Civil engineering structures such as bridges need to be periodically inspected in order to assess their current functional state, predict their future condition, and inform their maintenance and rehabilitation decision. Among various inspection procedures, visual inspection of surface defects is more prevalent because, considering the physical size and number of bridges, vision-based methods are simple and inexpensive. To facilitate the visual inspection procedure, computer vision techniques can be applied on the inspection images to extract defect information. More advanced methods can combine the deep learning techniques with a remote image capturing device to perform the task with minimal human intervention. However, most deep learning techniques were originally developed fo...[ Read more ]

Civil engineering structures such as bridges need to be periodically inspected in order to assess their current functional state, predict their future condition, and inform their maintenance and rehabilitation decision. Among various inspection procedures, visual inspection of surface defects is more prevalent because, considering the physical size and number of bridges, vision-based methods are simple and inexpensive. To facilitate the visual inspection procedure, computer vision techniques can be applied on the inspection images to extract defect information. More advanced methods can combine the deep learning techniques with a remote image capturing device to perform the task with minimal human intervention. However, most deep learning techniques were originally developed for common object recognition, direct adaptation of them in recognizing concrete surface defects that usually differ in appearance and data distribution from common objects may not be effective. Also, recent developments of remote visual sensing technologies tend to collect a large volume of complex visual data from many different viewpoints and distances. Such unfavorable conditions of image data lead to a questionable accuracy and inefficiency of existing vision-based inspection approaches.

This research aims to resolve these challenges by developing efficient and accurate techniques to detect, segment and quantify surface defects using inspection images taken under various conditions. For detecting multiple surface defects, a faster, simpler box-level defect detection approach is proposed based on a real-time common object detector. To enhance the detection accuracy, the original model is further improved by introducing a novel transfer learning method with fully pretrained weights from a larger dataset having the closest distribution of geometric attributes to the concrete defect dataset. Next, for segmenting each individual defect on an image, a fully convolutional model for simultaneous pixel-level defect detection and grouping is proposed. The proposed model integrates an optimized mask subnet with a box-level detection network in an efficient and fully convolutional sense, where the former is responsible for pixel-level defect detection and the latter groups detected pixels for each defect. Finally, for quantifying surface defect, a detailed quantification method focusing on concrete crack is developed. This method aggregates the identified crack information from a set of unordered images using 3D reconstruction and Bayesian data fusion. The output is a voxel-based 3D crack representation model with detailed properties of each crack segment in a complete crack.

This research develops efficient and accurate surface defect detection, segmentation and quantification techniques, which are able to significantly reduce the effort required by human experts in the visual inspection of concrete bridges. The outcome of this research will also greatly enhance the development of fully automated structural assessment technique. Such an advancement can further facilitate the implementation of modern technology in the civil engineering discipline.