Detection and Classification of Road Damage Using CNN with Hyperparameter Optimization

Abstract

Accidents in Indonesia are caused by various factors, including damage. To reduce these factors, manual surveys are carried out periodically to detect road damage, but manual surveys make road repairs hampered and uneven. This study looks at road damage detection systems to help classify and assess road conditions, as well as evaluation and assessment materials for use by relevant agencies. To determine the type of damage, the data will be processed using the Convolutional Neural Network (CNN) algorithm and the SSD V2 model. It is then transferred to the website via FTP (File Transfer Protocol) (FTP). Damage image, location, classification results, and degree of damage depending on the Surface Distress Index (SDI) approach is one of the data displayed on the website. The research was conducted using two methods, namely changing the hyperparameter values on MobileNet SSD V2 and comparing the two models MobileNet SSD V2 and MobileNet SSD V1. The test results demonstrate the best learning rate and model for detecting road damage. The MobileNet V2 model is the most accurate with tuning hyperparameters and by using step 500,000 and batch size 8 obtained optimal results and used in the detection of road damage.