Detecting road damage utilizing retinanet and mobilenet models on edge devices

Abstract

A particular form of road digitalization produces a system that detects road damage automatically and in real time, employing the device to detect road damage as an edge device. The application of RetinaNet152 and MobileNetV2 models for road damage detection on edge devices necessitates a trade-off between high system performance and efficiency. Currently, edge devices have limited storage. In this paper, we explore how tuning hyperparameters with batch size and several optimizers improves system performance on RetinaNet152 and MobileNet models, as well as how they are implemented on edge devices. After tuning hyperparameters in the batch size of the optimizer, the Adam optimizer displayed enhanced performance with mean average precision (mAP), average recall (AR), and F1-score. This implies a positive impact on overall model performance. The MobileNetV2 model's hyperparameter tuning technique significantly improves performance, resulting in faster inference times and overall system performance. This demonstrates that the MobileNetV2 model could be used directly on edge devices to identify road damage. However, the RetinaNet152 model has a lower inference time, which cannot be deployed directly to edge devices. The RetinaNet152 model can be deployed on edge devices; however, a technique for speeding up inference time is essential.