Autonomous Multiple Damage Detection and Segmentation in Structures Using Mask R-CNN

Abstract

AbstractThere are approximately 730,000 road bridges in Japan. As of 2020, more than 23% of them have been aged over 50 years or more and the percentage will go over 50% within the year 2030. Over the last decade there are numerous bridge accidents all over the world that have cost both monetary values and human lives significantly. Thus, regular inspection of bridges is very important to check the overall condition. However, the inspection is predominantly maintained manually which depends on person’s expertise and often the task is cumbersome, expensive and error prone. Over the years, different deep learning-based techniques, such as Convolutional Neural Network (CNN), have been utilized to detect the damages automatically. However, most of them concentrate on damage detection and often been used for single class of detection only. Instance segmentation is a method where each object is detected as separate instance and by adopting a Region-based CNN model, such as Mask R-CNN, the instances can be shown separately. Though instance segmentation has been applied extensively for the detection of common objects in the real world, the application for multiple structural damage detection is very limited so far. Specially, the training and testing of the R-CNN model for multiple structural damage detection is different and challenging than common objects. This study is a step towards the feasibility and application of instance segmentation for multiple damage detection in structures and to evaluate the feasibility for real time detection with complex background.KeywordsDeep learningDamage detectionInstance segmentationMask R-CNNCNN