(A) study on segmentation model using deep learning for effective crack detection in tunnel

Abstract

As technology develops, the utilization of tunnels is diversifying, and at the same time, the importance of maintenance and repair of structures is being emphasized. However, despite the recognition of its high importance, regular inspections are conducted by manpower, still require a lot of time, and rely on methods that lack objectivity. In order to solve this problem, many technologies using high-resolution image processing methodologies have been developed, and in recent years, researches using deep learning technologies have been conducted in various ways. Nevertheless, micro-cracks occurring on the tunnel surface are still challenging because their shapes vary, and their recognition performance varies depending on shooting conditions. In addition, more research has been focused on improving the crack detection recognition rate, and further studies are still needed to obtain information for crack diagnosis. This thesis proposes a new tunnel crack detection method using semantic segmentation and a new measurement method using image processing. A new segmentation algorithm using hierarchical convolutional neural networks is developed to improve the speed of crack detection, and a multi-loss update method is proposed to improve accuracy. The performance of the model is evaluated in terms of accuracy and speed. As a result, an improvement of 2.165% for intersection over union of crack (IoU of crack) and an increase of 1.97 times for inference speed is found. Also, measurements on the shape, area, length, and width of the crack are derived using a reference instrument and compared and evaluated in terms of precision and accuracy with the method using the field of view (FOV), which is existing technology. As a result, reducing the error of changes in working distance and shooting angle is derived. This achievement is expected to be utilized as an accurate inspection technology to improve the safety of tunnel maintenance in the future.