EP1281157A1 - Method and device for determining an object in an image - Google Patents

Abstract

For determining an object in an image, hierarchical partial areas and sub-partial areas are selected, which are recorded with different resolution on each hierarchical level and which are compared with features of the object to be identified. If the object is identified with a sufficient level of certainty, the object to be identified is output as an identified object. If this is not the case, an additional sub-partial area of the current partial area is selected, and information with an, in turn, increased local resolution is detected from said sub-partial area.

Description

METHOD AND ARRANGEMENT FOR DETERMINING AN OBJECT IN AN IMAGE

The invention relates to a method for determining an object in an image and arrangements for determining an object in an image.

Such a method and such an arrangement are known from [1].

In the procedure known from [1], information is recorded in a partial area of the image from an image recorded by means of a camera, in which an object to be recognized is contained. A feature extraction is carried out for the recorded information and, using a known pattern recognition method, the extracted features from the partial area are compared with previously extracted features that describe the object to be recognized.

If the similarity between the extracted features from the partial area and the specified features that describe the object to be recognized is sufficiently great, the method is ended and the recognized object for which the extracted features have been formed is output as a recognized object.

The method is carried out iteratively for different sub-areas of the image until the object has been determined or until a predetermined termination criterion is fulfilled, for example a predetermined number of iterations or the object to be recognized is recognized with sufficient accuracy. A disadvantage of this procedure is in particular the very large computing time required to determine an object in the image to be examined. This is due in particular to the fact that all partial areas of the image are treated in the same way, that is to say the local resolution is the same for all partial areas of the image in the context of the method for object detection.

Furthermore, a so-called two-dimensional Gabor transformation is known from [2] as a wavelet transformation. The two-dimensional Gabor transformations are basic functions that use local spatial bandpass filters to achieve the theoretical optimal overall resolution in the spatial and frequency ranges, that is, in the one-dimensional spatial range and in the two-dimensional frequency range.

Further transformations are known from [3] and [4].

The invention is based on the problem of determining an object in an image, the determination being able to be carried out with a statistically lower computing time requirement. Furthermore, the invention is based on the problem of training an arrangement capable of learning in such a way that it can be used in the context of determining an object in an image, so that less computing time is required to determine the object in an image using the trained arrangement capable of learning than with the known procedure.

The problems are solved by the methods, the arrangements, the computer program element and the computer-readable storage medium with the features according to the independent patent claims. In a method for determining an object in an image, information is acquired from the image with a first local resolution. A first feature extraction is carried out for the recorded information. On the basis of the first feature extraction, at least one partial area in which the object could be located is selected from the image. Information with a second local resolution is also acquired from the selected partial area. The second local resolution is larger than the first local resolution. A second feature extraction is carried out for the information that has been acquired with the second local resolution, and a check is carried out to determine whether a predetermined criterion regarding the features extracted from the information by means of the second feature extraction is fulfilled. In the event that the predefined criterion is not met, information is iteratively recorded from at least one sub-region of the selected sub-region, in each case with a higher local resolution, and it is checked whether the information recorded with the respectively higher local resolution fulfills the predefined criterion for as long as until the specified criterion is met, or a further partial area is selected from the image and information from the further partial area is recorded with a second local resolution. Alternatively, the process can be ended.

In the context of digital image processing, the information can be brightness information and / or color information that is / are assigned to pixels of a digitized image.

The invention achieves considerable computing time savings in the context of the determination of an object in an image.

The invention is clearly based on the knowledge that in the context of the visual perception of a living being Probably a hierarchical approach to the perception of individual areas of different sizes with different local resolution usually leads to the goal of recognizing a searched object.

The invention is clearly to be seen in that, in order to determine an object in an image, hierarchical sub-areas and sub-sub-areas are selected, each of which is recorded with a different resolution on each hierarchical level and is compared with features of the object to be recognized after feature extraction has taken place. If the object is recognized with sufficient certainty, the object to be recognized is output as the recognized object. However, if this is not the case, there are alternatively the options available either to select a further sub-area of the current sub-area and to acquire information from this sub-area with a further increased local resolution, or to select a different sub-area and then in turn to this according to the object to be recognized investigate.

In a method for training an adaptable arrangement that can be used to determine an object in an image, an image that contains an object to be determined is captured. The position of the object to be recognized within the image and the object itself are predefined. Several feature extractions are carried out for the object, each with a different local resolution. The learnable

Arrangement is trained with the extracted features for a different local resolution.

That in the invention can be implemented both by means of a computer program, that is to say in software, and by means of a special electronic circuit, that is to say in hardware. Preferred developments of the invention result from the dependent claims.

The further refinements relate both to the methods, the arrangements, the computer-readable storage medium and the computer program element.

The test can be used as a predetermined criterion as to whether the information recorded with the respective local resolution is sufficient to determine the object with sufficient accuracy.

The predefined criterion can also be a predefined number of iterations, that is to say a predefined number of maximum iterations, in each of which a lower part area is selected and examined with an increased local resolution.

Furthermore, the predefined criterion can be a predefined number of subareas to be examined or maximum subareas to be examined.

The feature extraction can take place by means of a transformation with different local resolutions.

A wavelet transformation is preferably used as the transformation, preferably a two-dimensional Gabor transformation (2D Gabor transformation).

By using the two-dimensional Gabor transformation, the image information is encoded in an optimal manner both in the spatial area and in the spectral area, that is to say an optimal compromise is achieved in the context of the reduction of redundancy information between the local area coding and the frequency area coding. Any transformation that meets the following requirements in particular can be used as a transformation:

• The aspect ratio of the elliptical Gaussian envelope should be essentially 2: 1; • the plane wave should have its direction of propagation along the shorter axis of the elliptical Gaussian envelope;

• Furthermore, the half-amplitude bandwidth of the frequency response should have approximately 1 to 1.5 octaves along the optimal direction.

Furthermore, the mean value of the transformation should have the value zero in order to ensure a permissible functional basis for the wavelet transformation.

Alternatively, the transformations described in [3] and [4] can also be used.

The transformation can take place by means of a neural network or a plurality of neural networks, preferably by means of a recurrent neural network.

By using a neural network, a very fast transformation arrangement that can be adapted to the object to be recognized or to the correspondingly captured image information is used in particular.

In a further embodiment of the invention, a plurality of partial areas is determined in the image, with a probability being determined for each partial area that the corresponding partial area contains the object to be recognized. The iterative procedure is carried out for detail areas in the order according to the falling probability of belonging to the object to be determined accordingly. This procedure results in a further reduction in the computing time required, since an optimal procedure for determining the object to be recognized is specified from a statistical point of view.

To further reduce the required computing time, it is provided in a further development of the invention to adapt the shape of a selected partial area essentially to the shape of the object to be determined.

In this way, a subarea or a subarea is examined, which essentially corresponds to the object to be determined. This avoids examining an image area in which the object to be determined is certainly not located, since the corresponding image area then already has a different shape.

At least one neural network can be used as an arrangement capable of learning.

The neurons of the neural network are preferably arranged topographically.

An embodiment of the invention is shown in the figures and is explained in more detail below.

Show it

Figure 1 is a block diagram in which the architecture of the

Arrangement for determining the object according to an embodiment of the invention is shown;

FIG. 2 shows a block diagram in which the structure of the module for carrying out the two-dimensional Gabor

Transformation from Figure 1 according to the Embodiment of the invention is shown in detail;

FIG. 3 shows a block diagram in which the recognition module from FIG. 1 according to the exemplary embodiment is shown in detail;

Figure 4 is a block diagram showing the architecture of the

Arrangement for determining the object according to an embodiment of the invention is shown, the determination of a priority map being shown in detail;

FIGS. 5a and 5b show sketches of an image with different objects from which the object to be determined is to be determined, the different recorded objects being shown in FIG. 5a and the recognition result having been determined in FIG. 5b at different local resolutions;

Figure 6 is a flowchart showing the individual steps of the method according to the embodiment of the invention.

1 shows a sketch of an arrangement 100 with which the object to be determined is determined.

The arrangement 100 has a visual field 101.

Furthermore, a detection unit 102 is provided, with which information from the image can be detected via the visual field 101 with different local resolution.

The detection unit 102 has one

Feature extraction unit 103 and a recognition unit 104. FIG. 1 shows in the acquisition unit 102 a multiplicity of feature extraction units 103, each of which acquires information from the image with a different local resolution.

Features extracted from the captured image information are fed to the recognition module, that is to say the recognition unit 104, as feature vector 105 by the feature extraction unit 103.

In the recognition unit 104, which is explained in more detail below, a pattern comparison of the feature vector 105 with a previously formed feature vector is carried out in the manner explained in more detail below.

The recognition result is fed to a control unit 106, from which it is decided which sub-area or sub-area, as will be explained in more detail below, of the image is selected, and with which local resolution the respective sub-area or sub-area is examined. The control unit 106 also has a decision unit in which it is checked whether a predefined criterion with regard to the extracted features is met.

Arrows 107 symbolically indicate that the individual detection units 104 are "switched" to acquire information in different detection areas 108, 109, 110, each with a different local resolution, depending on control signals from the control unit 106.

The feature extraction unit 103 shown in detail in FIG. 2 is explained in more detail below. If the two-dimensional Gabor wavelets are set up in such a way that the frequency range is arranged in a logarithmic division, each recorded frequency is referred to as an octave. Each octave is referred to below as a local resolution.

Each unit that performs a wavelet transformation at a given local resolution has an arrangement of neurons whose detection range corresponds to a two-dimensional Gabor function and is dependent on a specific orientation.

The output of the corresponding neuron is also dependent on the given local resolution and is symmetrical. Each feature extraction unit 103 has a recurrent neural network 200, as shown in FIG. 2.

In the following, a digitized image 201 with n * n pixels is assumed (according to this exemplary embodiment, n = 128, that is to say according to the exemplary embodiment, the image has 16384 pixels).

Each pixel is a brightness value I? ^ ¹ ^ between "0 *

(black) and "255 ^λ (white) assigned.

The brightness value I ° f ^ig denotes the

Brightness value that is assigned to a pixel, which pixel is located at the local coordinates denoted by the indices i, j within the image 201.

The image 201, that is to say the pixels which lie in the respective detection area, becomes an average brightness value DC,

the brightness values I ° ^{-g of} the pixels of the image 201 which lie in the detection range and the average brightness value DC is determined by a

Contrast correction unit 202 is subtracted from the brightness values I? J ^{lg of} each pixel.

The result is a set of brightness values that are contrast invariant. The contrast-invariant description of the

Brightness values of the pixels in the detection area are formed in accordance with the following regulation:

The DC-free brightness values are fed to a neuron layer 203, the neurons of which carry out an extraction of simple features.

The neurons in the neuron layer 203 have receptive ones

Fields that perform a two-dimensional Gabor transformation in accordance with the following regulation.

c ^{os Θ +} y ^s i ⁿ ©) ² + (- x sin Θ + y cos θ) ² )

K ^*

₅ iωo (x cos Θ + y sin θ)

O:

being with • (Dg is an angular frequency in radians per unit length, and

• Θ the orientation direction of the wavelet is designated in radians.

The Gabor Wavelet is at

x = y = 0 (4)

o centered and normalized using an L standard such that:

(Ψ, Ψ) = 1. (5)

The frequency bandwidth is determined with the constant K.

According to this embodiment

K = π (6)

used, which corresponds to a frequency bandwidth of an octave.

A family of a discrete 2D Gabor wavelet G] pg_ (x, y) can be determined by discretizing the frequencies, orientations and centers of the continuous wavelet function (3) according to the following rule:

Gkpql / y) = a ^"k Ψ _Θl (a ^{~ k} x - pb, a ^{~ k} y - qb), (7)

With

Ψ _Θl = ψ (x cos (lΘ ₀ ) + y sin (lΘ ₀ ), - x sin (lΘo) + y cos (lΘ ₀ )) (8)

and the base wavelet:

Designated according to this regulation

Θg = - the step size of the respective angle rotation, L

1 the index of rotation according to the preferred orientation Θ] _ = -, k the respective octave, and p and q the position of the center of the receptive field (c _x pba and Cy = qba)

For a given octave k, the maximum values of p and q result according to

and

n

Q = dl! ba '

where | _xj is the largest integer number that is less than x.

The activation of a neuron in the neuron layer 203 is also referred to as rj qi.

The activation rj qi depends on a certain local frequency, which is dependent on the octave k for a preferred orientation, which is determined by the rotation index 1 and an excitation at the center, determined by the indices p and q. The activation ηcpqi of the neuron in the respective

Neuron layer 203 is defined as the convolution of the corresponding receptive field and the image, that is to say that

Brightness values of the pixels, which results in the activation rj qi of a neuron according to the following rule:

nn rpql = ( ^G kpql ' ^τ ) = ∑ ∑ ^G kpql (i' j) ^{• x} ij ^• 9ij # ⁽¹²⁾ i = lj = l

where with gj_j a weight value for the pixel (i, j) of the

Detection unit is designated with the corresponding local resolution k.

It should be noted that the activation rjφq] _ of a neuron is a complex number, which is why the exemplary embodiment uses two neurons for coding a brightness value Ij_j, one neuron for the real part of a brightness value Iij and one neuron for the imaginary part of the transformed brightness information Iij.

The neurons 206 of the neuron layer 205, which detect the transformed brightness signal 204, generate a neuron output value 207.

A reconstructed image 209 is formed in an image reconstruction unit 208 by means of the neuron output signal 207.

According to this embodiment, the

Image reconstruction unit 208 neurons that perform a Gabor wavelet transformation.

For this purpose, the image reconstruction unit 208 has neurons that operate according to a feed-forward structure connected to each other that correspond to a Gabor-receptive field.

In other words, this means that the image is reconstructed according to the following rule:

KPQ Ll! Ij = ^C Σ Σ Σ Σ ^r kpql ^G kpql (i, ^j ) # ⁽¹³⁾ k = 0p = 0q = 01 = 0

where K is the maximum resolution.

A constant C denotes the density of the wavelet base used. Due to the non-orthogonality of the Gabor-Wavelet basic functions, regulation (13) and its linear superposition do not guarantee that a minimum of a reconstruction error E, which is formed according to the following regulation:

is achieved.

A correction of this regulation (14) can be obtained by dynamically optimizing the reconstruction error E by means of a feedback connection.

Furthermore, a feedback correction term r ^or ₁ ^{r is formed} for each neuron 206 of the neuron layer 205.

The dynamics of the recurrent neural network 200 are determined in such a way that a dynamic reconstruction error is formed in accordance with the following regulation:

(15)

The dynamic reconstruction error of the recurrent neural network 200 is minimized.

This is achieved by dynamically adjusting the correction

"corr

Ter s ^{according to the} following regulation:

= corr ij ^E ijGkpqlC j) = η (G _pq ι, E),

(16)

in which

KPQ L-lj ^ E iD ij ^"C Σ Σ Σ Σ feql + ^r kpql pkpqlfc ^j ) Ü7) k = 0p = 0q = 01 = 0

and η denotes a change coefficient (according to the exemplary embodiment η = 0.1).

The constant C is formed according to the following rule:

where max () denotes the maximum value of the respective values.

This dynamic described above can be interpreted clearly in the following way. If the reconstruction error signal E is fed back and folded with the same Gabor-receptive fields Gkpql, EJ, then the entire dynamic system converges to an attractor that corresponds to the minimum of the reconstruction error signal 214.

The reconstruction error signal 214 is formed by means of a differential unit 210. The contrast-free brightness signal 211 and the reconstructed brightness signal 212 are fed to the differential unit 210. By forming the difference between the contrast-free brightness value 211 and the respective reconstructed brightness value 212, a reconstruction error value 213 is formed, which is fed to the receptive field, that is to say the Gabor filter.

In a learning phase, a training method according to regulation (16) is carried out for each object to be determined from a set of objects to be determined, that is to say to be recognized, and for each local resolution in the feature extraction unit 103 described above.

This is done by extracting the corresponding 2D Gabor wavelet features for each object at any local resolution.

The recognition unit 104 stores the extracted feature vectors 105 in their weights of the neurons individually for each local resolution.

Different feature extraction units 103 are thus trained in accordance with each local resolution for each object to be determined, as is indicated in FIG. 1 by the different feature extraction units 103. The positions of the centers of the receptive fields are discretized and result for a local resolution of the degree k

c _x = pba ^k (18)

and

c _y = qba ^k . (19)

This clearly means that spatially closer wavelets are separated by smaller steps and more distant wavelets by larger steps.

According to this exemplary embodiment, the receptive fields cover the entire detection area in the same way at every local resolution, that is to say they always overlap in the same way.

A feature extraction unit 103 thus has the local resolution k

Gabor neurons.

The Gabor neurons are clearly identified by means of the index kpql and the activation ^], ql, which, as described above, are given by the folding of the corresponding receptive field with the brightness values I j of the pixels of the detection area.

Due to the procedure described above, a feature extraction unit 103, which is preferably used, quickly becomes one through the forward-looking Gabor connections a sufficiently good set of wavelet basis functions for the greatly improved coding of the brightness values is determined, which is formed by the recurrent dynamic analysis of the reconstruction error value 213, so that a smaller number of iterations is achieved in order to determine the minimum of the reconstruction error value 213.

The feedback reconstruction error E is used according to the exemplary embodiment in order to dynamically improve the forward-facing Gabor representation of the image 201 in the sense that the problem of redundancy set out above in the description of the image information is dynamically corrected on account of the non-orthogonality of the Gabor wavelets ,

The redundancy of the Gabor feature description has therefore been dynamically reduced considerably by improving the reconstruction in accordance with the internal representation of the image information.

This structure therefore makes it non-linear

Correction of the usual linear representation of a Gabor filter is achieved, whereby a more efficient predictive coding of the image information is achieved.

The number of iterations required to achieve optimal predictive coding of the image information can be further reduced by using an over-complete number of Gabor neurons for the feature coding.

A base that is thus complete allows a larger number of base vectors than input signals. For a feature extraction unit 103 of the local resolution K, according to the exemplary embodiment for the reconstruction of the internal representation of the Gabor neurons with wavelet

Characteristics corresponding to the octave at least the number of number given by the local resolution K is used.

According to the exemplary embodiment, six octaves, that is to say six feature extraction units 103 (N = 6) with eight orientations (L = 8), with b = 1 and a = 2 are used, so that all degrees of resolution are used

coding Gabor neurons are used

Since 16,384 pixels are contained in the image according to the exemplary embodiment, 174,080 coding Gabor neurons are used to form the overcomplete base.

The neurons of the neuron layer 205 are explained in detail below (see FIG. 3).

It is assumed according to the exemplary embodiment that for each neuron 206 (a neuron 300 is provided for a real part and a neuron 301 for the imaginary part of the Gabor transformation, as explained above, that is to say two neurons for a “logical” neuron) with the corresponding connections to the feature extraction unit 103 in each case as weight information, which the description is stored by means of feature vectors of an object for a specific local resolution and a specific position of the object in the detection area.

The neurons 206 of the neuron layer 205 are arranged in columns so that the neurons are arranged topographically. The receptive fields of the recognition neurons are set up in such a way that only a limited square detection area of the neuron input values is transmitted around a certain center area.

The size of the quadratic receptive fields of the recognition neurons is constant and the recognition neurons are set up in such a way that only the signals from neurons 206 of the neuron layer 205 which are located within the detection range of the respective recognition neurons 301, 302 are taken into account.

As part of the training phase, the center of the receptive field is in the brightness center of the respective object.

Translation invariance is achieved in that for each object to be learned, that is to say to be recognized in the application phase, of identical recognition neurons, that is to say neurons that share the same weights but have different centers, are distributed over the entire coverage area.

Rotation invariance is achieved by storing the sum of the wavelet coefficients along the different orientations at each position.

In summary, according to the exemplary embodiment, a separate number of recognition neurons is provided for each new object to be learned during the learning phase, which store in their weights the corresponding wavelet-based internal description of the respective object, that is to say the feature vectors that describe the objects.

For each local resolution, a recognition neuron is generated which corresponds to the respective internal description in accordance with the corresponding octave, that is to say the corresponding local one Resolution corresponds and the respective recognition neuron for all center positions is distributed in the entire detection area.

The recognition neurons are linear neurons, which output a linear correlation coefficient between its input weights and the input signal, which are formed by the neurons 206 of the neuron layer, which are located in the feature extraction unit 103.

3 shows the respective recognition neurons 305, 306, 307, 308, 309, 310, 311, 312 for different objects 303, 304. During the training phase, each object is clearly provided at a time in the detection area at a predetermined, freely definable position ,

The recognition neurons store the wavelet-based information in their weights. For a given PPoossiittion, that is, a center with the pixel coordinates \ Cχ _f Cyj, two recognition neurons are provided for each object to be learned, one for storing the real part of the wavelet description and one for storing the imaginary part of the internal wavelet description.

The internal description of the neurons after

Convergence of the recurrent dynamics as described above is stored according to the following two tensors:

^w kpq = + ^r k (p + ^r c _χ ) (q + cy) l (2i;

and

where Re () denotes the real part and Im () denotes the imaginary part and applies to the indices p and q:

p, q e [- R, R], (23)

where R is the width of the receptive field in recorded pixels.

According to the exemplary embodiment, R = 32 pixels is selected.

During the training phase, the center (c _x , Cyj is formed by the center of brightness of the respective object, which is given according to:

n c _x = (24) n

and

By forming the sum over all indices 1, a rotation-invariant description of the corresponding object is achieved.

Neurons that are activated due to excitation in another center are formed in the same way, with the same weights for recognizing the same object be used in a shifted position within the detection range.

The output of a recognition neuron in the context of the recognition phase is given by one

Correlation coefficient that describes the correlation between the weights and the output of the neurons 206 of the neuron layer 205.

According to the exemplary embodiment, the output of a recognition neuron in the recognition unit 104 at a local resolution k, based on the real parts of the neurons 206 of the neuron layer 205). the local resolution k and related to the center \ z _x ,, Zzy _y ]) given by:

The output of the corresponding recognition neuron for the imaginary part is given by:

Λ ^z χ / ^Z y] _

(A) denotes the mean value and σ _a the standard deviation of a variable a over the detection range, i.e. over all indices p, q.

It should be noted that the neurons are activated at every local resolution depending on the detection of the same object but also on the different positions, since the same weights are stored for different positions according to the object. According to the exemplary embodiment, the centers of the recognition neurons are arranged over the detection area in such a way that they completely cover the detection area and one neuron in each case overlaps half with the detection area of another neuron, that is to say for n = 128 and R = 64, nine centers become at the following positions arranged ((32, 32) (32, 64) (32, 96) (64, 32) (64, 64) (64, 96) (96, 32) (96, 64) (96, 96)).

During the detection phase, the different detection units 104 are thus activated serially by the control unit 106, as will be described below.

After the activation of the corresponding recognition unit 104, a check is carried out to determine whether a predetermined criterion is met or not, the activation of the recognition neurons with the greatest activation being determined in accordance with the octave, which is greater than or equal to the current octave, that is to say by taking into account only the activated ones Detection units 104 at the appropriate time.

In other words, a so-called winner-takes-all strategy is used in deciding which recognition neuron is selected in such a way that the selected recognition neuron, which is assigned to a specific center and a specific object, is analyzed by the control unit 106.

As will be explained further below, the control unit 106 can further decide whether the identification of the corresponding object is sufficiently precise or whether a more precise analysis of the object by selecting a smaller, more detailed area with a higher local resolution is required. If this is the case, then further neurons are activated in the further feature extraction units 103 or recognition units 104, so that the local resolution is increased.

As shown in FIG. 1, a priority map is formed by the recognition unit 104 for the detection area with the coarsest local resolution, individual priority areas of the image area being indicated by the priority map and the probability being assigned to the corresponding area areas, indicating how likely it is is that the object to be recognized is in the partial area (see FIG.).

The priority card is symbolized with 400 in FIG. A partial area 401 is characterized by a center 402 of the partial area 401.

The individual iterations in which different sub-areas and sub-areas are selected and examined with higher local resolution are explained in more detail below.

According to the exemplary embodiment, a serial feedback mechanism is provided for masking the detection areas, whereby successive others

Detection units 102 and feature extraction units 103 and detection units 104 are activated in accordance with the respectively selected increased resolution k, that is to say the control unit 106 regulates the positioning and size of the detection area in which visual information is received by the system and processed further.

In a first step, the entire image 201 is processed, but with the coarsest local resolution, that is to say only the first recognition unit and feature extraction unit with k = N is activated. With this rough local resolution, usually only the position of the object is practically recognizable and a very rough determination of the global shape of an object is determined.

Depending on the respective task, the control unit stores the result of the recognition unit as a priority map and selects a partial area of the image in which, as will be described below, image information is examined.

The corresponding selection of the partial area is fed back through the same feedback connections through the activated wavelet module.

The selection of the sub-area, that is to say the specification of which pixels are examined in greater detail with increased local resolution, is dependent on the pixels which describe the object of the last activated local resolution.

The corresponding pixels are selected on the basis of the pixels which enable a good reconstruction, that is to say a reconstruction with a small reconstruction error, and by pixels which do not correspond to a filtered black background.

In other words, the attention mechanism is object-based in the sense that only the areas in which the object lies are further analyzed in series with a higher local resolution.

This means that the corresponding lower octaves are activated serially, but only in the selected section.

The attention mechanism is described mathematically using a matrix G ^ j, the elements of which have the value "1 * if the corresponding pixels are to be taken into account and have the value "0" if the corresponding pixels are not to be taken into account.

At the grossest local dissolution within the framework of the

Object detection (k = N) the entire image 201 is analyzed, that is

gij = 1 Vi, j (28;

The priority map is generated and the control unit 106 decides which object is to be analyzed in more detail in a further step, so that only the pixels which lie in the image area, that is to say in the selected partial area, are taken into account in the context of the next higher local resolution.

According to the exemplary embodiment, two further conditions are assumed.

The first condition is that the reconstructed image has brightness values I j> 0 and the second condition is that the reconstruction error is not greater than a predetermined threshold, that is to say:

g jE j <α. (29)

The control unit 106 thus decides that the object is analyzed in more detail at a center (c _x , Cy) in the priority map, then the mask, given by the matrix Gij, is updated in accordance with the following regulations:

In general, the attention feedback between the local resolution k and the subsequent local resolution k - 1 (i.e. the increased local

Attention) for k> N is only regulated by the two conditions mentioned above.

A new matrix value G j is therefore defined in accordance with the exemplary embodiment for the activation of the next, increased local resolution k-1 in accordance with the following regulation:

The course of the different iterations of the examination of the individual sub-areas and sub-areas with different local resolutions is described for a concrete object recognition.

In this example, four types of objects are provided, as shown in Fig. 5a.

A first object 501 has a global form of an H and has object components of the form T as local elements, which is why the first object is called Ht.

The second object 502 has a global H-shape and also H-shaped components as local object components, which is why the second object 502 is referred to as Hh. A third object 503 has a global and also a local T-shaped structure, which is why the third object 503 is referred to as Tt.

A fourth object 504 has a global T-shape and a local H-shape of the individual object components, which is why the fourth object 504 is referred to as Th.

5b shows the recognition results of a device according to the invention for different local resolutions, in each case for the first object 501 (recognized object at first local resolution 510, at second local resolution 511, at third local resolution 512, at fourth local resolution 513).

5b also shows the recognition results of a device according to the invention for different local resolutions, in each case for the second object 502 (recognized object at first local resolution 520, at second local resolution 521, at third local resolution 512, at fourth local resolution 523).

5b also shows the recognition results of a device according to the invention for different local resolutions, in each case for the third object 503 (recognized object at first local resolution 530, at second local resolution 531, at third local resolution 532, at fourth local resolution 533).

5b also shows the recognition results of a device according to the invention for different local resolutions, in each case for the fourth object 504 (recognized object with first local resolution 540, with second local resolution 541, with third local resolution 542, with fourth local resolution 543). As can be seen from FIG. 5b, the respective object is already recognized with a very good, at least sufficient accuracy, at the highest local resolution.

The method for determining an object in an image is clearly explained again with reference to FIG.

In a first step (step 601) the for the pixels, that is, for the brightness values

Pixels, a feature extraction with a first local resolution j-1 is carried out on the captured image (step 602).

In a further step, a first partial area Tbi is formed from the image (step 603).

For each sub-area Tbi that is formed, a probability is determined that the object to be determined is in the corresponding sub-area Tbi. The result is a priority map that contains the respective assignments probability and partial area (step 604).

According to the priority card formed, a first one

Partial area Tbi with i = 1 is selected and the neurons are activated, so that the selected partial area is incremented by the value 1 in step 605, so that the selected partial area Tbi is examined with an increased local resolution (steps 606, 607).

In a test step 608, it is checked whether the object has been recognized with sufficient certainty (step 608).

If this is the case, the recognized object is output as a recognized object (step 609). If this is not the case, then in a further test step (step 610) it is checked whether a predetermined termination criterion has been met, according to the exemplary embodiment, whether a predetermined number of iterations has been reached.

If this is the case, the method is ended (step 611).

If this is not the case, it is checked in a further test step (step 612) whether a further lower part area should be selected.

If a further subsection area that is to be examined with an increased resolution is to be selected, this corresponding subsection area is selected

(Step 613) and the method continues in step 606 by incrementing the local resolution for the corresponding sub-area.

However, if this is not the case, a further partial area Tbi + 1 is selected from the priority map (step 614), and the method is continued in a further step (step 605).

The following publications are cited in this document:

[1] A. Treisman, Perceptual Grouping and Attention in Visual Search for Features and for Objects, Journal of Experimental Psychology: Human Perception and Performance, Vol. 8, pp. 194-214, 1982

[2] J. Daugman, Complete Discrete 2D-Gabor-Transforms by Neural Networks for Image Analysis and Compression, IEEE-Transactions on Acoustics, Speed and Signal Processing, Vol. 36, pp. 1169-1179, 1988

[3] D.J. Heeger, Nonlinear Model of Neural Responses in Cat Visual Cortex, Computational Models of Visual Processing, Edited by M. Landy and J.A. Movshon, Cambridge, MA, MIT Press, pp. 119-133, 1991

[4] D.J. Heeger, Normalization of Cell Responses in Cat

Striate Cortex, Visual Neuro Science, Vol. 9, pp. 181-197, 1992

Claims

claims 1. Method for determining an object in an image, In which information from the image is recorded with a first local resolution, In which a first feature extraction is carried out for the information from the image, In which at least a partial area in which the object could be located is selected from the image on the basis of the feature extraction, • in which information from the selected partial area is acquired with a second local resolution, the second local resolution being greater than the first local resolution, • in that for the information from the selected one Sub-area a second feature extraction is performed, • in which a check is carried out as to whether a predetermined criterion is fulfilled, • in which the method is ended or a further partial area is selected from the image and information from the further partial area is recorded with a second local resolution if the predetermined criterion is not fulfilled , • where iterative information from at least one The subsection of the selected subsection is recorded with a higher local resolution and a check is carried out to determine whether the information recorded with the respectively higher local resolution fulfills the specified criterion until the specified criterion is met. 2. The method of claim 1, wherein the criterion is whether the information captured with the second local resolution is sufficient to capture the information with sufficient accuracy 3. The method of claim 1, wherein the criterion is a predetermined number of iterations. 4. The method according to any one of claims 1 to 3, wherein the feature extractions are carried out by means of a transformation with different local resolutions. 5. The method according to claim 4, in which a wavelet transformation is used as the transformation. 6. The method according to claim 5, in which a two-dimensional Gabor transformation is used as the wavelet transformation. 7. The method according to any one of claims 4 to 6, wherein the transformation is carried out by means of a neural network. 8. The method according to claim 7, wherein the transformation is carried out by means of a recurrent neural network. 9. The method according to any one of claims 1 to 8, • in which a plurality of sub-areas are determined in the image, each of which contains the object to be recognized with a determined probability, • in which the iterative method is carried out for the sub-areas in the order in which the probability falls. 10. The method according to any one of claims 1 to 9, wherein the shape of a selected portion corresponds substantially to the shape of the object to be recognized. 11. A method for training an arrangement capable of learning, which is to be used for determining an object in an image, in which an image which contains an object to be recognized is captured, the position of the object to be recognized in the image and the object are given In which several feature extractions are carried out for the object, each with a different local resolution, • in which the arrangement with the extracted features is trained for a local resolution. 12. The method according to claim 11, in which at least one neural network is used as the arrangement. 13. The method according to claim 12, wherein the neurons of the neural network are arranged topographically. 14. Arrangement for determining an object in an image, with a processor which is set up in such a way that the following method steps can be carried out: A first feature extraction is carried out for the information from the image, • from the image at least a partial area in which the object could be located due to the Feature extraction selected, Information from the selected partial area is recorded with a second local resolution, the second local resolution being greater than the first local resolution, A second feature extraction is carried out for the information from the selected partial area, • it is checked whether a specified criterion is met, • The method is ended or a further partial area is selected from the image and information is extracted from the further partial area with a second local one Resolution recorded if the specified criterion is not met, • It is iteratively recorded information from at least one sub-area of the selected sub-area, each with a higher local resolution, and a check is carried out to determine whether the information recorded with the respective higher local resolution fulfills the specified criterion until the specified criterion is met. 15. Arrangement for determining an object in an image, with A capturing unit for capturing information from the image with several different local resolutions, a feature extraction unit for extracting features for the information captured by the capturing unit, A selection unit for selecting at least a partial area from the image in which the object could be located on the basis of the features extracted by the feature extraction unit, A control unit for controlling the detection unit, which control unit is set up in such a way that information from the selected partial area is detected with a second local resolution, the second local resolution being greater than the first local resolution, A decision unit in which it is checked whether a specified criterion with regard to the extracted features is met, • The control unit is further set up in such a way that the method is ended or a further partial area is selected from the image and information is acquired from the further partial area with a second local resolution if the predetermined criterion is not met, - Iteratively, information from at least one sub-area of the selected sub-area is recorded in each case with a higher local resolution and that it is checked whether the information recorded with the respectively higher local resolution fulfills the predefined criterion until the predefined criterion is met. 16. Computer-readable storage medium in which a computer program for determining an object is stored in an image, which, when executed by a processor, has the following method steps: Information from the image with a first local resolution is recorded, a first feature extraction is carried out for the information from the image, At least a partial area in which the object could be located is selected from the image on the basis of the feature extraction, information from the selected partial area is recorded with a second local resolution, the second local resolution being greater than the first local resolution, A second feature extraction is carried out for the information from the selected partial area, • it is checked whether a specified criterion is met, • The method is ended or a further partial area is selected from the image and information is extracted from the further partial area with a second local one Resolution recorded if the specified criterion is not met, • It is iteratively recorded information from at least one sub-area of the selected sub-area, each with a higher local resolution, and a check is carried out to determine whether the information recorded with the respective higher local resolution fulfills the specified criterion until the specified criterion is met. 17. Computer program element for determining an object in an image, which, when executed by a processor, has the following method steps: Information from the image with a first local resolution is recorded, A first feature extraction is carried out for the information from the image, At least a partial area in which the object could be located is selected from the image on the basis of the feature extraction, Information from the selected partial area is recorded with a second local resolution, the second local resolution being greater than the first local resolution, A second feature extraction is carried out for the information from the selected subarea, it is checked whether a predetermined criterion is fulfilled, The method is ended or a further partial area is selected from the image and information is recorded from the further partial area with a second local resolution if the predetermined criterion is not met, • It is iteratively recorded information from at least one sub-area of the selected sub-area, each with a higher local resolution, and a check is carried out to determine whether the one with the higher local Resolution captured information the specified criterion fulfilled until the specified criterion is met.