A novel approach for vehicle identification based on image registration and deep learning

Fine-grained vehicle type recognition using on-road cameras is amonginteresting topics in machine vision. It has several challenges like inter-class similarity,di erent viewing angles, and di erent lighting and weather conditions. This paper presentsa novel approach for vehicle classi cation based on a novel augmentation method anddeep learning. In the proposed smart augmentation, the vehicle images of each class areregistered on the reference vehicles of all other classes and then added to the training setof that class. In this way, we will have a lot of new images which are very similar toboth reference and target classes. This helps the Convolutional Neural Network (CNN)model to handle inter-class similarities very well. In the test phase, the input image isregistered on every reference image in parallel and applied to the model. Finally, thewinner is determined by summing up the provided scores of all models. The targeted dataaugmentation along with the proposed classi cation strategy has high recognition powerand is capable of providing high accuracy using small CNNs or any other classi cationmethod without the need for large datasets. The proposed method achieved a recognitionrate of 99.8% with only 150 K parameters. (c) 2024 Sharif University of Technology. All rights reserved.