Optimizing Traffic Sign Detection System Using Deep Residual Neural Networks Combined with Analytic Hierarchy Process Model

Abstract

This paper utilises image pre-processing techniques and deep residual neural networks to enhance the traffic sign detection system. A novel Analytic Hierarchy Process (AHP) model for performance evaluation has been proposed and utilised to determine the optimal parameter configuration of the learning models. Four evaluation metrics, namely accuracy, stability, response time, and system capability, have been defined for AHP measurements. The experiments were conducted using a comprehensive dataset, with VGG-16 and Google Net implemented for comparisons. Finally, the combination of ResNet-50 and the AHP model yielded the best results, achieving a 98.01% accuracy rate, 0.09% false alarm rate, and 1.28% undetection rate.