Traffic-Sign-Detection-using-CNN

Algorithms

1. Convolutional Neural Network (CNN)

• Core technique for image classification that forms the backbone of the system. CNN is ideal for image data.
• Architecture:
• Input layer to accept 32x32x3 RGB images as input to the network.
• Multiple convolutional layers for hierarchical feature extraction:
• First conv layer has 32 filters of size 3x3 to extract low-level features like edges, colors, gradients etc.
• Second conv layer has 64 filters to extract higher-level features like shapes, textures etc.
• Convolution operation captures spatial relationships between pixels in the image.
• Convolution is applied on the input and previous layer's output.
• Max pooling layers to progressively reduce spatial dimensions:
• Max pooling of filter size 2x2 applied after specific conv layers.
• Takes the maximum value in each 2x2 window to downsample feature maps.
• Reduces computations and parameters, avoiding overfitting.
• Fully connected layers for high-level reasoning:
• Flatten layer to convert 2D feature map matrix to 1D vector.
• Dense layer with 128 units learns non-linear combinations of low-level features.
• Output layer with 43 units and softmax activation to predict class probabilities.
• Training:
• Adam optimizer used for efficient training with adaptive learning rates for each parameter.
• Categorical cross-entropy loss function to quantify difference between target and predicted class probabilities.
• Accuracy metric on test set used to evaluate real-world performance of the model.
• Trained for 5 epochs due to early stopping to prevent overfitting.

2. Image Preprocessing

• Resizes all images to 32x32 pixels to standardize input dimensions for the CNN model.
• Normalizes pixel values between 0-1 by dividing by 255 to center input data for stable training.
• Augments dataset using rotations, shifts, zooms etc. to expand diversity of images for better generalization.

3. Transfer Learning

• Leverages knowledge from a CNN like InceptionV3 pretrained on large datasets like ImageNet.
• Allows model to converge faster with fewer training samples and epochs.
• Provides much better weight initialization for new task compared to random initialization.

4. Object Detection

• Scans input image for potential traffic sign regions using sliding window approach.
• Extracts sign candidates using techniques like Haar cascades, HOG feature matching etc.
• Applies trained CNN model on the extracted regions for accurate classification.

5. Optimization Algorithms

• Adam optimizer for efficient training with per-parameter adaptive learning rates.
• Dropout regularization randomly sets layer activations to zero during training. Improves generalization.
• Early stopping stops training when validation loss stops improving after an epoch's training.

6. Performance Metrics

• Accuracy, precision, recall etc. used to quantitatively evaluate model's real- world performance.
• Target accuracy of 97% aimed on unseen test set for acceptable performance.