MAFF: A Novel MobileNetV3 Attention Feature Fusion Network for Automatic Vehicle Classification
Muhammad Khalid Mumtaz, Bingcai Chen, Muhammad Usman Saeed, Muhammad Nadeem, Muhammad Altaf · 2023
Efficient traffic and information systems heavily rely on real-time traffic data for effective regulation and management. Surveillance cameras, deployed in recent times, play a pivotal role in monitoring and controlling traffic. In this context, video surveillance technologies and image processing techniques have been extensively studied to enhance traffic management. This research focuses on the utilization of surveillance cameras for the analysis of traffic data, providing advanced warning systems and real-time extraction of vital vehicle-related information. We introduce an innovative approach that seamlessly integrates an advanced attention feature fusion block within the MobileNetV3 architecture. This integration leads to the development of a novel deep learning framework for automatic vehicle classification. Our model's performance stands out as it surpasses existing models on four distinct datasets: TAU Vehicle, Stanford Car, Vehicle Rear, and Tiny and Low-Quality Images. It achieves remarkable accuracy with the TAU Vehicle dataset at 97.64%, the Stanford Car dataset at 97.59%, the Vehicle Rear dataset at 97.41%, and even the challenging Tiny and Low-Quality Images dataset at an impressive 96.29%. Our methodology significantly enhances accuracy across these diverse datasets, setting a new benchmark in image classification and outperforming prior state-of-the-art approaches.