PatentDe  


Dokumentenidentifikation EP0628190 10.10.2002
EP-Veröffentlichungsnummer 0628190
Titel VERFAHREN ZUR MASKENERZEUGUNG
Anmelder British Telecommunications p.l.c., London, GB
Erfinder SHACKLETON, Andrew, Mark, Suffolk IP4 5AH, GB;
WELSH, John, William, Suffolk IP2 9EF, GB
Vertreter Beetz & Partner, 80538 München
DE-Aktenzeichen 69332265
Vertragsstaaten DE, ES, FR, GB, IT, NL, SE
Sprache des Dokument EN
EP-Anmeldetag 28.01.1993
EP-Aktenzeichen 939024758
WO-Anmeldetag 28.01.1993
PCT-Aktenzeichen PCT/GB93/00179
WO-Veröffentlichungsnummer 0093015475
WO-Veröffentlichungsdatum 05.08.1993
EP-Offenlegungsdatum 14.12.1994
EP date of grant 04.09.2002
Veröffentlichungstag im Patentblatt 10.10.2002
IPC-Hauptklasse G06K 9/62
IPC-Nebenklasse G06K 9/46   

Beschreibung[en]

This invention relates to a method of forming a template of an image of an object, which template finds particular, but not exclusive, application to the recognition of objects in images for security and surveillance applications and in model-based image coding.

A template is a representation of an image of an object which aims to contain the image's salient features so that the template can be subsequently compared to other images, the goodness of fit of the template to another image providing a measure of the likelihood that this image is of or contains a sub-image of the object. For example, a series of templates may be formed from respective images of the faces of a group of people. An image of a person's face captured by a security camera could then be compared to the templates to determine if that person is one of the group.

A known approach to the use of templates is described in an article entitled "Feature Extraction From Faces Using Deformable Templates" by A L Yuille, D S Cohen and P W Hallinan, Harvard Robotics Laboratory Technical Report 33-2, 1988. In this approach a feature or object of interest is described by a geometric representation which is specified by a set of parameters which incorporate knowledge of the shape of the object. The article gives examples of two such templates; one for use in recognising an eye and the other a mouth. In the former case the eye template decided upon by the authors incorporated a circle corresponding to the outer boundary of the iris, boundary contours corresponding to the top and bottom of the eye comprising two parabolas, regions corresponding to the whites of the eyes, and two points corresponding to the centres of the whites of the eyes.

The template of Yuille et al is deformable in that the components of the template can move subject to notional structural, restraining forces between the components. When the template is matched to a test image, the components of the template are attracted to various valleys, edge and peak components of the image which attractive forces are generally resisted by the structural forces between the components of the template. An energy function is defined in terms of the attraction of the template components to the features of the image and the structural forces resisting deformation of the template, the template being deformed to minimise this energy function. This energy function gives a measure of the goodness of fit of the template to the image and can be used as a measure of whether the template has, in fact, been successfully matched to the image, ie that the image does indeed correspond to the template.

This approach to the forming of templates was devised so that a priori information available about the appearance of the features of the object to which the template was to be matched, in Yuille et al's particular case facial features, could be incorporated into the template so providing a mechanism for detecting edges or other features which comprise local information as global structures.

The Applicant has appreciated that there are, however, disadvantages associated with this approach to forming a template. Firstly, because the template design is based on a priori information, it is important that the information incorporated into the template is accurate. For example, a simplification of an edge so that it can be mirrored in the template as a simple geometric line may in fact hinder matching of the template to an image. It has been found, for example, that the edge of the lower eyelid can vary considerably between the populations of different countries so that a template designed to recognise a typically "Western" eye fairs badly when it is attempted to fit the template to a typically "Eastern" eye.

It is also a characteristic of such a known template that it will, in some cases, not describe accurately the feature in an image but rather what the human perceives to be the salient features of the object in the image. A further disadvantage is that these templates have to be designed "by hand" and individually coded which is time consuming and may be carried out inaccurately.

According to the present invention a method of forming a template of an image of an object includes the steps of:

  • a) detecting occurrences of at least one feature type within the image which meet a respective criterion;
  • b) for each such occurrence determining a feature position which is a point within the image at which the occurrence of the feature type is to be considered as being located; and
  • c) constructing a structural mesh of links between the feature positions.

The method of the present invention processes an image to form a template appropriate to a particular object (for example an eye or a face) directly from a training image of the object. The advantage of this method is that the template encodes the significant attributes of the object which are present in the image and does not attempt to encode possibly illusory details. Furthermore, the method can be carried out without any human input into the design of the template and can potentially be applied to any object presented in a training image. Once formed the template can be used in the manner of known deformable templates for use in image recognition and other applications.

The template is derived, in the method of the invention, directly from the training image with no guidance from a human operator and without prior knowledge of what the image represents being necessary by processing a training image.

In order to provide the method of the present invention with a reasonable opportunity of deriving a suitable template from an image, the object is preferably centred within the image, and is separated from other features, that is, only the object for which a template is sought should be present in the image. The conditions may be relaxed when the template is subsequently used for matching to a test image.

The nodes of the structural mesh of links between the feature positions can be considered as linked tokens in which each token consists of a feature type, possibly attribute values such as the strength and the extent of the feature type, and its position. The template formed by the method of the present invention can be considered as being constructed of a network of such primitive tokens that have been extracted from the training image by appropriate image processing operations.

The method of the present invention is selective as to which occurrences of a feature type are included in the template. For example, it is preferable that only feature types having sufficient strength in the image are considered as these are more likely to be salient image attributes so the criterion can be set to discard all but the desired feature types detected.

In general, a feature type will extend over a portion of the image. In the template of the present invention, the feature type is considered to be located at a point within the image, so that it can form a node of the structural mesh of the template. Thus, for an occurrence of a feature type meeting the required criterion a point has to be determined which is to be that point where the feature is deemed to be located. This could, for example, be the centroid of the feature type.

The structural mesh of links between the feature positions will incorporate structural forces between the nodes of the mesh to restrain the deformation of the template when being matched to test images.

The detection of an occurrence of a feature type within an image can be carried out directly by:

  • d) forming a representation image in which only those portions of the image which are of the feature type are retained;
  • e) thresholding the representation image to obtain an image bitmap;
  • f) locating areas of activity within the image bitmap; and
  • g) determining which of the located areas of activity meet the criterion to be detected as an occurrence of the feature type.

These steps are carried out for each class and orientation of feature type that is to be used to form a template. Once all the feature types have been identified then the structural mesh of links can be formed between them.

The feature types may include, for example, oriented edge, valley and peak images which may not be ideal but which have been found to provide a useful starting set for forming templates according to the present invention.

The chosen feature types can be identified by any appropriate means. For example, the edge images may be found by using simple oriented Sobel filters and the peak and valley images derived using orientation-sensitive morphological filters. Preferably to remove any overlap of responses between feature types of a given class of feature (eg edges or valleys) but of different orientations, only the maximum valued picture elements across all orientations are retained; the corresponding picture elements at all other orientations are set to 0. It is possible in this way to obtain a set of representation images for each orientation of each feature type.

The criterion that a feature type is to be regarded as occurring in an image is conveniently that the number of picture elements in an area of activity is greater than a threshold size. Those areas that are smaller are discarded, only the remainder being used to form the nodes of the structural mesh of the template.

The feature position of an occurrence of a feature type is preferably regarded as being the centroid of the corresponding area of image activity, or blob, in the modified image bitmap.

The present invention is not restricted to any particular type of structural mesh formed to interlink the feature types. One convenient structural mesh is formed by performing a known Delauney triangulation of all the feature types. The resulting triangulated mesh of feature types, or tokens; then forms the template for the object represented in the training image.

In some instances it may be preferable to delete from the template links greater than a predetermined length in order to allow greater freedom of movement of the template between areas which may not be closely linked physically.

A token associated with any feature type may include as additional attributes the size, principal orientation and aspect ratio of the activity area, for example.

It is possible to extract meaning from the final template. The template which has been derived can be displayed by individually interpreting each constituent token. As each token describes an expected image attribute, the corresponding attribute can be displayed in the appropriate position on the image and so allow the template to be viewed, as a whole, relative to it. It is then possible to assign meaning to the parts of the template by visual inspection and labelling key points of interest which need not themselves actually be coincident with any of the tokens. For example, it may be appropriate that a key point is maintained with respect to an image to which the template is being fitted, for example for a face the nominal centre position of a nose could be labelled which may not have any coincident image activity. This position could then be obtained by its context within the fit of the template to any face.

A template formed by analyzing a single training image may be particularly suited to identifying when that image re-occurs by matching the template to it. However, there are many applications where it is necessary to identify the presence of an object which has not appeared specifically in a training image. For example, if it is necessary to recognise, in general terms, the presence an eye in an image then a template formed from analyzing one example of an eye may not be ideally suited to this general recognition problem as the particular eye on which the template was formed may contain feature types which are not a common characteristic of eyes generally. In these circumstances it has been found beneficial to form a generalised template formed by combining templates formed from several images each being a different example of the type of object to be recognised. So, for example, a series of different eyes can be used to form a respective series of templates. These templates are then combined to form a generalised template.

A number of possible approaches to the problem of combining the templates to form a composite template have been considered by the applicant. These included a graph matching approach, for example as described by J Segan in an article entitled "Model Learning and Recognition of Non-Rigid Objects" Computer Vision and Pattern Recognition, pages 597-602, 1989; and an energy minimisation comparison technique, similar to that used in fitting templates to images, but used between templates.

However, the approach which appears to be most promising at the time of filing this application is to use a genetic algorithm. Genetic algorithms are based on the principle of natural selection and evolution found in biological systems. The Applicant has devised and applied a genetic-type algorithm to combine templates. A set of possible solutions, in this case templates, are maintained as a population which is adapted by combining the best members whilst others are discarded to generate increasingly successful templates better able to fit faces, for example.

An embodiment of the present invention will now be described, by way of example, only with reference to the accompanying drawings in which:

  • Figure 1 is a photograph constituting an image of a man's head and shoulders;
  • Figures 2(a) to 2(d) are photographs of representation images obtained from the image of Figure 1, each of which retains occurrences of a different feature type;
  • Figures 3(a) to 3(d) are photographs of image bitmaps obtained from the representations images 2(a) to 2(d), respectively;
  • Figures 4(a) to 4(d) are photographs of those areas of activity larger than a predetermined threshold size and so are regarded as occurrences of a feature type in the image bitmaps of the Figures 3(a) to 3(d), respectively;
  • Figure 5 is a photograph of a template formed from the feature types of Figures 4(a) to 4(d) shown superimposed on the image of Figure 1 from which the template was derived;
  • Figure 6 is a schematic diagram showing a chromosomal representation of a template for a genetically based, template combining algorithm;
  • Figure 7 is a diagram of an image divided into zones for representation as a chromosome;
  • Figure 8 is a representation of the information associated with each gene of the chromosome of Figure 6;
  • Figure 9 is a diagram showing a method of dividing an image into zones by a recursive method;
  • Figure 10 is a graph showing the subdivision recursion tree associated with Figure 9;
  • Figures 11 and 12 are diagrams showing how two parent chromosomes produce two child offspring chromosomes by a crossover mechanism; and
  • Figure 13 is a diagram showing how a chromosome can be modified by inverting a gene string.

Referring first to Figure 1 there is shown an exemplary image of a man's head and shoulders for which a method of forming a template according to the present invention will now be described. The image was captured by means of a TV camera coupled to a image storing device which stored the values of the constituent picture elements. In this particular example four feature types were considered, namely vertical and horizontal edges and edges running at 45° to the vertical. The image was therefore processed in known manner to form four representation images in which only those portions of the image of a respective feature type were retained. The resulting representation images for these features are shown in Figures 2(a) to 2(d).

Consideration will now be restricted to the processing of the Figure 2(a) representation image to describe how feature types were extracted from a processed representation. The same process was also applied to the other three representation images of Figures 2(b) to 2(d).

First, the representation image of Figure 2(a) was thresholded to obtain an image bitmap as illustrated in Figure 3(a) which retained only those portions of the image appropriate feature type. The picture elements of the image bitmaps of different orientations of a given feature type were then compared and the maximum valued picture element only retained. That is the image bitmap containing the maximum valued picture element retained that pixel whereas the equivalent picture elements in the other image bitmaps with which it has been compared are deleted. This provides a series of image bitmaps for a given feature type without any overlap between the areas of activity for the various orientations of that feature type. The size of the individual areas of the image bitmap of Figure 3(a) were then sized by counting the number of picture elements in the particular feature. This gave a measure of the size of the feature which was then tested against a threshold. Those which were too small were deleted from the image bitmap to obtain the modified image bitmap of Figure 4(a).

The areas of activity that remained in the modified image bitmap of Figure 4(a) were then detected as an occurrence of the feature type, all the other instances being ignored. The feature position was calculated by weight averaging the picture element constituents of the feature, ie the centroid of the feature, which position was considered to be the point of the image at which the feature was located. Associated with this feature was a token which represented the feature type and its position within the image.

Any attributes that may be of interest according to the particular application to which the template is going to be put could also form part of the token. For example, it could contain information about the size of the feature and its strength.

The result of this processing was a series of tokens which characterised the feature type content of the original image. In a similar fashion a set of tokens was obtained for the other feature types that were to be used in forming the final template. The full set of tokens for all the feature types was then constructed which set formed the basis of a template of the original image obtained by triangulating the feature type positions.

Referring now to Figure 5 there is illustrated a template constructed by forming a structural mesh of links between the feature positions of all the tokenised feature types of Figures 4(a) to 4(d). In this case the mesh was formed by connecting the feature positions with a triangulated mesh using Delauney triangulation of all the feature positions. The template so formed was the desired template representation of the original image.

The photograph at Figure 5 shows the template superimposed on the original image with the feature positions being denoted by node positions which are linked by straight line segments.

The choice of which feature types are to be used in the formation of a templare is not fixed. The set selected should, however, be chosen so that its members are distinct, informative in the sense of picking out key object details, and as independent of changes in illumination as possible. An article entitled "Pre-Attentive Processing in Vision" by Anne Treisman, in Computer Vision, Graphics, and Image Processing vol 31, 156-177 (1985) defines a possible set of such feature types.

In general, the different processed images will not be shown visually on a visual display unit during processing of an image to form a template, rather the steps of the invention will utilise stored representations of the training image and its processed forms.

A template according to the present invention can be used to both recognise and describe the shape of novel instances of the object type upon which it has been trained by initialising it in the vicinity of an object in an image and allowing template feature types to be attracted to corresponding characteristics in the target image whilst at the same time deforming its geometric shape within limits according to how closely image detail can be matched. This fitting process is analogous to that used by traditional deformable templates of the type described by Yuille et al, referenced above. If a template is to be suitable for fitting and so describes the shape of the majority of novel instances of a given object which may be subsequently encountered then it has been found by the Applicant, to be greatly beneficial that templates are formed from more than one image of different members of the class of object which is to be recognised and combining the templates to form a generalised template. This scheme, which also constitutes an aspect of the present invention, provides a generalised template which incorporates the salient features appropriate to the class of object rather than the specific features for a particular member of the class. This provides a template which is more suitable for classifying objects as opposed to identifying a particular member of a class of objects where a template solely based on that one image will be preferable.

There will now be described a particular method of combining templates to form such a generalised template according to the present invention.

The method of forming a generalised template presently preferred by the Applicant is to use a genetic algorithm to determine which of the template features are most valuable in recognising an object in the class on which it is trained. The basic algorithm requires a genetic representation analogous to the chromosomes in living creatures, a way of interpreting this representation to form a population member and a method of establishing the fitness of each population member.

The algorithm itself starts with an initial, possibly poor or unfit, population and iteratively improves the fitness by preferentially selecting fit members to be the parents of the next generation' s members. Fitter members thus tend to contribute preferentially to the gene pool of the next generation, and it can be shown that genes which contribute to the above average fitness tend to increase exponentially in frequency from generation to generation.

The power of genetic algorithms does not lie merely in the survival of the fittest strategy but also in the way in which the two population members are combined. The chromosomes from any two parents should both contribute to the chromosome of their children (usually two are produced) by means of crossover. This is a phenomenon observed in animal genetics in which two chromosomes align with one another and contiguous strands of genetic material are exchanged. Thus new combinations of genes from fitter population members are constantly being tried together to obtain an overall fitness improvement.

In the context of the present invention, the set of templates to be combined should be based on a set of training images which are representative of objects that are likely to be encountered in the future. In particular, the set of image characteristics described by the derived templates should contain all the relevant elements to describe the appearance of most new objects likely to be presented to an image recognising apparatus.

The genetic algorithm is then applied to swap elements between chromosomes from generation to generation which has the effect of producing new chromosomes representing templates which contain new mixtures of elements taken from various different initial templates. A part of a template good at matching for example eyes in one template may be co-joined with a part of another template good at matching a mouth thereby producing a composite template good at doing both tasks.

The details of an implementation of such a genetic algorithm will now be described in more detail.

Each template that is going to contribute to the composite template must be encoded as a chromosomal representation for the genetic algorithm to operate upon.

As well as a chromosomal representation of a template there must also be a method of interpreting chromosomes in order to derive the corresponding template individual. In the context of the present invention one must also be able to create a chromosome given the template description in order to initialise the starting population given the training set of derived templates.

There are certain properties that it is desirable a chromosome representation exhibits due to constraints laid down by genetic algorithms in general as well as practical considerations. It is preferable to have a fixed length chromosome description consisting of a fixed number of gene units as this helps during mating and crossover of chromosomes. Collections of genes which provide advantageous attributes in individuals are said to be co-adapted, that is they work well together. It is preferable that such co-adapted genes congregate locally on the chromosome such that they tend to be preserved during crossover. Each gene should preferably represent an element from an alphabet of low order, ie it should only be able to take on a small range of values.

The chromosomes described to date in the scientific literature consist of linear strings of genes. Although a 2-D chromosomal representation may seem to be closer to the 2-D nature of the templates, the present embodiment uses the more traditional linear mapping of genes for which the theory is now reasonably well developed albeit in other applications. Alternative chromosome representation strategies may, of course, be adopted if they prove to be advantageous. The linear structure of the genes within a chromosome should not present any problems, however, since during initialisation the feature types which are in close proximity spatially are represented by genes close together on the chromosome. The property exhibited by genetic algorithms is to provide clustering of co-adapted genes on the chromosome which should help to prevent such clusters from being broken up during crossover.

Referring now to Figure 6 a chromosome representation of a template used comprised a set of n genes Gi, i=0 to n-1, each gene Gi representing the feature type in one area of the image. In order to fix the length of the chromosome the image area was quantised into a fixed number of zones (square regions) Zi as shown in Figure 7. Each zone Zi could contain one or more instances of the recognisable feature types but only one was coded in the chromosome.

Figure 8 shows the information encoded by a gene Gi of a chromosome. It consisted of three gene fields, F1, F2 and F3. Gene field F1 contained the number of the zone Zi which it encodes - in this case zone number 34. Gene field F2 contained a flag indicating whether the feature type within that zone is active. Gene field F3 denoted the feature type - in this case a top right to bottom left, 45° edge.

Since the genes Gi contained a reference to the zone Zi of the image to which they corresponded the zone number Zi and the position of a gene on a chromosome were therefore independent so they could change their relative chromosome location freely, as desired. The genes were restricted to elements from a relatively low order alphabet consisting of the number of allowable feature types.

As stated above, the zone Zi to which a gene Gi corresponded was indicated by a zone number. The actual zone details such as zone position and adjacency information were held in a separate location in order to minimise the extraneous detail which had to be coded by each gene. A compact representation is more efficient which was important during implementation of the algorithm.

The connectivity of the feature types within the template structural mesh was not encoded at all within the genetic representation but was created automatically during expression of the chromosome to produce the corresponding template as will be described below.

Although genetic algorithms are often initialised randomly and then allowed to search for a solution, it is preferable to initialise the population from a set of base templates derived from a set of training images of the object of interest. It should be noted that these images were very crudely normalised such that the centres of the eye and mouth in each image also align across all the images in order to assist the genetic algorithm in combining templates during later stages.

The first stage of initialisation was to convert the base templates into their equivalent chromosome representations. This consisted of determining within which zone each template feature type lay and then assigning a gene for that feature type. When more than one 9 feature type falls within a single zone, an arbitrary one was kept the rest being discarded. The only remaining variable in the initialisation was the ordering of the genes on the chromosome. Since it was not known at initialisation which feature types were related, and hence which genes were co-adapted, the optimal ordering of genes was not known. However, a strategy was adopted which located genes containing spatially close tokens in similar proximity on the chromosome as it had been found that local feature types are often related to one another.

The strategy used in the present embodiment to decide which genes go where on the chromosome consisted of a recursive spatial subdivision process applied to the space of the entire image.

Referring to Figures 9 and 10, a full image 62 was recursively split once, evenly, into four sub-regions S1 to S4 about the midpoint of the image, which sub-regions were then randomly ordered and split again until the sub-regions Sijk reached the specified zone size. Each time the recursion reached a minimum sized zone sub-region the corresponding gene was added onto the end of the growing list of genes on the chromosome. Spatially proximate zones were generally close in the recursion tree and so mapped onto nearby genes.

Although there is no perfect mapping from 2-D to 1-D which preserves locality, this process provided orderings which exhibited a high degree of locality whilst introducing some variability into the ordering as between the different initial chromosomes.

The chromosomal description of an individual is known as the genotype and this can be expressed or interpreted to form the individual itself known as the phenotype. It is necessary to be able to express the chromosome in terms of the phenotype before one can go on to evaluate the fitness of the given individual. The phenotype consists of a template in the present instance.

The process of expressing the chromosome was relatively straightforward in the present case. Each gene Gi of the chromosome was examined and if marked as active a corresponding feature type was created in the generalised template structure of the type specified by the field F3 of the gene Gi. The gene Gi also implicitly codes the position of the feature type in terms of the zone index which could be cross referenced to the list of zone details to provide the feature position information.

The template nodes had to be triangulated as previously described to form the template and for fitting to a given test image. The triangulation was performed as a post-processing stage just prior to fitting the template. No node connectivity information was encoded within the genetic representation itself.

In order that the mechanism of selection could proceed it was necessary to have a means of evaluating a measure of merit or fitness for a given template offspring arising out of the combination of two template chromosomes. The individual template was first obtained from the corresponding chromosomal representation as described above. It then had to be evaluated for fitness.

It is first necessary to define what is meant by a good template for a given image of an object. The basic requirement is that the template should achieve a good match to the image data of a given test image as well as achieving a geometric fit with small deformation. The fitting process used to test the fitness of a given template will correspond to the fitting process to be used on the final composite template when matching test images.

Given the above requirements it can be seen that a trivial solution for obtaining a maximal fit and image response is a template which saturates a test image with all possible, feature types to be matched. Since it is also required that a template description is as concise as possible, that is without an instances excess of feature types, a cost is associated with the template which increases with increasing numbers of occurrences of feature types and so prevents oversaturation of template nodes.

Bringing these elements together one can formulate an objective measure of merit for a given template applied to a given test image where greater values indicate more desirable templates subject to our criteria. The merit equation used in this embodiment was the merit value G given by:

where the variables are:
G
unnormalised number of merit
n
number of nodes
N
desirable number of nodes
It
image response representation types t
ti
type of node i
xi
position of feature type i
num_edges
number of inter-node connections
lij
current distance between connected nodes
L0ij
original distance between connected nodes
Gs
Global mesh scale
p
excessive node penalty constant
ki
empirical determined weighting factors

The final template had to work well for most of the novel faces likely to be encountered so it was necessary to assess the template' s merit measure against a range of faces. An environment of face images was therefore used which is separate from the initial training set. To evaluate each member of the population at a given generation a sample of test face images was randomly picked from the environment set. Each individual template was then fitted to each of the test set of images using a small number of iterations on the above measure of merit and a single average merit value calculated per template. These values formed the final unnormalised measures of merit for each template.

The merit value G described above includes both negative and positive values. It is necessary to form a measure of fitness for use by the genetic algorithm which is both non-negative and exhibits an even spread of values across the population during any given generation so the better members can be effectively discriminated for. The measure G was normalised to achieve these criteria and yield a fitness value that was useable.

There are several possibilities for the normalisation of the objective merit measure to form a more useful fitness measure. Normalisation was performed such that an average performing member of the population received a fitness of unity, with better members having greater fitness values and worse members having lesser but always non-negative fitness values. A possible normalisation scheme is to calculate the normalised fitness f using the equation f=1 + (G - µ) / ((k.σ)) where the mean µ and standard deviations σ of the values G across the population are first calculated and then the normalised fitness is calculated for each population member. "k" is a small constant, typically less than 3, which governs the spread of resulting fitness values. In a few instances where the resulting fitness f will be negative, it is set to 0.

Genetic algorithms select members from the existing population to breed and produce the next generation. This selection is performed in a probabilistic fashion depending on the relative fitness of each member. The actual method used in this embodiment was stochastic remainder selection which has been found to perform better than purely random procedures. This method guarantees that the better than average members will be appropriately represented in the succeeding generation by deterministically assigning part of the breeding population prior to random selection of the remainder of the breeding population.

Crossover is the most powerful of the genetic operators employed in this genetic algorithm and will now be explained with references to Figures 11 and 12. Crossover provides the means for useful attributes of one individual to combine with other useful attributes from a second individual to produce a child which hopefully exhibits the strengths of both. Many combinations are tried together with poor combinations being weeded out by fitness selection. Crossover was applied with a high probability of about 0. 6.

To effect crossover, two cut points 70 and 71 were randomly chosen using a uniform probability distribution over the length of two chromosomes GA and GB having gene strings GA1 to GAn and GB1 to GBn, respectively. The cut points 70 and 71 define two regions of the parent chromosomes GAc to GAd and GBc to GBd, respectively. These were exchanged to create two children chromosomes GC and GD as shown in Figure 12.

There was a minor problem which had to be resolved to maintain the integrity of both child chromosomes. It would often happen that the set of zones represented by the incoming genes for one child would not match those coded for by the outgoing genes as the gene sequences varied. In such a case some zones will be multiply represented and others completely unrepresented. This was remedied by matching surplus genes in the first child GC outside the crossover region with surplus genes in the second child GD also outside the crossover region and exchanging these between the two strings. In this way both children GC, GD ended up with a full complement of genes representing all the possible tokens zone positions as we require. The integrity of the sequence of genes in the crossover region was always maintained. This type of approach had been successfully applied previously to similar problems where gene ordering can vary such as genetic algorithms applied to the travelling salesman problem.

A further technique employed in the genetic algorithm of the present invention was mutation, a genetic operator which acts on a single gene to randomly change it in some way. The mutation operation was applied with a very low probability as a background operator to the algorithm as a whole, its function being to reintroduce genetic variation which may have been lost, or new useful variations. The principal mechanisms of this genetic algorithm is the combination of the crossover operator and the survival of the fittest selection.

As mutations occur relatively infrequently we adopted the following strategy rather than keep checking to see whether a mutation will occur for every gene: we decide in advance which genes will mutate. Assuming probability p of single gene mutating, the cumulative probability c of a single mutation in a gene string is given by: c=1-(1-p)n

It was therefore possible to randomly select a cumulative probability c in the range [0, 1) and calculate the next gene to mutate as being the gene n gene-positions ahead in the gene string. The value of n was calculated by rearranging the above equation to give: n=log(1-c/log(1-p))

When the mutation occurred within a gene in the model, the activity of the gene and hence the corresponding token was toggled. If the gene was active and so contained a token it was marked inactive, otherwise it was marked active and the corresponding feature type field F3 assigned a random feature type from those allowable. As there was a gene for every zone within an image, a feature type could, potentially, have become active with any zone Zi at any of the available image locations.

The inversion operator was a further type of mutation operator also included in the algorithm of the present invention which operated at a fairly low probability. It operates on a single chromosome and its affect is to completely invert the sequence order of a randomly chosen string of genes on the chromosome as shown in Figure 13.

The gene string of a chromosome GE to be inverted is chosen by randomly selecting two end points 72, 73 based on the uniform probability function over the entire length of the chromosome. This gene string is then reversed in order within the chromosome GE to form a new chromosome GF.

It is to be noted that inversion does not affect the template corresponding to the chromosome during the current generation as the same tokens are coded regardless of the gene order. The main effect of inversion is to allow genes to change position on the chromosome such that co-adapted genes tend to become clustered over time. Such clusters of genes will tend to be preserved during crossover as the cut point is less likely to fall between co-adapted genes if they are physically close on the chromosome. As an example, genes coding for feature types good at matching an eye may cluster and then usefully propagate through the population and eventually be combined with a similar co-adapted gene cluster relevant to matching a mouth to thereby improve the template as a whole.

The algorithm described above was carried out on a Meiko transputer array consisting of an array of 64 individual transputer based processing nodes and some special nodes consisting of graphics display cards. The actual implementation of the algorithm described above is arbitrary to the extent that it only affects the time taken to compute the generalised template from the set of individual templates.


Anspruch[de]
  1. Verfahren zum Erzeugen einer Schablone eines ersten Bildes für den nachfolgenden Vergleich mit einem zweiten Bild, das umfaßt:
    • Erfassen von Auftritten von einem oder mehreren in dem ersten Bild vorhandenen, im voraus definierten Bildmerkmalstypen;
    • Zuordnen eines Ortes innerhalb des ersten Bildes zu jedem solchen Auftreten;
    • Zuordnen eines Schablonenknotens zu jedem solchen Ort;
    • für jeden solchen Auftritt Zuordnen des Auftrittsmerkmalstyps zu dem dem Auftrittsort zugeordneten Knoten; und
    • Zuordnen von Schablonenverknüpfungen zu Schablonenknoten, um so eine Schablone des ersten Bildes zu erzeugen.
  2. Verfahren nach Anspruch 1, das ferner für jeden erfaßten Auftritt das Zuordnen weiterer Auftrittsdaten zu dem Knoten, der dem Auftrittsort zugeordnet ist, umfaßt.
  3. Verfahren nach Anspruch 1 oder Anspruch 2, bei dem der Schritt des Zuordnens von Schablonenverknüpfungen zu Schablonenknoten die Verwendung der Delauney-Triangulation umfaßt.
  4. Verfahren zum Erzeugen eines Ähnlichkeitsmaßes zwischen einer Schablone eines ersten Bildes und einem zweiten Bild, bei dem die Schablone gemäß einem vorhergehenden Anspruch einschließlich der Erfassung des Auftritts von einem oder mehreren im ersten Bild vorhandenen, im voraus definierten Bildmerkmalstypen erzeugt wird, wobei das Verfahren umfaßt:
    • Erfassen von Auftritten von einem oder mehreren in dem zweiten Bild vorhandenen, im voraus definierten Bildmerkmalstypen;
    • Verformen der Schablone des ersten Bildes in Abhängigkeit von einer im voraus definierten Energiefunktion, um eine verformte Schablone zu erzeugen, derart, daß die Schablonenknoten zu den Auftritten ihrer jeweiligen zugeordneten Merkmalstypen angezogen werden, während die Schablonenverknüpfungen die Schablonenverformung einschränken; und
    • Erzeugen eines Ähnlichkeitsmaßes zwischen der Schablone des ersten Bildes und dem zweiten Bild in Abhängigkeit von der Differenz der Energiezustände, die der Schablone des ersten Bildes und der verformten Schablone zugeordnet sind.
  5. Verfahren zur Bilderkennung, das umfaßt:
    • Erzeugen eines Ähnlichkeitsmaßes zwischen einer Schablone eines ersten Bildes und einem zweiten Bild nach Anspruch 4; und
    • Veranlassen, daß das zweite Bild als eine Instanz des ersten Bildes erkannt wird, falls das Ähnlichkeitsmaß einen im voraus definierten Schwellenwert übersteigt.
  6. Verfahren zum Erzeugen einer Schablone einer Menge von Bildern für den nachfolgenden Vergleich mit einem weiteren Bild, das die folgenden Schritte umfaßt:
    • (a) Erzeugen einer Schablone für jedes Trainingsbild einer Trainingsbildmenge gemäß dem Verfahren nach Anspruch 1;
    • (b) Erzeugen eines Datensatzes, der jeder solchen Schablone entspricht; wobei jeder Datensatz eine lineare Reihe von Einträgen besitzt, wobei jeder Eintrag einem der Knoten der jeweiligen Schablonen entspricht und wenigstens einen zugeordneten Knotenmerkmalstyp und eine zugeordnete Knotenadresse repräsentiert;
    • (c) Erzeugen einer Schablone, die jedem Datensatz entspricht; Verwenden jedes Eintrags, um einen entsprechenden Schablonenknoten des zugeordneten Merkmalstyps zu erzeugen, der sich an einem Schablonenort, der der zugeordneten Adresse entspricht, befindet; und Zuordnen von Schablonenverknüpfungen zu Schablonenknoten, um jede Schablone zu erzeugen;
    • (d) Zuordnen eines Gütemaßes zu jeder Schablone für jedes Bild der Bildmenge;
    • (e) Zuordnen eines Paßmaßes zu jeder Schablone für die Bildmenge in Abhängigkeit von den im Schritt (d) erzeugten Gütemaßen;
    • (f) Auswählen einer Untermenge der Datensätze in Abhängigkeit vom Paßmaß ihrer im Schritt (e) erzeugten zugeordneten Schablonen; und
    • (g) Anwenden genetischer Operatoren wenigstens auf die Untermenge, um neue Datensätze zu erzeugen.
  7. Verfahren nach Anspruch 6, bei dem die Datensatzeinträge, die Knoten entsprechen, denen Auftritte von in einem Bild nahe beieinander liegenden Bildmerkmalstypen zugeordnet sind, in der entsprechenden linearen Reihe nahe beieinander angeordnet sind.
  8. Schablonenerzeugungsvorrichtung zum Erzeugen einer Schablone eines ersten Bildes für den nachfolgenden Vergleich mit einem zweiten Bild, die umfaßt:
    • Mittel, die so beschaffen sind, daß sie Auftritte eines oder mehrerer in dem ersten Bild vorhandener, im voraus definierter Bildmerkmalstypen erfassen;
    • Mittel, die so beschaffen sind, daß sie jedem Auftritt in dem Bild einen Ort zuordnen;
    • Mittel, die so beschaffen sind, daß sie jedem Ort, dem ein Auftritt zugeordnet ist, einen Schablonenknoten zuordnen;
    • Mittel, die so beschaffen sind, daß sie für jeden Auftritt den Auftrittsmerkmalstyp dem dem Auftrittsort zugeordneten Knoten zuordnen; und
    • Mittel, die so beschaffen sind, daß sie den Schablonenknoten Schablonenverknüpfungen zuordnen, um eine Schablone des ersten Bildes zu erzeugen.
  9. Bilderkennungsvorrichtung, die umfaßt:
    • Mittel, die so beschaffen sind, daß sie eine Schablone eines ersten Bildes speichern, wobei die Schablone gemäß einem der Ansprüche 1-3 erzeugt wird, umfassend das Erfassen von Auftritten von einem oder mehreren in dem ersten Bild vorhandenen, im voraus definierten Bildmerkmalstypen;
    • Mittel, die so beschaffen sind, daß sie Auftritte des einen oder der mehreren in dem zweiten Bild vorhandenen, im voraus definierten Bildmerkmalstypen erfassen;
    • Mittel, die so beschaffen sind, daß sie die Schablone des ersten Bildes in Abhängigkeit von einer im voraus definierten Energiefunktion verformen, um eine verformte Schablone zu erzeugen, derart, daß die Schablonenknoten zu Auftritten ihrer jeweiligen zugeordneten Merkmalstypen angezogen werden, während die Schablonenverknüpfungen Schablonenverformungen begrenzen; und
    • Mittel, die so beschaffen sind, daß sie ein Ähnlichkeitsmaß zwischen der Schablone des ersten Bildes und dem zweiten Bild in Abhängigkeit von der Differenz der Energiezustände, die der Schablone des ersten Bildes und der verformten Schablone zugeordnet sind, aufbauen; und
    • Mittel, die so beschaffen sind, daß sie veranlassen, daß das zweite Bild als eine Instanz des ersten Bildes erkannt wird, falls das Ähnlichkeitsmaß einen im voraus definierten Schwellenwert übersteigt.
  10. Schablonenerzeugungsvorrichtung zum Erzeugen einer Schablone einer Menge von Bildern für den nachfolgenden Vergleich mit einem weiteren Bild, die umfaßt:
    • Mittel, die so beschaffen sind, daß sie Schablonen jedes Trainingsbildes einer Trainingsbildmenge speichern, wobei jede Schablone gemäß einem der Ansprüche 1-3 einschließlich der Erfassung von Auftritten von einem oder mehreren in dem ersten Bild vorhandenen, im voraus definierten Bildmerkmalstypen erzeugt wird;
    • Mittel, die so beschaffen sind, daß sie eine Menge von Datensätzen erzeugen; wobei jeder Datensatz einer Schablone entspricht; wobei jeder Datensatz eine lineare Reihe von Einträgen besitzt, wobei jeder Eintrag einem der Knoten der jeweiligen Schablone entspricht und wenigstens einen zugeordneten Knotenmerkmalstyp und eine zugeordnete Knotenadresse repräsentiert;
    • Mittel, die so beschaffen sind, daß sie eine Schablone erzeugen, die jedem Datensatz entspricht, wobei jeder Eintrag verwendet wird, um einen entsprechenden Schablonenknoten des zugeordneten Merkmalstyps zu erzeugen, der sich an einem Schablonenort befindet, der der zugeordneten Adresse entspricht; und wobei Schablonenverknüpfungen zu Schablonenknoten zugeordnet werden, um jede Schablone zu erzeugen;
    • Mittel, die so beschaffen sind, daß sie jeder Schablone ein Gütemaß für jedes Bild der Menge von Bildern zuordnen;
    • Mittel, die so beschaffen sind, daß sie einer Schablone ein Paßmaß für die Bildmenge in Abhängigkeit vom zugeordneten Gütemaß zuordnen;
    • Mittel, die so beschaffen sind, daß sie eine Untermenge von Datensätzen in Abhängigkeit vom Paßmaß ihrer entsprechenden Schablonen auswählen; und
    • Mittel, die so beschaffen sind, daß sie genetische Operatoren wenigstens auf die ausgewählte Untermenge anwenden, um eine neue Menge von Datensätzen zu erzeugen.
Anspruch[en]
  1. A method of creating a template of a first image for subsequent comparison with a second image comprising:
    • detecting occurrences of one or more predefined image feature types present in the first image;
    • associating a location within the first image with each such occurrence;
    • associating each such location with a template node;
    • in respect of each such occurrence, associating the occurrence feature type with the node associated with the occurrence location; and
    • associating template links with template nodes as to create a template of the first image.
  2. A method as claimed in claim 1 further comprising, in respect of each detected occurrence, associating further occurrence data with the node associated with the occurrence location.
  3. A method as claimed in claim 1 or claim 2 in which the step of associating template links with template nodes comprises utilising Delauney triangulation.
  4. A method of establishing a measure of similarity between a template of a first image and a second image, in which the template is created according to any preceding claim including detection of occurrences of one or more predefined image feature types present in the first image, the method comprising:
    • detecting occurrences of the one or more predefined image feature types present in the second image;
    • in dependence upon a predefined energy function, deforming the template of the first image as to create a deformed template such that the template nodes are attracted to occurrences of their respective associated feature types whilst the template links restrain template deformation; and
    • establishing a measure of similarity between the template of the first image and the second image in dependence upon the difference in energy state associated with the template of the first image and the deformed template.
  5. A method of image recognition comprising:
    • establishing a measure of similarity between a template of a first image and a second image according to claim 4; and
    • in the event that the measure of similarity exceeds a predefined threshold value, causing the second image to be recognised as an instance of the first image.
  6. A method of creating a template of a set of images for subsequent comparison with a further image including the steps of:
    • (a) creating a template of each of a set of training images in accordance with the method of claim 1;
    • (b) creating a data record corresponding to each such template; each data record having a linear string of entries, each entry corresponding to one of the nodes of the respective templates and representing at least an associated node feature type and an associated node address;
    • (c) creating a template corresponding to each data record, utilising each entry to create a corresponding template node of the associated feature type and located at a template location corresponding to the associated address; and associating template links with template nodes as to create each template;
    • (d) associating with each template a measure of merit with respect to each one of the set of images;
    • (e) associating with each template a measure of fitness with respect to the image set, in dependence upon the measures of merit established in step (d);
    • (f) selecting a subset of the data records in dependence upon the measure of fitness of their associated templates established in step (e); and
    • (g) applying genetic operators to at least the subset to create new data records.
  7. A method as claimed in claim 6 in which data record entries corresponding to nodes associated with occurrences of image feature types located closely together in an image are located closely together in the corresponding linear string.
  8. A template creation apparatus for creating a template of a first image for subsequent comparison with a second image comprising:
    • means arranged to detect occurrences of one or more predefined image feature types present in the first image;
    • means arranged to associate a location within the image with each occurrence;
    • means arranged to associate each location associated with an occurrence with a template node;
    • means arranged to associate, in respect of each occurrence, the occurrence feature type with the node associated with the occurrence location; and
    • means arranged to associate template links with template nodes as to create a template of the first image.
  9. An image recognition apparatus comprising:
    • means arranged to store a template of a first image, the template created in accordance with any one of claims 1-3 including detection of occurrences of one or more predefined image feature types present in the first image;
    • means arranged to detect occurrences of the one or more predefined image feature types present in the second image;
    • means arranged to deform the template of the first image in dependence upon a predefined energy function as to create a deformed template such that the template nodes are attracted to occurrences of their respective associated feature types whilst the template links restrain template deformation; and
    • means arranged to establish a measure of similarity between the template of the first image and the second image in dependence upon the difference in energy state associated with the template of the first image and the deformed template; and
    • means arranged to cause the second image to be recognised as an instance of the first image in the event that the measure of similarity exceeds a predefined threshold value.
  10. A template creation apparatus for creating a template of a set of images for subsequent comparison with a further image comprising:
    • means arranged to store templates of each of a set of training images, each template created in accordance with any one of claims 1-3 including detection of occurrences of one or more predefined image feature types present in the first image;
    • means arranged to create a set of data records; each data record corresponding to a template; each data record having a linear string of entries, each entry corresponding to one of the nodes of the respective template and representing at least an associated node feature type and an associated node address;
    • means arranged to create a template corresponding to each data record, utilising each entry to create a corresponding template node of the associated feature type and located at a template location corresponding to the associated address; and associating template links with template nodes as to create each template;
    • means arranged to associate with a template a measure of merit with respect to each one of the set of images;
    • means arranged to associate with a template a measure of fitness with respect to the set of images, in dependence upon associated measures of merit;
    • means arranged to select a subset of data records in dependence upon the measure of fitness of their corresponding templates; and
    • means arranged to apply genetic operators to at least the selected subset to create a new set of data records.
Anspruch[fr]
  1. Procédé de création d'un gabarit pour une première image en vue d'une comparaison ultérieure à une seconde image, comprenant :
    • la détection d'occurrences d'un ou plusieurs types de caractéristiques d'image prédéfinis présents dans la première image,
    • l'association d'un emplacement à l'intérieur de la première image et de chaque telle occurrence,
    • l'association de chaque tel emplacement et d'un noeud du gabarit,
    • en fonction de chaque telle occurrence, l'association du type de caractéristique de l'occurrence et du noeud associé à l'emplacement de l'occurrence, et
    • l'association de liaisons de gabarit avec des noeuds de gabarit de façon à créer un gabarit de la première image.
  2. Procédé selon la revendication 1, comprenant en outre, en fonction de chaque occurrence détectée, l'association d'autres données d'occurrence et du noeud associé à l'emplacement de l'occurrence.
  3. Procédé selon la revendication 1 ou la revendication 2, dans lequel l'étape d'association des liaisons de gabarit à des noeuds de gabarit comprend l'utilisation d'une triangulation de Delaunay.
  4. Procédé d'établissement d'une mesure de similarité entre un gabarit d'une première image et d'une seconde image, dans lequel le gabarit est créé selon l'une quelconque des revendications précédentes, comprenant la détection des occurrences d'un ou plusieurs types de caractéristiques d'image prédéfinis présents dans la première image, le procédé comprenant :
    • la détection d'occurrences des uns ou plusieurs types de caractéristiques d'image prédéfinis présents dans la seconde image,
    • suivant une fonction d'énergie prédéfinie, la déformation du gabarit de la première image de façon à créer un gabarit déformé tel que les noeuds du gabarit soient attirés vers des occurrences de leurs types de caractéristiques associés respectifs alors que les liaisons de gabarit restreignent la déformation du gabarit, et
    • l'établissement d'une mesure de similarité entre le gabarit de la première image et la seconde image, suivant la différence d'état d'énergie associé au gabarit de la première image et au gabarit déformé.
  5. Procédé de reconnaissance d'image comprenant :
    • l'établissement d'une mesure de similarité entre un gabarit d'une première image et une seconde image selon la revendication 4, et
    • dans le cas où la mesure de similarité dépasse une valeur de seuil prédéfinie, en amenant la seconde image à être reconnue en tant qu'instance de la première image.
  6. Procédé de création d'un gabarit d'un ensemble d'images en vue d'une comparaison ultérieure avec une autre image comprenant les étapes consistant à :
    • (a) créer un gabarit de chaque ensemble d'images d'apprentissage conformément au procédé selon la revendication 1,
    • (b) créer un enregistrement de données correspondant à chaque tel gabarit, chaque enregistrement de données comportant une chaîne linéaire d'entrées, chaque entrée correspondant à l'un des noeuds des gabarits respectifs et représentant au moins un type de caractéristique de noeud associé et une adresse de noeud associée,
    • (c) créer un gabarit correspondant à chaque enregistrement de données, en utilisant chaque entrée pour créer un noeud de gabarit correspondant du type de caractéristique associé et situé à un emplacement du gabarit correspondant à l'adresse associée, et en associant des liaisons de gabarit à des noeuds de gabarit de façon à créer chaque gabarit,
    • (d) associer à chaque gabarit une mesure de qualité en ce qui concerne chaque image de l'ensemble d'images,
    • (e) associer à chaque gabarit une mesure d'adaptation en ce qui concerne l'ensemble d'images, suivant les mesures de qualité établies à l'étape (d),
    • (f) sélectionner un sous-ensemble des enregistrements de données suivant la mesure d'adaptation de leurs gabarits associés établie à l'étape (e), et
    • (g) appliquer des opérateurs génétiques au moins au sous-ensemble afin de créer de nouveaux enregistrements de données.
  7. Procédé selon la revendication 6, dans lequel les entrées des enregistrements de données correspondant aux noeuds associés à des occurrences des types de caractéristiques d'image localisés de façon rapprochée dans une image, sont localisées de façon rapprochée dans la chaîne linéaire correspondante.
  8. Dispositif de création de gabarit destiné à créer un gabarit d'une première image en vue d'une comparaison ultérieure à une seconde image, comprenant :
    • un moyen agencé pour détecter des occurrences d'un ou plusieurs types de caractéristiques d'image prédéfinis présents dans la première image,
    • un moyen agencé pour associer un emplacement à l'intérieur de l'image à chaque occurrence,
    • un moyen agencé pour associer chaque emplacement associé à une occurrence et un noeud du gabarit,
    • un moyen agencé pour associer, en fonction de chaque occurrence, le type de caractéristique d'occurrence au noeud associé à l'emplacement de l'occurrence, et
    • un moyen agencé pour associer des liaisons de gabarit à des noeuds de gabarit de façon à créer un gabarit de la première image.
  9. Dispositif de reconnaissance d'image comprenant :
    • un moyen agencé pour mémoriser un gabarit d'une première image, le gabarit étant créé selon l'une quelconque des revendications 1 à 3, comprenant la détection des occurrences d'un ou plusieurs types de caractéristiques d'image prédéfinis présents dans la première image,
    • un moyen agencé pour détecter des occurrences des un ou plusieurs types de caractéristiques d'image prédéfinis présents dans la seconde image,
    • un moyen agencé pour déformer le gabarit de la première image suivant une fonction d'énergie prédéfinie de façon à créer un gabarit déformé tel que les noeuds du gabarit soient attirés vers des occurrences de leurs types de caractéristiques associés respectifs alors que les liaisons de gabarit restreignent la déformation du gabarit, et
    • un moyen agencé pour établir une mesure de similarité entre le gabarit de la première image et de la seconde image suivant la différence de l'état d'énergie associé au gabarit de la première image et au gabarit déformé, et
    • un moyen agencé pour amener la seconde image à être reconnue comme une instance de la première image dans le cas où la mesure de similarité dépasse une valeur de seuil prédéfinie.
  10. Dispositif de création de gabarit destiné à créer un gabarit d'un ensemble d'images en vue d'une comparaison ultérieure à une autre image, comprenant :
    • un moyen agencé pour mémoriser des gabarits de chaque image d'un ensemble d'images d'apprentissage, chaque gabarit étant créé selon l'une quelconque des revendications 1 à 3, comprenant la détection des occurrences d'un ou plusieurs types de caractéristiques d'image prédéfinis présents dans la première image,
    • un moyen agencé pour créer un ensemble d'enregistrements de données, chaque enregistrement de données correspondant à un gabarit, chaque enregistrement de données comportant une chaîne linéaire d'entrées, chaque entrée correspondant à l'un des noeuds du gabarit respectif et représentant au moins un type de caractéristique de noeud associé et une adresse de noeud associée,
    • un moyen agencé pour créer un gabarit correspondant à chaque enregistrement de données, en utilisant chaque entrée pour créer un noeud de gabarit correspondant du type de caractéristique associé et situer à un emplacement du gabarit correspondant à l'adresse associée, et en associant des liaisons de gabarit aux noeuds du gabarit de façon à créer chaque gabarit,
    • un moyen agencé pour associer à un gabarit une mesure de qualité en ce qui concerne chaque image de l'ensemble d'images,
    • un moyen agencé pour associer à un gabarit une mesure d'adaptation en ce qui concerne l'ensemble d'images, suivant les mesures de qualité associées,
    • un moyen agencé pour sélectionner un sous-ensemble d'enregistrements de données suivant la mesure de qualité de leurs gabarits correspondants, et
    • un moyen agencé pour appliquer des opérateurs génétiques au moins au sous-ensemble sélectionné afin de créer un nouvel ensemble d'enregistrements de données.






IPC
A Täglicher Lebensbedarf
B Arbeitsverfahren; Transportieren
C Chemie; Hüttenwesen
D Textilien; Papier
E Bauwesen; Erdbohren; Bergbau
F Maschinenbau; Beleuchtung; Heizung; Waffen; Sprengen
G Physik
H Elektrotechnik

Anmelder
Datum

Patentrecherche

Patent Zeichnungen (PDF)

Copyright © 2008 Patent-De Alle Rechte vorbehalten. eMail: info@patent-de.com