Build a Traffic Sign Recognition Project with TensorFlow 2.0
More examples can be found at TensorFlow tutorials
Raw data from Kaggle: GTSRB - German Traffic Sign Recognition Benchmark. Each image in raw downloaded data has different sizes, thus they were processed to have a consistent shape. After processes including: resizing, splitting into training and validation, conversion to grayscale and normalization.
- Number of training examples = 31367
- Number of validation examples = 7842
- Number of testing examples = 12630
- Image data shape = (32, 32, 1)
- Number of classes = 43
My final model consisted of the following layers:
| Layer | Description |
|---|---|
| Input | 32x32x1 Grayscale image |
| Convolution 5x5 | 1x1 stride, valid padding, outputs 28x28x6 |
| RELU | |
| Max pooling | 2x2 stride, outputs 14x14x6 |
| Convolution 5x5 | 1x1 stride, valid padding, outputs 10x10x16 |
| RELU | |
| Max pooling | 2x2 stride, outputs 5x5x6 |
| Fully connected | Input = 400. Output = 120. |
| Fully connected | Input = 120. Output = 84. |
| Dropout | |
| Fully connected | Input = 84. Output = 43. |
| Softmax | Output layer |
My final model results were:
- validation set accuracy of ~ 0.98
- test set accuracy of ~ 0.92