Traitement en cours

Veuillez attendre...



Aller à Demande


Note: Texte fondé sur des processus automatiques de reconnaissance optique de caractères. Seule la version PDF a une valeur juridique

[ EN ]

A learning method for an image analysis system for use in the analysis of an object as well as uses of the method

The invention relates to a learning method for an image analysis system for use in the analysis of an object, wherein the object is compared with a reference object, and comprising the steps of:

a. capturing a video image of the reference object which is represented by a plurality of pixel positions, each pixel having its own identification.

b. dividing parts of the video image into a plurality of subareas which represent a plurality of classes.

The invention moreover relates to uses of the method.

Such a method is known e.g. from EP Patent Application No. 574 831. The method of this European publication comprises capturing a video image of an object and defining image areas of the video image on a video screen. Each of these areas is assigned to a specific class. Once the classes have been determined, an iterative process is initiated, comprising mapping the entire video image by means of the selected classes. In other words, the areas of the video image which have not been selected for a specific class, will be assigned to a class which depends on the colour value which the current pixel in the video image has. In the method known from the above-mentioned EP publication the process time is rather long, the reason being that the value of a Hermite polynomial having 64 coefficients has to be calculated for each class/component in each colour space position, i.e. pixel position. This additionally results in a restriction in the resolution, since 64 coefficients mean that the col- our value of a pixel in RGB colour coordinates can only comprise resolutions of 4 x 4 x 4, which may be insufficient for the classification of patterns complicated m terms of colour, such as chequered fabrics.

Accordingly, the object of the invention is to provide a method of the type stated in the introductory portion of claim 1, allowing the process time to be reduced considerably, and allowing complicated patterns to be ana-lyzed.

The object of the invention is achieved by:

c. setting up a user-defined table of classes, where a plurality of identifications is assigned to each class in the table, each class in the able being formed by the user by his selection of a plurality of pixels in the reference image,

d. setting up a special zero class in the table which contains pixel identifications which have not been assigned to any class by the user, and

e. setting up a special conflict class which contains the pixel identifications which have been assigned to more than one class by the user.

It is ensured in this manner that the teaching of the image analysis system can take place with a low process time, while currently monitoring whether the classes of the class table are selected expediently.

When, as stated m claim 2, all pixels belonging to the zero class or the conflict class are assigned to the class having a pixel identification which is closest to the identification of the pixel concerned, the entire im- age face will receive pixel values which belong to one of the selected classes, so that the entire image has now been converted into an image which is composed of the symbolic colour values occurring for the selected classes.

It is noted in this connection that the expression "which is closest to" should be interpreted m view of the selected pixel identification. If, e.g., this ldentifica-tion is expressed as a set of parameters m the form of a vector, the expression "which is closest to" may be interpreted as the nearest vector distance.

When, as stated in claim 3, the assignment takes place m several stages, each stage comprising processing pixel identifications which adjoin pixel identifications which have already been assigned to a class, it is ensured that the colour composition of the "taught" system has colour change areas which correspond to the original image, but now with more distinct details. Thus, the teaching of the system takes place by adding a growth to the original image, an identification being allocated to all pixels in the image .

For the purpose of using the system for analysis proper, i.e. use of the system after learning, it is an advantage if several reference images are used in the learning, since this allows provision of an image which is representative of an average of objects to be analyzed.

Owing to the flexibility of the image analysis system m the learning it is an advantage, as stated in claim 6, that a user-selected class may be suppressed, which may be desirable if too many items of colour information have been allocated to the selected class.

Other expedient embodiments of the invention are defined m the remaining dependent claims.

As mentioned, the invention also relates to uses of the method. These uses are defined in claims 10-15.

Several of these uses contain very complicated image compositions, but thanks to the very flexible class table structure according to the invention it is possible to provide extremely useful analyses which have been impossible to perform till now without using much more complicated analysis systems.

The invention will now be explained more fully below with reference to an example which is shown m the drawing and illustrates the principles of the invention. In the drawing :

fig. 1 shows an analysis set-up for use in the method ac-cording to the invention,

fig. 2 shows a user display for use m the teaching of the system according to the invention,

fig. 3 shows the use display after an image has been classified,

fig. 4 shows the user display of fig. 3 after a further teaching stage of the image analysis system,

fig. 5 is a block diagram of a process stage in the taechmg of the image analysis system of fig. 3, and

figs. 6A-6B schematically show several stages m the taechmg of the image analysis system of fig. 4.

In fig. 1, the numeral 1 designates an object of a certain complexity, the numeral 2 designates a camera which can capture an image of the object 1 which is sent to a so-called video grabber 3, which is moreover connected to a processor 4. The processor 4 has a keyboard 5 and a mouse 6, as is well-known in connection with image analysis systems. Further, the processor 4 is connected to a display 7 which shows a video image 8 of the object 1.

Fig. 2 generally shows a display which may e.g. be a screen for a PC adapted to receive instructions from the user for use when performing the principles of the invention. The numeral 9 designates an image, shown here as a test pattern for DR-TV which is to be used as a reference image m the teaching of the image analysis system. The numeral 8 designates a first class 1 whose identification at 10 is shown as a background. A further position, for which the colour green has been chosen, may be selected below the position 8. Each of these colours corresponds to a class which may be selected arbitrarily by the user, who, as shown by the reference numeral 11, has the option of changing colour m the class concerned. As additionally shown by the display of fig. 2, a symbol field "show error" is shown at 12, said "show error" field being adapted to show how the selected classes will transform the image 9. The numeral 13 designates a functional field 13 that may be activated by means of a mouse, which initiates the next stage of the learning, as will be explained in connection with fig. 6. Further, the numeral 14 designates a functional field m which all classes may be reset at once, if the classification made by the user is not expedient. Finally, the numeral 15 designates a functional field which is adapted to remove one or more of the classes selected previously.

The system of fig. 2 operates in the following manner:

The user selects areas on the image 9 which are defined by means of a line or a polygon. All the pixels encircled by the line or the polygon are then assigned to a class, e.g. No. 1, and a colour, which may e.g. be the back-ground 10, is then selected for the associated class. What now happens is that all pixel identifications comprised by the line or the polygon will be assigned to the class 1. It is noted in this connection that identifications are taken to mean a set of one or more numerical values which describe properties of the pixel concerned. Then a new class is set up by drawing another line or polygon. In this case a class with a number of pixels will likewise be set up, but pixels previously assigned to class 1, will now be removed from the new class as well as from the old class and be transferred to a so-called conflict class. The procedure is repeated with a suitable number of classes, which may be up to 8 in the shown display of fig. 2, which, however, is not to be regarded as a restriction. Once all the classes have been selected, the functional key 12 is activated, and an image like the one shown in fig. 3 will appear. A selection of four classes is shown m fig. 3, viz. a first class, where grey, white, black, palo green, magenta and other colours have been assigned to the first class, which is called white, while the next class has been given the colour grey, formed by the shades red, blue plus some other colours. Finally, the colours green, yellow plus some other colours have been assigned to the class pale grey, while unknown colours, i.e. colours not found in any other class, have been given the shade black. Activation of the key 12 in fig. 2 provides an image which has the appearance shown in fig. 3. It will now be seen that the image has been distorted somewhat, as only pixels contained m the classes indicated m fig. 3 are included in the image.

Fig. 4 shows a next stage m the teaching process of the image analysis system according to the invention. For the understanding of fig. 4 it is expedient to have a look at figs. 6A-6I. Figures 6A-6B thus show a first stage m the teaching of the image system according to the invention, more particularly the process stages occurring m connection with the matter shown in figs. 2 and 3. In fig. 6A the starting point is a blank screen, and fig. 6B shows the selection of two classes, where class 1 has been given the designation 22, while class 2 has been given the designation 23. These classes have been produced e.g. by drawing a line or a polygon on the screen. As will additionally be seen from fig. 6B, the numeral 24 designates 1/2, the reason being that the polygon or line se-lected has hit a pixel value which belongs to both classes. This is removed in fig. 6C so that both classes are unambiguously defined on the screen image. It should be noted that there are some areas which it has not been possible to assign to any class, the reason being that they contain pixel identifications which are not contained in any of the two classes .

To arrive at the result m fig. 4, the following stages shown m figs. 6D-6E take place. In fig. 6D, positions of pixels disposed m the vicinity of class 2 are filled by pixel identifications which correspond precisely to class 2. Correspondingly, pixel identifications encircled by pixel identifications m fig. 6C, which have the class 1, have been assigned to the class 1. The process continues in fig. 6E, and, as will be seen from the subsequent figures, the entire image field is filled with pixel values which may be allocated either to class 1 or to class 2. This may be expressed m the manner that once the classes have been selected, it is decided to allow the image to "grow" in stages, following which all pixels m the entire image are given a pixel identification. It should be noted that time-consuming calculations may be avoided by dividing the image in this manner.

It is noted that the teaching of the image analysis sys-tern according to the invention, as shown in figs. 6A-6I, is highly simplified. Thus, of course, nothing prevents the use of several classes, and the stages shown in figs. 6D-6I from taking place in several dimensions.

Fig. 5 shows a block diagram of a routine in the division of a reference image for use in the teaching of the image analysis system according to the invention, in the first stage. As will be seen, a first reference image is shown initially and is processed by means of the marking field 12 in fig. 2, which results in the display of a non-grown image. If it now turns out that the image is much too uniform or does not contain the desired information, a new class may be set up, which is shown by the feedback loop. Alternatively, a class may be removed, which may perhaps be necessary if the selected class contains too many pixel identifications. Classes may be set up and be removed in this manner until the resolution required for the purpose of analysis is achieved.

A system as described above offers an incredible number of applications because it is flexible, both as regards analysis of very complicated patterns and as regards analysis of very simple patterns.

Examples of possible applications include:

• Exposure (image processing like e.g. in a photo shop)

• Analysis of bacteria
• Analysis of cells/tissue
• Evaluation of wood
• Quality control in the textile industries • Quality control in industrial laundries (stains, holes and discoloration)
• Analysis of cement
• Geological analysis
• Metallurgy
• Analytical chemistry
• Grading of fruit, vegetables, berries or the like
• Analysis of plastics
• Quality control in the food industries, e.g. of cuts of meat.

Although the invention has been described in connection with very flexible teaching of an image analysis system, nothing prevents the principles of the invention from be-ing used in other connections.

The principles may e.g. be generalized as follows:

Each of a plurality of objects N has properties charac-terized by an identification in the form of sets of parameters Xi, X2...Xn = X,-. A plurality of objects y, where y<N, each object y having the identification X, is now given the same identification Xni, which is called class 1. Similarly, another plurality of objects Z is picked out, where 2<N, and these objects are given the identification Xn,2, which is called class 2. The process continues until a suitable number of classes has been created. Then, all objects are localized in N which have not been assigned to a class, or which have been assigned to more than one class. These objects are assigned to a class whose identification is closest to the current identification of the object concerned, which may be done e.g. by using the vector distance between the parameter sets of the individual objects and the identification of the classes. This provides a system which is generally useful in connection with automatic classification of objects .