CNN-Based Hyperparameter Optimization Approach for Road Pothole and Crack Detection Systems

Poorly maintained roads contribute to the number of fatal auto accidents that occur each year. The condition of damaged roads in Indonesia reached around 2500 miles or more than 8% of the total national roads. The higher the number of road damages, the probability of a traffic accident due to road damage also rises. In order to avoid this, the roads need to be repaired in a comprehensive and long-term way. However, the way to check for road damage is still based on inefficient methods. In this paper, we propose the optimization of the detection of potholes and cracks using a deep learning convolutional neural network with a pre-trained SSD MobileNet V2 model by adjusting the hyperparameter. The optimization was carried out on our previous mobile road inspection system. The effectiveness is confirmed through experiments with the optimal mAP and loss values determined by the model parameter testing process.