Research on Road Damage Recognition Based on VanillaNet Neural Network

Abstract

This paper presents a highly efficient and accurate model for identifying road conditions, specifically distinguishing between normal roads and potholes in images. Given the challenges posed by potholes, such as their prevalence in shaded areas, cluttered backgrounds, difficulty in annotation, and diverse shapes, we initiate the process by normalizing the dataset. We select the VanillaNet neural network as our foundational training architecture due to its elegant and simplistic design, which maintains exceptional performance in computer vision tasks. VanillaNet commences with several layers incorporating nonlinear activation functions, gradually eliminating these layers as training progresses, facilitating easy integration while preserving inference speed. This network surpasses contemporary models in both efficiency and accuracy, making it an ideal choice for our analysis. The quantitative results demonstrate promising outcomes, laying a solid foundation for future research endeavors. Subsequently, we train the developed VanillaNet model, employing deep training techniques to enhance its performance. The model is comprehensively evaluated based on accuracy, speed, and generalization capabilities to ensure it meets our requirements. In the object detection experiments, the model's performance is assessed using four key metrics.