Evaluation of Deep Learning Architectures for Vehicle Recognition in Cloud-Based Systems

Abstract

This study aims to develop an accurate and scalable cloud-based vehicle image recognition system by leveraging advanced deep learning architectures. The primary objective is to evaluate and compare the performance of three prominent convolutional neural network models—InceptionV3, Xception, and MobileNetV2—for effective vehicle classification using diverse and high-resolution datasets. To achieve this, a comprehensive methodology was adopted, starting with data collection from the Stanford Cars and VehicleID datasets, supplemented with verified web-scraped images to ensure coverage of various vehicle types and conditions. A novel preprocessing pipeline was implemented, incorporating YOLOv8-based vehicle detection for region cropping, adaptive Gaussian filtering for noise reduction, and a hybrid augmentation strategy combining traditional techniques with GAN-generated images. Transfer learning was applied using pretrained weights, and a smart layer-freezing strategy was introduced to optimize fine-tuning. The training utilized a hybrid loss function combining categorical cross-entropy with center loss to enhance class separability. Evaluation metrics such as accuracy, precision, recall, F1-score, and inference time were computed to assess model performance on a stratified test dataset. Results indicated that all three models achieved over 95% test accuracy, with InceptionV3 attaining the highest accuracy (96.84%) and F1-score (96.85%), albeit with higher computational cost. Xception offered a balance between accuracy and resource efficiency, while MobileNetV2, though slightly lower in accuracy, excelled in inference speed and minimal model size, making it suitable for edge applications. Overall, the proposed framework demonstrates high accuracy, scalability, and adaptability for real-world smart traffic and transportation systems, proving its effectiveness in intelligent vehicle recognition.