Determining image shares relevant for image classifier decisions with binary masks

Abstract

Method for measuring a share (2a) of an input image, wherein an image classifier (1) makes its own decision about the assignment of the input image (2) to one or more classes of a predefined classification depending on said share, having the steps of providing a binary mask (4) which specifies which pixels of the input image and / or of an intermediate product (3) formed in the image classifier are considered relevant, evaluating said binary mask with a quality function (5), wherein the quality function is a measure for how strongly at least one classification score provided by the image classifier about at least one target class changes if the pixels of the input image or of the intermediate product which are relevant depending on the binary mask are changed, determining the sought share (2a) of the input image which is relevant for the decision of the image classifier in dependence on a combination of the binary mask and the respective evaluation thereof by the quality function.

Description

Technical Field

[0001] This invention relates to examining the behavior of trainable image classifiers, which can be used, for example, for quality inspection of mass-produced products or for at least partially automating driving vehicles. Background Technology

[0002] In mass production, continuous quality inspection is typically required. The goal is to identify quality problems as quickly as possible to eliminate the root cause and minimize the loss of excessive units of the corresponding defective products.

[0003] Optical inspection of the geometry and / or surface of a product is rapid and non-destructive. WO 2018 / 197 074A1 discloses an inspection apparatus in which an object is subjected to a variety of lighting conditions, wherein an image of the object is recorded using a camera under each of these lighting conditions. The topographie of the object is evaluated from these images.

[0004] Product images can also be directly assigned to one of several pre-defined categories (Klasse) using an image classifier based on an artificial neural network. Based on this, the product can be assigned to one of several pre-defined quality categories. In the simplest case, this classification is binary ("OK" / "Not OK").

[0005] When driving vehicles in at least a partially automated manner, a trainable image classifier is also used to evaluate traffic conditions or at least study the amount of objects they contain (Gehalt). Summary of the Invention

[0006] Within the scope of this invention, a method has been developed for measuring the following share of an input image, wherein an image classifier makes its own allocation decisions regarding the input image based on the share.

[0007] The method begins by providing binary masks that specify which pixels of the input image and / or the intermediate products formed in the image classifier are considered relevant. These binary masks can be drawn, for example, from a random distribution.

[0008] Binary masks are evaluated using a quality function that measures how strongly a classification score (Klassifikations-Score) for at least one target category is changed if the relevant pixels of the input image or intermediate product are changed according to the corresponding binary mask.

[0009] The desired share of the input image relevant to the decision of the image classifier is determined based on the combination of the binary mask and the corresponding evaluation of the binary mask through the quality function. If the binary mask affects the intermediate product of the image classifier and thus causes a combination of multiple such masks and their evaluations to primarily affect the decision-related share of the intermediate product, then this share can be transformed, for example, through interpolation or other upsampling, into the desired decision-related share of the input image.

[0010] Evaluation using a quality function allows for the meaningful aggregation of, in particular, multiple randomly selected binary masks into a desired share relevant to the decision of the image classifier for the input image or intermediate product. Therefore, the sum of binary masks, weighted individually using the evaluation of the binary masks by the quality function, can be determined as the decision-related share for the input image or intermediate product. This means that even binary masks with lower or poor evaluations by the quality function are not necessarily discarded, but rather all masks studied in the overall dataset can be included in the final result. Finally, the aggregation of several randomly selected binary masks also results in a further averaging of the random component within the masks.

[0011] Furthermore, this can be considered more precisely through the specific design of the quality function: the same change in a pixel can have very different effects depending on the circumstances. If, for example, a pixel value is set to zero or another constant value, it is not guaranteed that this merely removes information from the input image or intermediate product. On the contrary, the opposite effect may occur, i.e., new information may be added. Thus, for example, setting a pixel to zero in the input image can add additional gradient information to the input image. If a particular pixel in the intermediate product represents the characteristic that a particular image region has a low intensity value, then similarly, setting it to zero does not remove any information from the intermediate product. Instead, the information is merely altered such that the image region in question no longer has a low intensity value but has a high intensity value.

[0012] Precise consideration of the aforementioned effects results in the final outcome being the ability to draw conclusions about the decision-related share of the input image based on a relatively small number of binary masks. Therefore, these conclusions can be obtained quickly overall.

[0013] Furthermore, the evaluation of the binary mask using the quality function does not mandate that the input image or intermediate product be tested with a specific pre-given interference in the pixels determined by the mask. As shown below, the sensitivity of the classification score to changes in the input image or intermediate product can also be determined by calculus.

[0014] In a particularly advantageous design, an intermediate product is chosen, which is generated from the input image through processing in one or more convolutional layers. Such an intermediate product has a significantly reduced dimension relative to the input image. Therefore, the search space for binary masks with respect to this intermediate product is also significantly smaller than the search space for binary masks with respect to the input image. Correspondingly, only a smaller number of binary masks are needed to obtain the desired conclusions regarding the decision-related regions. The intermediate product may, for example, comprise a large number of feature maps, which have been generated, respectively, by applying filter kernels to the input image or to the intermediate product previously generated from the input image.

[0015] It is particularly advantageous to choose an intermediate product that, in the image classifier, is mapped from the classifier layer to at least one classification score. This intermediate product is therefore not further convolved before being fed to the classifier layer. The classifier layer can be, in particular, a fully connected layer. If the intermediate product being examined is indirectly fed to the classifier layer, the effect of changes in the intermediate product can be optimally and quantitatively examined and incorporated into the evaluation performed by the quality function. Furthermore, the processing path that the intermediate product changes must traverse within the image classifier is thus the shortest. Conversely, if the selected intermediate product is first fed to other convolutional layers and then to the classifier layer, the spatial resolution of such an intermediate product is better.

[0016] The terms "convolutional layer" and "classifier layer" can specifically refer to layers, for example, in which a neural network used as an image classifier is divided into "convolutional layers" and "classifier layers." However, other types of image classifiers can also be organized using such layers.

[0017] In a particularly advantageous design, the strong correlation between the at least one classification score and the pixels related to the input image or intermediate product in terms of the binary mask affects the value of the quality function in an improved manner. This means that changes in the pixels are particularly suitable for changing the classification of the input image from the category originally assigned by the image classifier to one or more other categories. For example, if a defect or damage to the product becomes unrecognizable due to a change in the input image, then an image of a product already classified as "NOK" based on its defect or damage, within the scope of quality inspection, might be classified as "OK".

[0018] Advantageously, the strong correlation between the at least one classification score and pixels that are unrelated to the input image or intermediate product according to the binary mask affects the value of the quality function in a degraded manner. The correlation between the classification score and these pixels can be considered as a kind of "background". The greater the change in the classification score by regions that are unrelated to the binary mask (in an improved or degraded manner), the greater the change in the classification score by regions that are related to the binary mask, i.e., "in a diluted manner". Therefore, the correlation between the classification score and pixels that are unrelated to the binary mask affects the value of the quality function in a degraded manner. The value of the quality function thus yields a characteristic of the signal-background spacing.

[0019] Similarly, at least one derivative and / or gradient of the classification score of at least one pixel in the input image or intermediate product that is relevant according to the binary mask can advantageously affect the value of the quality function in an improving manner. This is unrelated to whether the change in the relevant pixel causes an improvement or a deterioration in the quality function. The important thing is simply that the region relevant according to the binary mask changes the strength of the classification score (in a better or worse way). Therefore, this strength should be evaluated as "improved" by the quality function. Here, the gradient and / or derivative of the classification score of at least one pixel in the input image or intermediate product that is unrelated according to the binary mask can then affect the value of the quality function in a deteriorating manner. If the intermediate product being examined consists of multiple feature maps, then the contributions of multiple feature maps, especially for example, in forming the derivative or gradient, can be summed.

[0020] Suppose L is, for example, an image classifier generated from an input image I and consisting of multiple feature maps L. j The intermediate products are composed of [variable name]. Assume m is a binary mask, where a pixel value of 1 in the binary mask means that the corresponding pixel of the intermediate product L is relevant, and a pixel value of 0 means that the corresponding pixel of the intermediate product L is unrelevant. Assume f is a classification score provided by an image classifier. An exemplary quality function R [function name]. I,f (m) is therefore written as:

[0021] .

[0022] This footnote passes through all feature maps L j Valid. The greater the gradient intensity in the region of the intermediate product L relevant to the binary mask m, and the smaller the gradient intensity in the region of the intermediate product L unrelated to the binary mask m, the stronger the mass function R. I,f The higher the value of (m), the better.

[0023] In the mass function R I,f In (m), optionally, the difference in the external summation is cut off at 0 if it is otherwise negative. According to L jThe gradient can be a simple gradient, but it can also be, for example, a processed product of the gradient obtained by averaging or smoothing multiple easily induced versions of the intermediate product L. Alternatively, only such contributions to the inner summation can be selected, with the corresponding gradient having a uniform sign (positive or negative) for said contributions.

[0024] Furthermore, the feature map L j The contribution can be summarized into the mass function R not only through addition but also in other ways. I,f The value of (m). Therefore, the average of these contributions or the maximum of these contributions can be used, for example. The reason for the latter is that the output of the single filter is already able to show strong recognition of the decision-related features, so that the mask thus earns through the quality function R. I,f (m) gave a good evaluation.

[0025] Advantageously, in response to the absolute value of at least one gradient being below a pre-given threshold, the at least one gradient is left unconsidered. In this way, it can be considered that the gradient actually measures an infinitesimal change in the classification score, thus allowing only a very imprecise determination of direction based on gradients with small values.

[0026] Advantageously, at least one gradient and / or at least one contribution of a pixel to such gradient is discretized in such a way that the gradient and / or the contribution is set to a first value in response to itself being above a pre-given threshold, and the gradient and / or the contribution is set to a second value in response to itself being below a negative pre-given threshold. In this way, only the direction is extracted from the gradient, while the potentially imprecise magnitude of the gradient is discarded.

[0027] Therefore, for any pixel λ, for example through the discretized version G, it is possible to... L (λ) is used to replace the gradient (d(f) / dL)(λ) in other processing chains, wherein the discretized version is derived from this:

[0028] if ,but

[0029] if ,but

[0030] if ,but , where t is the threshold.

[0031] In another advantageous design, if the input image or intermediate product is modified in terms of the relevant pixel aspects according to the binary mask as follows:

[0032] If a pixel's contribution to the gradient of a classification score based on related pixels exceeds a pre-given threshold, the pixel is changed to a first pre-given alternative value L. max ;and

[0033] If a pixel's contribution to the gradient of a classification score based on the relevant pixel is below a pre-given threshold of negativity, the pixel is changed to a second pre-given alternative value L. min ,

[0034] The quality function then includes at least one classification score, wherein the image classifier provides the classification score.

[0035] Thus, on the one hand, the gradient is discretized, and on the other hand, the perturbation applied to the relevant region for the purpose of evaluating the quality function is related to the gradient direction. The alternative value L used for this perturbation... max and L min It can be adapted to (zuschneiden auf) corresponding applications.

[0036] In another particularly advantageous design, the following statistics are determined: how each pixel of the input image changes with input images from a pre-given calibration set (Kalibrier-Menge), or how each pixel of the intermediate product changes if all input images from said calibration set are successively fed to the image classifier. The maximum pixel value of this statistic is set to a first pre-given surrogate value L for that pixel. max The minimum pixel value for this statistic is set to a second pre-given surrogate value L for that pixel. min This allows for the adaptation of alternative values ​​to the current application.

[0037] This can be described, for example, in the form of a mass function:

[0038] ,

[0039] in:

[0040] if ,but ;

[0041] if ,but ;

[0042] if ,but .

[0043] The mass function The interference of the intermediate product L is identified based on the binary mask m. The classification score f is equipped with an apostrophe. It should show: compared to targeting The formula is different; instead of feeding the modified input image I to the classifier at this position, a modified intermediate product L is fed.

[0044] Alternatively, L can be replaced by other statistical characteristics formed by the correction set, such as the 90th or 10th percentile. max and L min .

[0045] In a particularly advantageous design, the sum of the binary masks, each weighted by an evaluation of the binary mask using a quality function, is determined as the decision-related share of the input image or intermediate product. Thus, the decision-related share S can be approximated, for example, as:

[0046] ,

[0047] Where, m i E[m] represents N distinct binary masks drawn from a random distribution, such as a Bernoulli distribution. i ] is the mask m extracted from it. i The expected value of the distribution.

[0048] In a particularly advantageous design, images of mass-produced products are selected as input images. The classification categories then represent the product's quality assessment, such as "OK," "NOK," "indecisive," or any intermediate level. Of particular importance in this regard is the explanation of why the image classifier assigns the input image to the category "NOK." Using such an explanation, not only can the image classifier itself be improved, but also insights into the possible causes of quality problems identified during manufacturing can be gained. If, for example, a large number of small, localized defects on the product are decisive for being classified as "NOK," and these defects are related to specific physical conditions during product manufacturing (such as high temperature or high pressure), then targeted efforts can be made to correct these conditions so that a larger proportion of future product samples can be classified as "OK."

[0049] The determined share of the input image upon which the image classifier bases its decisions can be compared, for example, with the share of the input image that has been determined to be relevant to the quality assessment of the product by means of observations of the same product using other imaging modalities. Based on the result of this comparison, the quality assessment used for the image classifier can be determined. For example, the hyperparameters of the image classifier can be optimized with the objective of optimizing the quality assessment.

[0050] In another particularly advantageous design, images of traffic conditions recorded from vehicles are selected as input images. The classified categories then represent an evaluation of the traffic conditions, and the future behavior of the vehicle is planned based on this evaluation. In this application area, image classifiers and downstream warning systems, driver assistance systems, or systems for at least partially automated driving can typically only be trusted if the decisions of the image classifier are interpretable. This method happens to provide these interpretations.

[0051] The determined share of the input image upon which the image classifier bases its decisions can be compared, for example, with a known share of the input image relevant to the assessment of traffic conditions. Based on the result of this comparison, a quality assessment for the image classifier can then be determined. Similar to an application example of quality inspection in mass production, the hyperparameters of the image classifier can be optimized, for example, with the objective of better evaluating the quality of the image classifier thereafter.

[0052] In particular, these methods can be implemented entirely or partially by a computer. Therefore, the present invention also relates to a computer program having machine-readable instructions that, when executed on one or more computers, cause the one or more computers to perform one of the described methods. In this sense, control devices for vehicles and embedded systems for technical devices, which are also capable of executing machine-readable instructions, should also be considered as computers.

[0053] The present invention also relates to machine-readable data carriers having computer programs and / or downloadable products. Downloadable products are digital products that can be transmitted over a data network, i.e., downloaded by a user of the data network, and such digital products may, for example, be sold in online stores for immediate download.

[0054] In addition, computers may be equipped with computer programs, machine-readable data carriers, or downloadable products. Attached Figure Description

[0055] Other measures to improve the invention are shown in more detail below, together with the description of preferred embodiments of the invention according to the accompanying drawings.

[0056] Figure 1 An embodiment of a method 100 for measuring the decision-related share 2a of an input image 2 for an image classifier 1 is shown;

[0057] Figure 2 The diagram illustrates two binary masks 4 that are evaluated as being good to different degrees by the quality function 5 within the scope of method 100. Detailed Implementation

[0058] Figure 1 A schematic flowchart illustrates an embodiment of a method 100 for measuring the following share 2a of an input image 2, wherein an image classifier 1 makes its own decision regarding assigning the input image 2 to one or more categories of a pre-given classification based on the share. The input image 2 may optionally be an image of a mass-produced product according to block 106 or an image of traffic conditions according to block 107.

[0059] In step 110, binary masks 4 are provided, which indicate which pixels of the input image 2 and / or the intermediate product 3 formed in the image classifier 1 are relevant according to the corresponding mask 4. These binary masks 4 may be drawn, for example, from a random distribution.

[0060] These binary masks 4 are evaluated in step 120 using a quality function 5. This quality function 5 is a measure of how strongly the at least one classification score provided by the image classifier 1 for at least one target category is altered if those pixels of the input image 2 or intermediate product 3, relevant to the binary mask 4, are changed. Based on the combination of the binary mask 4 and its corresponding evaluation 4a performed using the quality function, the desired share 2a of the input image 2 relevant to the decision of the image classifier 1 is determined in step 130.

[0061] According to block 111, these binary masks 4 may, in particular, relate to an intermediate product 3, which is generated from the input image 2 through processing in one or more convolutional layers. This intermediate product may, according to block 111a, be specifically an intermediate product that is mapped to at least one classification score by a classifier layer in the image classifier 1.

[0062] According to block 121, the strong correlation between the at least one classification score and pixels that are related to the input image 2 or intermediate product 3 according to the binary mask 4 can affect the value of the quality function 5 in an improved manner. According to block 121a, the strong correlation between the at least one classification score and pixels that are not related to the input image 2 or intermediate product 3 according to the binary mask 4 can then affect the value of the quality function 5 in a deteriorated manner.

[0063] According to block 122, at least one derivative and / or gradient of the at least one classification score of the pixels related to the input image 2 or intermediate product 3 according to the binary mask 4 can affect the value of the quality function 5 in an improved manner. According to block 122a, at least one derivative and / or gradient of the at least one classification score of the pixels unrelated to the input image 2 or intermediate product 3 according to the binary mask 4 can then affect the value of the quality function 5 in a deteriorated manner.

[0064] According to block 122b, in response to the absolute value of at least one gradient being below a pre-given threshold, the at least one gradient is kept unconsidered. Alternatively, or in conjunction with it, according to block 122c, at least one gradient and / or at least one contribution of a pixel to such gradient can be discretized in such a way that the gradient and / or the contribution is set to a first value in response to itself being above a pre-given threshold, and the gradient and / or the contribution is set to a second value in response to itself being below a negative pre-given threshold.

[0065] For example, the quality function may include at least one classification score, wherein the image classifier provides the classification score for a test image generated from an input image or intermediate product. This test image is modified relative to the input image or intermediate product, specifically regarding pixel aspects relevant to the binary mask 4. According to block 122d, in response to a pixel's contribution to the gradient of the classification score based on the relevant pixel exceeding a pre-given threshold, the pixel is changed to a first pre-given alternative value L. max According to block 122e, in response to a pixel's contribution to the gradient of a classification score based on the relevant pixel being lower than a pre-given threshold of negativity, the pixel is changed to a second pre-given alternative value L. min .

[0066] In particular, according to block 122f, statistics can be determined, for example, regarding how each pixel of the input image 2 or intermediate product 3 changes with the input image 2 given a pre-given set of corrections. According to block 122g, the maximum pixel value of this statistics is then set as a first pre-given substitution value L for that pixel. max According to block 122h, the minimum pixel value for this statistic is set to a second pre-given surrogate value L for that pixel. min .

[0067] According to block 131, the sum of binary masks can be determined, for example, as a decision-related share 2a of the input image 2 or intermediate product 3, wherein these binary masks are weighted using an evaluation 4a of these binary masks 4 by a quality function 5. The decision-related share of the intermediate product 3 can be transformed into the desired decision-related share 2a of the input image 2 by interpolation or other upsampling, regardless of its acquisition method.

[0068] Once block 106 has determined the decision-related shares 2a based on the input image 2 displaying the mass-produced products, these shares 2a can be compared in step 140 with shares 2b of the input image, determined to be relevant to the quality evaluation of the product by means of observations of the same product using other imaging modalities. The quality evaluation 1a for image classifier 1 can then be determined in step 150 based on the result 140a of this comparison 140.

[0069] Once the decision-related shares 2a have been determined based on the input image 2 displaying traffic conditions according to block 107, these shares 2a can be compared in step 160 with the known shares 2b of the input image that are relevant to the assessment of traffic conditions. Then, in step 170, the quality rating 1a for the image classifier 1 can be determined based on the result 160a of this comparison 160.

[0070] Figure 2Two masks, m1 and m2, are illustrated, illustrating which pixels of the input image 2 should be considered relevant. As explained above, it is particularly advantageous to work in the latent space using masks in the space of the intermediate product 3; however, masks in the space of the input image 2 are more readily understood to clarify this effect.

[0071] Input image 2 shows the following road, and the image classifier 1 is trained based on the recognition of the road. Mask m1 evaluates pixels uniformly distributed across the entire input image 2 as relevant. Conversely, pixels considered relevant according to mask m2 are more strongly concentrated within the area of ​​road 10. Mask m2 is evaluated as better than mask m1 by quality function 5. If many masks 4 are randomly selected, evaluated separately using quality function 5, and aggregated according to the evaluation 4a obtained therein, then the image region composition (herauskristallisieren) with road 10 constitutes the decision-related share 2a of input image 2.

Claims

1. A method (100) for measuring the components of an input image (2), wherein an image classifier (1) makes its own decision regarding assigning the input image (2) to one or more categories of a pre-given classification based on the components, the method comprising the steps of: Provide a binary mask (4) that indicates which pixels of the input image (2) and / or the intermediate product (3) formed in the image classifier (1) are considered relevant; The binary mask (4) is evaluated using a quality function (5) that measures how strongly at least one classification score for at least one target category provided by the image classifier (1) changes if the pixels of the input image (2) or the intermediate product (3) relative to the binary mask (4) are changed. Based on the combination of the binary mask (4) and the corresponding evaluation of the binary mask through the quality function (5), the desired components of the input image (2) relevant to the decision of the image classifier (1) are determined. The value of the quality function (5) is influenced in an improved manner by at least one derivative and / or gradient of at least one classification score of the pixel associated with the input image (2) or the intermediate product (3) according to the binary mask (4).

2. The method (100) according to claim 1, wherein an intermediate product (3) is selected, wherein the intermediate product is generated from the input image (2) by processing in one or more convolutional layers.

3. The method (100) according to claim 2, wherein an intermediate product (3) is selected, the intermediate product being mapped to at least one classification score by a classifier layer in the image classifier (1).

4. The method (100) according to any one of claims 1 to 3, wherein the correlation between the at least one classification score and the pixels of the input image (2) or the intermediate product (3) in relation to the binary mask (4) affects the value of the quality function (5) in an improved manner.

5. The method (100) according to claim 4, wherein the correlation between the at least one classification score and the pixels of the input image (2) or the intermediate product (3) that are unrelated to the binary mask (4) affects the value of the quality function (5) in a degraded manner.

6. The method (100) according to claim 1, wherein at least one derivative and / or gradient of at least one classification score of a pixel that is irrelevant to the input image (2) or the intermediate product (3) according to the binary mask (4) affects the value of the quality function (5) in a deteriorating manner.

7. The method (100) of claim 6, wherein the at least one gradient is kept unconsidered in response to the absolute value of at least one gradient being below a pre-given threshold.

8. The method (100) according to claim 6 or 7, wherein at least one gradient and / or at least one contribution of a pixel to such gradient is discretized in such a way that the gradient and / or the contribution is set to a first value in response to itself being above a pre-given threshold, and the gradient and / or the contribution is set to a second value in response to itself being below a pre-given threshold of negativity.

9. The method (100) according to claim 6 or 7, wherein if the input image or intermediate product is modified in terms of the relevant pixel aspects with respect to the binary mask (4) as follows: If a pixel's contribution to the gradient of a classification score based on related pixels exceeds a pre-given threshold, the pixel is changed to a first pre-given alternative value L. max ;and If a pixel's contribution to the gradient of a classification score based on the relevant pixel is below a pre-given threshold of negativity, the pixel is changed to a second pre-given alternative value L. min , The quality function (5) then includes at least one classification score, wherein the image classifier provides the classification score.

10. The method (100) according to claim 9, wherein The following statistics are determined: how each pixel of the input image (2) or the intermediate product (3) changes with the input image (2) given a pre-given set of corrections; The maximum pixel value in the statistics is set as a first pre-given substitution value L for the pixel. max ;and The minimum pixel value in the statistics is set as a second pre-given surrogate value L for the pixel. min .

11. The method (100) according to any one of claims 1 to 3, wherein the sum of binary masks (4) is determined as a decision-related component of the input image (2) or the intermediate product (3), wherein the binary masks are weighted by the evaluation of the binary masks (4) by the quality function (5).

12. The method (100) according to any one of claims 1 to 3, wherein the decision-related components of the intermediate product (3) are transformed into the decision-related desired components of the input image (2) by interpolation or other upsampling.

13. The method (100) according to any one of claims 1 to 3, wherein an image of a mass-produced product is selected as an input image (2) and wherein the category of the classification represents a quality evaluation of the product.

14. The method (100) according to claim 13, wherein the determined components of the input image (2) on which the image classifier (1) makes its own decisions are compared with the components of the input image (2) that have been determined to be relevant to the quality evaluation of the product by means of other imaging modalities, and wherein the quality evaluation for the image classifier (1) is determined based on the result of the comparison.

15. The method (100) according to any one of claims 1 to 3, wherein an image of traffic conditions recorded from a vehicle is selected as an input image (2) and wherein the category of the classification represents an evaluation of the traffic conditions (50), wherein the future behavior of the vehicle is planned based on the evaluation of the traffic conditions.

16. The method (100) according to claim 15, wherein the determined components of the input image (2) on which the image classifier (1) makes its own decisions are compared with known components of the input image (2) that are relevant to the assessment of the traffic conditions, and wherein a quality assessment for the image classifier (1) is determined based on the result of the comparison.

17. A computer program product comprising machine-readable instructions that, when executed on one or more computers, cause the one or more computers to perform the method (100) according to any one of claims 1 to 16.

18. A machine-readable data carrier or downloadable product having a computer program product according to claim 17.

19. A computer having a computer program product according to claim 17 and / or a machine-readable data carrier or downloadable product according to claim 18.

Images