Traffic Sign Classification with Neural Networks

Project Overview

This project involves developing a neural network to classify traffic signs using TensorFlow. The goal is to accurately identify different types of traffic signs from images to support self-driving car technologies. The dataset used for this project is the German Traffic Sign Recognition Benchmark (GTSRB), which includes thousands of images of 43 different road signs.

Dataset

The dataset is available at GTSRB Dataset and consists of images categorized into 43 directories, each representing a different road sign category. The images vary in size and are preprocessed to be 512x512 pixels.

Installation

1. Clone the repository:

   git clone https://github.com/rujeetjahagirdar/traffic_sign_detection.git
   cd traffic-sign-classification

2. Install dependencies:

   pip install -r requirements.txt

Usage

1. Download the data:

   • Download the dataset from the provided link and place it in the gtsrb directory inside the project directory.

2. Run the script:

   python traffic.py gtsrb

   This command will train the model on the GTSRB dataset and evaluate its performance.

Implementation Details

1. Data Loading: The load_data function reads and preprocesses images from the dataset, resizing them to a standard size and splitting them into training and testing sets.

2. Model Building: The get_model function constructs a neural network using TensorFlow, including layers such as convolutional, pooling, and fully connected layers. The model is trained to classify road signs into 43 categories.

3. Experiments:

   • Various architectures and hyperparameters were tested, including different numbers of convolutional layers, filter sizes, and dropout rates.

Experimentation Details

Configuration Variation:

Various configurations of convolutional and pooling layers, filter sizes, and hidden layers were tested to identify the optimal architecture for the CNN.

Observations:

Experimentation revealed that increasing the number of hidden layers and adjusting dropout rates played a crucial role in improving model accuracy. Filter sizes and pooling layer variations were observed to have different effects on feature extraction and model robustness.

Resume Functionality:

The project features a resume function, allowing the continuation of experiments from the last completed iteration. This helps save time and resources during experimentation.

Experimentation Steps

1. Data Loading and Preprocessing:

   • In this project we have worked on the German Traffic Sign Recognition Benchmark (GTSRB) dataset, which consists of different images of the traffic signs categorised into 43 categories.

2. Model Architecture:

   • For this project we have used CNN model. We have experimented with various architecture of the CNN network.
   • Convolutional Layers: Extracts hierarchical features from input images using learnable filters, capturing spatial patterns crucial for image understanding.

   Pooling Layers: Downsamples feature maps, reducing dimensionality and preserving essential information while enhancing computational efficiency.

   Hidden Layers: Enables the network to learn complex hierarchical representations, facilitating the abstraction of high-level features for accurate classification.

3. Hyperparameter Tuning:

   • In this project we have exprimented with various hyperparameters such as filter sizes, pool sizes, and dropout rates.

4. Training and Evaluation:

   • We have used 'ADAM' optimizer along with categorical_crossentropy loss function.

5. Experiment Results:

   • We have observed that, the dropout rate of the model is affecting the result significantly. As we increased the dropout rate, the model seems to not perform well.

What Worked Well

• Choosing different hyperparameters like 3 convolution layers of various size (32, 64, 128), and dropout rate of 0.4 seems to work well.
• This was the result of one of the experiment that we have performed.

What Didn't Work Well

• We have observed that increasing the dropout rate reduceses the performance of the model.

Input and Output

• Experiment Input File: The project reads experiment parameters from an input file (experiment_parameters.txt) in the format specified.

• Results Output File: The results of each experiment, including accuracy, are written to a log file (experiment_log.txt) for future reference.

Report

A detailed report on the experimentation process, model performance, and results is included. The report discusses various approaches, what worked well, and areas for improvement.

Contributing

Feel free to fork the repository and submit pull requests with improvements or additional features.

License

This project is licensed under the MIT License. See the LICENSE file for details.

References

1. TensorFlow Keras Overview
2. OpenCV-Python Documentation
3. German Traffic Sign Recognition Benchmark