Traffic Sign Classifier

A project that stretched my understanding of Neural Networks especially CNN’s and their successful use in Computer Vision. I built the network on Keras framework running on top of Tensorflow backend, trained on Google Colab using thousands of images in the training set and attained an accuracy of 97.8% against the test set. I further tested the model on some independent external data and it performed well.

The training source code is available in a notebook file containing all the source code, visualizations and notes: https://github.com/ItsCosmas/Traffic-Sign-Classification/blob/master/Traffic_Sign_Classification.ipynb The trained model file is a .h5 file named my_model.h5

• Keras
• Python

Python Libraries

• Open CV
• Numpy
• Pandas
• Matplotlib
• and more ...

Training

I trained this model in an online cloud instance on Google Colab

Key points

I made use of Histogram Equalization technique to standardize lighting in all our images. I used an ImageGenerator to help show different angles and views of the same data set to the model for it to better identify the features.

Building the Network

I used and tweaked a leNet model to attain desired results:

def better_model():
  model = Sequential()
  # add the convolutional layer
  #filters, size of filters,input_shape,activation_function
  model.add(Conv2D(60,(5,5), input_shape= (32,32,1), activation = 'relu'))
  model.add(Conv2D(60,(5,5), input_shape= (32,32,1), activation = 'relu'))
  #pooling layer
  model.add(MaxPooling2D(pool_size = (2,2)))
  # add another convolutional layer
  model.add(Conv2D(30, (3, 3) , activation = 'relu'))
  model.add(Conv2D(30, (3, 3) , activation = 'relu'))
  # pooling layer
  model.add(MaxPooling2D(pool_size = (2,2)))
  
  #model.add(Dropout(0.5))
  
  #Flatten the image to 1 dimensional array
  model.add(Flatten())
  #add a dense layer : amount of nodes, activation
  model.add(Dense(500, activation = 'relu'))
  # place a dropout layer
  #0.5 drop out rate is recommended, half input nodes will be dropped at each update
  model.add(Dropout(0.5))
  # defining the ouput layer of our network
  model.add(Dense(num_classes, activation = 'softmax'))
  
  
  #Compile Model
  # we use Adam optimizer with a learning rate 0f 0.01
  # A categorical_crossentropy'
  
  model.compile(Adam(lr = 0.001), loss = 'categorical_crossentropy', metrics = ['accuracy'])
  
  return model

Overfitting and Underfitting

Reduced the number of training epochs to achieve standard results, I repeatedly redisgned and tweaked the model until I attained the right overfitting and underfitting metrics.

Testing the model

If you wish to test this model locally, please refer to this jupyter notebook file: https://github.com/ItsCosmas/Traffic-Sign-Classification/blob/master/Test%20Model%20Using%20local%20image%20in%20OpenCV.ipynb

If you wish to test this model with images from the internet, please refer to this notebook: https://github.com/ItsCosmas/Traffic-Sign-Classification/blob/master/Test%20Model.ipynb