(A) study on the deep learning-based concrete surface crack detection and measurement automation system

Abstract

consequently, a portable equipment that can accurately measure the size of cracks without contact is proposed.

A robot-based automatic inspection system using technology capable of measuring the size of concrete damage is proposed. In the future, the economically active population is expected to decrease owing to aging and population decline. Therefore, the maintenance labor for the structure will inevitably decrease. Nevertheless, the level of demand for quality infrastructure services remains unchanged. Thus, an automatic inspection system for damage inside tunnels was developed comprising an inspection robot and a management system. The inspection robot autonomously drives inside the tunnel and automatically analyzes the damage in three dimensions. The management system can remotely control the robot and monitor its status. Consequently, this robot system can automatically inspect several tunnels by a small number of managers.

The results of this study are based on deep learning-based core technology that can detect and measure cracks that occur inside a concrete tunnel and analyze the damage in three dimensions through the implementation of an automatic inspection system. Finally, it has significance in that the proposed automatic system was verified by field tests.

Concrete structures have significantly contributed to improving the quality of infrastructure services. Additionally, these structures have reliably performed their functions for a long period of time. However, because most of these structures were constructed during periods of economic expansion, they are already deteriorating. To make matters worse, the number of structures that have been in operation for more than 30 years has steadily increased, and by 2036, the proportion of these structures is expected to reach approximately 44\%. These old structures have been subjected to considerable damage from external forces and environmental factors for long periods. Cracks, peelings, and spalling are some of the defects that occur on their surfaces and can impair the intrinsic function of the structure

thus, they are considered a risk factor that can result in accidents.

Periodic inspections and accurate diagnoses are fundamentally important for maintaining the safety of concrete structures. When the current condition of the structure is accurately diagnosed, the priority of repair can be determined, and the appropriate repair can be performed. Therefore, an accurate inspection of the damage is required. In the traditional inspection method, the operator measures the size of the damage using a crack ruler or a crack microscope. In this case, the measurement value changes according to the operator’s experience, and the result of the state diagnosis is inconsistent. Therefore, the demand for new alternatives to visual inspection methods that depend on subjective judgment has gradually increased.

To perform an objective inspection, a crack-measurement technique using images is proposed in this study. This technology comprises a deep learning algorithm to detect cracks and a stereovision-based triangulation technique to measure their size. First, a hierarchical deep neural network is proposed to develop a crack-detection algorithm. The proposed neural network is more efficient than neural networks used in other fields owing to better crack-detection performance and lower computational amount. Moreover, using adversarial learning as a type of semi-supervised learning, a new fusion with a hierarchical deep neural network is presented and contributes to improving the recognition performance. In addition, a method is proposed to solve the problem wherein the recognition performance deteriorates when the deep learning algorithm is applied to a new environment owing to high data dependence. Through the application of the ensemble technique to adversarial learning, the deep learning detection algorithm increased the applicability to various crack images. Subsequently, a technique for measuring the size of cracks using stereo vision was developed based on a crack-detection algorithm. Because the diagnosis of the condition of a structure is determined based on the physical size of the damage, measurement accuracy is important. Therefore, a triangulation technique was chosen