Traffic Sign Classification using CNN

This project implements a Convolutional Neural Network (CNN) to classify traffic signs using TensorFlow and Keras.

Dataset

The model is trained on the German Traffic Sign Recognition Benchmark (GTSRB) dataset.

• Classes: 43 different traffic sign classes.
• Structure: Contains Train, Test, and Meta folders.

Dependencies

Ensure you have the following libraries installed before running the project: pip install numpy pandas matplotlib pillow scikit-learn tensorflow keras

Preprocessing

Before feeding the images into the neural network, the following preprocessing steps are applied:

• Resizing: All images are resized to 32 × 32 pixels.
• Grayscale: Images are converted to grayscale.
• Normalization: Pixel values are scaled to the range [0, 1] by dividing by 255.
• Label Encoding: Labels are one-hot encoded using Keras' to_categorical.

Usage

1. Download and extract the GTSRB dataset into the notebook's directory.
2. Open traffic sign classifier.ipynb in Jupyter Notebook.
3. Execute the cells sequentially.

CNN Architecture & Model

(Note: Replace images/model_architecture.png with the actual path to your model's image) The CNN model takes an input of shape 32 × 32 × 3 and consists of the following layers:

• Block 1:
  • Conv2D (32 filters, 3 × 3 kernel, same padding) — Output: (None, 32, 32, 32)
  • MaxPooling2D (2 × 2 pool size) — Output: (None, 16, 16, 32)
• Block 2:
  • Conv2D (64 filters, 3 × 3 kernel, same padding) — Output: (None, 16, 16, 64)
  • MaxPooling2D (2 × 2 pool size) — Output: (None, 8, 8, 64)
• Block 3:
  • Conv2D (128 filters, 3 × 3 kernel, same padding) — Output: (None, 8, 8, 128)
  • MaxPooling2D (2 × 2 pool size) — Output: (None, 4, 4, 128)
• Classifier:
  • Flatten — Output: (None, 2048)
  • Dense (256 units) — Output: (None, 256)
  • Dropout
  • Output Dense (43 units, Softmax activation) — Output: (None, 43)

Model Summary

Layer (type)	Output Shape	Param #
conv2d_12 (Conv2D)	(None, 32, 32, 32)	896
max_pooling2d_12 (MaxPooling2D)	(None, 16, 16, 32)	0
conv2d_13 (Conv2D)	(None, 16, 16, 64)	18,496
max_pooling2d_13 (MaxPooling2D)	(None, 8, 8, 64)	0
conv2d_14 (Conv2D)	(None, 8, 8, 128)	73,856
max_pooling2d_14 (MaxPooling2D)	(None, 4, 4, 128)	0
flatten_4 (Flatten)	(None, 2048)	0
dense_8 (Dense)	(None, 256)	524,544
dropout_4 (Dropout)	(None, 256)	0
dense_9 (Dense)	(None, 43)	11,051

• Convolutional Layers Parameters: 896 + 18,496 + 73,856 = 93,248
• Dense Layers Parameters: 524,544 + 11,051 = 535,595
• Total / Trainable Parameters: 628,843
• Non-trainable Parameters: 0

Training Configuration

• Optimizer: Adam
• Loss Function: Categorical Crossentropy
• Metrics: Accuracy
• Epochs: 15
• Batch Size: 32
• Data Split: 80% Training / 20% Validation

Results

The model achieved the following accuracy scores:

• Training Accuracy: 0.9976
• Validation Accuracy: 0.9967
• Test Accuracy: 0.9345

The learning curves indicate effective learning and generalization without significant overfitting.