Optimizers Impact on RetinaNet Model for Detecting Road Damage on Edge Device

Roads are a form of infrastructure that has an important role in supporting land transport and supporting equitable development in an area. There are numerous forms of road damage and varied sizes of road damage in the road damage detection system. As a result, in order to build a detection system with high accuracy and performance, a detection system with high robustness is required. It is critical to identify the design configuration and training approach for road damage classification using the CNN architecture. One of them is the choice of hyperparameters related to network structure and training. The research uses the RetinaNet-152 pre-trained CNN model to develop a road detection system. It also uses an optimizer selection and tuning hyperparameter optimizer that selects learning rates. According to our testing, the Adam optimizer has the lowest loss, high recall, mAP, and 70 Mb model size. The RetinaNet152 applies it to an edge device, resulting in an inference time of 19.14
$s$
and an FPS of 0.05. This demonstrates that the RetinaNet152 model can detect road-damaged objects.