Title of Invention

A METHOD AND A DEVICE FOR DETECTING DEFECTS IN TEXTILE WEBS

Abstract The invention relates to a method and a device for detecting defects in textile webs. In order to rapidly adapt devices of this type to widely varying textile webs and to be able to operate such devices simply, brightness values are determined from the web and are supplied directly to a filter constructed as a neural network. Figure 1
Full Text

Method and device for detecting defects in textile webs.
The invention relates to a method and a device for detecting defects in textile webs.
Proposals are known from the Textile Research Journal 63(4), pages 244-246 (1993) and 66(7), pages 474-482 (1996) under the title: "Assessment of Set Marks by Means of Neural Nets" and "Automatic Inspection of Fabric Defects Using an Artificial Neural Network Technique", according to which neural networks can be used for the detection of defects in textiles. In this respect, the method is such that particular input values are firstly determined for the network. Such input values are, for example, the distance between threads in the fabric at a given site or the mean value of this distance over the entire fabric, the standard deviation from values for the said distance, the yarn mass or intensity values, which have been derived from a fabric image subjected to Fourier transformation. These are all measurement values which must firstly be obtained from values from the fabric by way of more or less extensive calculations. '
A disadvantage of methods of this type consists in that they are not very flexible, so that the detection of defects in different fabrics requires calculations which need to be carried out in advance. Thus, it is not possible to derive or deduce input values from the web for a defect detection system which are adequate for all possible types of web texture. If an approximation of this method is nevertheless to be achieved, then a very large number of different measurement values must be determined, resulting in a correspondingly high calculation outlay. High speed and high cost computers are required to this end.
The invention, as characterised in the claims, therefore attains the object of providing a method and a device which

can be rapidly adapted to widely varying textile webs and is simple to operate.
This is attained by way of the skilful use of modern, cost-effective computers operating in parallel. The web is scanned in known manner, for example line-by-line, by a camera which supplies a store. Values for the brightness or intensity of scanning points or partial areas of a web are Stored in the store. In this manner, the store eventually contains an image of a section of the web. Values from connected are&S are then removed in parallel from the store and supplied in parallel to a neural network, which was previously trained to recognise defects. The neural network indicates in the form of a result whether there is a defect in the examined area. This result is read into a further store, which stores this result taking into account the position of the area on the web. As the examined areas gradually cover the entire width of the web and therefore also cover the web over a section of its longitudinal direction, conclusive data regarding defects in the examined section is eventually available. In accordance with the invention, a neural classifier known per se is used as a non-linear filter and operates, not with additional measurement values, but directly with brightness values from a relatively large environment (e.g. 10 x 100 pixel) as input values for the neural network. The environment is displaced pixel-by-pixel over the surface of the web, so that a filtering is carried out. At the output of the classifier a filtered binary image is produced, in which defects in the web are clearly shown. By means of a learning process, both the filter structure and the filter parameters are automatically determined and in this manner adapted to any type of textured and small-patterned surfaces. The learning process can be effected by the presentation of approx. 20 to 100 image patterns, which contain defects and the same number of image patterns containing no defects. By dividing the filter into two

neural networks for input environments which are oriented in the warp or weft direction in the case of wovens, the distinction between warp and weft defects can be further supported.
The advantages attained by way of the invention are to be seen in particular in that a device of this type can be constructed from cost-effective, simple computers which operate in parallel and are optimised for neural networks. As a result of the parallel processing of all input values, very high computer capacities (e.g. several Giga MAC (multiply accumulate)) are attained, so that the result of the examination can also be continuously determined even at high product web velocities. Computers of this type can be extensively integrated in a single silicon chip and used in the form of add-on boards in personal computers. Examples of circuit boards of this type would be the PALM PC board made by the company Neuroptic Technologies, Inc and the CNAPS PC board made by the company Adaptive Solutions. In this manner, high inspection speeds of, for example, 120 m/min are possible.
The learning process can be effected very simply with the aid of a section recognised as defect-free by the eye and defective sections of the web. In addition, the sensitivity of the defect detection can be increased by the particular form and orientation of the areas from which input values are derived. Using the simple learning process, a high degree of adaptability to differently textured webs is possible. No specially trained personnel are required for the simple operating procedure. The invention can be used for textured and patterned surfaces.
The invention will be explained in further detail in the following by way of an example with reference to the attached drawings, in which:

Figure 1 shows part of a textile web on which different features are schematically indicated,
Figure 2 is a schematic illustration of a non-linear filter operation,
Figure 3 is an image of the web with defect markings,
Figure 4 is a schematic illustration of a device according to the invention, and
Figure 5 is a schematic illustration of part of the device.
Figure 1 shows part of a web 1, in this case a woven, for example, which is formed of warp threads 2 and weft threads 3, of which only a few are illustrated here. In addition, a plurality of lines 4 are shown, as can be covered for example by a line camera, which scans the web 1 in such a manner that the entire web is covered. Lines 4 of this type can also overlap so that no gaps are left between the lines. In addition, areas 5 and 6 can be seen, which are formed by 72 partial, areas 7 and 56 partial areas 8 respectively. Areas 5, 6 of this type are only defined for a given period of time and are therefore defined for other periods of time in the same form and size, but in a different position. 5a, 5b and 6a, 6b indicate further such areas in other positions, a plurality of areas 5, 5a, 5b and 6,- 6a, 6b defined for successive intervals overlapping. These areas preferably extend with time in the direction of an arrow 9 over the width of the web 1 in such a manner that successive areas 5, 5a, 5b and 6, 6a, 6b are offset relative to one another by one partial area 7, 8.
In a plane 13, Fig. 2 schematically shows a store content with input values 14a, 14b, 14c etc., which represent the brightness or a grey value of the web as detected by a sensor or a camera. In a plane 15, signals are illustrated

as output values or results, only one signal 16 being visible in this case, which can preferably indicate two possible conditions, namely: defect or no defect. Arranged between the planes 13 and 15 is a non-linear filter operation if this drawing is viewed in terms of function. However, the drawing can also be viewed as showing the structure of a device. In this case, 17 designates an intermediate computer and 16 an output computer. The input values 14 can also be seen as input neurones, the intermediate computers 17 as concealed neurones and the output computers 16 as output neurones of a neural network.
Fig. 3 is an enlarged view of an image 10 of a section of the web 1. Two regions 11 and 12 containing defects are marked on the image 10. These regions 11, 12 are composed of partial areas according to Fig. 1, so that, as shown in the drawing, a plurality of partial areas are occupied by a defect signal and together produce the regions 11 and 12.
Fig. 4 is a schematic illustration of the layout of a device according to the invention. The latter comprises a camera 21 arranged directly adjacent the web 20, e.g. a CCD camera or more generally a photoelectric converter, which is connected to a store 22. Signals from a plurality of adjacent lines 4 are stored in the store 22 for a given period of time. These signals and lines are processed in the store 22 according to the FIFO principle. The store 22 is connected to a non-linear filter 23, which can be constructed for example as a computer, in which a corresponding filter program is loaded. The filter program is designed according to the principles of a neural network, so that the filter 23 exercises the function of a classifier. The latter is connected to a store 24, in which defect signals (or just no-defect signals) are stored with their allocation to areas on the web. Also in this case, the defect signals remain stored in the store 24 for a given period of time and the defect signals are also processed

according to the FIFO principle. The store 24 is conneccea via a connection 25 to a distance recorder or length encoder 26, so that data relating to the instantaneous position of the camera 21 along the web 20 can be fed into the store 24. In order to display the results of the examinations of the textile web 20, a display unit 27 is connected to the store 24, which can be constructed for example as a printer or monitor. However, a processing unit, e.g. a computer, can also be provided in place of the display unit 27, which processing unit subjects the content of the store 24 to a further classification, namely so that defect regions such as the regions 11 and 12 from Fig. 3 can be compared with given criteria, so that they can be associated with different types of defect. For example, in the case of v/ovens, the defects can be classified into weft and! warp defects. The iregion 11 in Fig. 2 would therefore indicate * weft defect and the region 12 a weft defect.
Fig. 5 shows a section of a non-linear filter 23 (Fig. 4), the filter being constructed in this case as a neural . network. It comprises processors 30 arranged in a first layer and processors 35 arranged in a second layer. In contrast to Fig, 2, the processors 3 0 are to be regarded as exemplary embodiments for the intermediate computers 17 and the processors 3 5 for the output computers or output neurones 16. The processors 30 are constructed from a plurality of multipliers 31 with associated stores 32, which are all connected to an adder 33. This is in turn connected at its output to a processing stage 34, which has a nonlinear characteristic curve. The multipliers 31 are connected to the store 24 for receiving input values 14a, 14b, 14c etc. The processors 35 are constructed in like manner, although the processing stages 34 of the processors 30 are connected to the multipliers 31 of the processors 35. The latter comprise an output 16 for output values. The illustrated arrangement, in which the processors 3 0 of the first layer are acted upon by all input values of an area,

is realised in this case as a parallel computer, wnicn comprises solely like processors 30, 35.
The method of operation of the method and device according to the invention is as follows: In relation to the web 1, areas 5, 6 are firstly defined in the store 22 in that instructions are preset in the store or in the filter 23 connected thereto, which determine from which store locations in the store 22 values are taken and supplied as input values for the filter 23. On the one hand, such areas 5, 6 should have sides lying parallel to the lines 4 recorded by the camera 21 from the web 1. On the other hand, the areas should preferably also have a main direction which lies parallel to the texture features of the web 1. In this case, the area 5 lies with its main direction parallel to the weft threads 3 and the area 6 parallel to the warp threads 2.
A learning phase then follows in order to adjust the filter coefficients or filter parameters. In this phase, the camera 21 is aimed alternately at areas containing no defects and areas containing a defect. The result which should be displayed by the filter 23 is predetermined in each case. In this respect, the computer, which acts as that filter, is operated in a mode in which it does not transmit results but adapts its coefficients and parameters from the results and the input values. The coefficients and parameters are firstly predetermined as output values, for example as values in the stores 32 or as parameters of the non-linear characteristic curve of the processing stage 34, and are adapted by the learning process according to given rules, so that the filter receives a specific transmission function. This process is preferably not only carried out once but is repeated each time a new web 1, 20 is presented.
Once the learning phase is complete, the mode in the computer is changed and the detection of the defects can be

carried out on a web 1 which is moved in a direction perpendicular to the arrow 9. This means that the camera now passes over the web 1 in a manner known per se and therefore not illustrated in further detail in the direction of the arrow 9 and thereby optically scans lines 4. The recorded values for the brightness or colour intensity are supplied to the store 22, which also stores said values in lines, for example. The values for all partial areas 7, 8 from areas 5, 5a, 5b, 6, 6a, 6b etc. are supplied in parallel from the store 22 to the filter 23, which for each area 5, 5a, 5b, 6, 6a, 6b transmits art output value, result or signal 16 (Fig. 3). This signal, which is preferably of binary nature, is read into the store 24 together with data relating to the position of the area from which the signal is derived/ and is stored for a period of time required by the camera 21 in order to cover a plurality of lines 4. Thus, an occupation of storage location by signals is produced in the store 24, which can correspond to an image 10 as is known from Fig. 3. Within this image 10 signals 16 are recognisable, which, since they are usually not isolated but occur in groups, are combined to form regions 11, 12 ' , indicating a defect in the web 1. This image 10 can also be made'visible on a display unit 27.
If a processing unit is provided instead of the display unit 27, the processing unit is constructed as a computer which can carry out an image segmentation in order to combine individual pixel to form regions according to a suitable method, as described for example in "Rafael C. Gonzalez and Paul Wintz: Digital Image Processing, Addison-Wesley Publishing Company, Reading Massachusetts, 1987".
If the non-linear filter 23 has a construction according to Fig. 5, then input values 14a, 14b, 14c etc., selected according to the areas 5, 6 are all supplied to each of the processors 30 of the first layer. Each processor 30 therefore comprises the same number of multipliers as the

number of partial areas in the area. In the multipliers, the input values 14 are multiplied by factors which are stored in the stores 32 and then added in the adder, so that , a mixed value is produced, which is composed of all input values of an area. This mixed value is further changed by the non-linear characteristic curve of the processing stage 34. The adapted mixed values are in turn supplied to the processors 35 of the second layer, where they are processed in the same manner as in the processors 30. An output value for each area is produced at the output 16. These output values are supplied to the store 24 where they are distributed as illustrated in Fig. 3.
Although the invention is explained here in particular by way of example;of a woven, it is equally possible to use the invention in knitted or similarly textured webs. In that case, particular attention must be paid to ensure that the areas 5 and 6 are aligned so that their main axes lie parallel to prominent lines in the pattern or knitting. In this respect, it is also possible to arrange the main axes of the areas 5, 6 in any manner (not at right angles) and to select a direction for the progression or displacement of the areas other than that according to the arrow 9.












1. A method for detecting defects in textile webs (1), characterised in that, for a plurality of partial areas (7, 8) of an area (5, 6), values for the brightness in each partial area are determined, said values are supplied in parallel as input values (14) to a non-linear filter operation and a signal (16) is transmitted as the result of the filter operation, indicating whether there is a defect in the area.
2. A method according to claim 1, characterised in that all partial areas (7) of an area (5) form input values and further areas (5a, 5b) , which overlap with the previously-formed areas (5), are formed at intervals.
3. A method according to claim 1, characterised in that the filter operation is carried out in a neural network capable of learning.
4. A method according to claim 1, characterised in that the areas (5, 6) comprise a main direction lying parallel to texture features (2, 3) of the web.
5. A method according to claim 1, characterised in that input values are determined from a plurality of differently oriented areas (5, 6).
6„ A method according to claim 1, characterised in that filter coefficients for the filter operation are initially predetermined at random and are then modified in a learning process.
7. A device for carrying out the method according to claim 1, characterised by a series connection of an image recording device (21), a store (22) for values of a

plurality of partial areas, a non-linear filter (23) and a store (24) for values associated with defects.
8. A device according to claim 7, characterised in that a neural network which operates as a classifier is provided as a non-linear filter (23).
9. A device according to claim 8, characterised in that the neural network is constructed as a parallel computer.
10. A device according to claim 9, characterised in that
the parallel computer is formed by a plurality of like
processors (30, 35).


Documents:

1629-mas-1997- abstract.pdf

1629-mas-1997- assignment.pdf

1629-mas-1997- claims duplicate.pdf

1629-mas-1997- claims original.pdf

1629-mas-1997- correspondence others.pdf

1629-mas-1997- correspondence po.pdf

1629-mas-1997- description complete duplicate.pdf

1629-mas-1997- description complete original.pdf

1629-mas-1997- drawings.pdf

1629-mas-1997- form 1.pdf

1629-mas-1997- form 26.pdf

1629-mas-1997- form 3.pdf

1629-mas-1997- other documents.pdf

abs-1629-mas-1997.jpg


Patent Number 207675
Indian Patent Application Number 1629/MAS/1997
PG Journal Number 26/2007
Publication Date 29-Jun-2007
Grant Date 20-Jun-2007
Date of Filing 22-Jul-1997
Name of Patentee M/S. USTER TECHNOLOGIES AG
Applicant Address WILSTRASSE 11, CH-8610 USTER.
Inventors:
# Inventor's Name Inventor's Address
1 ROLF LEUENBERGER HERMATSWIL, CH-8990 PFAFFIKON.
PCT International Classification Number G01N21/84
PCT International Application Number N/A
PCT International Filing date
PCT Conventions:
# PCT Application Number Date of Convention Priority Country
1 NA