Image Classification for Vehicle Type Dataset Using State-of-the-art Convolutional Neural Network Architecture

Abstract

Fast development in Deep Learning and its hybrid methodologies has led diverse applications in different domains. For image classification tasks in vehicle related fields, Convolutional Neural Network (CNN) is mostly chosen for recent usages. To train the CNN classifier, various vehicle image datasets are used, however, most of previous studies have learned features from datasets with a single form of images taken in the controlled condition such as surveillance camera vehicle image dataset from the same road, which results the classifier cannot guarantee the generalization of the model onto different forms of vehicle images. In addition, most of researches using CNN have used LeNet, GoogLeNet, or VGGNet for their main architecture. In this study, we perform vehicle type (convertible, coupe, crossover, sedan, SUV, truck, and van) classification and we use our own collected dataset with vehicle images taken in different angles and backgrounds to ensure the generalization and adaptability of proposed classifier. Moreover, we use the state-of-the-art CNN architecture, NASNet, which is a hybrid CNN architecture having Recurrent Neural Network structure trained by Reinforcement Learning to find optimal architecture. After 10 folded experiments, the average final test accuracy points 83%, and on the additional evaluation with random query images, the proposed model achieves accurate classification.