This is project I did on🚦 Traffic Sign Classifier Using Deep Learning (TensorFlow) if you are interested in code its here Project github repo.

📝 Project Overview

The goal of this project was to identify road traffic signs using computer vision and convolutional neural networks (CNNs). The model was trained on the German Traffic Sign Recognition Benchmark (GTSRB) dataset using a modified LeNet CNN architecture.

This project is crucial for autonomous driving systems and ADAS (Advanced Driver-Assistance Systems), where precise identification of road signs is essential for safety and navigation.

Project highlights

  • Designed and trained a deep learning model for traffic sign recognition, achieving 95.1% accuracy using a modified LeNet CNN architecture.
  • Addressed class imbalance by augmenting underrepresented categories, significantly improving classification consistency.
  • Implemented grayscaling & normalization, reducing training time while maintaining model performance.
  • Optimized model generalization by integrating dropout layers (70% in Conv1 & FC1), reducing overfitting by X%.
  • Tuned Adam optimizer hyperparameters (learning rate = 0.001, batch size = 120) to achieve peak performance.
  • Analyzed Softmax probabilities to interpret model confidence, identifying edge cases like misclassification of Stop & Yield signs due to viewing angles.
  • Deployed data augmentation techniques (rotation, brightness adjustment, scaling) to enhance model robustness for real-world scenarios.

🏗 Project Architecture

The Traffic Sign Classifier consists of the following key steps:

1️⃣ Dataset & Preprocessing

  • Problem: The initial CNN model was overfitting due to uneven image distribution across different classes. Some traffic signs had very few images, making it difficult for the model to generalize.
  • Solution:
    • Data Augmentation: Added additional images to underrepresented classes, which significantly improved model accuracy.
    • Grayscale Conversion: Initially trained on normalized color images, then switched to grayscale images. While accuracy remained similar, grayscale images reduced training time.

📊 Before and After Data Rebalancing

Training data before rebalancing Distribution after rebalancing

2️⃣ Model Architecture

The CNN used a modified LeNet architecture with:

  • Two Convolutional Layers
    • Conv1: 6 filters, 5x5 kernel, ReLU + Max Pooling
    • Conv2: 16 filters, 5x5 kernel, ReLU + Max Pooling
  • Three Fully Connected Layers
    • FC1: 120 neurons, ReLU
    • FC2: 84 neurons, ReLU
    • Output: 43 classes (Softmax activation)

📉 Challenges & Solutions:

  • Initially, using only max pooling led to 91% accuracy, but top-5 probability analysis showed misclassification due to overfitting.
  • Solution: Introduced Dropout layers in Conv1, FC1, and FC2, which helped reduce overfitting and improved accuracy to 95%.
  • The model required 32x32x1 grayscale images, which simplified computations.

📊 Visualization of Model Architecture

graph TD A[Input: 32x32 Traffic Sign Image] -->|Grayscale + Normalize| B[Preprocessing] B -->|Data Augmentation| C{CNN Feature Extraction} C -->|Conv Layer 1: 5x5, ReLU| D[Max Pooling 1: 2x2] D -->|Conv Layer 2: 5x5, ReLU| E[Max Pooling 2: 2x2] E -->|Flatten Layer| F[Dense Layer: 120, ReLU] F -->|Dense Layer: 84, ReLU| G[Softmax Output 43 Classes] G --> H[Model Prediction]

3️⃣ Model Training

To optimize the training process, several hyperparameters were tested:

🔍 Optimization Approach:

  • Optimizer: Adam (performed better than gradient descent)
  • Learning Rate: 0.01 → 0.001 (best at 0.001)
  • Batch Size: Tested 100, 150, 200, 250 (best at 120)
  • Dropout Rate: Experimented with 50% to 70%, found 70% best for Conv1 and FC1.

🛠 Challenges & Fixes: ✅ Overfitting Issue: Resolved using dropout layers (70%)
✅ Slow Training Speed: Switched to grayscale input, reducing training time
✅ Learning Instability: Fine-tuned learning rate & batch size

Optimization

[!info] Reduced overfitting via 70% dropout (Conv1, FC1) and fine-tuned Adam optimizer (learning rate=0.001, batch size=120). what this mean is this real

What this mean?

  • Overfitting: Occurs when a model performs well on training data but fails to generalize to new data.
  • Dropout (70%): Dropout randomly “switches off” a portion of neurons during training, forcing the network to learn more robust, generalized features. While 70% is on the higher side (typical ranges are 30–50%), it can be effective in certain setups to combat heavy overfitting.
  • Conv1, FC1: Applying dropout in the first convolutional layer and the first fully connected layer helps ensure both early and later-stage representations don’t memorize training data too narrowly.
  • Fine-tuned Adam: The Adam optimizer is widely used in deep learning. Setting a learning rate of 0.001 and a batch size of 120 is perfectly valid—these hyperparameters are often adjusted experimentally to balance convergence speed and stability.

4️⃣ Performance Analysis

Final Model Performance:

  • 95% Accuracy on test dataset after 30 epochs
  • Softmax Probability Analysis for misclassified signs

New Image Performance:

  • Overall accuracy: 83%
  • Stop sign misclassified (likely due to angle distortion)
    🛠 Possible Fix: Perspective Transform to standardize image angles

📊 Softmax Probability Insights:

  • Yield, No Entry, and Caution Signs predicted with high confidence
  • “Dangerous Curve to Right” sign had 65% confidence, but 35% for “Slippery Road” and 5% for “Children Crossing”.
    • Reason: All three signs are triangular, causing confusion.
  • “60 km/h Speed Limit” was predicted with 80% confidence, significantly higher than other probabilities.

📊 Softmax Prediction Visualization

[image

🏆 Final Thoughts & Key Takeaways

✔️ Data Augmentation helped improve class balance
✔️ Grayscale normalization reduced training time without accuracy loss
✔️ Dropout layers helped prevent overfitting
✔️ Hyperparameter tuning led to optimal performance
✔️ Softmax analysis revealed areas for improvement (angle correction, perspective transform)

🚀 Future Improvements:

  • Implement perspective transformation for better real-world performance
  • Test model on real-world dashcam traffic images
  • Deploy model in an embedded ADAS system for real-time traffic sign recognition