The purpose of the repository is to create a LeNet convolutional neural network to classify images of road signs.
The steps are:
This neural network primarily uses keras models and optimizers. It also uses pickle and pandas for working with the data.
The labelled image data is imported from the link at the end of this README. This is a dataset of 34799 images over 42 different categories of traffic signs, 10 of which are shown below.
Figure 1: Training Images
The next step is to preprocess the road signs. This includes grayscale conversion, histogram equalization, and normalization.
Figure 2: Preprocessed Image
To improve the robustness of the model and augment our dataset, we perform image augmentation techniques including translation, zoom, and rotation.
Figure 3: Augmented Images
The next step is to create the LeNet model. This includes four convolutional layers with relu activation, two pooling layers, and two Dense Layers with relu and softmax activations, respectively.
The LeNet model takes about 1 minute per epoch to train, which means a total of 10 minutes for 10 epochs.
Shown below are the Loss and Accuracy Plots of the training data and validation data. The result is a model with 97.5% accuracy on unseen testing data.
Figure 4: Training and Validation Loss
Figure 5: Training and Validation Accuracy
To ensure that the model works on other images, we test that it works on different images found on the internet. It successfully identified all of them. Below is an example of the Bike Crossing image that was correctly classified.
Figure 6: Non-Training Image
Figure 7: Preprocessed Non-Training Image
https://bitbucket.org/jadslim/german-traffic-signs
This code is intended to be run on Google Colab as a Python 3 Notebook, as the model fitting is faster with GPU processing.
Press Runtime> Change runtime type > Hardware Accelerator > GPU to improve the runtime signficantly.