Вы находитесь на странице: 1из 4

507

SURFACE DEFECT AND TEXTURE IDENTIFICATION IN CONSTRUCTION MATERIALS

A Boyd, C Connolly, A Starr.

University of Huddersfield, UK.

INTRODUCTION

3. The texture or the defects in each sample is then identified.

This paper describes an investigation into the assessment of inconsistencies in texture of concrete materials, using image processing. This work is being carried out in collaboration with a major manufacturer of construction materials. The materials should have a fairly rough but uniform surface (Figure 1). Rejected samples exhibiting surface openness and cracks respectively, are shown in Figure 2 and Figure 3. Texture assessment is complicated by the fact that the materials are a random mix of several different colour concretes. The materials are inspected on wooden boards. This is currently done by eye but gives inconsistent results. Inspectors tend to do local subjective comparisons between samples rather than comparing each object to a target, or objective measurements. In addition, the repetitive nature of the work makes consistent inspection tedious. This often leads to misclassification of samples, which can result in either a reduction in the quality or an increase in the scrap materials. Both of these cases directly effect costs. For these reasons an automated system of inspection is desirable. The production rate is 48 items every 16 seconds, so for the required 100% inspection the application sets a considerable challenge. Figure 1 Smooth Texture

Figure 2 Open Texture

INVESTIGATION METHOD

The approach used can be broken down into 3 main stages:

1. Finding the location of each of the items in the image. The production variables mean that the items can move as much as five centimetres in a horizontal plane.
2. A region of interest is then extracted for each sample of size 48 x 100 Pixels. By processing a smaller

region of interest, means that time is not wasted processing unwanted information.

Figure 3 Cracked Surface

IPA97, 15-17 July 1997, Conference Publication No. 443 0 IEE, 1997
Authorized licensed use limited to: CSIR Madras Complex. Downloaded on January 12, 2010 at 00:01 from IEEE Xplore. Restrictions apply.

508

Location of the Samples

Two different methods of locating the items were tested: colour segmentation and a method which uses the pattern of the position of the items. Since the items produced should be within a specified range of colours, colour segmentation can be performed to separate the objects from their background. Such a method is adopted by Sobottka and Pitas (1) to locate skin-like regions for face location. In their approach Sobottka and Pitas use the HSV colour space to segment the image, identifying a range of hue and saturation values which encompass the skin colours. An alternative colour thresholding technique was tested using the CIE 1976 L* a* b* equation. The CIELAB colour space was used by Tominaga in image processing (2). It was found that since the objects under study are predominately red. The a* value varied with the exact colour of the samples but the L* and b* values remained fairly constant. Therefore segmentation could be achieved by thresholding using limits set for the L* and b* values. However the boards the items are made on, tend to pick up colour pigment on them. Combined with the wear and tear they are subjected to during the manufacturing process, make this method unsuitable. An alternative method of identifying the objects makes use of the fact that the items are set out in rows and columns. A monochrome image is used and a threshold level is selected such that all the items are white and the shadows cast between them are black. This enables the positions of the 20 items in the centre of the board to be found. From these co-ordinates, the positions of the others items can easily be found, since all the items are evenly spaced. This method does not suffer from problems previously outlined.

interest. Then by selecting various features different objects can be easily distinguished. A classic example is that of an image of a nut and bolt, The bolt can be identified as the blob with the greater length measurement and the nut as the blob with a hole in it. It can also be used in this application, by careful selection of a suitable threshold level the granularity or defects on the surface can be highlighted. It has been found that various geometric features, including area and compactness, can then be used to make a decision about whether an object passes or fails.

CLASSIFICATION METHOD

Initial investigations were carried out on 12 samples of different textures and a mix of colours which had been classified by a inspector on the shop floor. An image was captured using a colour camera in near uniform lighting conditions. The following things were discovered: Open texture samples tend to have more blobs than smooth texture samples. Samples with defects of a large size are failed A compactness measure is useful in identifying cracks in the samples. Research by Julesz (4) shows that the human ability to recognise texture quickly is based on textons, which are elongated blobs. A standard measure of blob compactness is derived from the perimeter (P) and area (A) of the blob:
D2

Compactness = -L-(474
Texture and Defect Identification A circular blob is the most compact and is defined to have a compactness measure of 1.0; more convoluted shapes will have a larger value. This measure does not describe how dense the blobs are or their orientation, but it does detect the presence of elongated blobs in the
image.

Karayalannis and Stourattis (3) suggest a texture inspection system for the classification of fabrics. This method is far less computationally expensive than the popular spatial grey level dependence matrix, often used for texture problems. In this method the mean and the variance of the non defective material are computed and are compared to the material under inspection. This method was tested for this application, however it was found that it was unable to identify cracks in otherwise smooth texture items. Blob analysis is very commonly used to identify and to distinguish between objects within an image. The image is first binarized to identify the objects of

A decision-making rule has been formulated based upon the number, size and shape of the blobs in each item. This rule has been empirically derived.

1. Samples with more than 50 blobs are failed


2. Samples with blobs of an area greater than 20 pixels are failed

Authorized licensed use limited to: CSIR Madras Complex. Downloaded on January 12, 2010 at 00:01 from IEEE Xplore. Restrictions apply.

509

3. Samples with more than 2 blobs with a compactness measure greater than 2.0 are failed. 4. Samples not having any blobs with a compactness measure greater than 2.0, are passed provided that they have met all the previous conditions.
5. If the sample has not been classified by rules 1-4, a crack identification test is performed on it.

images are extracted from the colour image, the results of the texture and defect identification from each image are compared. The two monochrome images are: the L* image from the CIELAB colour space, and the luminance image from the HSL (Hue Saturation, Luminance) colour space. The location of each of the samples was found by the method previously described and the region of interest (48 x 100 pixels) was extracted within the centre of each object. The variance of the grey levels was calculated for each region of interest, these were plotted and compared. Both plots showed peaks at samples 7, 9,34,43. This is shown in Figure 5 (L* plot). Items 7, 34, 43 were agreed unanimously to be bad texture. Item 7 is borderline and was classified as bad texture by the shop floor but was identified as just acceptable by the expert panel.

Crack Identification Test

To identify the direction and the density of the blobs in the remaining items, slices of the binary image of the item are extracted, centred on the least compact blob, first in a horizontal direction and then in a vertical direction, This is shown in Figure 4. The number of highlighted pixels in each region is then calculated. Expressed as a percentage, this value can be used to determine the nature of the fault. If the percentage figure for the X direction is twice or more than that for the Y direction then the sample is failed due to a horizontal fault. Likewise if the percentage figure for the Y direction is twice or more than that for the X direction then the sample is failed due to a vertical fault. If neither of these cases are satisfied, then the sample is classified as pass.

0 1

1 0

2 3

J)

40

50

Sample Number Figure 5 : Plot of the Variance Results for the L* Image

Fig. 4 Binary image showing crack identificationtest

The blob analysis method of texture identification was tested on both of the monochrome images. The correlation between the expert panel, the shop floor and the decision from L* and L images has been calculated and is expressed in Table 1, below. It can be seen that the highest correlation is between the shop floor decision and the L image results. If however we take the expert panel decisions to be the most accurate, the image with highest correlation to that is the L* image.

RESULTS
Investigations have been carried out on 48 items of different textures and blend of colour. All the items have been previously classified on the shop floor and also by a panel of experts. Images of the items are captured using a colour camera in near uniform lighting conditions. Two monochrome

TABLE 1-Table showing correlation between classification results

I Decision I ExDert I ShoD I

L*

Authorized licensed use limited to: CSIR Madras Complex. Downloaded on January 12, 2010 at 00:01 from IEEE Xplore. Restrictions apply.

510

Figure 6 shows the classification results for each item. A score of -4 indicates that the items has been identified as bad texture by all parties, likewise a score of +4 indicates that all the classifiers were in agreement and passed the item as good texture. Out of the 48 items 23 (48%) were agreed by all groups to be either bad or good texture. Agree Pass

REFERENCES

1. Sobottka & Pitas, 1996. Face Localization and facial feature extraction based on shape and colour information. IEEE International Conference on Image Processing. 2. S Tominaga. 1992. Colour Classification of Natural Color Images. Color Research & Application. 17. No. 4. 3. Karaylannis & Stourattis. 1993. Fault Classification of fabrics based on statistical measures. Conference on efficient texture analvsis methods - advanced methods. 72 - 8 1.
4. Julesz. 1981. Textons, the elements of texture perception, and their interactions, Nature. 290. 9 197.

. ...

.. a

.. .
m ..

-3 -4

I .... .

. .. .

Agree Fail

Sample Number

Figure 6 : Classification results of each item.

CONCLUSION

Using the techniques outlined of successive elimination of failed items, combined with the simple blob analysis approach, image capture and classification of 48 items can be achieved in under 10 seconds. This makes this method suitable for 100% inspection at the present production rate. It was found that the variance figures alone provided limited classification of the items, regardless of which of the 2 colour spaces were used. (HSL or CIELAB) If the L* image is used for the blob analysis procedure, this provides results that are more in agreement with the expert panel decision. From the 48 items that were classified 23 of them had textures that were agreed by all to be either good or bad. By identifying textures that most human inspectors are in agreement. The system could be trained, possibly by using a neural network to respond to individual demands and markets.

ACKNOWLEDGEMENT

The work reported in this paper is part of a PhD project supported by a Bursary from the School of Engineering at the University of Huddersfield.

Authorized licensed use limited to: CSIR Madras Complex. Downloaded on January 12, 2010 at 00:01 from IEEE Xplore. Restrictions apply.

Вам также может понравиться