DE102020210732A1 - Method for measuring the relevance of image areas for decisions of an image classifier - Google Patents

Abstract

Method (100) for measuring the extent to which a first subset A of the pixels of the specified input is used for the decision of a specified image classifier (1) on the assignment of an input image x to one or more classes (3a-3c) of a specified classification (3). -image x, and/or an intermediate product x' generated by the image classifier from the input image x, is more relevant than a second subset B of the pixels of the same input image x, or of the same intermediate product x', with the steps:• a large number of modifications x* of the input image x, or of the intermediate product x', are generated (110), these modifications x* being identical to the input image x, or with, at least for all pixels of the first subset A the intermediate product x', and are changed at least for all pixels of the second subset B compared to the input image x, or compared to the intermediate product x';• the modifications x* are assigned by the image classifier (1) to assignments (3* ) e.g u mapped (120) to one or more classes (3a-3c) of the specified classification (3); these determined assignments (3*) are combined (130) to form a measure Sf(A)B for how relevant the subset A\B of the pixels of the input image x, or of the intermediate product x', is for the assessment of the input image x by the image classifier (1).

Description

The present invention relates to the control of the behavior of trainable image classifiers, which can be used, for example, for the quality control of mass-produced products or also for the at least partially automated driving of vehicles

State of the art

In the series production of products, it is usually necessary to continuously check the quality of the production. The aim is to identify quality problems as quickly as possible in order to be able to remedy the cause as soon as possible and not lose too many units of the respective product as scrap.

Optical inspection of a product's geometry and/or surface is fast and non-destructive. the WO 2018/197 074 A1 discloses an inspection apparatus in which an object can be exposed to a plurality of lighting situations, images of the object being recorded with a camera in each of these lighting situations. The topography of the object is evaluated from these images.

Images of the product can also be directly assigned to one of several classes of a given classification using an image classifier based on artificial neural networks. On this basis, the product can be assigned to one of several specified quality classes. In the simplest case, this classification is binary ("OK"/"not OK").

In the case of at least partially automated driving of vehicles, image classifiers that can be trained are also used in order to evaluate traffic situations or at least to examine their content of objects.

Disclosure of Invention

As part of the invention, a method was developed for measuring the extent to which a first subset A of the pixels of the specified input image x, and/or of an intermediate product x' generated by the image classifier from the input image x, is more relevant than a second subset B of the pixels of the same input image x, or of the same intermediate product x'.

The input image x can be, for example, a camera image, a video image, a radar image, a lidar image and/or an ultrasound image. The intermediate product x' can be, for example, a feature map that generates a convolutional layer in a neural network of the image classifier. Such convolution layers can be used in an image classifier, in particular for the extraction of features from the input image x. There is a more or less pronounced spatial correspondence to the input image x. This means that, for example, a top right corner of a feature map provided by a convolutional layer depends more heavily on the top right corner of the input image x than on other parts of the image. This spatial correspondence is typically broken up in an image classifier when a feature extraction that has arisen with the participation of convolution layers is fed to a global pooling layer or a fully networked layer in order to arrive at a class assignment.

A large number of modifications x* of the input image x, or of the intermediate product x', are generated as part of the method. These modifications x* are identical to the input image x, or to the intermediate product x', at least for all pixels of the first subset A. However, at least for all pixels of the second subset B, the modifications x* are changed compared to the input image x, or compared to the intermediate product x′.

The modifications x* are mapped by the image classifier to assignments to one or more classes of the specified classification. These determined assignments are combined to form a measure S _f (A) _B of how relevant the subset A\B of the pixels of the input image x, or of the intermediate product x', is for the assessment of the input image x by the image classifier .

It was recognized that a high relevance of the first subset A for the decision of the image classifier has an effect on a large number of the modifications x* formed while retaining this first subset A.

If the content of the first subset A, for example, already speaks very strongly for a certain evaluation of the input image x by the image classifier, in the remaining area of the input image x, or the intermediate product x' generated from it, "come what may". , without affecting the decision of the image classifier. For example, during the quality control of mass-produced products, a certain visually recognizable defect or damage can be so serious that the product must definitely be classified as "not OK = NOK". Irrespective of what other features the input image x shows, these additional features can do not "undo" the defect or damage.

If, on the other hand, the image classifier is trained in such a way that the features characterized by the first subset A only lead to the classification of the input image x in a specific class in conjunction with other features, the class assignment can be changed if in the modifications x* these further Features are sufficiently obscured. Depending on how much the modifications x* change the input image x or the intermediate product x′, a larger or smaller proportion of the modifications x* is mapped by the image classifier to a changed class assignment. It is therefore not sufficient to merely fix the pixels in the first subset A and to change all other pixels. This could lead to the wrong conclusion that the pixels in the first subset A are not relevant.

• • Abwandlungen x*, in denen eine zufällige Auswahl von Pixeln des Eingabe-Bildes x, bzw. des Zwischenprodukts x', unverändert bleibt, einerseits und
• • Abwandlungen x*, in denen zusätzlich Pixel der ersten Teilmenge A unverändert bleiben, andererseits

ist im Mittel über alle Abwandlungen 0.If the class assignment by the image classifier does not significantly depend on the features characterized by the first subset A, it makes no difference on average over many modifications x* whether the image classifier sees the pixels in the first subset A unchanged or whether they have changed are. That is, the difference between the classification scores for

• • Modifications x*, in which a random selection of pixels of the input image x, or of the intermediate product x', remains unchanged, on the one hand and
• • Variants x*, in which additional pixels of the first subset A remain unchanged, on the other hand

is 0 on average over all modifications.

The fact that the pixels of the second subset B are changed in all modifications x* ensures that the importance of the first subset A is evaluated without any involvement of pixels in the second subset B.

The procedure is somewhat comparable to a representative "survey" in a "population" of variants x* with the question of whether the first subset A is more relevant than the second subset B. These variants x* can differ in particular, for example, in that to what extent pixels of the input image x or of the intermediate product x', which belong neither to the first subset A nor to the second subset B, are changed. The “poll” then specifically excludes “participants” who know from the outset that they are biased in relation to the question asked.

In a particularly advantageous embodiment, a large number of further modifications x** of the input image x or of the intermediate product x' are additionally generated. These modifications x** are identical to the input image x, or to the intermediate product x', at least for all pixels of the second subset B. At the same time, they are changed at least for all pixels of the first subset A compared to the input image x, or compared to the intermediate product x'. In comparison to the modifications x*, the roles of the two subsets A and B are reversed.

The modifications x** are mapped by the image classifier to assignments to one or more classes of the specified classification. These determined assignments are combined to form a measure S _f (B) _A of how relevant the subset B\A of the pixels of the input image x, or of the intermediate product x', is for the assessment of the input image x by the image classifier .

S _f (B) _A is therefore a measure of the extent to which the opposite hypothesis to the hypothesis to be tested is true, namely that the second subset B of the pixels of the input image x, or of the intermediate product x', is more relevant for the decision of the image classifier is as the first subset A. By comparing the measures S _f (A) _B and S _f (B) _A , and/or by taking a difference between these measures S _f (A) _B and S _f (B) _A , an overall assessment of how relevant the first subset A is compared to the second subset B can thus advantageously be obtained. This overall assessment is less sensitive to changes caused by different physical recording conditions under which input images x were taken. It is therefore better comparable between different images.

This is somewhat analogous to the fact that in signal processing, especially in the field of communications engineering and high-frequency technology, an absolute signal strength (usually in dB or dB related to a specific power unit) is often given by the additional specification of the noise background for a signal-to-noise distance is extended. This signal-to-noise ratio is decisive for establishing a successful connection.

Measuring in this way whether the first subset A is more relevant for the decision of the image classifier than the second subset B is significantly more sensitive than a mere visual assessment, for example of a "heat map" of the activations of certain features on a feature map. The activations of features in the subset A can be so slightly increased compared to the environment of this subset A that no significant optical contrast is visible, and at the same time this subset A can already prove to be very relevant for the decision of the image classifier with the method described here. This is particularly pronounced when the heat map has a few very high peaks that "outshine" the rest of the distribution.

The generation of at least one modification x*, x** can in particular include, for example, a corresponding pixel value for all pixels of the input image x or of the intermediate product x′ that do not belong to a set M of pixels not to be changed to add the interfering image P. A multiplicity of modifications x*, x** can then be generated which are based on the same interference image P and differ only in terms of which pixel values are changed. These modifications x*, x** are particularly easy to compare with one another, so that class assignments to which these modifications x*, x** are mapped can be made in a well-motivated manner to measure S _f (A) _B or S _f (B) _A can be merged.

If the modifications x*, x** are generated directly from the input image x, for example, then the interference image P can be chosen in particular, for example, such that the entropy of the modification x*, x**, and/or the associated received class assignment, is maximized. The interference image P then largely only has the effect that the modification x*, x** has reduced information compared to the input image x. By contrast, the entry of new information (such as gradient information) into the variant x*, x** through the use of the interference image P is reduced to a minimum.

If the modifications x*, x** are obtained from an intermediate product x', such as from a feature map, then an interference image P completely covered with uniform intensity, for example zero intensity, can advantageously be selected. In a feature map created by applying one or more convolution layers, pixel zeroing has the effect of removing certain detected features from the modifier x*, x** without, in return, introducing significant additional information into the modifier x*, x** coming. Alternatively, however, an interference image P can also be used to generate modifications x*, x** from an intermediate product x', such as a feature map, which has the entropy of the modifications x*, x**, and/or the associated received class assignments, maximized.

In a further advantageous embodiment, the generation of at least one modification x* includes the set M of pixels that are not to be changed compared to the input image x, or compared to the intermediate product x', as a union of the first subset A and one of the second subset B to form a disjoint configuration Z of further pixels and to change all pixels of the input image x, or of the intermediate product x', that do not belong to M. Analogously, the generation of at least one modification x** can include the set M of pixels that are not to be changed compared to the input image x or compared to the intermediate product x′ as a union of the second subset B and a configuration disjoint from the first subset A to form Z further pixels and to change all non-M pixels of the input image x, or of the intermediate product x'.

und $$S_{ƒ}{\left(B\right)}_A={\displaystyle \underset{Z}{\int }{ƒ}_c\left(\text{x}**\left(Z\right)\right)={\displaystyle \underset{Z}{\int }{ƒ}_c\left(p\left(x,m_{B\bigcup Z\backslash A}\right)\right)}}.$$

This formalism makes it possible to give closed formulas for S _t (A) _B and S _f (B) _A : $$S_{ƒ}{\left(A\right)}_B={\displaystyle \underset{Z}{\int }{ƒ}_c\left(\text{x}*\left(Z\right)\right)={\displaystyle \underset{Z}{\int }{ƒ}_c\left(p\left(x,m_{A\bigcup Z\backslash B}\right)\right)}}$$

and $$S_{ƒ}{\left(B\right)}_A={\displaystyle \underset{Z}{\int }{ƒ}_c\left(\text{x}**\left(Z\right)\right)={\displaystyle \underset{Z}{\int }{ƒ}_c\left(p\left(x,m_{B\bigcup Z\backslash A}\right)\right)}}.$$

sowie $$m_{B\bigcup Z\backslash A}=1\text{f}\ddot{\text{u}}\text{r}u\in B\bigcup Z\backslash A\mathrm{,0}\text{sonst}$$

eine binäre Maske. p(x, m) ist eine Kurzschreibweise für die Anwendung des Stör-Bildes P vermittelt durch diese binäre Maske m: $$p\left(\text{x},m\right)=\text{x}\cdot m+\text{P}\cdot \left(1-m\right).$$

Here, f _c are classification scores determined by the image classifier for variants x* and x**, respectively. These modifications x* and x** depend in turn on the respective configuration Z of the pixels that are not to be changed, which are combined with the first subset A and with the second subset B, respectively. Furthermore, for any pixel u $$m_{A\bigcup Z\backslash B}=1\text{f}\ddot{\text{and}}\text{right}\mathrm{and}\in A\bigcup Z\backslash B\mathrm{,0}\text{otherwise}$$

as $$m_{B\bigcup Z\backslash A}=1\text{f}\ddot{\text{and}}\text{right}\mathrm{and}\in B\bigcup Z\backslash A\mathrm{,0}\text{otherwise}$$

a binary mask. p(x, m) is a shorthand notation for the application of the perturbation image P mediated by this binary mask m: $$p\left(\text{x},m\right)=\text{x}\cdot m+\text{P}\cdot \left(1-m\right).$$

With this, S _f (A) _B and S _f (B) _A can in principle be calculated directly, with the number of possible configurations Z growing exponentially with the number of pixels in the input image x or in the variant x'. There are approximate solutions for applications in which the direct calculation would take too long.

Such an approximate solution consists in sampling a large number of configurations Z from a random distribution and averaging classification scores f _c , onto which the modifications x*, x** generated from this are mapped by the image classifier. This averaging is found, for example, in the formulas given above in the integration over the configurations Z.

• • zwei Kandidaten-Konfigurationen Z zu einer neuen Kandidaten-Konfiguration Z zu verrechnen (d.h. zu „kreuzen“),
• • Kandidaten-Konfigurationen Z in zufälliger Weise abzuwandeln und/oder
• • neue zufällige Kandidaten-Konfigurationen Z in den Pool einzuführen.

Alternatively or also in combination with this, new candidate configurations Z can be generated from a plurality of candidate configurations Z using a genetic algorithm. The goal can be pursued of achieving the best possible values with regard to at least one optimality criterion expressed in S _f (A) _B , S _f (B) _A and/or S _f (A) _B -S _f (B) _A . Starting from an existing pool of candidate configurations Z, a genetic algorithm can contain, for example,

• • to calculate (ie to “cross”) two candidate configurations Z to form a new candidate configuration Z,
• • Modify candidate configurations Z in a random manner and/or
• • introduce new random candidate configurations Z into the pool.

For example, an image of a mass-produced product taken as part of the quality control of that product can be selected as an input image. The classes of the classification can then represent a quality assessment of the product, for example "OK", "not OK = NOK" and optionally any intermediate levels.

There are then various possibilities for specifying the first subset A of the pixels that are not to be changed in the variants x*, x**. For example, this first subset A can be selected by observing the same product with a different imaging modality. If, for example, an ultrasonic image shows a contrast at a certain point of the product, this can be an indication that the corresponding point in an optical image of this product is also important.

However, it is also possible, for example, to determine subset A using an image area with a feature that is relevant to the quality assessment of the product. Such features can be, for example, optically visible defects or damage such as scratches or cracks. The procedure can then be used to examine whether the decision to classify the product as "not OK = NOK" was actually made on the basis of visible defects or damage to the product. The second subset B can then represent, for example, features that are guaranteed not to be relevant to the quality assessment, such as features in the image background or on a holder that holds the product in the camera.

For example, an image of a traffic situation taken from a vehicle can also be selected as the input image. The classes of the classification can then represent assessments of the traffic situation, on the basis of which the future behavior of the vehicle is planned. The assessment of the traffic situation can, for example, include an analysis of the objects visible in the input image. However, the assessment of the traffic situation can also include, for example, a classification of the traffic situation as a whole (e.g. “pedestrian enters the roadway”), a risk potential emanating from the traffic situation and/or a vehicle reaction that is appropriate in view of the traffic situation.

For example, at least one image area can then be identified as important for the assessment of the traffic situation and defined as the first subset A, independently of the image classifier. For example, by analyzing the optical flow between consecutive images or by additionally using an event-based optical sensor, it can be determined where dynamic events take place in the image, and the area of this dynamic event can be defined as the first subset A. Furthermore, for example, an image area can be identified as not being important for the assessment of the traffic situation and can be defined as the second subset B of the pixels to be changed in the modifications x*.

In principle, the subsets A and B can also be defined as partial areas of the input image x if the modifications x* and x** are not generated from the input image x, but from the intermediate product x'. Due to the spatial correspondence between the input image x and the intermediate product x', the specification at the level of the input image x can be translated to the level of the intermediate product x'.

In particular, the method can be fully or partially computer-implemented. The invention therefore also relates to a computer program with machine-readable instructions which, when executed on one or more computers, cause the computer or computers to carry out the method described. In this sense, control devices for vehicles and embedded systems for technical devices that are also able to execute machine-readable instructions are also to be regarded as computers.

The invention also relates to a machine-readable data carrier and/or a download product with the computer program. A downloadable product is a digital product that can be transmitted over a data network, i.e. can be downloaded by a user of the data network and that can be offered for sale in an online shop for immediate download, for example.

Furthermore, a computer can be equipped with the computer program, with the machine-readable data carrier or with the downloadable product.

Further measures improving the invention are presented in more detail below together with the description of the preferred exemplary embodiments of the invention with the aid of figures.

exemplary embodiments

• 1 Ausführungsbeispiel des Verfahrens 100;
• 2 Veranschaulichung

It shows:

• 1 embodiment of the method 100;
• 2 illustration

1 10 is a schematic flow chart of an embodiment of the method 100.

In step 105, for example, an image of a mass-produced product 11 can be selected as the input image x. There are then various reasonable possibilities to select a subset A of the pixels in the input image x, in order to then use the method 100 to test the hypothesis whether this subset A is more relevant than another subset B. According to block 105a, the product 11 can be observed with a different imaging modality, and from this observation the subset A can be determined. Alternatively or in combination with this, the subset A can also be determined according to block 105b using an image area obtained from another source with a feature 12 that is relevant for the quality assessment of the product 11 .

In step 106, for example, an image of a traffic situation taken from a vehicle can be selected as the input image x. According to block 106a, subset A can then be determined for testing the hypothesis as to whether this subset A is more relevant than another subset B, based on at least one image area that is relevant for assessing the traffic situation.

In step 110, a multiplicity of modifications x* of the input image x or of an intermediate product x′ generated therefrom in the image classifier 1 is generated. These modifications x* are identical to the input image x, or to the intermediate product x', at least for all pixels of the first subset A, and are identical to the input image x, or to the intermediate product, at least for all pixels of the second subset B Intermediate x', altered.

In step 120, the modifications x* are mapped by the image classifier 1 to assignments 3* to one or more classes 3a-3c of the specified classification 3. These assignments 3* are in turn combined in step 130 to form a measure S _f (A) _B of how relevant the subset A\B of the pixels of the input image x, or of the intermediate product x', is for the assessment of the input image x is 1 by the image classifier.

Analogously, in step 140, a large number of further modifications x** of the input image x, or of the intermediate product x', are generated. In comparison to the modifications x*, the roles of the subsets A and B are reversed. In step 150, the modifications x** are mapped by the image classifier to assignments 3** to one or more classes 3a-3c of the specified classification 3. These assignments 3** are in turn combined in step 160 into a measure S _f (B) _A for how relevant the subset B\A of the pixels of the input image x, or of the intermediate product x', is for the assessment of the input image image x by the image classifier 1.

Finally, in step 170, the two measures S _f (A) _B and S _f (B) _A are combined by comparison and/or by forming a difference to form an overall assessment 2 of how relevant the first subset A of the pixels of the input image x, or the intermediate product x', compared to the second subset B of the pixels of this input image x, or this intermediate product x'.

According to block 111 or 141, the generation of at least one modification x*, x** can include all pixels of the input image x or of the intermediate product x′ that do not belong to a set M of pixels not to be changed, respectively to add a pixel value of an interference image P corresponding thereto.

Alternatively or in combination with this, according to block 112 or 142, further pixels from a configuration Z can be added to a respective quantity M of pixels that are not to be changed, and according to block 113 or 143, in the modifications x* or x** all cannot pixels included in the set M are changed. Configurations Z can be sampled from a random distribution according to block 114 or 144, whereupon classification scores f _c can then be averaged according to block 131 or 161, onto which the modifications x*, x** generated from this are mapped by the image classifier 1 .

Configurations Z, for example candidates for random sampling, can be obtained according to block 115 or 145, for example with a genetic algorithm.

If the input image x is an image of a mass-produced product 11, according to block 121 or 151, the classes 3a-3c of the classification 3 can represent a quality assessment of the product 11. If the input image x shows a traffic situation, according to block 122 or 152, the classes 3a-3c can represent an assessment of the traffic situation.

2 illustrates the course of the method 100 using the example of an input image x from quality control. The input image x shows a brake disk 11 with a crack 12. The image classifier 1 is designed to classify the input image x into one of several quality classes 3a-3c (such as “OK” or “not OK=NOK”). in the in 2 In the example shown, the method 100 is used to test the hypothesis as to whether a subset A of the pixels in the image x, which shows the crack 12, is more relevant for the decision of the image classifier 1 than a subset B of the pixels from the image background.

The selected subsets A and B decide on the formation of the modifications x* and x**. These modifications x* and x** can each be created from the input image x or from an intermediate product x' formed during processing in the image classifier 1 and then instead of the input image x or instead of the intermediate product x' , be further processed in the image classifier 1.

The class assignments 3* formed from the variants x* and the class assignments 3** formed from the variants x** become dimensions S _f (A) _B or S _f (B) _A in steps 130 or 160 of the method 100 further processed. These measures S _f (A) _B and S _f (B) _A are then combined in step 170 of the method 100 for the overall assessment 2, how relevant the first subset A of the pixels of the input image x, or of the intermediate product x', im compared to the second subset B of the pixels of this input image x, or of this intermediate product x'.

The method 100 can thus be viewed somewhat analogously to an electrical measuring instrument whose test tips A and B can be placed at different locations in the input image x. This can then be used to measure how relevant one point in the input image x is compared to the other.

QUOTES INCLUDED IN DESCRIPTION

This list of documents cited by the applicant was generated automatically and is included solely for the better information of the reader. The list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions.

Patent Literature Cited

• WO 2018/197074 A1 [0003]

Claims (15)

Method (100) for measuring the extent to which a first subset A of the pixels of the specified input is used for the decision of a specified image classifier (1) on the assignment of an input image x to one or more classes (3a-3c) of a specified classification (3). -image x, and/or an intermediate product x' generated by the image classifier from the input image x, is more relevant than a second subset B of the pixels of the same input image x, or the same intermediate product x', with the steps: • a large number of modifications x* of the input image x, or of the intermediate product x', are generated (110), these modifications x* being identical to the input image x, or with, at least for all pixels of the first subset A are the intermediate product x' and are changed at least for all pixels of the second subset B compared to the input image x and compared to the intermediate product x'; • the modifications x* are mapped (120) by the image classifier (1) to assignments (3*) to one or more classes (3a-3c) of the specified classification (3); • These determined assignments (3*) are combined (130) to form a measure S _f (A) _B for how relevant the subset A\B of the pixels of the input image x, or of the intermediate product x', is for the assessment of the input image x by the image classifier (1). Method (100) according to claim 1 , where in addition • a large number of further modifications x** of the input image x or of the intermediate product x' is generated (140) which, at least for all pixels of the second subset B, are identical to the input image x or with are the intermediate product x' and are changed at least for all pixels of the first subset A compared to the input image x and compared to the intermediate product x'; • the modifications x** are mapped (150) by the image classifier to assignments (3**) to one or more classes (3a-3c) of the specified classification (3); and • these determined assignments (3**) are combined (160) to form a measure S _f (B) _A for how relevant the subset B\A of the pixels of the input image x, or of the intermediate product x', is for the Evaluation of the input image x by the image classifier (1). Method (100) according to claim 2 , where an overall assessment (2) of how relevant the first subset A of the pixels of the input image x, or of the intermediate product x', is compared to the second subset B of the pixels of this input image x, or of this intermediate product x' , is, from a comparison of the dimensions S _f (A) _B and S _f (B) _A , and/or as a difference between the dimensions S _f (A) _B and S _f (B) _A , is formed (170). Method (100) according to any one of Claims 1 until 3 , wherein the generating (110, 140) includes at least one modification x*, x**, one for each pixel of the input image x, or of the intermediate product x', which do not belong to a set M pixels not to be modified to add the corresponding pixel value of an interference image P (111, 141). Method (100) according to claim 4 , where an interference image P for generating modifications x*, x** from the input image x is chosen such that the entropy of the modification x*, x**, and/or a classification score f _c , is maximized, and/or wherein a spurious image P completely covered with uniform intensity is selected for generating modifications x*, x** from the intermediate product x'. Method (100) according to any one of Claims 1 until 5 , where • the generation (110) includes at least one modification x*, the set M of pixels not to be changed compared to the input image x or compared to the intermediate product x' as a union of the first subset A and one of the second subset B to form (112) a disjoint configuration Z of further pixels and to change (113) all pixels of the input image x, or of the intermediate product x', that do not belong to M; and/or • the generation (140) of at least one modification x** includes the set M of pixels that are not to be changed compared to the input image x, or compared to the intermediate product x', as a union of the second subset B and a to to form the first subset A disjoint configuration Z of further pixels (142) and to change (143) all pixels of the input image x, or of the intermediate product x', that do not belong to M. Method (100) according to claim 6 , where a large number of configurations Z are sampled from a random distribution (114, 144) and where classification scores f _c , onto which the variations x*, x** generated from this are mapped by the image classifier (1), are averaged (131, 161). Method (100) according to any one of Claims 6 until 7 , wherein new candidate configurations Z are generated from a plurality of candidate configurations Z using a genetic algorithm (115, 145) in accordance with at least one optimality criterion with regard to S _f (A) _B , S _f (B) _A and/or S _f (A ) _B -S _f (B) _A . Method (100) according to any one of Claims 1 until 8th , wherein an image of a mass-produced product (11) is chosen as the input image x (105) and wherein the classes (3a-3c) of the classification (3) represent (121, 151) a quality assessment of the product (11). Method (100) according to claim 9 , the first subset A of pixels not to be modified in the variants x*, x** being selected (105a) based on the observation of the same product (11) with a different imaging modality, and/or this first subset A based on an image area is determined (105b) with a feature (12) that is relevant for the quality assessment of the product (11). Method (100) according to any one of Claims 1 until 8th , wherein an image of a traffic situation taken from a vehicle is selected as the input image x (106) and the classes (3a-3c) of the classification (3) represent assessments of the traffic situation (122, 152), on the basis of which the future behavior of the vehicle is planned. Method (100) according to claim 11 , the subset A of the pixels not to be changed in the modifications x* being determined (106a) on the basis of at least one image region which is relevant for the assessment of the traffic situation. Computer program containing machine-readable instructions which, when executed on one or more computers, cause the computer or computers to perform the method (100) according to any one of Claims 1 until 12 to execute. Machine-readable data carrier with the computer program Claim 13 . Computer equipped with the computer program according to Claim 13 , and/or with the machine-readable data medium Claim 14 .

Images