Damage Evaluation of RC Deck by Pattern Recognition

In this study, the damage of Reinforced Concrete (RC) bridge decks is evaluated with the aid of digital photos and pattern recognition (Gonzales and Woods, 2002; Yagi, 2000). In general, the procedure for extracting the characteristics of cracks showing up on concrete decks through digital images and classification based on the damage levels are used in the typical pattern recognition system. Figure 1 shows the procedure of such a pattern recognition system. First, the input data to


M. Pandey et al. (eds), Advances in Engineering Structures, Mechanics & Construction, 81-86 © 2006 Springer. Printed in the Netherlands.

Fig. 1. The procedure of pattern recognition system.

this system consists of the digital images of the concrete decks taken by a digital camera. Next, the linear pattern of cracks is extracted from the digital images of the concrete decks through image processing techniques. Moreover, the characteristics of cracks such as the projection histograms are extracted. Finally, the digital images of cracks are classified into different damage levels based on the extracted characteristics through the TAM (Topographic Attentive Mapping) network (Seul, 2001).

To obtain the test material, digital images of concrete decks taken by a digital camera are used. If input data that can be acquired in low resolution and by using common digital camera is used, the costs for the assessment of integrity can be reduced and input data can be acquired easily. The total number of digital images is 47 and each image is scanned with the resolution of 360 pixels per inch in both directions. In this resolution, each image is normalized to the 768 x 480 pixel rectangle and converted to greyscale image. The digital images used in this study are obtained by marking the cracks with white chalk. The damage levels for all digital images are classified into three categories by an expert. Some examples for each damage level are shown in Figure 2.

Cracks in a digital image of concrete deck are detected in accordance with the following pro­cedure: First, the digital image undergoes geometric transformation to extract a rectangular part containing a crack zone. The binarization is a method for transforming greyscale image pixels into either black or white pixels by selecting a threshold. Because the crack zone existed in only a small part of digital image, and also the brightness is not uniform through the crack zone due to the uneven lighting, the extracted rectangular part is divided into smaller blocks. The method proposed by Ohtsu (Yagi, 2000) is applied to the block unit to determine the threshold for binary-coding processing. Then, each block is divided into sub-blocks and the binary-coding processing is applied to each sub-block. These binary images are reduced some noise such as spots and holes after the binary­coding processing. The aim of thinning processing is to reduce the crack zone pixels to lines one pixel width. The crack pattern can be easily recognized by such a thinning processing. Finally, the smoothing processing such as the reduction of insufficient points and the addition of missing line is implemented. After all of these processing, the crack pattern is obtained and used for extracting the characteristics of digital images. The procedure is shown in Figure 3.

In this study, characteristics are extracted based on four criteria; continuity, concentration, dir­ectionality (unidirectional or bi-directional), and types (hexagonal or linear) of cracks. The crack pattern of thin lines can be considered a set of directional linear elements and hence characteristics extraction by the projection histogram (Seul, 2001; Sakai, 2002; Duda et al., 2001) would be ef­fective. Because the characteristics of projection histogram of a crack pattern provide information

(a) Images of damage level I

(b) Images of damage level II

(c) Images of damage level III

Fig. 2. Examples of images associated with different damage levels in concrete decks.

on the positions and the quantities of cracks, they can be used as the quantitative characteristics representing the continuity and the concentration of cracks, for the classification of crack patterns. The histograms projected on two directions are computed for extracting a crack pattern; one is the horizontal direction and the other is the vertical direction. The projection histograms are data structures used to count the number of crack pixels when the image is projected on the vertical and horizontal axes.

The characteristic values in each dimension are the number of crack pixels in row for the hori­zontal histogram, in column for the vertical histogram, and are the quantum numbers in accordance with the dimensionality of characteristics vectors. Figure 4 shows an example of horizontal and vertical projection histograms extracted from a crack pattern.