Bridges, Tunnels and other Transportation Structures
Full Session with Abstracts
The surface crack in concrete structures is an important indicator for structural damage and concrete durability. For the structural safety, concrete cracks are visually inspected, monitoring crack-related information such as location, length, and width. Even conducted by well-trained inspectors, the visual inspection is often considered to be insufficient and ineffective in terms of cost-effectiveness, inspection time, safety issues, and assessment accuracy. Recently, digital image processing has been introduced to more efficiently and objectively obtain crack information from concrete surface images. An important challenge for the use of the image processing in practice is how we can automatically identify cracks from a large amounts of concrete surface images containing actual cracks as well as crack-like noises such as dark shadows, holes, stains, and lumps. In addition, the ordinary RGB cameras provide only 2-dimensional RGB information, which is insufficient for calculating crack sizes when the cameras are not perpendicular to the concrete surface. Thus, the previous studies have assumed that images are obtained from the cameras placed perpendicularly to the surface. However, it is expected that the camera and concrete surface are often difficult to be aligned in the inspection of full-scale civil structures.
This study proposes a machine learning approach for concrete crack identification that can distinguish the actual cracks from the crack-like noncracks objects and determine crack sizes from the concrete surface images taken from cameras at arbitrary angles. This study employs an RGBD camera that provides RGB and depth information for determination of crack existence, location, and size. In the training stage of the proposed approach, a wide variety of concrete surface images are acquired, from which the regions of crack candidate are extracted utilizing image binarization. A convolutional neural network is trained using the crack candidates containing actual cracks and crack-like noncracks. Subsequently, crack identification is conducted using concrete surface images to validate the crack detection performance of the proposed approach. Once cracks are determined, the crack size is calculated using the RGBD information.