The algorithm has been tested on real crack images of various sizes. In additions to results shown earlier, other results are presented here.
The questions still needing to be answered at this point are: 1) Should multi-orientation filtering be used or a single structuring element is sufficient to enhance the crack patterns? 2) Which shape and size of the structuring element should be used? 3) Which automatic thresholding algorithm is more suitable? 4) Should variable thresholding be used and, if it should, what should be the grid size? Difficulties occur in answering these questions on the basis that the process is entirely dependent on human observation (i.e. no strong analytical method).
Refering to Figure 3.8, it can be seen that multi-orientation filtering is effective when the crack patterns possess near-similar orientation properties as the structuring element. However, the vast variety of crack patterns makes this technique less practical, since, in order to represent more orientations, more structuring elements have to be used and this increases the computational load. To make matters more difficult, the enhancing ability of the multi-orientation filtering approach does not really overcome the problem of the single structuring filtering approach when applied on somewhat circular crack patterns (see Figure
3.9) and the same goes for rectangularly-arranged crack patterns (see Figure 3.10). As the crack patterns are not intended to be enhanced orientation-wise, the multi-orientation filtering approach offers no clear advantage over single structuring element filtering.
As for the shape size/dimension of the structuring element, the disk-shaped structuring element with a dimension of 5x5 is used solely based on continuous observations. Getting near-perfect accuracy is not a high priority, since the main target in the end is to classify
patterns and it is believed, as long as the stronger features of the crack patterns are detected, there will not be significant loss of information.
Going into the segmentation stage, the significant difference between the Otsu and the SIS technique is that the former consistently produces higher threshold values compared to the latter on the majority of occasions. Consequently, this raises concerns as to whether it is better to have a high threshold or a low threshold. In principle, high threshold might remove some information while low threshold can produce unwanted noise in the segmented image. Again the choice of technique is based on observations and Figures 3.15,3.16 and
3.17 visualise some of the detected crack patterns using the Otsu and SIS techniques.
(a)
(b) (c)
Figure 3.15: First sample comparison between crack detected using automatic threshold
techniques: (a) original image; (b) Otsu technique; (c) SIS technique.
Inconsistent illumination and contrast can cause major difficulties in segmenting crack patterns if appropriate strategies are not taken as visualised in Figure 3.2. Four different approaches are compared and Figures 3.18, 3.19 and 3.20 visualise the comparisons for some selected images. The approaches are:
i. straightforward thresholding using Otsu technique,
ii. CTH operation followed by segmentation using the Otsu technique,
(a)
(b) (c)
Figure 3.16: Second sample comparison between crack detected using automatic thresh-
old techniques: (a) original image; (b) Otsu technique; (c) SIS technique.
(a)
(b) (c)
Figure 3.17: Third sample comparison between crack detected using automatic threshold
iv. CTH operation followed by a variable thresholding using Otsu technique with grid size 64.
Although in principle, thresholding alone in some cases can be used to segment crack patterns, it has been proven that the CTH operation contributes significantly in enhancing the suspected crack patterns, as evident in the sample results, with approaches (ii) and (iv) producing the better results. There is no doubt at this point that the CTH operation will be a vital component in the crack detection approach in this thesis. However, it is still necessary to determine how effective variable thresholding is.
The variable thresholding approach is seen as a good step in dealing with illumination inconsistency as demonstrated in Figure 3.21, using Otsu technique with grid size (G) 64. The dotted rectangle highlights the region which global thresholding failed to segment properly, but which was dealt with effectively by the variable thresholding approach.
However, variable thresholding can cause oversegmentation in areas where there are no cracks, especially if the grid size is too small. In this case, “cracks” will be forced to appear, as can be seen in Figure3.22.
These circumstances raised a dilemma in terms of choosing between global thresholding and variable thresholding. The tradeoff between the two approaches is the amount of noise. Global thresholding works well for small images but in very large images, some features will be left out, while variable thresholding tends to introduce unwanted noise especially in areas not affected by cracks.
3.8
Summary
This chapter has shown how crack patterns can be detected using the morphological top- hat operation. It is a widely used technique for extracting ridge-valley structures. The importance of the technique in detecting cracks has been shown and the top-hat operation is considered as a vital component.
An automatic segmentation strategy has also been presented, using the Otsu technique, which has been demonstrated and proven to be useful in separating enhanced crack patterns. Several known obstacles in performing the task have been also outlined particularly those related to contrast and illumination inconsistency within a crack image. The issue regarding the size of the crack patterns has also been raised. This matter is not discussed in detail,
since it is less important. The main interest here is to capture the dominant features in a crack pattern and the approach undertaken is seen to be sufficient.
Multi-orientation filtering was also attempted to see the effect it has in enhancing crack patterns. Cracks are elongated structures which in a way match the shape of a rectangular (one-dimensional) structuring element. Experiments were conducted to demonstrate the effectiveness. The key point in multi-orientation filtering is angle resolution where in order to represent more angles, more structuring elements must be used. The drawback is higher computational load caused not only by a bigger structuring element size, but also by the extra number of structuring elements used. Furthermore, based on the experiments, multi- orientation filtering does not offer a significant advantage over the single structuring element approach.
Another critical point raised is the use of variable thresholding in segmenting the enhanced crack patterns. Oversegmentation occurs when variable thresholding is used particularly affecting areas without cracks. A portion of Chapter 4 will discuss a technique to further eliminate unwanted elements within the detected crack image.
(a) Original image.
(b) Method (i). (c) Method (ii).
(d) Method (iii). (e) Method (iv).
Figure 3.18: First example of crack patterns detected using four methods, namely (i)
straightforward thresholding, (ii) CTH operation followed by thresholding, (iii) variable thresholding using grid size 64, and (iv) CTH operation followed by variable thresholding
(a) Original image.
(b) Method (i). (c) Method (ii).
(d) Method (iii). (e) Method (iv).
Figure 3.19: Second example of crack patterns detected using four methods, namely (i)
straightforward thresholding, (ii) CTH operation followed by thresholding, (iii) variable thresholding using grid size 64, and (iv) CTH operation followed by variable thresholding
(a) Original image.
(b) Method (i). (c) Method (ii).
(d) Method (iii). (e) Method (iv).
Figure 3.20: Third example of crack patterns detected using four methods, namely (i)
straightforward thresholding, (ii) CTH operation followed by thresholding, (iii) variable thresholding using grid size 64, and (iv) CTH operation followed by variable thresholding
(a) Original image.
(b) Cracks segmented using manual thresholding.
(c) Cracks segmented using variable thresholding.
(a)
(b) (c)
(d) (e)
Figure 3.22: Figure showing the oversegmentation effect when “cracks” are forced to
appear, as the variable thresholding technique assumes the existence of cracks in each square region. The rectangles highlight the regions affected by oversegmentation, clearly spotted in the image produced from approach (iv): (a) the original image; (b) cracks detected using approach (ii); (c) cracks detected using approach (iv); (d) the thinned
Craquelure Representation
4.1
Introduction
Having segmented the crack contours from the background, the next stage involves repre- senting the contours in a different form such that analysis can be made effectively. This chapter presents a stage-by-stage process of converting the crack contours from an image- based representation into a numerically structured representation, with the main aim to provide a platform from simple, effective and flexible data manipulation for use in the later stages. An approximation scheme using conservative shapes is also introduced to facilitate both feature extraction and area of interest determination. Finally, a crack pruning stage is presented which aims at removing noise and insignificant crack patterns.