!FREE! Crack Application Mover 4 3
In past years, inspection of cracks has been done manually by careful and experienced inspectors, a method that is subjective and scarcely efficient. Besides, the poor lighting conditions in the tunnels make it hard for inspectors to see cracks from a distance. Therefore, developing an automatic crack detection and classification method is the inevitable way to solve the problem. The work presented herein endeavors to solve the issues with current crack detection and classification practices, and it is developed for achieving high performance in the following three aspects:
Crack Application Mover 4 3
The above three requirements are the principles for developing the automatic crack detection and classification method. First of all, to guarantee high detection rate, the captured tunnel images should be able to present cracks as much as possible, thus the captured images should have acceptable resolutions. The high-speed CMOS line scan cameras [1] are able to shrink both the space required and the cost involved in securing superior speed, reliability, and image quality. Hence, they have been installed in the subway train and have been used as image sensors to perceive the images of tunnel surfaces. With the obtained high-quality images, image processing techniques become the main issue to be solved.
To be efficient, the image processing techniques used here must avoid complex computations as a prerequisite for high detection rate and accuracy. In gray-scale images, cracks present themselves as dark regions with local minimum gray-level components. Morphological image processing operations have an advantage in segmenting relevant structures without complex iterative calculations. For this reason, they can be used to detect local dark regions in the original gray scale image, in which most irrelevant objects with high gray levels or small regions can be easily eliminated by thresholding operations. To automatically separate the cracks, parameter setting involved in the crack segmentation process is an important subject to be studied. Thus, how to find the optimal parameter settings is another issue to be solved, and it can be answered by experimental results from practical applications.
After the crack segmentation process, there may still be many misidentified objects that appear as cracks. In attempting to distinguish between cracks and unexpected irrelevant objects, feature extraction becomes the key problem to be solved. Distance transformation is usually used in shape segmentation, which inspired us to use distance to map spatial shape into a probability sequence by distance. Thus, a shape descriptor called a distance histogram is proposed to perceive the difference between cracks and irregular objects. The standard deviation of the distance histogram is an effective criterion for describing the degree of irregularity of a spatial shape. Along with the standard deviation of the distance histogram, two additional numerical features are used as the basis for classifying the cracks. With a pattern recognition algorithm or a thresholding classification operation, unexpected objects will be removed and cracks will remain. In conclusion, the main contributions of the proposed methodology include the following three aspects:
Mathematical morphology is another flexible image processing technique. In [11], cracks in digitized paintings are detected by a top-hat transform, and the misidentified thin dark strokes are excluded by hue and saturation. Still, many irrelevant objects are misidentified as cracks. Miwa et al. [12] uses watershed segmentation to find the linear shape of cracks, but the result distorts the original pattern of the cracks. In [13], a morphological image processing and thresholding method is applied to detect and classify pipe defects. While this method is simple and easily implemented, some dark, thin, or small noises are still misidentified as cracks. In [14], a ranged image morphology-based crack detection for steel slabs is presented, in which over 80% of cracks are classified by the automated on-line detection system.
where ND is the pixel number in region D and xn are all the pixels in D. D can be square or circular, and its window size can be adjusted according to the resolution of the image. A too large window size of D will eliminate the details of cracks and a too small window size of D is not able to effectively smooth the original image. Therefore, the window size of the average filter should be initialized properly. As shown in (b), the unnecessary peaks and valleys are eliminated by average filtering while preserving the details of cracks. Note that (a) is not specific, because almost every original image is polluted by these salt-pepper like noises. In the following section, positive effects of the average filtering will be illustrated by examples.
where w0, w1, μ0 and μ1 denote the background occurrence probability, objective occurrence probability, background mean levels, and objective mean levels, respectively [22]. The threshold for separating cracks is slightly modified as
where ta > 0 denotes the acceleration parameter and its unit is a gray level. A suitable value of ta removes the pixels with gray levels lower than T, thus improving the crack detection efficiency. However, an overlarge value of ta will remove too many pixels with gray levels lower than T, decreasing the crack detection accuracy. Hence, it should be initialized properly. The pixel x in binary representation of fBTH can be expressed as
The obtained binary image is expected to give a relatively clear presentation of the cracks compared with the original gray-scale image. Figure 5 is an example illustrating the comparison of the obtained binary images, and its size is 1669 1019: (a) is the original gray-scale image, which it contains a Y-shaped crack; (b) is the blurred gray image with an averaging filter window size of four. Their corresponding binary images are (c) and (d), respectively. The binary image (d) obtained from smoothed image (b) clearly represents the cracks. However, cracks in (c) are hard to distinguish. Apparently, average filtering eliminates unnecessary details in the original images, which improves the performance of the crack segmentation process.
In Figure 5, many local dark regions or pixels are misidentified as cracks and most of them are noises with small regional sizes. Therefore, a morphological algebraic area opening operation [23], which is an adaptive filter removing the connected components less than a certain number, can be utilized to filter the irrelevant noises. The area opening operation can be expressed as
where NA denotes the number threshold value and X represents the connected set of fT. NB(X) is the pixel number of the connected set X distributed in SE B. NA can be set according to specific applications, usually it can be set from 300 to 700. Examples of the obtained binary sub-graphs from Figure 5 are shown in Figure 6.
Figure 6 shows the segmentation results of the Y-shaped crack with different parameter settings, (a), (c), and (e) are the binary images obtained from Figure 5c, (b), (d), and (f) are binary images obtained from Figure 5d. Apparently, binary images obtained from the non-smoothed images are heavily polluted by the unnecessary details, which makes cracks hard to be separated using the thresholding operation. Compared with (a), (c), and (e), crack details in (b), (d), and (f) are clearly presented to us. These results prove that the averaging filtering process is able keep the crack details when removing noises from the original images.
As shown in Figure 7a, the 24 candidate objects are obtained after the previous area opening process. Nos. 1, 2, 3, 6, 8, 16, 19, 22, 23, 24 are irrelevant irregular objects that need to be removed; the rest are cracks. The shape distance histogram standard deviations of the 24 candidate objects are shown in Figure 6b. Note that sizes of these 24 candidate objects have been adjusted for better presentation and the shape distance histogram standard deviations are calculated from their original shapes. The standard deviations of irrelevant objects are all larger than one, which is larger than the objects of the cracks. Apparently, the statistical results show that the standard deviation of shape distance histogram is an effective criterion for crack classification.
In practice, an ELM is used for both its universal approximation and classification capabilities. And also, it is easy to implement without iterative calculations because its input weights are randomly generated and independent from any specific applications [29]. With adequate training samples, the ELM is able to classify newly observed objects. Given a candidate object, the output is
where fL (X) denotes the sub-graph after thresholding operation, Tσ and Tg are the thresholds of shape distance standard deviation and average gray level, respectively. Note that the pixel number is not used here to avoid removing too many crack-like objects.
Example of the classified cracks, in which (a) represents the original image; (b) the detected cracks before classification; (c) the classification results by ELM; and (d) the classification result by thresholding operation.
where ni denotes the number of the pixels judged as cracks in the transversal of mi. Admittedly, the crack quantification is not able to show the real lengths of the cracks when these cracks are detected as discontiguous parts. The accuracy of the crack quantification method with different λ and Tw will be presented in the experimental section.
where LDC and LD denote the length of detected cracks and length of all detected objects, respectively. LDC and LD can be obtained from the automatic detecting results. Loc represents the length of the original cracks, which is measured by on-the-spot investigation. Before the experiments, a tunnel section (50 m) is determined and its images of lateral surfaces are captured and stored. Then the DR and DA are calculated according to the automatic detected results and the on-the-spot inspection results.