German Traffic Sign Recognition with VGG16

This repository contains a fine-tuning of the VGG16 pre-trained model to solve a classification problem using images of German traffic signs. The goal is to predict the type of traffic sign in the images. This project is part of a competition hosted on Kaggle.

Important note: You cannot submit the code to the Kaggle competition without first modifying the prediction output.

Dataset

The dataset used for this project can be downloaded from Kaggle. @ Kaggle Competition: GTRSR - German Traffic Sign Recognition

The dataset consists of:

• Training set: Images and labels for training the model.

• Validation set: Images for evaluating the model.

• Classes: 43 different classes of stop signs

• Data/: Contains the training images and their respective CSV files.

• GTSRB_Classification_Using_VGG.ipynb: Jupyter notebook containing the entire workflow from data preprocessing to model evaluation.

Required Libraries

• pandas: For data manipulation and analysis.
• numpy: For numerical operations.
• matplotlib: For data visualization.
• tensorflow and keras: For building and training the CNN model.

Usage

1. Download the dataset from Kaggle.

2. Open the Jupyter notebook GTSRB_Classification_Using_VGG.ipynb and run all the cells to execute the following steps:

   • Load and preprocess the data.
   • Visualize the data distribution.
   • Build and compile the CNN model.
   • Train the model.
   • Evaluate the model.
   • Make predictions on the test set.

Model Architecture

The CNN model is built using TensorFlow and Keras. The architecture consists of:

• Base Model: VGG16 pre-trained on ImageNet.
• Custom Layers:
  • Global Average Pooling layer.
  • Dense layer with softmax activation for the output layer to classify the traffic signs.

Results

The model is trained for 20 epochs (as a start) with data augmentation. The training and validation accuracy and loss are plotted to visualize the model's performance.

Important note

Model accuracy isn't good after training it for 20 epochs because of the complex model architecture. So, we have two options.

• First is to increase the number of the epochs and change slightly in the model (adding drop out to avoid overfitting) -> will be expenses
• We can choose another model like MobileNetV2 or EfficientNet-B0 -> we will try this soon

Acknowledgements

Special thanks to Kaggle for providing the dataset and hosting the competition. This project is an excellent opportunity to apply deep learning techniques to real-world image classification tasks.

Kaggle Competition: GTRSR - German Traffic Sign Recognition