Multi-Column Convolutional Neural Network for Vehicle-Type Classification

Vehicle type classification is one of the major important applications in field of road security. In this paper, we present a deep learning-based approach to classify vehicles into six categories using Multi-Column Convolutional Neural Network (MCCNN). Instead of using fixed scale convolutional layer, the model we used is able to extract features from different scale of an image using variant filters to improve classification performance especially in vehicles with similar appearance like Suv and Sedan. The feasibility of the proposed approach is checked through an experimental investigation conducted on BIT dataset which includes 9850 frontal-view images of different types and high-resolution. The MCCNN model achieves a classification accuracy rate of 95.48% using the whole unbalanced dataset outperforming previous studies. An additional evaluation using 200 images from each class is performed yielding to an average accuracy rate of 88.37%.