Bridge-surface panoramic-image generation for automated bridge-inspection using deepmatching

Visual inspection is important for the efficient maintenance of bridge structures and has recently been supplemented with the use of image-processing techniques that can localize and quantify damages using images captured from bridges. A series of overlapping bridge images can be combined for constructing a panoramic bridge-surface image in which the locations and sizes of the damages can be noted. Despite the excellent performance of image-processing techniques, generating panoramic images from a series of bridge-surface images is challenging as bridge-surface images may not possess distinct patterns or patterns that can act as reference feature points for stitching adjacent images. To address this issue, this paper presents a general method for stitching bridge-surface images using Deepmatching, which determines a pixel-wise correspondence between an image pair in comparison with conventional feature-wise matching methods. To employ Deepmatching for panoramic-image generation, (1) image matching pair search using 2D Delaunay triangulation, (2) parametric model for optimal image stitching were developed, and (3) field validation was conducted in this study. First, possible image matching pairs are organized using the two-dimensional Delaunay triangulation, and then Deepmatching is used to determine the matching points between possible image pairs. The developed parametric model refines the valid image matching pair, which is used for obtaining optimal global homographies for panoramic-image generation. For the validation of the proposed method, a lab-scale experiment on a flat concrete wall and a field experiment on a concrete bridge were conducted. The experimental validation demonstrates that the proposed method successfully identifies dense matching points between image pairs and generates a panoramic image while minimizing the occurrence of ghosting and drift.

Automated summary

Visual inspection is crucial for the maintenance of bridge structures, now enhanced by image-processing techniques that can localize damages. However, generating panoramic images from a series of bridge-surface images is challenging. This study introduces a method using Deepmatching for stitching bridge-surface images, with successful validation in lab-scale and field experiments.