Developing an image processing algorithm for detection of deformations of tunnel walls

Masato UKAI Senior Engineer, Signaling Systems Group, Transport Systems Development Division, Technological Development Department

Structural soundness of railway tunnels is currently inspected by visual monitoring of deformations such as cracks on the wall. However, work environment is so bad and monitoring area is so extensive that it is very difficult to follow all deformations developing on the wall to the last crack. Recently, inspection system using continuous scan image (CSI) taken by a linear sensor camera, has reached a practical application stage. This system has a possibility to enable us to diagnose soundness and durability of tunnel walls more accurately.
We have developed an algorithm for detection of deformations of tunnel walls from CSI images. We applied various techniques in this algorithm. One is to distinguish cracks from the CSI image of tunnel walls. Actually there are several noise elements in the obtained image. We must separate real cracks from other noises. The other is to measure various dimensions such as length or width of cracks. These data make it possible to get a ranking of the deformations, and accordingly to get a prognosis of their growth.
As tunnel walls have curved surfaces, it is difficult to illuminate the walls evenly. Thus, lighting inconsistencies are inevitable in the images taken. It is necessary to correct these lighting inconsistencies and the luminance differences of tunnel the wall surface itself. For the correction, we perform a dynamic binarization which changes the threshold level in linkage with partial area features.
Considering the possible workstation indication size with no reduction or shifting, the smallest block is set as 864~864 pixels. By composing crack images taken from each block after processing, the long crack in the tunnel wall surface will be processed. As the size of the tunnel wall image to be processed is as large as 16,000 ~1,000,000 pixels, we are working for improving the processing speed and for making adjoining information processing smoother.
Our algorithm consists of processing in dynamic binarization, dilating and eroding for crack junction, eliminating particles after the crack junction, and analyzing particles for a crack prospective sampling. Regarding the particle analysis, we measure the length of direction X, direction Y, and Feret's occupancy rate. Each process has a different parameter, and we have tried various parameters through test programs and decided the best value by examining the situations for juncture, and crack and noise sampling.
In order to improve the detecting rate, we investigated ways of distinguishing joints from cracks on concrete walls. Promoting the detection of cracks with restraint on horizontal joints was made possible by utilizing the following: spatial frequency filters which separate the widths of horizontal joints and cracks, and the characteristic that the position changes in the direction in which the cracks grow larger than that in the joint construction direction. Furthermore, by applying a spatial frequency filter with multiple directions and bandwidths, we found that prospective cracks could be sampled in stages. Nevertheless, there is a risk of picking up horizontal joints with similar widths as cracks. In actuality, it is difficult to 100% automatically pick up the cracks alone. Taking this fact into consideration, we are examining to provide our present system with a function by which people can make corrections in sampling results through interaction.
We verified the possibility of detecting typical cracks by using the proposed algorithm mentioned above. Especially comparison in time series, including old and new images, is so significant for inspection of facilities that we have been developing such function. But there remain several problems in this algorithm to be solved. It is necessary to upgrade the ability of detection, to shorten processing time on a workstation, to cope with a wider variety of uses and to develop a measuring method of crack size.