The purpose of this project is to develop a model for traffic sign recognition. The project involves the following steps:
- Loading the data set
- Exploring, summarizing, and visualizing the data set
- Designing, training, and testing a model architecture
- Using the model to make predictions on new images
- Analyzing the softmax probabilities of the new images
- Summarizing the results in a written report
In this section, I will address each of the rubric points individually and describe how I have implemented them in my project.
Here is a link to my project code.
To calculate the summary statistics of the traffic sign data set, I used the numpy library in Python. Here are the details:
- The size of the training set is
34,799 - The size of the validation set is
4,410 - The size of the test set is
12,630 - The shape of a traffic sign image is
(32, 32, 3) - The number of unique classes/labels in the data set is
43
To visualize the data set, I created a bar chart that shows the distribution of classes in the entire data set. This provides an overview of how the data is distributed among the different traffic sign categories.
As a preprocessing step, I normalized the images using the cv2.normalize method from the OpenCV library. Additionally, I converted the normalized RGB images to grayscale before using them for training. The preprocessed image size is (32, 32, 1).
The final model architecture consists of the following layers:
| Layer | Description | Output Shape |
|---|---|---|
| Input | 32x32x1 Grayscale image | |
| Convolution 5x5 | 2x2 stride, same padding, 32 filters | 28x28x32 |
| Tanh activation | ||
| Average pooling | 2x2 stride, valid padding, outputs 14x14x32 | |
| Convolution 5x5 | 1x1 stride, valid padding, 64 filters | 10x10x64 |
| Tanh activation | ||
| Average pooling | 2x2 stride, valid padding, outputs 5x5x64 | |
| Flatten layer | Output shape 1600 1D | |
| Dense layer1 | Output shape 400 1D | |
| Tanh activation | ||
| Dense layer2 | Output shape 120 1D | |
| Tanh activation | ||
| Dense layer3 | Output shape 84 1D | |
| Tanh activation | ||
| Dense layer4 | Output shape 43 1D | |
| Softmax |
Total parameters: 85,631 Trainable parameters: 85,631 Non-trainable parameters: 0
To train the model, I used the Adam optimizer with the categorical cross-entropy loss function. The model was trained for 20 epochs. I included a TensorBoard callback for visualizing the model architecture and training progress. The validation set was used to monitor the training progress.
The model achieved a training set accuracy of 99.59% and a validation set accuracy of 93.11%. The test set accuracy is 92.44%.
I used five German traffic sign images found on the web to test the model's predictions. Here are the images:
The model correctly predicted all five traffic signs, resulting in an accuracy of 100%. This performance compares favorably to the accuracy on the test set, which is 92.44%.
For each of the new images, I analyzed the top 5 softmax probabilities predicted by the model. The model displayed high confidence in its predictions. Here are the results for the first image:
| Probability | Prediction |
|---|---|
| 0.99999321 | General caution |
| 0.00000630 | Traffic signals |
| 0.00000033 | Pedestrians |
| 0.00000005 | Wild animals crossing |
| 0.00000004 | Right-of-way at the next intersection |
Here are some suggestions for further improvement:
- Increase the size of the training data set to improve the model's ability to generalize to different traffic sign datasets from various countries.
- Include a validation set with a different distribution to better evaluate the model's performance on unseen data.
- Address any potential overfitting issues by using a lighter model architecture or applying regularization techniques.