Efficient Vehicle Recognition and Classification using Convolutional Neural Network

Convolutional Neural Network (CNN) is arguably one of the best deep learning techniques commonly employed for image recognition and classification. The aim of this work is to develop an efficient recognition and classification system for vehicle using CNN, detect types of vehicle commonly found on the road for database collection purpose and improve the existing vehicle recognition for advanced application. A total of 1881 images from 4 different databases including ideal, non-ideal front-view and side-view, as well as non-vehicle images are used to train the neural network. The flow of the program is library import, image processing, one hot encoding, training, and validation of the neural network for the results. Significant results such as graph of accuracy and loss function against the number of epochs, confusion matrix, misclassified image, and influence of the number of convolutional layer on accuracy are obtained and analyzed. The accuracy of neural network on ideal front-view database is 98.75%. However, the addition of non-vehicle images reduces the prediction accuracy to 94.99% and the extension of non-ideal images decreases the prediction accuracy to 91.92% while the further inclusion of side-view images boosts the overall prediction accuracy to 96.55%.