A Deep Learning-Based Approach for Road Surface Damage Detection

Timely detection and elimination of damage in areas with excessive vehicle loading can reduce the risk of road accidents. Currently, various methods of photo and video surveillance are used to monitor the condition of the road surface. The manual approach to evaluation and analysis of the received data can take a protracted period of time. Thus, it is necessary to improve the procedures for inspection and assessment of the condition of control objects with the help of computer vision and deep learning techniques. In this paper, we propose a model based on Mask Region-based Convolutional Neural Network (Mask R-CNN) architecture for identifying defects of the road surface in the real-time mode. It shows the process of collecting and the features of the training samples and the deep neural network (DNN) training process, taking into account the specifics of the problems posed. For the software implementation of the proposed architecture, the Python programming language and the TensorFlow framework were utilized. The use of the proposed model is effective even in conditions of a limited amount of source data. Also as a result of experiments, a high degree of repeatability of the results was noted. According to the metrics, Mask R-CNN gave the high detection and segmentation results showing 0.9214, 0.9876, 0.9571 precision, recall, and F1-score respectively in road damage detection, and Intersection over Union (IoU)-0.3488 and Dice similarity coefficient-0.7381 in segmentation of road damages.