Tantussi, G.; Lanzetta, M. and Santochi, M.

This article shows the results of a preliminary study developed in order to solve one of the most important problems in the field of marble, determined by the need of automatically select the tiles as they are extracted from a slab.

The study has been divided in the following phases:

- to single out suitable sorting criteria;

- to overcome the lighting and images acquisition and pre-processing problems;

- to choose statistic parameters to optimize the selection process.

The analysis of the results on a sufficiently wide collection of samples has shown a very good agreement between the automatic selection and the manual one carried out accordingly with the new criteria.

Therefore a selection in homogeneous classes with similar characteristics is needed.

Actually this operation is manually performed by specialized staff, since the untrained human eye is not able to appreciate, at the production speed, the belonging of tiles to the different classes requested by the market.

This selection method is severely subjective due to the following factors:

- lack of coded criteria;

- different workman’s experience and skill;

- workman’s stress;

- different personnel in different shifts.

An automatic system based on objective sorting criteria may produce the following advantages:

- possibility of increasing productivity according to the hardware power;

- labor cost reduction;

- reduction of errors due to lack of attention or stress for the process monotony.

In addition, by increasing the number of classes the final product value is improved since more homogeneous choices imply a better aesthetic result. This also implies further economic advantages since the economic value of a choice is determined by the worst quality elements.

An automatic selection system involves the following indirect advantages:

- reduction of customer's claims;

- integration with the warehouse management information system;

- statistics on the produced material percentage in the different classes in order to determine the raw material need to get tiles belonging to a certain class;

- graphic database management and computer aided virtual layout to simulate the effects of tiles combinations.

According to principles commonly applied in the quality control in similar fields, such as wood [1], [2], [3], textile [4], etc., in order to automatically select tiles, it has been carried out a preliminary study of an artificial vision based selection system to be exploited in an industrial environment.

At this early stage the attention has been focused to Carrara marble tiles which contain only different gray - levels. An example is shown in fig. 1.

To get full advantage from the possibilities of an automatic selection system, it has been necessary to establish at the beginning, objective selection criteria [5].

Moreover, to improve the final product value, the number of classes has been increased in comparison with the four actually on the market.

To deal with physical data, a set of 49 300X300 mm sized samples, identified by a progressive number has been examined.

The features characterizing the examined tiles are:

- the background, defined as the brighter shaded region;

- the spot, jargon word to indicate the presence of widely different shaped dark elements.

Working in tight connection with specialized staff charged with the selection task, two basic kinds of tiles have been identified: the "veined" and the "clouded" ones. The first ones contain sufficiently wide background areas and more or less thin shaped darker areas (vein or "spot"); for the second ones the bright and dark areas are evenly distributed on the whole surface.

For the "veined" tiles the following parameters have been taken into consideration:

- the background brightness;

- the spot area.

The manual selection has singled out four classes according to increasing values of the background darkness, from A to D, and seven classes for the spot region, from 0, minimum area, to 6, maximum area. The 43 "veined" tiles have been therefore arranged in the grid in fig. 2.

Not all the boxes have been filled, since on the workman's opinion more intermediate grades were possible among those available in the considered set.

The same operation has been achieved for the "clouded" tiles. The parameters taken into consideration are:

- the mean gray - level of the whole tile surface;

- the contrast among brighter and darker areas.

Three classes have been singled out for the mean value (NA, NB, NC) and three classes for the contrast (0, 1, 2). The 6 "clouded" tiles have been located in the grid shown in fig. 3.

At this research stage the sorting is still subjective but it's based on explicit criteria.

The aim of the next step has been the choosing of the most suitable statistic parameters to characterize the features described earlier and to relate the ranges of those parameters to each class.

The images acquisition has been obtained by means of a non professional TV camera and acquisition board controlled by an IBM-compatible Personal Computer.

A proper tile positioning and lighting system has been designed and realized as well.

It consists of an optically isolated tunnel, lighted by pentaphosphorus fluorescent lamps able to ensure the following performances: short and long term time-invariant characteristics, minimum heat generation, absence of reflections on the tile glazed surface and of direct radiation to the TV camera lens.

Before dealing with the analysis of data, it has been necessary to pre-process images to overcome the following problems due to the particular acquisition hardware:

- zero - average random noise introduced by the analog to digital converter;

- unequal lighting distribution;

- time and space sensitivity variation of the TV camera CCD;

- the automatic gain control cannot be disconnected.

To get the final image, the following operations are made [6].

To reduce the noise effects each image is obtained as the average image of 4 acquisition.

The effects of unequal lighting and of the space sensitivity of the TV camera CCD are corrected analyzing a sample tile made of white Plexiglas whose image has been previously divided in 8x8 sectors. This number represents a compromise between spatial resolution distribution and computing time.

Analyzing the sample image, the coefficients to correct the corresponding areas are calculated in order to get a uniform output from the system. This way of partitioning is used both in the pre-processing phase and for the next investigations about the spatial background and spot distribution.

The effect of the automatic gain control is to reduce the differences among different brightness levels tiles. Since it cannot be disconnected, the images have been corrected utilizing two white Plexiglas stripes placed along the tile sides. The dimensions of the stripes are such as to entirely fill the TV camera field of view with an aspect ratio of 4:3 between the two sides.

The whole image of the tile is then corrected with a further coefficient given by the ratio between mean lightness of the stripes and the reference white value conventionally fixed at level 230.

Finally the two white stripes and a slight border containing some elements outside the tile from the positioning system are eliminated from the image.

The data in RGB format can then be converted to monochrome without loss of information because of the absence of color components.

For the tile sorting, algorithms and procedures extensively reported in literature have been employed [6], [7], [8], [9], [10], [11].

The pre-processed images have the following features: 416x416 pixel on 256 gray - levels (lightness). Each image is divided in 64 square sectors.

In order to locate the background and the spot, the gray - levels histograms of each sector are analyzed by means of the following statistic parameters:

- mean value;

- skewness;

- mode;

- variance.

The mean value and the mode represent the gray - level of each sector.

The skewness expresses the possible presence of dark elements on an uniform background.

The variance expresses the spot presence and contrast.

To maintain the information about the spot distribution, the analysis is operated on each sector of the tile.

Fig. 4 - Tile n. 20: histograms of the 16 sectors

In fig. 4 the histograms of each sector of tile n. 20 are represented. For a clearer representation, the tile has been divided in just 4x4 sectors.

To find the sectors containing a prevalence of background, the variance is compared with a "threshold". The " threshold " is determined from the variance of a white tile with less than 2% spot.

In the three dimensions graph of fig. 5 each sector variance of the same tile is shown. The "threshold" is shown as well to indicate the sectors containing only the background.

If the sector number is limited to few (£ 6), the image represents a "clouded" tile whose belonging class is related to the whole surface mean value and variance, else the tile is a "veined" one.

Fig. 5 - Tile n. 20: variance of the 64 sectors

In this case, operating with the background sectors pixels lightness, the mean Mf, representing the background level and the standard deviation SDf are calculated.

Furthermore, to find the spot, a threshold value is calculated by the formula

threshold value = Mf - 3 x SDf

The spot percentage is expressed by the ratio between the number of pixels under the threshold value and their total number.

The result of the automatic classification of the tiles is synthetically shown in fig. 6 and 7.

Fig. 6, related to the "veined" tiles, associates the background mean gray - level and the spot area respectively to the letters from A to D and to the numbers from 0 to 6 as already shown in fig. 2.

It is to be noticed that the spot percentage associated to the same number may vary in relation to the background value, as shown by the solid separation lines among the classes.

In the figure the tiles position according to the automatic classification is represented as well.

With reference to the results of the manual sorting the disagreements indicated in tab. I can be noticed.

This discrepancies, only concerning neighboring indexes, are probably due to a subjective and [lightly] incorrect [inexact, inaccurate] manual selection.

Fig. 7, related to the "clouded" tiles, shows the correspondence between the mean gray - level and the variance parameters of the whole tile with the alphabetical and numerical indexes NA to NC and 0 to 2 already displayed in fig. 3. The graph also indicates each tile position after the automatic classification. In this case a perfect agreement with the manual sorting has been achieved.

This method utilized for the automatic sorting of the examined marble tiles has shown to be easily and cheaply implemented.

This study could be pushed forward finding further sorting parameters that could be more effective operating on different marbles. It is even possible to better characterize the spot kind, position or orientation finding, for instance, the preferential directions of veins, [the] texture, etc.

A relevant part of the work is due to the particular low cost hardware utilized for this preliminary study.

In spite of the slight resources [low cost means], the achieved results show the high potentiality coming from the introduction of the artificial vision in the quality control in the field of marble. The hardware limitations have been overcome through pre-processing thus severely affecting the computation time. For this reason the selection time resulted in less than 40 s. In order to align the computation time to the production speed, about 5 s, various options are possible. In particular:

- to improve the hardware in order to reduce the pre-processing time;

- to utilize several systems on different lines;

- to parallel - compute distributing different tile sectors to several boards;

- in the case in which the color information becomes relevant, to parallel - compute each color component.

An alternative option consists in making use of conveniently configured and trained neural networks. For this goal several tests have been performed to detect the sectors containing the background or the spot with favorable results.

This system has the benefit to be easily trained by non - specialized staff even if experience and performing many trials is needed in order to find the better configuration for each kind of material.

The authors wish to thank ing. Marco Pardini for his contribution to this work and Industria Marmi Pietrasantina S.r.l. for supplying the tiles and their technical support.

[1] Ciccotelli, J.; Portala, J.-F.; 1992, Applications of artificial vision in the wood industry, Industrial Metrology, 2, 3-4: 185-194

[2] Cho, T.-H.; Conners, R. W.; Araman, P. A.; 1991, A computer vision system for analyzing images of rough hardwood lumber, Proceedings - International Conference on Pattern Recognition, 1: 726-728

[3] Conners, R. W.; Cho, T.-H.; Ng, C. T.; Drayer, T. H.; Araman, P. A.; Brisbin, R. L.; 1992, Machine vision system for automatically grading hardwood lumber, Industrial Metrology, 2, 3-4: 317-342

[4] Erenyi, I.; Pongracz, J.; 1991, Quality control in textile industry via machine vision, Microprocessing and Microprogramming, 32, 1-5: 807-813

[5] Ersboll, B. K.; Conradsen, K.; 1992, Automated grading of wood slabs. The development of a prototype system, Industrial Metrology, 2, 3-4: 219-236

[6] Sanz, J. L. C.; Petkovic, D.; 1988, Machine Vision Algorithms for Automated Inspection of Thin-Film Disk Heads, IEEE Transaction on Pattern Analysis and Machine Intelligence, 10, 6: 830-848

[7] Ersboll, B. K.; Conradsen, K.; 1993, Strategy for grading natural materials using a two-step classification procedure, Proceedings of SPIE - The International Society for Optical Engineering, 1821: 318-329

[8] Russ, J. C.; 1992, The Image Processing, CRC Press

[9] Gonzales, R. C.; Wintz, P.; 1977, Digital Image Processing, Addison-Wesley, Massachussetts

[10] Castleman, K. R.; 1979, Digital Image Processing, Prentice-Hall, New Jersey

[11] Baxes, G. A.; 1994, Digital Image Processing, J. Wiley, New York