobileNetV2 was selected for the final application due to its superior efficiency.
-
Core: Python, TensorFlow, Keras
-
Data Processing: Numpy, Pandas, OpenCV
-
Visualization: Matplotlib, Seaborn
-
Interface: Gradio