be-5638 : assist : automatic system for surface inspection
Transcript
be-5638 : assist : automatic system for surface inspection
SYNTHESIS REPORT FOR PUBLICATION CONTRACT No: BRE2-CT9241260 PROJECT N“: TITLE: ASSIST (Automatic System for Surface Inspection and Sorting of Tiles) PROJECT 20-ORDINATOR: Bull HN Italia PARTNERS: I-T One Tile Thomson B University of Surrey - DEEE University di Geneva - DIBE Thomson TCS [Associated to Thomson B) STARTING DATE: PROJECT FUNDED BY THE EUROPEAN COMMUNITY UNDER THE BRITEIEURAM PROGRAMME DURATION: 43 months ASSIST: Automatic System for Surface Inspection and Sorting of Tiles Report Coordinators: VSSP Department of Electronic and Electrical Engineering University of Surrey Guildford GU2 5XH, United Kingdom 1. H’UTRODUCTION The ceramic tiles industrial sector has taken significant advantage of the advances in the world of automation in recent years. Ail production phases have been addressed through various technical innovations, with the exception of the final stage of the manufacturing process, namely the product hspecfion. This is still performed manually and is concerned with the sorting of tiles into distinct categories or the rejection of the tiles found with defects and pattern faults. The research effort expended upon the problem of objectively inspecting, analysing and characterizing ceramic tiles is easily justified by the commercial and safety benefits to the industry: I ■ automation of a currently obsolete and subjective manual inspection procedure 8 100% product inspection ■ more robust and less costly inspetilon m higher homogeneity within sorted classes of products ■ increased processing stability and improved overall production performances through the removal of a major bottleneck continuation and consolidation of the leadership currently enjoyed by the European Community in this area [n this paper we describe the integrated system developed under the ASSIST project, for the detection of defects on colour tiles and for the colour grading of defect-free tiles. In Section 2 an overview of the sensor developed specifically for this project is presented. In Section 3 the algorithms developed for the detection of the defects will be described, while the coiour grading framework is discussed in Sectkm 4. In Section 5 the various types of defects are quantified geometrically and statistically. In Section 6 the system integration is presented, and finally in Section 7 the performance evaluation of the system is discussed. 2. THE SYSTEM SENSORS To meet the ASSIST demands, a multi-camera vision approach has been developed in order to cover many different kinds of defects. The acquisition system which has been designed for the ASSIST project, is mainly made up of three CCD-based cameras. 2.1 Colour CCD image sensor The first camera designed and developed for the project is a colour camera (TH7821 C), based on a high resolution line array device sensor. This is a trilinear array with up to 3 x 8640 pixets, featuring very high performance in terms of linearity, dynamic range, and low noise response. It uses high quality on-chip colour filters to obtain full colour information from a single chip. The following topics highlight the electro-optical performances that make this sensor the ideal device to satisfy the ASSkST demands. Signal to noise ratio The TH7821 C image sensor was especially designed to provide very high saturation output voltage (3 V) in spite of its small pixel size (7x7 micron). By measuring temporal noise in darkness based on correlated double sampling, an analogue signal processing technique employed allows the suppression of the reset noise and the reduction of low frequency noise contributions. The F\MS value corresponds to 32 electrons, mainly contributed by the sensor output amplifier. The high saturation voltage and very low temporal noise make the TH7821 C a fri-CCIY colour image sensor with the smallest pixel size and hig’best dynamic range in the world, featuring 15000 (83 dB) signal to noise ratio (relative to RMS noise in darkness). Charge transfer efficiency The typicai value of the total charge transfer efficiency is about 0.98 and the charge transfer inefficiency deduced for one CCD output register stage is equal to 5x1 C@. These results indicate a very good charge transfer for 6 cm iength (XD shift registers. Lag This important parameter corresponds to the charge transfer efficiency from the pixel storage region to the CCD shift register. For a colour image sensor, lag effects induce error in colour detection, together with reduced resolution in the scanning direction. To obtain a good colour restitution and therefore to avoid false colour phenomena, a new pixel technology was introduced in the TH”7821 C image sensor to minimise the lag effect. The lag effect was measured for an average signal from 50 mV to 600 lmV and the results obtained for this tricolour CCD image sensor show excellent transfer efficiency {cO.3%) as compared with a conventional linear CCD image sensor (>5%). Responsivity A linear image sensor with photodiodes and 100% pixel aperture features high quantum efficiency providing a maximum sensitivity with an improved responsivity at 400 nm (blue band). [n order to further improve this sensor, new colour filters with high transmission and low crosstalk have been developed. The primary ccdour filter transmission curves were optimised for colour restitution and the crosstalk between Blue/Green and Green/Red were reduced to suppress coiour mistakes due to colour mixing. Linearity Linearity is mainly dictated by the output amplifier characteristics. For a dynamic range of 20mV to 2500mV the linearity deviation is less than 2%. This value is better than 0.5% from 20mV to 500mV which fully corresponds to the ASSIST needs. These very good [inearities are required for RGB channels to ensure perfect colour restitution, allowing then good defect recognition in order to perform reliable colour shade grading of ceramic tiles. 2.2 CoIour line-scan camera, . ? ~ # a ! s The colour line-scan camera is used for the high precision colour grading of the tiles. It is fitted with the integrated 3 X 8640 element-per-line, colour (2CD image sensor described in Section 2.1. It provides three -bit digitised” colour outputs, one for each of the RGB channels. The circuitry within the camera comprises logic and driver controls of the CCD image sensor, the sensor itsellf, video preprocessing and processing blocks, the analog to digital converters and the line memories. An infra-red reject optical filter and a lens mounting adaptor are included in the enclosure. To achieve fast scanning it includes a ve~ high resolution trilinear architecture array sensor and functionalities for the remote control of the operational parameters such as gain and black level adjustments of each RGB video channel. These parameters will be tuned by the computer platform depending on the light conditions detected and the colour of the tiles. Since each colour has different sensitivity and responsivity characteristics, it is necessa~ to tune each colour in the camera for the maximum dynamic range of all colours for a given light source or coloured tile. The trilinear CCD image sensor ow’ttains 3 coloured pixel arrays (red, green, and blue) which are implemented in parallel. Each line array is separated from the next by a space equal to 42 lines (294 micron). This separation between RGB lines shows that at a given time, each pixel array does not capture the same image. A digital system mainly made up of FIFO memories and EPLD devices to controt read and write FIFO operations has been designed for the automatic restoration of the image geometry starting from 3 image lines acquired in different spatial positions. Three FIFO banks are acthg as digital delay lines to store respectively 15 red lines, 8 green lines and only 1 blue line, assuming that the red line is captured first. Once these lines’ have been stored, the 3 memory banks corresponding to each colour are simultaneously read, providing 3 superposed images. 8 ‘it I 2.3 BNV cameras There are two Black and White cameras used in the system and they are both fitted with linear CCD arrays. Three boards are currently used in these cameras: a sensor driver board, a timing generator board, and a digital video 3 processing board. Each key component such as the sampling circuit to improve the CCD signal to noise ratio, the low noise video amplifier, the high speed drivers and the fiexible programmable logic devices have been carefidly tested and hence selected to improve and to exploit CCD performances. A major concern is the pairing effect which is caused by signal difference (gain and DC level misalignment) between odd and even CCD outputs. This produces a kind of fixed pattern noise at half the pixel frequency. In order to cancel this effect, the two CCD channels have been carefully matched from dark to white level so that the pairing effect is less than 1%. In order to minimise transfer inefficiency and a dark signal component that increases with increasing temperature, the drivers and the trilinear colour CCD need to be thermally coupled to the camera front end. For each 80C step, the dark current doubles, reducing the signal to noise ratio. High performance silicon grease ensures an ef[cient thermal exchange between CCD as well as CCD drivers and the camera front end fitted with the lens mounting system. 3. TILE DEFECT DETECTION The expected result of a defect detection module is the identification of defective regions (defect localisation) and the quantification of defect parameters (shape, extension, etc.) to be used for further classification purposes, Depending on the number of defects and their dimensions, the tiles are grouped into: m First Class (none or very few acceptable defects) z Second Class (few but still acceptable defects) M Waste (unacceptable defects) I u 1“ Some of the most common and anti-aesthetic defects found on both dain and textured tiles can be cateaorised as cracks, bumps, depressions, pin-holes, dirt, drops, undulations, and colour and texture defects. These are pr&ented in more detail in Table 1. : :%Efissuresandwts tenthsofmm osomecms : E!==== ‘ ’ depressions holes dirt drops water drops undulations ‘ F;y :fiousm o ~ ‘ ’ ’ reliefs of the glazed surface with few tenths to some mms respect to the planar surface of tile general depression of glazed surface radius usually <5mm with a circular shape pin-tips, small bubbles, holes, craters minimum 0.25mm dust particles or dried glaze residuals few tenths to some mms drops during glaze app[i@ion few tenths to very low mms irregular shapes (condensation drops) few mms to cms longitudinal bars, rills and imperfect widths and lengths of few cms ’ u rs blobs discemib[e from regular texture few tenths of mm to cms few tenths of mm to cms Table 1: Typical Ceramic Tile Defects ., We now provide an overview of the most innovative approaches developed by the consortium for detecting different types of defects in tile images. # 3.1 Inspection of the edges of the tile The border area of a tile is a critical area for physical defects since it is the most fragile. The inspection of tile borders, however, is particularly difficult because, due to the geometry of both material and acquisition systems, the image of the tile border suffers from the presence of shadows, reflectance, and discontinuities. 4 The problem was tacklecf in two ways; initially, by adopting appropriate sensing and lighting techniques to reduce as much as possible the acquisition problems (shading, reflections and non-uniform illumination), and subsequently by preprocessing the border area to produce a set of data compatible with the techniques used for the inner zone. This last operation was performed by using an algorithm made up of three steps: 1. Identification of the extericr border of the tile 2. Compensation of the luminance variation in the border area 3. Extension of the image by a pre-defined number of image lines (mirroring). 3.2 Crack detection in uniformiy-coloured tiles Cracks are linear structures contrasting with the rest of the tile. Our method [1] consists of two 1 D convolutions, in the horizontal and vertical directions respectively with line detection filters [2]. Local maxima in the output indicate the possible presence of a line, and trigger the hypothesis that a line is present. The shape of the output signal around a local maximum is compared with the expected shape if a line were present in order to confirm or reject the hypothesis. 3.3 Spot-like defect detection in uniformly-co[oured tiles On light-coloured plain tiles, small, :;pot-like faults are of reasonably high contrast against the background. However, due to various sources of noise, e.g. non-uniform illumination, a simple threshold will not serve as an adequate solution to their detetilon. Thus, an adaptaticm of the line filter method from Section 3.2 was developed for spot-like defects [1]. The only difference is that the tile image is convolved with only one filter which is optimised for spot profiles. The spot peaks thus enhanced are extracted by thresholding. Figure 1 shows a plain and a lightly-textured tile with detected defects. ● 8 1 Figure 1: A plain white tile and a lightly-textured tile with small crack and spot defects 3.4 Crack detection on textured tiles: A quick solution Morphological techniques were used for the identification of cracks [3,4]. In particular, we used the difference between a Closing and an Opening for the identification of thin structures. On the basis of these operations, a criterion called Anomaly Presence Degree (APD) has been proposed for the identification of the defective parts of the image. During the anaiysis phase the image is subdivided in blocks of dimension L x L, and for each block the APD parameter is 5 evaluated and compared with a threshold: if the threshold is exceeded, the block is classified as anomalous, otherwise it is considered regular. The goal in an adaptation algorithm used, if applied to a normal image used for training, is to minimise the maximum APD value present in the image. 3.5 Crack detection on textured tiles: A more sophisticated approach We use the cojoint spatial and spatial frequency representation of the Wigner Distribution [1,5,6] to enhance pattern separability as the crack and texture patterns have disjoint support regions in the cojoirrt representation. According to this method, at each pixel position (x,y) we calculate the Fourier transform of a non-linear combination of pixel values within a window of size N x N centred at pixel (x,y). I e The Wigner distribution is a real function and its components constitute the feature vector at each pixel position. All local Wtgner spectral components are normalised by their corresponding dc component, so that only the general shape characteristics of each spectrum are captured. These components form the features that describe the texture pattern. The statistical distribution of these features is computed from the defect-free image during an off-line training phase. In the testing stage, the Mahaianobis distance of the feature vector of each pixel from this distribution is calculated. The values of this distance are used to form a residual map image. This image is subsequently processed by the optima! linear filter mentioned in Section 3.2 to detect the cracks. Figure 2 shows the app]i~~on of our approach to !&@ner distribution for the detection of (synfheticaliy produced) cracks_on highly textured tiles — ) Figure 2: Rossini textured tile with short crack defect and Donizetti textured tile with long crack defect 3.6 Blob-detection: a quick approach The image is subdivided into square blocks which are classified on the basis of their rank functions or morphological properties. The generalised rank function provides exactiy the same information as the histogram but has the advantage of making it possible to define an efficient distance measure that complies with the three following requirements: {~ to be zero if and oniy if the two histograms are exactly equal; (ii) to be propotilonal to the distotilon both for the concentration and for the values; (iio to weight the point differences in relation to their distances from the mean value. If a uniform texture is partitioned into sufficiently large blocks, each block is representative of the whole texture: the problem is then to define a prototype of the texture from the analysis of the block histograms. By using the rank distances, our strategy is to define as a prototype the rank funtilon computed on the block, and when testing an unknown texture to evaluate the distance between the rank function of each block and the prototype ~6,7,8J. 6 3.7 Blob-detection: a more sophisticated approach Using a perfect tile during the training stage, the various colour categories present in the defect-free tile can be identified with the aid of K-means clustering in RGB space. The number of these clusters is chosen to be high so that over-segmentation into chromatic classes is obtained, thus minimizing (and eliminating) the under-segregation error. Next, these clusters are transformed into CIE-LUV unifoml colour space for perceptual merging, i.e. merging of small clusters into super-clusters. Thus, the image is segregated into chromatic categories which are perceptually uniform. The image can then be split into a stack of binary images one for each chromatic category. We perform morphological smoothing on each binary image to remove noise before characterizing the structure of the left-over blobs. For each blob we compute as structural features its area, perimeter fractality, elongatedness, and some spatial information about the distribution of other blobs around it, and model the distribution of these attributes. During testing, the image pixels are classified into the chromatic categories defined during the training stage using the nearest neighbour rule. Any unclassified pixels are rejected and considered as colour defects. The accepted pixels are used to form the stack of binary blob images again. The structural features of each resulting blob are then computed and any blob-like texture defects are identified by means of the Mahalanobis distance function using the structure statistics saved in the training phase. Figure 3 shows the application of our technique to a real defect in a Rlain tile and a simulated defect on a highly textured tile. e P Figure 3: Plain white tile with large water drop defect and Puccini textured tile with blob defect 3.8 Regular Pattern Fault Detection The research activity carried out in the field of regular tile pattern analysis was mainly concerned with two problems: first, the registration of the acquired data (i.e., the digitised image of the tile to be analysed) with a template (target pattern); second, the identification of faults (relevant differences between actual and target patterns). The analysis has to be performed in full colour, due to the possibility of errors in colour layer superposition. Due to the very high computational requirements of such analysis it was decided to reduce the resolution of the sensor used for performing the acquisition of these tiles: an RGB coiour camera at 24 bpp with resolution 1024xI 024 pixeis has been judged suficient to achieve a good quaiity evaluation for tiles up to the maximum standard format. As it was verified that non-negligible shifts and rotations can happen among different layers of the same inspected tiie (generated by successive applications of siik-screen coiour drawings) it resuited that a reliabie registration of inspected tiies with the prototype (tempiate) can oniy be obtained if the matching is performed for each iayer separately. Such iocai shifts can in fact produce apparent errors (imperfect superposition between current and reference tile) which have 7 not to be detected as real flaws. Therefore, a layer splitthg strategy was studied (based on K-Means clustering approach) and further analysis was performed only at the layer level. The registration algorithm developed takes into account two primary objectives: to maintain the distance between template and inspected pattern within one pixel accuracy and to minimise the computational requirements. The solution adopted consists of the use of a fast roto-translation procedure, that allows a fast optimisation of the angle and shift parameters. The mean absolute difference between corresponding pixels was selected as a measure of correct pattern reproduction, because it showed the best trade-off between speed and reliability. Testing results showed that defects with total dimension as low as 3 pixel (at a resolution of 1024x 1024 pixel per tile) can be successfully detected by applying the proposed strategy. 4, COLOUR SHADE GRADING The scientific advances of the method developed consist of: ■ the development of an automatic scheme that can cope with spatial and temporal variability of illumination ■ the development of methodology that allows the replacement of the spectral blurring of a colour, caused by the electronic sensor, by the spectral blurring caused by the human colour sensory system. - The developed methodology allows the identification of coiour grades that correspond to the threshold of human colour perception which are discriminated by human inspectors wcrldng at the peak of their performance. To achieve this, our methodology had to be able to measure colour differences at least one order of magnitude smaller than the various types of noise involved in the process of colour recording. I ‘8 4.1 Spatial and temporal illumination correction The spatial variation of the illumination manifests itself as a low frequency variation over the profile of the tile. Assuming for the moment that the inspected tile is of uniform colour, we fit the grey level function of the tile in a single colour band with a low order polynomial using the least squares error method. This polynomial fitting is equivalent to identifying the low frequency component of the function. At each location then, the actual value of the grey level function is divided by the corresponding value of the fitted polynomial and multiplied with some reference value. This way, the low frequency variation due to variable illumination is removed and all values are referred to the reference illumination value chosen. This process is repeated for all colour bands separately. Thus, the multiplicative noise due to the spatial variation of the illumination is taken care of. n To take care of the temporal variation of the illumination, we have to image each tile with the same reference surface, another tile which plays the role of a pallet. When the system is operatio~al, the difference in the observed variation of grey value that can be attributed to the temporal illumination variation is removed with the hehI of the reference surface and only the difference that remains is re”corded as genuine variation. In other words, all ‘the ~ecorded signals are referred to the same reference signai recorded at a specific time [9,1 O]. After the above corrections take place, the tiles are graded by comparing their colour histograms with a reference histogram and using the correlation coefficient as a measure of similarity [1 1,12,13]. 4.2 Using the results of Psychophysics research All the above described methodology was applied to several series of ceramic tiles which had been colour graded by human inspectors. In most of the cases the tiles could be clustered in clusters identical to the cfusters the human experts had clustered them into. In some cases, however, there was confusion and disagreement with the human classification. That disagreement was removed when the difference in the electronic and the human sensor response was taken into consideration in the way described below. The visible part of the e[ectro-magnetic spectrum was discretised and represented by the values at 31 equidistant wavelengths. The sensitivity of each colour sensor and of the cones of the human retina were thus expressed by vectors of length 31. The spectrum of the illuminating source was also a similar known vector. The intrinsic reflectance 8 of the inspected sutface was expressed by an unknown vector of the same length. As we only know the recordings of the three colour sensors, we have an under-determined system of equations which cannot be solved for the intrinsic reflectance values, as we have 28 degrees of freedom. We make the following assumption: The transformation between the three recorded intensities by the electronic sensors and the three intensities the human sensors would have had recorded, is affine. We created a large number of corresponding triplets of recordings using a Monte Carlo approach. We then solved in the least square error sense to identify the e[ements of the affine transformation between the two sets of recordings. This transformation can be used to predict what the human sensor would have recorded, given what the electronic sensors have recorded. Then we transform the data into L.UV coordinates, where we know that the Euclidean distance between any two points in this space is proportional to the perceived colour difference behveen the two colours represented by these two points. When the above process was applied to several series of tiles graded by human observers, almost total agreement behiveen the clusters created by the human observem and the computer was achieved [14,15]. When one deals with finely textured surfaces, however, the dominant effect is not so much the difference in spectral blurring between the eye and the sensor, as the dlferenf pok?f spread function of the electronic sensor and the eye. The latest Psychophysical experiments have shown that humans perceive texture in a different way for different colours and there are blurring masks that have been derived that quantify the way we biur colour textures viewed from different distances. To take this effect into account when grading crdour textures, we first measured the spatiai blurring of our sensor for each of the three RGB channels, and performed a Wiener filtering method to restore each chancel. We then converted the coiour image into a set of new colours, called opponent colours where we applied the spatial blurring performed by the human eye at the distance of 40cm. The corrected tiles were then graded by comparing their colour histograms again. This time, however, the separation of ‘the grades the humans had defined was much better, and where some clusters were overlapping before, we had no overlapping grades this time at all [16, 1~. 5. DEFECT QUANTIFICATION Defect quantification concerns the extraction of numerical (quantitative) information from the defetifve reaion identified. in order’ to achieve a better knowledge of the defect cttaracteristi~’. The output data are used both” for statistica~ purposes and for allowing a feed-back on the production process. 8 The computation of numerical parameters on the defects required the definition and algorithmic implementation of a number of measures on the geometry and content of the flaws. A number of image processing algorithms were selected taking into account the strong real-time requirements. A list of the different defect types was defined, discriminating them on the basis of their visual characteristics. Consequently, a [ist of the more significant parameters was formulated, by taking into account the requirements of the following defect classification subsystem. Two main classes have been defined: defects representing a thin structure and defects representing a circular structure. For each of these categories, several geometrical and statistical measures have been proposed, listed in Table 2. Defect Thin Structures Other defects Geometry total length number of branches centroid diretilon [ineatity area centroid perimeter length minimum bounding rect. shape eccentricity aspect ratio moments Statistks mean value standard deviation variance of the length of branches mean value variance light variations optical centroid Table 2: Geometrical and Statistical measures for defects I Specific algorithms have been designed and implemented in software to compute the great part of the above mentioned measures, working on hi-level patterns simulating real defects after segmentation. The eticiency and the computational complexity of these algorithms have been tested and some improvements carried out in order to achieve peak performances. The calculation of such parameters required the development and optimisation of several computer graphics algorithms like: filling, thinning, skeletonization, oontour following, finite-state automatas, and so on. 6. SYSTEM INTEGRATION The integrated vision system is schematically presented in Figure 4. The tiles are first inspected for bumps and depressions using a laminated light source at an angle for maximum sensitivity to surface imperfections. This means that each line across the tile reeeives more or less light of the same possible intensity. Nexl, the tile is inspected for other structural faults like cracks and holes, using diffuse lighting. Finally, the tile is inspected for colour defects and colour grading using the tri-linear colour scanner and again, diffuse lighting. ~ I ........ 1 Figure 4: CMeralt ASSIST system structure The system architecture (Figure 5) has been designed according to a modular approach. Two main functional modules have been identified: the Host !vlodule and the Data Acquisition and Processing Module (DAPM). Selection control color ~ camera ~ b~w 1024~ camera blw3100 camera Host Module J Data AcquMtion Ramote Host Module ~ Local and Processing Module Host Module ~ 4 / 1 $ 5 VME Bus ● m Ethernet LAN Figure 5: Overall ASSIST system architecture The DAPM Module has been designed to contro[ image grabbing, to process the image, to classify tiles and to drive the proper 1/0 module for tile sorting. The Host Module is in charge of the interface with the human operator, output data collection and defect diagnosis. This module is split into a Local l-lost Module dedicated to provide 110 funtilonality (file access, screen and keyboard functions, etc.) to the image acquisition and processing sub-system and into a Remote Host Modu}e providing the graphical man-machine interface (MMI) and high-level data processing functionality (storing, statistics, and diagnosis). The DAPM Module and the Local Host Module are mapped onto the same machine, a rugged VME-bus machine with embedded PC able to operate in industrial environments, while the Remote Host Module is implemented on a standard PC. The DAPM and Local Host communication is based on VME bus while a IAN interface (Ethernet &pe) is chosen for inter-machine communications. This architecture, while allowing the use of rugged and costly machine only where needed, has the great advantage of being easily expandable : the Remote Host could control and collect data coming from more than one Local machine. 6,1 Hardware ”Architecture The system has been realised using two machines, a “local” one placed near the sorting line and a “remote” one that controls operations and collects outputs. While a standard PC is used for the remote machine, both standard and special hardware are used for the local one, A VME cabinet has been chosen to host both the Local Host, a standard PC, and the DAPM module, implemented using both off-the-shelf and specially designed boards. The DAPM module can be divided into three wmponents: Camera Control Modules, Camera Adapter Modules, and the DSP Network. Camera Control has been designed to receive synchronisation signals coming from an enmder and photocells and to generate line scar! camera timing (exposure time, pixel c?ock, etc.) white the Camera Adapter is needed to directly interface camera outputs to 13SP ports. Due to the high input data rate and to the high processing power required, the processing sub-module uses a network of high performance Digital Signal Processors based on the Texas TM S320C40 family. 6.2 Software Architecture A multi-tasking operating system has been chosen both for the Local Host and for the Remote one: Microsoft Windows for Workgroup has the features needed for multiprocessing and IAN communication requirements. An i10 Server daemcm has been implemented into the Locai Host in order to provide the DAPM moduie with ito functionality. The Remote Host is made up of the foilowing moduies: Supervisor Module, Output Data Collector Module, Statistics Module, Diagnostic Moduie. The DAPM Module is based on a DSP network ; a quick parallel kemei was chosen in order to satisfy real-time requirements. This kernel was then extended by developing its routing and distributed fiie system functionality. Message Router is in charge of routing message packets among the processors of the network while Link File System provides a mechanism for data exchange in the form of standard fiie operations (open, read, write, etc.) regardless of processes allocation. s 7. PERFORMANCE EVALUATION The integrated system was extensively evaluated on the factory floor, mainiy with respect to defect detection. in particular, the quick methods used for defect detection were extensively tested with respect to two major kinds of tiles: uniformiy coloured tiies and textured tiles. Performance is quite iike that of human operators in the case of uniformiy coloured tiles, as Figure 6(a) shows. It must be noted that the main difference in the resuits between the prototype and human classifications is due to defects beyond the ASSIST specifications (i.e. smalier size fauits than originally specified by the manufacturer), or due to minor problems (illumination or thresholds settings) which were iater solved. tom tm !ELl (a) r -! i a i -— (b) Figure 6: ASSIST vs. Human Accuracy in (a) Helsinki tiles (a white plain tile) and (b) Ecate tiles (a grey textured tile) t Textured tiles include a great variety of products, with different technologies and aspects . The main kinds are flamed tiles, randomly textured tiles (i.e. dropped tiles), and pseud~randomly textured tiles (i.e. marble like tiles). While flamed tiles (tiles with shadings caused by flame exposure) and dropped tiles [tiles coloured by drops, granite like) have been successfully tested, marble like tiles were beycmd the limits of effectiveness of the quick approaches. More preciseiy, defects caused by problems in glazing by means of silk-screen rolls, proved very difficult to be detected. Figure 6(b) shows the comparison between automatic and human classification in the case of a marble-like textured product. As a final evaluation, we can say that the quick methods developed are quite effective in defect detection in uniformly ccdoured tiles and in some kinds of textured tiles whereas they are less reliable in pseudo-random textures. 8. SUMMARY Abll) CONCLUSIONS We described here an integrated system for the inspection and colour grading of ceramic tiles. Although not all of the methods proposed are feasibie for real-time implementation with the existing technology, it is beiieved that perhaps in 5 years time real-time implementation wiii be possible. it is hoped that the considerable advance achieved in overali production through the automation of ceramic tiie inspection wiii eliminate an estimated 70-80%customer complaint rate regarding product quaiity. Furthermore, the spinoffs of the findings of this project can have an impact in other industrial fields presenting simiiar problems; for instance in the textde industry for defect detection, [oose threads detection, and colour shading classification on fabrics, the agro- 12 food industry for visual analysis of crops such as apples/orangeslpears/etc, the wood industry for texture and colour classification, and in a number of other industries. Acknowledgments Assist consortium is grateful to the European Community for the economic and trws.tful support. References il] [2] J. KHtler, R. Marik, M. Mirmehdi, M. Petrou, and J. Song. Detection of defects in colour texture surfaces. /APR Pmt. of Machine Vision Appficatiorrs 94, pages 558-567, December 1994. M. Petrou. Optimal convolution filters and an algorithm for the detection of linear feature. fEE Proceedings-1, 140(5):331-339, 1993. [3] R. Fioravanii, S. Fioravanti, P. Giusfo, and Vemazza G. Real-time texture defect detection. [n Proceedings of Phare-accord Workshop on Eii?cienf Texfure Analysis, volume 1, pages 57-66, 1994. [4] D. Corso, R. Fioravanti, and S. Fioravanti. Morphofogicd analysis of textured images for identification of thin structures. In Proceedings of LEEE kf. Conf OR Acoustic, Speech, and Signal Processing, 1995. [5] J. Kittler, R. Marik, M. Mirmehdi, M. Petrou, and K.Y. Song. The detection of local abnormalities in random macro textures. Technical report, University of &mey b’s5p-T??-4@5, fgg~. . [6] S. Fioravanti, R. Fioravanti, F.G. DeNataie, R. hfarik, M. Mkrnehdi, J. Kittler, and M. Petrou. Spectral and rank order approaches to texture analysis. Eurvpean Transactions on Telecommunications, 6(3):287-300, 1995. go 9 E 9 & m R. Fioravanti and S. Fioravanti. Texture description through rank-order functions, in Proceedings of JEEE Workshop on Nonlinear Signs/ and Image Processing, 1995. [8] F. G. B. DeNatale. Rank-order functions for image texture discrimination. h’rternational Journal of Pattern Recognition and Artificial Intelligence, To appear, 1996. [9] C. Boukouvalas, J. Kittler, R. Marik, and M. Petrou. Automatic grading of ceramic tiles using machine vision. in iEEE M Symposium on Industrial Electronics, pages 13-18, 1994. [1 o] C. Boukouvalas, J. Kittler, R. Marik, and M. Petrou. Automatic coiour grading of ceramic tiles using machine vision. IEEE Trans. Industrial Electronics, To appear February 1997. [11] C. Boukouvalas, J. Kittler, R. Marik, and M. Petrou, Automatic grading of textured ceramic tiles. In Proceedings of the SPIE, volume 2423, pages 248-266, 1994. [12] C. Boukouvalas, J. Kittler, R. Marik, and M. Petrou. Automatic colour grading of randomly textured ceramic tiles using machine vision. In IASTED M. Conf on Computer Applications in k?dusfry, pages 121 –1 23, 1995. [13] C. Boukouvatas, J. Kittler, R. Marik, and M. Petrou. Colour grading of randomly textured ceramic tiles using colour histograms. IEEE Trans. Industrial Electronics, Submitted, 1996. 11 , [14] M. Petrou and C. Boukouvalas. Model based cofour grading. In IEEE and LSTIMDSP Workshop, pages 110-111, 1996, [15] C. Boukouvaias and M. Petrou. Perceptual correction for co[ow grading using sensor transformations and metameric data. IEEE Trans. hdustry Applications, Submitted, 1996. [16] M. Petrou and C. Boukouvalas. Colour quantification for industrial inspection. In [EEE [WSIP, pages 6f 1-614, 1996. [4 q C. Boukouvalas and M. PeWou. Perceptual correction for colour grading of random textures. IEE Vision, Image and Signal Processing, Submitted, 1996. 8 I t 9 I , ~ 9 B ‘9 IS II 14 . EXPLOITATION REPORT CONFIDENTIAL CONTRACT N*: BRE2-CT92-0260 PROJECT N*: BE5638 TILE: ASSIST (Automatic System for Surface Inspection and Sorting of Tiles) COORDINATOR: Bull HN Italia PARTNERS: I-T One Tile Thomson B University of Surrey - DEEE Universit5 di Geneva - DIBE Thomson TCS [Associated to Thomson CSF) REFERENCE PERIOD: from ‘1 March,1993 to 30 September, 1996 STARTING DATE: March 1993 DURATION: 43 months PROJECT FUNDED BY THE EUROPEAN COMMUNITY UNDER THE BRITEIEURAM PROGRAMME pag 1 oflf) 1. Description of results A prototype system able to select ceramic tiles by means of image grabbing, storing and processing has been developed and tested with good results. The system, is currently able to operate at reduced speed (less than final system requirements) cm plain and textured tiles. Tiles are classified in up to four classes depending upon relevance of physical defects detected, in addition color shade classification can be performed. Recently, prototype of systems addressing the same problem have been presented to the market, but none of them seems to be completely adequate to the purpose. Some of them have been also installed in production plants, but not yet in Italy where the majority of the european tile production is located and where quality requirements for tiles are more stringent. We think that ASSIST may be well considered representative of the state of the art in vision technology applied to ceramic tiles inspection. Besides addressing its main goal, the development of this precomtpetitive system has brought to the availability of many building blocks and tools reusable in similar areas of application. The main reusable objects that can be identified are the following: 1. Advanced linear cameras, blacklwhite and color. The color camera is based on new 2. 3. 4. 5. 6. high resolution three linear sensors. A special controller for the direct correction of linear cameras to DSP modules, based on an innovative architecture which has drastically simplified the image acquisition channel. There is no need for a specific frame grabber module, but the data are feed directly into the DSP module. A mu[ti DSP software platform that can be use on systems with one to twenty 13SP modules, and can be easily expanded to support even more complex architectures. A sophisticated DSP algorithm library for image processing and acquisition An image analysis workbench integrated into a software development environment which simplifies the evaluation and development of new Visual Inspection solutions. A Methodology for visual inspection problems approach, from economical to technical evaluation of-new opportunities. Some of these have already been successfully tested by Bull in V[sual Inspection concerning different areas of application. Related Subject Descriptors The following codes applies to the scientific field of this project El 5 inspection, testing Technology for contact less, high speed, real time products inspection E37 Quality assurance System for acquisition and recording of quality data on 100% of production E32 Process control pag20f10 Real time products inspection allowing for immediate feedback on production process. Furthermore collected data are extremely useful for evaluation of Overall Equipment Efficiency (OEE) E04 Ceramic, glasses The main objective of the project is in the area of quality control and sorting for ceramic tile production D32 Signal processing Visual Inspection technology is based on fast signal detection, via linear or arrays cameras, and processing based on paralle[ DSP architecture 2. Industrial applications To date there is a growing interest in the manufacturing sector for automatic quality measuring systems strictly integrated with the production process. Visual inspection, by means of computer vision, is probably one of the most growing technology in the area of quality control. The value of this solutions is not only in the final sotiing process of good and defective products, where they are directly operating. But also in their capability of collecting real time data about the overall production process, the basis for successful plant maintenance and effectiveness improvement. Besides its final objective, which will require further development to move from the results of this precompetitive “research to the development of an industrial product, ASSIST has brought to the deve~opment of many applications taking advantage of the know how and building blocks of this project. identified areas of application, some of them already under exploitation, are the following: m 1. Automatic quaiity control of TV monitors 2. Inspection of silk-screen prints on electronic components 3. Inspection of packaging for alimentary products 4. Quality control of security printings 5. Quality control of wooden panels, raw or decorated 6. Quality control of other kind of plastic laminated materials 7. Inspection of plastic molded objects (2 D) Related market sectors The following codes apply to the market area of application of ASSIST results E32 wood products inspection of wooden panels produced in large mills, some of them using recycled materials like pallets or wooden boxes. The production system is fully automated and integrated. Quality information is of high importance to altow immediate feedback in case of abnormally defective products, or to identify possible problems cause. pag30f 10 E19 packaging Automatic inspection of packaging, mainly for the pharmaceutical and alimentary industry, petformed both at the packaging line or at the package production plant, is of paramount importance for product integrity and human safety. These production processes are so fast (10-20 parts per second) that 100% human inspection is absolutely impossible, while computer vision technology can be successfully used. El 5 manufacturing control systems In general a computer vision system can be directly connected to a plant information system providing real time data concerning 100 % product quality and plant performance. E23 printing, publishing The printing process is to date extremely fast (up to 6 m /see) and is sometimes used to print important. information (i.e. serial number on a check or a note). Human control of this process is impossible, while on time detection of errors may save the production of useless documents. E03 ceramic products The main target of ASSIST is the tile production industry that cannot implement a continuous production process only due to the need for manually inspecting each single tile produced. A significant economic impact is expected by the automation of this process. From the previous list if industrial sectors that can make use of computer vision technology we can see that many benefits might derive from it in two different but related areas: 100% product quality control and real time equipment efficiency monitoring. The vision technology is expected to spread in the manufacturing industry starting from those sectors for which one of this two aspects is more important (which means convenience to invest at to date cost of vision technology, ). With the expected cost reduction of these systems in the coming years more companies and even other sectors wili find convenient their usage. From a point of view of companies involved in bringing vision solution to the market, we think (and have also experienced) that many companies wii[ be involved in this process: ● Modules suppliers 1. Companies specializing in optics and illumination techniques 2. Companies producing CCD sensors and DSP processors 3. Companies producing cameras ● Research institution developing innovative algorithms ● System Integration Companies providing the integrated vision system (or better “vision engine”) ☛ Factory automation companies pag40f10 capable of integrating vision systems with the specific automation required for each target sectors (tiles, wood panels, plastic, print) having a good level of competence in that market . End users (production plants) in different sectors All these kind of companies will be involved in the process of bringing eticiently to the market vision systems. 3= Market analysis The tile production in 1995 had a positive performance of plus fl 0 ‘A in quantity over 1994. This represents a significant increase in line with predictions. A more detailed ana!ysjs indicates a different trend for the various kind of products: Type of tile on tot. prod. 62.60 16.68 14.71 6.01 0/0 single-firing double firing vitrified stoneware (gres) other trend stable new kilns 16 notes - -1.5Y0 +12,62% ~? 14 constant increase Production distribution in 1995 (Italy) , year ?989 1990 1991 ? 992 A 993 fi 994 ~ 995 quantity million mq 434 446 432 434 458 510 562 variation +72.47% +2.92% -3.19?40 +0,51$4 +5.53!40 +1 1% +10% Italian tile production from 1989 Even the forecast for the next year production is positive and is estimated in the range of O +lO /O. pag50f~0 Capital investments One further important aspect to consider is the amount of capital investments which the tile industry is typically affording. The kalian tile industry has invested 400 million ECU’S in 1995, with an increase of 44% over 1994. These investments are mainly in the area of technological innovation and improvement of products and services. For 1996 the amount of capital investments is expected to be significantly lower, due to less favorable fiscal condition. I Prod uct target cost Tile production costs have again analyzed in order to understand the value of an ASSIST system, i.e. the price that the industry may be willing to pay. Tile production cost per unit (square meter) may be considered equally divided among three areas: raw materials, energy and man power. ASSIST addresses the last element which results to be the one were significant changes and improvements are still possible. A production line may be considered as build around a kiln, which is the unit defining the production speed. H has been calculated that for a typical modern production line the saving, due to automation of the selection process, will be in the range of 120,000 ECU per year. Assuming an amortizing period of two years it seems reasonable to consider that an investment of 240,000 ECU may be industrially acceptable to fully automate one production line (1 kiln). Beside the absolute vaiue, it has been also calculated that ASSIST, with full functionality, will reduce production costs of apx. 4-5Y0. AS a resuit of all previous consideration we may assume a target ASSIST price of 120,000 ECU per system. H has-to be noted that this price cc)nsideration does not take into account any other beneficial effect that ASSIST may provide like: better quality, objective control, continuous production process, equipment efficiency monitoring. These will be considered as plusses that the system will provide beside the direct economic benefit. Potential market Concerning the potential market, we have estimated that there are 1500 selection lines currently operating in Europe and America which may be interested in. a system like this. For a rough business evaluation we may assume that one third of them may integrate ASSIST in 5 years since its introduction and that they are going to be mainly concentrated in Italy where high quality products are produced. This results in 100 systems per year, with possible revenues of 12 million ECU’s per year. This analysis brings to the conclusion that: the direct business associated with ASSIST sales is interesting and that the economic effort required to the tile industry is absolutely in line with the typical trend of investments. As far as ASSIST industrialization, production, distribution and maintenance channels are concerned it results more convenient to develop an agreement with one of the various automation companies already specialized in the tile industry. Contacts with some of them have already been established. ** production data in this section are from 16th Halian statistical analysis - Assopiastrelle pag60f 10 4. industrial and intellectual property riahts Up to date it was decided not to apply for patents or other protections of inteiiectuai property rights, even though some of the ideas on which software and hardware moduies are based may be suitabie for protection. The reason for this choice is that algorithms and other software soiutions, for the time being, are well protected due to the hardware piatform on which they run. The issue of protection wiii be reconsidered during the exploitation and industria[izat,ion phase, and at that moment, before deiivery of first product to the market, proper protections will be implemented. 5. Exploitation and marketing p [an The exploitation pian of ASSiST is based on three main assumptions: ● ● ● Some of the obtained test resuits are positive, even though more tests and further development in some areas is thought necessary. We have build up an important know how on tiie inspection. An industrial partner, from the tiie automation industry, must be invoived in the exploitation phase, for many reasons iike: - know how on automation aspects (product industriaiization) - deep knowiedge of market - customer base - product lunch strategy MODULES SUPPLJERS V.1. m’s NTEGRATOR AUTOMAllON suPPilERs SPEC(FIC ViSION TECHNOLOGY EXPLOITATION CHANNELS pag70f10 Based on above assumptions, next steps of the exploitation process are: 1. presentation of current results to “potential partners” from fl 2-96 to 3-97 2. demo of ASSIST to selected interested partners 3. definition of preliminary agreement, MOU or statement of interest 4. joint Product Functional Specification definition 5. verify at which extent current results are satisfactory for a first level product 5.1. or identify R&D effort still required to obtain a satisfactory product 5.2. or identify classes of products that can be addressed by actual system 6. definition of a joint development plan with the objective of bringing a first product to the market, for a beta testing phase 6.1. Expected delivery of a beta level system: end of 1997 6.2. Expected delivery of product: mid 1998 Step 5.2 is extremely important because we think that it can be possible to start delivery of a product addressing a relatively simpler class of products, while at the same time further research might take place. e.g. it may be noticed that in 1995, 17 new kilns were installed just for GRES products, which present a relatively simpler texture. And this kind of product has been constantly growing in the past 10 years, 1997 G F M A M J J A S O N D G F M A M J J A S O PRESENTATION TO SELECTED PARTNERS DEFINITION OF PRELIMINARY AGREEMENT - MOU PFS ANALYSIS MARKET ANALYSIS - IDENTIFY FIRST TRAGST SEGMENT m DEVELOPMENT OF BETA PRODUCT BETA TEST PHASE PRODUCTION AND DISTRIBUTION ADDITIONAL RESEARCH EXPLOITATION IN OTHER TARGET SEGMENTS (TILES) Exploitation time-table pag80f10 Objective and interests of each partner in the direct exploitation program are: For Bull Italia To supply the vision engine, constituted by processing system hardware and software, to the company responsible for the delivery of a complete product to the end user. For Thomson To suppiy cameras thru T(X, or Jicence to build cameras, to be integrated in the image acquisition system. See annex 2 for a better understanding of Thomson position. For IT One Tile, and the Marazzi group to which it beiongs, to be among the first users of the system with all the possible returns that this may imply (economic and commercial advantages). i i For DEEE and CNBE Further involvement in research activities necessary for completion and/or evolution of the product. See annex 1 concerning CIEEE exploitation strategy Besides these objectives strictly related with the direct exploitation of ASSIST, each partner will exploit in other areas the know how and the building blocks made available during this program. 6. w Communication strategy As stated in previous sections, we think that the best strategy for delivery of vision solutions to the various markets is through cooperation with automation companies well known in each market segment. This strategy directly aff~cts the communication channels used to access, not the end users, but the companies who can successfully bring a fully integrated product to their customer base. We have already setup direct contacts with some of the major companies operating in the tile automation industry. Furthermore Bull has identified companies specialized in each industrial sector that it intends to target and has aiready established good bases for the definition of exploitation agreements. After the definition of agreement with these companies, most likely it will happen that the developed products will be presented in cooperation with them at specialized fairs of each sectors. e.g. Bull is analyzing the possibility of presenting a vision solution for plastic molded objects at the PLAST fair next may in Italy, in cooperation with a medium company of this sector. pag90f 10 7. Annexes 1- Exploitation strategy from University of Surrey 2- Exploitation plan from Thomson ‘B 2- Abstract from 16th Italian statistical analysis - Assopiastrelle I I e 9 pag 10of10 ANNEX 1 Exploitation: UNIVERSITY OF SURREY Exploitation of’ Results The algorithms for colour shade grading of plain, two colou.r and textured tiles and the crack detection and chrornatos~cturzd inspection algorithms developed as part of the project are applicable tocolour texture surface inspection and grading in other sectors. These algorithms advance the state of the art in stiace inspection and grading which, up to now, was largely limited to plain surfaces. The possibility of exploiting these techniques for visual inspection in textile industries and in particular in porcelain and earthenware industries+vill be explored. This currently involves an evaluation of the techniques on inspection problems in these areas in conjunction with existing research projects. These applications have considerable economic potential in terms of automating the inspection process and therefore aiding in the enhancement of the product quality. In addition, the inspection process can be used to control the parameters of the production process itself to reduce waste. Exploitation Plan im The University of Surrey will expIoit the results of research and development in a number of different ways. 1. The main role of the university is to provide education and to offer the best quality professional and vocational training for its students. The research results (inspection methodology) obtained as a result of project ASSIST will be incorporated into the syllabus of its courses, in particular at the MSC level, which wiH aid the European industry by providing graduate engineers with up to date knowledge in surface inspection and related areas. This wilI contribute to the competitiveness of the European industry and its ability to continue to innovate industrial technology and bring new products involving machine vision on the market, L . A second, and equally important role of the University is to carry out research and advance the frontier of science and engineering. The research results will be exploited in the mu-rent and i%ture research projects and through consuha.ncy. A particular effort will be made to attract projects horn industry. With this objective the University has appointed in the Centre for Vision, Speech and Signal Processing, a Research Support Off~cer with the main responsibility for technology transfer. This appointment represents a considerable investment to facilitate research result exploitation. 3. The University assumes that the main drive for industrial exploitation of the results of the project will be applied through the two industrial partners of the project. In this context the University will provide all the necessary help within its means to bring an inspection and colour texture grading system based on the research accomplished under the ASSIST project on to the market. One specific aspect of the exploitation plan which WH be addressed by the Universi& is a further evaluation of the inspection algorithms developed. This will be camied out as part of a natiomdl y fimded project entitled “Visual Inspection of Colour Texture Surfaces: Operational Integration Perfomumce Evaluation”. As part of the project a large number of tiles will be inspected to establish statistical] y significant petiormance characteristic of the machine inspection algorithms. 4. The University will also seek to exploit the research results directly in other industrial sectors via other industrial companies subject to any cosntraints imposed by the existing collaboration Agreement, ‘c 9 . t ● . e . , ~ O N lNlRi3DLlCTK3N Since the beginning of the ASSIST project in 1993, many changes have uccured in the organisation and goak$ uf the THOMSON Research Gentre located in Rennes (France) and this is the main justification for certain variations in the intentions of THOMSON with regard tu the industrial exploitation of ASSIST results. 1 ~ 1 v’ It is not worth coming back again in details here on ali these changes, but it should on[y be pohted out that, by now, the ASSIST camera arid the associated team are part of THOMSON rrwitimedia Corporate inrmvdkm and Research (W@, whose main goai is to make advanced oorpfxate studie$ and development for paving the way tithe future consumer digital video products af THOMSON. This mainly means that: a as ~l&R has n~ Me objective, nor the organisation to seil ~roducts by itself (as it couId have been the case with the ‘LB% structure’ at the beginning of the project), no ‘direct’ industrial ‘external’ exploitation of A&31ST resu!ts can now be planned; E ~in TJ-K)MSON, but ~ufside THOMSON multimedia Cl&R, exploitation Of the ASSiST oamera could be ernkaged, especially in T(X (THOMSON COMPOSANT$ SPECIFIQUES) which was our Associated Partner in the framework of the ASSET project. JNTEM3ED EXPLCW$TIC2N OF THE ASSIST CAMERA WITHi~ THOMSOt$ 1. Contaots have been iritiaiised with TCS tbr trarwfening a limnce of the camera to them, but, up to now, discus~on$ have not reached a final end, mainly because a patent problem has not been ck?ar’ed out. 2. Anyway, C}&R is open to discussions to seiling nun-exclusive Iicences of the ASSET oamera to any third-parties which could be interested in this equipment, provided that this would not create any conflict with basic industrial interests of the whole THOMSON group. ANNEx 3 9 ‘8 ,D part of 16° INDAGINE STATISTICA NAZIONALE I I s $ . .*/ ‘$ ASSOPIASTRELLE — J3ndaginc Statistic Nticwde & realizzata attraverw il censirnento di tutm k ;C m“ende itakuw produtici di piastrelle di I eeramica e eostittdsce cmmai imo strumfxto conuscitiw cmsolidato per 1a Winizicme dei l – dati strutturaIi de~ settcre. La 16aedizione ddl’hkg{ne si & svolta nei , 8 prfrni mesi de! 1996 con riferimento N’intem arm 1995. Intewistatwi oppcxtunarnente istruiti hanno visitato tutte k swiet& operami I S in ItaIia, ottenendo la compilazitme di w queshwio appositamepte predisposto. ‘ Eorgantiatione gtheral~, Wabcmzicme I < dei dati e la stesura del eommento smm stati curati direttamente dal Cen&u Stud AssopiaE .1 strek < . n :SQ - NOTE METODOLO GKHE ~ 11 @ Per una eorretta analisi e iriterpretazkme dei risultati presentati DWQ tabelk alIegatc, & utik Elustrare il metodo di rlhwazicme dei dati c ia twmincdogia adottata. 1) p e r k pmtkibne: a) quelIa diretta dde aziendc, retdizzata con proprio rnarchio; b) queNa per eonto terz~ vale a dire per canto di altra m“enda produttrke @o societi -comma%de, con marchio dell’uienda ~m. mittente. La smmw (u + b) fomkce fl qwmtifmivo male di>rodttrwne del setture. 21 per i% wmiiiie: a) que!le di attiviti?i diretta (Italia ed export), relative ckk alla produzione effettuaia con marchio proprio; 13) quek relativeak qrantit~prodotte per conta Wzi (Italia ed export), cio$ alla produzicme commissionata (eon proprkj marchio) da akra aziendaproduti~ e/o datma societ3 commercial; c) la cmnmercial-one deIie azitmde produttrici (Ralia ed export), ciok Ie vendite di un’azienda,di materiali fatti produrre (eon il pmprio marehio) da un’altra azierida produitric-e. pruduzhm delk aziende censite &stata ‘a *Lasuddivisa in aJ&itA diretta e attkitiii canto 1s r &rzJ. Per quanta riguarda lewndite, tstata evidcn- ziata la “Cf3mrnercia ~zione” ‘ di materiali prodotti da teti Quests eonfigmxzkme ha permewm di icientificarix 9 1 It 1 - , . .. . 9 B . U LEAZIJHVD E E I DIPEN DE NTI o AI1a data disvolgimento dell’indriginesono state registrate in ltalia 340 a.ziende produttrici di piastrelle di ceramiw. RisgMtto aikmna precedence, smm risukate 5 um”ti in meno, come saldo ti chiusure di aziend~ fusioni e/o incorpm-aziuni e ingresw nel settore di pkcde unitk ““~ micq costituziorie o ,.. non ctmsite in precedem, produdici principrdmente di pezzi spec@i e caedi. I dipendenti, al 31/3241995, sono ns~ki 32386 (+5,22%) conferm~do Mnversionedi * tendenza inizka riel 1994, dopo tie anni di >. ~ntimm cdo degli o&upati, 1 D s — @ 1= . . . :. in iinea con l’andarnenta delki produzione di tali comparti, came piti avanti descritto. Le variazioni def fomi relative al)e akre tipdogie productive trovano ~&’meri~ntrO negli andamenti delle rispettive pruduzion.i. 12utiltio dells capaciti produttiva k stata delJ’85,46% contro 1785,X% deI 1994. Da se.gnalare che [’ineremento del numem dci ford k je”riibtie teenck@c hanno mmpo~m 1a creaiiofietii tma rrmggiore capacit& produttiva the, kfatti, nel 1995 & aumentata de~ 10% circa rispett~ al 1994. M LA Pltmuzmm 12 1 FORM E L’UTILIZ ZO DELLA 9 C!APACIT~ PRC)DUTTNA ??el 1995 la piaduzicme comple~iva ha confermato ii &end di crescita che aveva earatterbto in”partiedwe Mh-ncI tritmnio; arrivando ad un nuovo record determinate soprattutto dal positivo andamento deU’export. WI 1995 sono stati pmdotti 562,2 mdioni di metri quadrati, pari ad un inereniento del 10,19% rkpetto ali’armo precedence. Al kmnine del k$)$ il mmxxo dei forni .. kstalkti & risultato di 396 uni~a .,, Rispetto alla precedence rilevazione si 5 registrato m awnentu ,del nurnero dei forni pad a 47 unit~ dovutci principaIm=te aWineremento degIi impianti per gres porcell%ato (+13, per mwmeottti chkra (+16) e ‘per corredi (battiscopa e altri pwzi speciaii + 14) i t 8 a @ sow SO.(M k \ \ , 8 1 8 2 \ \ \ 1 , U 8 4 W M m 8 E 8 9 . . ,. ---- q m u b u i * S U 91 1 19 2 i W W 9 15 =k?ulC “ .. 4ooq. ” 1 d 350 300 - “ 1 - 20(1 / h 50- -/ - / \ ’ “ — * . * — - - - m ~ d - = = * a m,a -*----.* . 1 , . w e 9 a 0 * . - . gres pmdlanEm s , * x & — - alcri prodoti ●✍ - . - & * ” * I i 80 al 82 m: M 8586.” 8?:”88 89 90-91 92 93 94 95 La ripartiz!onedellaprodtioneper tipo di prodotto non ha wbito sostanziali rnorii&@: stabile la morwottw cm il 62,60YQ” dell~ produtirme eompkssiva (era il 62$1% nel 1994’); la bicottura ha visto cake la propria quora std totqle daI 18,10% del, 1994. d 16,68% registiato nel 1995. 11 gres porcekrtato ha cmtinuato rid inerernentare i suoi wkmi &to a rappresefttare d 14,71% (era i] 10,81% nel 1992, ii X?Jl% nei 1993 e il 12,62% nel W34). 11 cotto rustieo, do@ d cah eorisisteote del 1993, ha ripreso,a erescere ma ha. p~po quota M totale il ~159Z eontto ii 2J9Ye dell’ann~ precedence, .b produtione dei””corredi (battkmpa cd akri pezzi Speciali) 6 stata di 12~ milioni di metri quadrati, 1 mikme in piii rispetta al 1994, pari ad w incrernento del 9,12%; rappresentarm, come neI 1994, i1~20% della produzicme cornplessiva. Tale andamerlto conferrna la crescente teridenza delle aziende a eompletare la prapria gamma di pradotti per offrim setvizi sempre piii rispondenti alle richieste di una clientela sdisticata ed esigente, 1. aflivif.d direfta e altivitd conto teni Nel 199S le aziende che harm prodotto per conto terzi, vale a dire con rnarchio di un’aitra azknda pmduttrice e[a socktk com rnerciale, hanno realizzato tm volume di pr dwkme pari a 90,8 milioni di metri quach (il 16,15% dells produzkne eomplessiva d 1995). Urmremento suli’armo precechte, dete rninato da 16,1 rnilioni rli eonto terzi Prorh in piii riqmtto al 1994,5 stato de] 21~57& ❑ w m-ckzi a!tdfrcmi Questo conferma, da Un latr+ Ja crwxnt i.ntegrazione tra aziende di produzione dall$altro il rualo sernpre piii impormte su mercato di soeit%i cornmerciali can tnarchi proprio, laeui attivk~ k ingrado di influenzu e rendere axmora piti eompkssa la struttuq dells distrjbitzjone sia in Italia &e Wester come. wdrema nci eapitoli sueeessivi. c) Ilnpa.$!!o per terzi 2 zone ifi kwdkmzwneprndut tiva Nel 1995 Ie provincie di Modena e Reggio Exr@a harmu pmdotto i~ 78,70% del totaie nazkmale (in particdare harmo pmdotto 1’80,73% & attivith:~etta ed,iI 68,14% di qttiVit& c/terzi). II tikrzi rappresenta in quests zona il 14% circa “deka produzjcme cornpksSim II resto Emilia Romagna detiene 1’8,57% del totale nazknale (la stessa pereentttale riguarda sia ajl’attivitil diretta che iI clterzi), Nel rests dWIta.Iia ? stata pradatto il residuo 1~74%: piii in specifico ii 10,69% di attiyit~ diretta e il X$34% di attVlt& e&xk Da scdtolineare la partiwlarith che nel resto Itaiia ia produzione c/terzi ,,mppmsenta quasi il.30% ‘ deIla produzione @.rqplessiva deIla zgrqi eun‘ tro iI 14!% CM Cornkensario, mme wpm riportato. Questo giustifiea d ado in terrnini I t di rappreseritativiti pwxentuak ck ii ComI prensoria ha registrato rispctto allo scorsc) a.mml passmdo da~ 79,21% al ?8,70%. 10 - ‘9* 1“ r f .- La produzione di impasto per terzi, pari a 1,6 milioni di trmnel]ate, si ~ portata su valori piil equiIibratl rkpetto agli armi pissati, Le vendite pari a 1,4 milioni di tonnellate sorm aumentatc deI 27,1296 (8&17~ deila produzione). In particcdare sono wmentate deI ~ 34,66% k vendite di ~ irnpasto atomizzato chiaro, in iinea con l’incrmrmto delhi mono--eottura chiara. 1.4 LEVENIMTE AIwhe ,Ie vendite totali hanno fatto registrar krcmenti positivi nel eorso de! 1995, d~ti escIusivamente Wandmnento delI’expart. Complcssivarnente sono stati wmduti ~ 5424 xnilioni di metri quadrati, pari ad un aumento del 7,58% rispetto al 1994, L atfiviiti direffi.z f? cfxrmzercia&z.azbe Le vmdite cmispondenti all’attivith S37% 12.74% E l ❑ rcsto Entilia R ❑ rcstolka[ia I I Ii) Suppati per ti~i Nel 1995 la produzione di supporto per terzi ha manterwto i Iivelli CM 1994 (?1 miiioni di metri quadrati). Le vendite di supportcI in ftrdi~ pari a 60,2 milkmi di metri quadmri, cmrispondono sostanzialmente, eselusi @i searti di kworaziorle, ai volumi di produzione delk smaltcrk (59,4 milioni di rnetri’quadrati). direttri pari a 452# milkmi di metri quadrati (+5,41% wl 1994), eostitukemm H13~O% ckk verkdik corrqdessive (erano 1’85,22% he! 1994). Le veridite di nmteriale commerciaIizzata, cioi+ Ie quantitii ck Ie stesse aziende hanrm fatto produrre con marchio proprio ad ahre aziende, sono state 42 rniiioni di rnetri quadrati, pari al 7,75% deHe vendite tutali. Mine, circa 47$ rrdioni di rnetri quadrati so no stati destinati a societ?s commerciaIi sia italiane che estere che operano SUI rmxcata con marchio proprio. Qu=t~ Ultimo fmmxmo ha registmto negli ultimi anni forti ritmi di creseita: dai 13 mfiioni del 1992 si A passati agli altre 4’7,5 attuali (+264%). 2. njxx!ogii dipmdotto . Dal punto di vista delle tipcdogie di prodotto, la monocottura rappresenta iI 62,72% delle vendite tota~i, q 340,2 tioni di metri quadrati venduti n&l*95 [+7,62%). Labi@tu~ con 90,1 milioni di metri quadrati, k passata dal Y&44%4” 16,62%,&Ile ,.. vendite (- 3,03%). ,: ‘.”+ ~~. ~n bucm hxrerncnto di ve~diti e,~ peso percentuaie SUI to@gdelJe &sse & stita registrato dal gresporcellanato, che passa dai 62,3 rnilioni dirnetriquadrati del 1994 ai 78 del ’95 (+25,17%). Ora iI gres porcelhmato rappresenta D 14,3$92 delle vendite cornpkssive deI settcw, ccmfermando una eostante ascesainiziata eon 1a second;~eth deg{~ &mi ’80 (e& 3 Z?% neI 1984, quasiiI.6%nei 1981i,I’8J?% qql 1990, 0 10,8% d 1992 e il Q4 nel 1994]. Al gres porcdanato seguono, in ordine di iiripartanza per vohirni vendu& i cotti rustici che eon 11,6 milio@ di mctri qua&i h,tinb mgistrato un inerenk?nto del 6,93%. In calo il kk.ker cm@3 mil.k~ dimetri quadmti (-&18) e iI gr~ rosso e COkatO Cbe, eon qwui 4,3 milioni di tietri quadrati (-3,S% sul 1994), rappresenta ma lotl,’79% del totak vendutg..~ ~::, ,. : Intmssante l’incre~e.nto iieIIe vt@ite di “akri prodotti” da 11,9 a 12# mi.bni di rnetri q u a d M i ( + 8 $ S % Sld 1 9 9 4 ) . Quests mtegmia include quasi mciusiwrmmte tipcdogie di eorredo (list~w grad&i~ bat?kqwq corrirnaqp: e=) non xientranti nelle altre ckssificazioti. Hncrernento di vendite di questi prodotti & cormesso allo” sforzo della quasi totalit~ dek aziende di cornpletare la propria gamma con tm vasto corredo di prudotti estetieamnte e/G fu.nzionahnente compkmentari alle piastreile di fondct. 3. deminazbn.e ven.dte Italti e Yendlk kapoti b a.ziende produttrici harum indicato di aver venduto 2123 mikmi di metri quadrati in, Italiq (-(),1270 rispetto al ’94) cd cOmpoi St : a) .@L66/$37Q ~q, vendite di attkiti diretta b) 17.7?5.676 ‘req. commereializzxdmm c) “11.863.206 mq. vendite di soueti commer– Ciay. ~n Inarchio proprio (k&fmma dib+c cmiqfmdedle wnfiifediaffii.fti Clrt?my M’esters ~cmo stati destinati, sempre semndo le indicazbni deJle aziende di produziwm,. 330,1 milioni di rnetri quadrati {+13,20%) eosl eompcxti a) 27&264.lfj8 rnq. vemiite di attivith direb b) 24.2+2.733 ~. cormnereiahzzazione c) 35589.6.15 mq; vendkc &c societi comrnmeirdi con marchio propriQ (lascmma dib+c umis-omfe&vena&dia uivitii dw!tzi) - Isil IB! o M & k @ w ~ N a z i m d e `.""" ="'"' ``.. ".". ".`..'" Vcndkc kdia Ventm upon V&ndk t4t&li ``. " . . . . . . . . . . . . . . . . . . . . . . ..i . . . . . . . . . . . . . . . . . . . . . . . . . . . . . .pagi? s $“’,i W 199S I’expoti” complessivb dell’iffriustria ceramiea -.Iie”gistrato ‘” @a:una specifim indagine condott~ &@wopiastrel]e - i ~tato n di 361,4 milicmi di r &.ri qu&~ti (~11,24% . rispetto al 1994). Quest: ciati meritano akune cansiderazioni. Ckxxxre rk.ordare cheia prexmte Indagine fornisee dati secondo le indicazioni dck irnprese produ~ici, La fonte aziendale rileva pertarm come “’vendite Ealia” anck quek i-iferite’ a- piastrelk cedute a suci@&comm”eddi o ad a.ltri cIieriti chc a loro vojt’d” effettiianb” operazid di espm-tazlone. “ ~ ‘ CZuesto confernm, seconcio quanto dieevamo primi+ iI ruolo’~empre rnaggiom ch+ k societ~ commerchdi “svcdgoxm rispetto aIla rete distributive. Tali’ societh infatti oltm a distribuire i quantitative faxi produrre cm rnarchio proprio, krnettono sxd mercato italiano ed estero anche prodotti delk &iende di produzione: : - lI”data expo~ fiiiki ‘kl 1992 di fonte Istat e successivarnente di ‘“kotite diretta Assc@astrelk, rileva invece tutte b quantitii che passano attrawxscs Je fiwntiere e quindi anche le quantiti vendute all’&&o dagli intermediary commerciaii italiani.” ~ Pertanto, i dati ripotiati nella prestmte indagine sono vaKdi dal puntm di vista delle aziende di produzion~ da] r?twto di Vi?S#a “clef BUUOIQ gli sfessi *gm essere r-icomposti can k rikvazimd ‘iIi.,.fonte dketta Assopiastrelk. E dwtque$ terkndo conto ~he k vendite cmnplessive sono pari a 542,4 miiioni cii metri quadrati, e IG ,vendite export 361,4 itki[ioni dL metri quadrati (oltre i! 6670 delk vendite Maii), il niereato ink-no ha assorbito, p e r d i f f e r e n m , ~ d i m e t r i ~ dH&(~Qn w parziale recupero di circa m punto percfmtuale sul 94). .... .;..”, : - ~.:..~ J&’ W@l &e&&fi JLJ &@ ’ ,.. . .... $ostanzialmente stabile 6 rirnasta Ia ripartizicme ddk vendite MIia per aree geografiche convenziwxd.k il Sud e klsOIerappresentarm iI 32@% delle vtmdite, seguiti dal Nerd-Est eon H 25,23%, dal Ncxd-Ovest eon il 22,33%, cd Mine ckd Centru con 11 20,06%. Le ;endite “Alla produzimw wmto terz.i si comentrarm, nOII a cam, I@ Nwd-Est (eke il 95% dkl tbtaie); paich~ proprici in quests area M sede la maggior park dek imprese di produzkme e di ccmmrerciakzazkme dek piastrelle di cerarnka. L5 IX GLM2ENZE DI MAG4ZZlN0 Le gkxnze di magazzhm allafine de] 1995 ““B.w$&&T****:l Www%w%-yiy$ V,q +4+w+i%. ~ . L k W. . . ..+w.xcw. . . . . . ,.. .J Soo ii , v “ %’ “g , 50 z m 90 ,.19a4a6 , Smm rimdtate pari a quad 1~1 mtiioni di metri quadrati, 20 miiiani in piii rispetta aI1’anno precedence, e rappresentano B 26,86% dells produziane dell’anno 1995 contro i! 25,69% deil’armo precedence. L6 IL FATTURATO ?1 fatturato relativo al prodotto finito & amrncmtato nel 1995 a 8.471 rniliardi di !ke ,’,.$ (con UKI increinento del 16,16% WI 1994) cd ripartiti 2334 rniIiardi di Ike iI fattumto ItaIia (+ 11,48% d 1994) e S887 rnilkrdi di Iire d fatturato export (+18S4% SUI19M). La cifra d%ffari raggiunta eorriplessimente d.ti.settore, quasi 8.500 mhrdi, rapp~esenta d nuovo piwo stori~, come giii rilevato per ia p~dticm e per le wmdite. - Annows ‘ p - ~ ~ i ~ Lh?hrJq. { Mer=to nazionde Export 7Male ., ‘..: ,. ! viz % 9S/94 14.276 “ +m,43 16.2/37 . + 6 , 3 8 15.616 +7,97 p==i speciali (un miiione in piil nel 199.5 rispettci al 1994), a riportare i prezzi me.diri valori piti ccmgrui e piti equilibrate rispetto alh.ndamento cmscente dei costi produttiti. Uncreme~to dei prezzi medi registrato sui mereati esteri, tenendo eonto deIIa debok~ dells Era nei confronti delk principdi valute straniere, non 8 invece da ritenersi deI tuttb .mddisfacente: a tionte infatti di maggiori vohmli esportati non si scum eonsegtiti corn: spondenti e cangrui increment di prezzo, ..- H fatkato compiessivo del supporto per terzi & ammontato nel 1995 a circa 198 rniikrdi dl I~e e queUQ deI1’impasto pez terzi a 154,7 rnik~di lire: da@ qwst~ da smmrmre d fatt~rato relativo al prodotto ilnito gw tma piil puntuale .vahizione del giro di affari .. aggregato deI setiore. 1.7 GLI INVESTTMEN’’I’I I 4 fronte di un increment del 758% .ciei volwni verlduti, i prbi medi smm saliti iel 1995 del 7,w?% rispetto dhrmo precedence. In particolare i preizi scm aumentati de] 10,43% in Italia E del 638% rdl’estero. h Italia le azkmde somy riuscite, grazie all’hmakamento del mix venduto dovuto’ prkcipahnente alI’awnentato vohune . dei A 8 R 9 i * WI 1995 I’industria eeramica ha investita okre 814 miiiardi di he in beni capitali (-+44,!3% rispetto d’annoprccedente) c482 m~i~di pr,evede hi irwestirne nel 1996. ~~~z@~~ivalofi degji ~~~mentidel 95 Sorlo .*Q .destinat prineipalmente all’incrementa del nurnero degii impkm~ al rinnovo tbologico, al rnigIiorarnento deI1’a.mbiente i .-$! Izl IllB El Skui f& Emu.@Q R&l famamto CzpQa famfrato ttmdc ed estemo, al rnigliormento del.l’effl=nza COmpkssiva attraverso I’uiterkn-e qua~cazione dei pr[dc$@ c d~i servi.zi. offe~i. Contemporae&tite sono .eontinyati j pro,’.: J. ; @ ~ irkte~on~ aziendak (a~uisizion.i dipaecheti azionar~ fusioni, immrporazioni). In particuke & possibile oggi affermare che grazie a questi i.nvestimenti I’apparato teamIogico e Ia qualit~ delle aziende nel ko complesso ha.tma raggiunto liveili & asxduta eCcellenza. b prevision di investfiento relative al 96 Cbrotano un atteggiamento piii cohtenuto e prudenziale che risp.ecchia sia Ie mutate mndizicini di rnercatmehe ii venir meno.di agevokzkmi fkcali Iegate”’aI reinvestimento degli inttxrm ?l~ in azienda. ‘t- 1.8 LE PREvIsION PERIL 1996 . . I-e previ.sioni formulate dd.le azknde hanno indicato un incrernento del 10,21% dells pmduzione COtnplessiv% come evidtmte prdezione dei~andamento del 1995. T;li indkaziuti non harm perh tenure conto dell’invenione di tendem deIIa doma@~ ski in ItaIia che all’estero, che si e’ man.ifestata con evidenza nei primi mesi de[ 1996. Tidi mutate condizionl richiedono infatti un :kihnerki&bento dei programti, giil visibile M cxmtrcdlo operato sullaprodwi~ne at~avexxo i] ricorso aUa CaS.sa Integrazione e aIle fene anticipate da parte di un certo mimero di aziende per contenere i liveI!i dei magazzini. Infatti & pib che mai neeessatio che il liwdlo cWa capacit& produttin e la pmduzione siano oggetto di approfondite rifiessioni da parte ddle aziendeche dwranno esercitare k massima attenzione aH’equiIlktia del rapporto ha doma.@a e offerta del mercato. . .. . ,