Optimization of Deep Neural Networks for Vehicle Classification

This paper addresses the challenges in vehicle recognition for optimizing results and reducing efforts in traffic management. The proposed solution focuses on developing solutions for a market leader of traffic cameras in Brazil, aiming to classify vehicle properties such as brand, segment, subsegment, and color. The proposal analyzes model portability and optimization techniques, demonstrating experimental results that significantly reduce computational costs without compromising accuracy. We applied pruning techniques and optimization approaches using metaheuristics to reduce approximately 40% of the model’s weights with minimal impact on accuracy. We also developed two inference and training scripts and a C++ executable packet.