Determination of the decision-relevant image components for an image classifier by targeted modification in the latent space

1. A method for measuring components of an input image on which an image classifier bases its decision regarding an assignment of the input image to one or multiple classes of a predefined classification, the method comprising the following steps:

processing the input image by the image classifier into an intermediate product using one or multiple convolutional layers;

mapping the intermediate product by the image classifier on a classification score with respect to at least one target class;

ascertaining a perturbation in a space of the intermediate product from one or multiple counter images which is preferentially assigned by the image classifier to at least one class other than the target class;

providing at least one binary mask, which has the same number of pixels as the intermediate product;

creating, from the intermediate product, at least one modification, in which pixels of the intermediate product established by the binary mask are replaced with pixels of the perturbation corresponding the pixels established by the binary mask;

mapping the modification by the image classifier on a classification score with respect to a predefined class; and

ascertaining from the classification score, using a quality function, to what extent the binary mask indicates sought-after decision-relevant components of the input image.

2. The method as recited in claim 1 , wherein the intermediate product is selected which is mapped on the at least one classification score in the image classifier by a classifier layer.

3. The method as recited in claim 1 , wherein the perturbation is formed of at least one intermediate product to which the image classifier processes the one or multiple counter images.

4. The method as recited in claim 3 , wherein the perturbation is formed by averaging or forming another summarizing statistic over multiple intermediate products into which the image classifier processes different counter images.

5. The method as recited in claim 3 , wherein at least one counter image is selected from the one or multiple counter images, for which an intermediate product formed by the image classifier comes closest to the intermediate product formed of the input image in accordance with a predefined distance dimension.

6. The method as recited in claim 1 , wherein the at least one binary mask includes a plurality of binary masks, and the sought-after decision-relevant components of the input image are ascertained from all of the plurality of binary masks and associated assessments by the quality function.

7. The method as recited in claim 6 , wherein at least one counter image of the one or multiple counter images is randomly selected for the assessment of each of the binary masks.

8. The method as recited in claim 6 , wherein each decision-relevant component of the input image is evaluated from a sum of the binary masks, which are each weighted with the assessments of these binary masks by the quality function.

9. The method as recited in claim 1 , wherein the predefined class for which the classification score is ascertained from the modification is the target class, the quality function including comparison of the classification score to a classification score ascertained for the intermediate product.

10. The method as recited in claim 1 , wherein each decision-relevant component of the intermediate product ascertained from the one or multiple binary masks is transferred into the sought-after decision-relevant component of the input image by interpolation or other upsampling.

11. The method as recited in claim 1 , wherein an image of a series-manufactured product is selected as the input image, and the classes of the classification represent a quality assessment of the product.

12. The method as recited in claim 11 , wherein the ascertained components of the input image on which the image classifier bases its decision are compared to a component of the input image which was ascertained to be relevant for a quality assessment of the product based on an observation of the same product using a different mapping modality, a quality assessment for the image classifier being ascertained from a result of the comparison.

13. The method as recited in claim 1 , wherein an image of a traffic situation recorded from a vehicle is selected as the input image, the classes of the classification representing assessments of the traffic situation, based on which a future behavior of the vehicle is planned.

14. The method as recited in claim 13 , wherein the ascertained components of the input image on which the image classifier bases its decision are compared to a component of the input image which is known to be relevant for the assessment of the traffic situation, a quality assessment for the image classifier being ascertained from a result of the comparison.

15. A non-transitory machine-readable data medium on which is stored a computer program including machine-readable instructions for measuring components of an input image on which an image classifier bases its decision regarding an assignment of the input image to one or multiple classes of a predefined classification, the machine-readable instructions, when executed by one or more computers, causing the one or more computers to perform the following steps:

processing the input image by the image classifier into an intermediate product using one or multiple convolutional layers;

mapping the intermediate product by the image classifier on a classification score with respect to at least one target class;

ascertaining a perturbation in a space of the intermediate product from one or multiple counter images which is preferentially assigned by the image classifier to at least one class other than the target class;

providing at least one binary mask, which has the same number of pixels as the intermediate product;

creating, from the intermediate product, at least one modification, in which pixels of the intermediate product established by the binary mask are replaced with pixels of the perturbation corresponding the pixels established by the binary mask;

mapping the modification by the image classifier on a classification score with respect to a predefined class; and

ascertaining from the classification score, using a quality function, to what extent the binary mask indicates sought-after decision-relevant components of the input image.

16. A computer configured to measure components of an input image on which an image classifier bases its decision regarding an assignment of the input image to one or multiple classes of a predefined classification, the computer configured to:

process the input image by the image classifier into an intermediate product using one or multiple convolutional layers;

map the intermediate product by the image classifier on a classification score with respect to at least one target class;

ascertain a perturbation in a space of the intermediate product from one or multiple counter images which is preferentially assigned by the image classifier to at least one class other than the target class;

provide at least one binary mask, which has the same number of pixels as the intermediate product;

create, from the intermediate product, at least one modification, in which pixels of the intermediate product established by the binary mask are replaced with pixels of the perturbation corresponding the pixels established by the binary mask;

map the modification by the image classifier on a classification score with respect to a predefined class; and

ascertain from the classification score, using a quality function, to what extent the binary mask indicates sought-after decision-relevant components of the input image.