Determining image shares relevant for image classifier decisions

Abstract

Method of measuring a share of an input image, wherein an image classifier bases its decision about assigning the input image to one or more classes of a pre-given classification on said share, the method comprising: processing the input image by the image classifier through a convolutional layer into an intermediate product; mapping the intermediate product by the image classifier to a classification score with respect to at least one target class; determining an interference in the space of the intermediate product by an anti-image, wherein the image classifier preferentially assigns said anti-image to other classes compared to the target class; providing a binary mask having the same number of pixels as the intermediate product; creating a variant from the intermediate product in which pixels set by the binary mask are replaced by corresponding pixels of the interference; mapping said variant by the image classifier to a classification score with respect to the pre-given classes; determining with a quality function from the classification scores: how much the binary mask accounts for the sought decision-related share of the input image.

Description

Technical Field

[0001] This invention relates to inspecting 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-given categories using an image classifier based on an artificial neural network. Based on this, the product can be assigned to one of several pre-given 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 (Anteil) of an input image, wherein an image classifier makes its own decision regarding assigning the input image to one or more categories of a pre-given classification based on the share.

[0007] In this method, the image classifier first processes the input image into an intermediate product through one or more convolutional layers. Compared to the input image, this intermediate product has a significantly reduced dimensionality and illustrates the activation of features that the convolutional layers have gradually identified in the input image up to this point. The more convolutional layers involved in creating the intermediate product, the more complex the features become, with the intermediate product exhibiting activation of these features. The intermediate product may, for example, include a large number of feature maps generated, either by applying a filter kernel to the input image or to an intermediate product previously generated from the input image.

[0008] Therefore, the intermediate product belongs to the "latent space" within the image classifier. The image classifier maps the intermediate product to classification scores about at least one target category.

[0009] The spatial interference of the intermediate product is determined by one or more inverse images (Gegen-Bild), wherein the image classifier preferentially assigns the inverse image to at least one other category compared to the target category. At least one binary mask is now provided, having the same number of pixels as the intermediate product. Each of these pixels in the mask can only have two distinct values, i.e., 0 and 1, or "true" and "false".

[0010] At least one variant (Abwandlung) is created from the intermediate product. In this variant, the pixels defined by the binary mask are replaced by the corresponding interfering pixels. For example, all pixels with a value of 0 or "false" in the mask can be replaced by the interfering pixels in this variant.

[0011] The image classifier maps the variant to a classification score about a pre-given category. Based on the classification score, a quality function (Gütefunktion) is used to determine how the binary mask represents the desired decision-related share of the input image. By appropriately selecting the quality function and the categories involved in the classification score of the variant, different aspects of the decision-related share of the input image can be studied.

[0012] If the pre-given category for which the classification score is determined according to the variant is, for example, the target category, then the quality function may include a comparison of the classification score with the classification score determined for the intermediate product. For example, a mask can now be provided that leaves a determined small region of the intermediate product unchanged, while simultaneously applying interference to the intermediate product. It can then be investigated, for example, whether the unchanged region is so important for assigning the intermediate product to the target category that applying interference to other regions of the intermediate product can no longer shake that assignment.

[0013] However, it's also possible to specifically inquire: through which changes in the intermediate product can the allocation of the intermediate product be switched from the target class to another category? If, for example, a product image recorded within the scope of product quality inspection has been classified as "NOK," then a corresponding mask can be used to inquire what changes must be made to the image to classify it as "OK." Ideally, these decision-relevant areas of the image should correspond to specific defects or damage. This is similar to what is expected of a human inspector: that the inspector points out the specific reasons for a product being rejected, rather than simply stating: the product is "somewhat unsatisfactory."

[0014] If the pre-given category is not the target category, it is possible to investigate, for example, which changes to the intermediate product can switch its assignment from the target category to a specific category among several possible other categories. In the context of the relevant application, misassignment to different categories can lead to different serious consequences. If, for example, a system for at least partially automated driving vehicles classifies traffic signs, misidentifying a "30 km / h" sign as a "50 km / h" sign is less serious than misidentifying the same "30 km / h" sign as sign 301, which grants right-of-way at the next intersection or street. If a vehicle incorrectly assumes it has the right-of-way, the risk of a collision increases significantly.

[0015] Of particular advantage in these and other studies is the determination of perturbations by one or more inverse images. It has been recognized that the effectiveness of this type of perturbation depends less on the specific characteristics of the model used in the image classifier compared to the case of commonly used perturbations. Two different image classifiers will certainly assign the same variant to different categories. Nevertheless, the results obtained for these different image classifiers can be meaningfully and quantitatively compared with each other. In this way, for example, analyzing which regions are decision-relevant for different image classifiers can provide criteria for the selection of one of several possible image classifiers.

[0016] Simpler perturbations, such as setting pixels to zero or other uniform values ​​in image classifiers using the ReLU activation function, depend on the model's effectiveness. Therefore, it's not only the presence of defects or impairments that might lead to classifying a product as "NOK," which manifest as additive features in the image, that can cause a product to be categorized as "NOK." The presence of such a feature in the image manifests as activations other than zero in the intermediate products, such that setting these activations to zero removes the corresponding feature. Conversely, the absence of a specific feature can also provide an incentive to classify a product as "NOK." For example, the manufacturing process of a router for home use might specify equipping the device with a tag containing the password to be used for initial setup. Therefore, during final inspection, it is important that the tag is actually present on the device, as the device is unusable without it. Therefore, if the activation of the feature “label” is set to zero when applying interference to intermediate products, it does not remove information from the intermediate products as intentionally would, but rather adds very important information about product defects.

[0017] Advantageously, intermediate products are chosen that are mapped to at least one classification score by the classifier layer in the image classifier. In the case of such intermediate products, how setting the activation to zero or other uniform values ​​has an effect depends on the model in a manner described.

[0018] Interference can be formed, for example, by at least one intermediate product, into which the image classifier processes one or more inverse images. This can lead not only to the activation of additional features but also to the deactivation of features in variations of the intermediate product. Interference can also be formed, for example, by forming or averaging other summary statistics about a plurality of intermediate products, into which the image classifier processes different inverse images.

[0019] In another advantageous design, at least one inverse image is selected from a plurality of inverse images, wherein for said at least one inverse image, the intermediate product formed by the image classifier is closest to the intermediate product formed by the input image according to a pre-given distance metric. For example, the cosine distance between vectors containing the pixel values ​​of the corresponding intermediate product can be used as the distance metric. The perturbation is thus more realistic and contains fewer semantic errors, such as an object floating freely in the air after being removed from the table it was on before.

[0020] In another particularly advantageous design, multiple binary masks are provided. The desired decision-related share of the input image is determined by the total number of said masks and the associated evaluation performed through the quality function. For example, this could involve assigning an evaluation number (Bewertungszahl) to each pixel of the input image, the evaluation number indicating how decision-related that particular pixel is. A large number of binary masks can be drawn, for example, from a random distribution, such as a Bernoulli distribution.

[0021] In this method, even binary masks with low or poor ratings from the quality function are not necessarily discarded. Instead, all masks studied in the overall population are incorporated into the final result. Finally, the aggregation of several randomly selected binary masks also results in the averaging of the randomness in the masks.

[0022] Advantageously, at least one inverse image is randomly selected for evaluating each mask. This makes the final aggregation result independent of the selection of a particular inverse image.

[0023] In another particularly advantageous design, the decision-related share of the input image is evaluated based on the sum of binary masks, wherein each binary mask is weighted using an evaluation of the binary mask by a quality function.

[0024] For example, assuming x is the input image, f L (x) is an intermediate product in the latent space generated by the image classifier. Therefore, using a binary mask m and an interference P in the latent space set according to the input image x... L (x) can be modified into x' of the intermediate product. L For example, it can be written as: .

[0025] Quality evaluation of each mask m R x,f (m) can now correspond, for example, to the classification score f. c The image classifier assigns the classification score to the variant x'. L :

[0026] .

[0027] If N binary masks mi are extracted from the space of the intermediate product, then the evaluation number is... It can be written, for example, as

[0028] .

[0029] Here, footnote f denotes an image classifier.

[0030] Typically, the decision-related shares of the intermediate product, determined according to one or more binary masks, can be transformed into the desired decision-related shares of the input image through interpolation or other upsampling. In this case, the spatial correspondence between the intermediate product and the input image is fully utilized (Korrespondenz).

[0031] 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." Here, the image assigned to the category "OK" by the image classifier is used as the inverse image. In mass production, in principle, all images—not only "OK" images but also "NOK" images—are very similar to each other because the manufacturing process typically provides highly reproducible results. Correspondingly, the activations in the intermediate products obtained from these images are also very similar to each other. The method presented here provides a very specific explanation of the differences between "OK" and "NOK" images.

[0032] Using such an interpretation, not only can the image classifier itself be improved, but also insights can be gained about the possible causes of quality problems identified during the manufacturing process. If, for example, a large number of small, localized defects on a 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 the manufactured product samples can be classified as "OK" in the future.

[0033] 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.

[0034] 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.

[0035] For example, a pedestrian recognition system can identify whether a traffic situation in an input image contains pedestrians (output 1) or does not contain pedestrians (output 0). If, for example, the input image does not contain pedestrians according to the image classifier, "pedestrian activations" can be added to the intermediate products created from that input image at different locations to study how the behavior of the image classifier changes.

[0036] In this regard, it is particularly advantageous to generate these variants by modifying the intermediate product rather than by modifying the input image itself. Adding "pedestrian activations" to the latent space of the intermediate product is significantly simpler than generating variants that include additional pedestrians and still appear realistic by modifying the input image pixel-by-pixel. Therefore, the method proposed here is often a simpler way to test novel data points generated using Generative Adversarial Networks (GANs) or other generative models.

[0037] 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.

[0038] 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.

[0039] 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.

[0040] In addition, computers may be equipped with computer programs, machine-readable data carriers, or downloadable products. Attached Figure Description

[0041] 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.

[0042] Figure 1 illustrates an embodiment of a method 100 for measuring the decision-related share 2a of an input image 2;

[0043] Figure 2 The diagram illustrates the problem of how deactivation of features in the latent space can also add information. Detailed Implementation

[0044] Figure 1 shows a schematic flowchart of an embodiment of a method 100 for measuring the following share 2a of an input image 2, wherein the 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.

[0045] In step 110, the input image 2 is processed into an intermediate product 3 by the image classifier 1 through one or more convolutional layers. Here, the intermediate product 3 can advantageously be selected according to block 111, which is mapped to at least one classification score 7 by the classifier layer in the image classifier 1. The intermediate product 3, in particular, cannot be processed by other convolutional layers until it has been processed by the classifier layer.

[0046] In step 120, the intermediate product 3 is mapped by the image classifier 1 to a classification score 7 for at least one target category.

[0047] In step 130, interference 6 in the space of intermediate product 3 is determined by one or more inverse images 5a-5c, wherein image classifier 1 preferentially assigns the inverse images 5a-5c to at least one other category compared to the target category. In step 140, at least one binary mask 4 is provided, the binary mask having the same number of pixels as intermediate product 3. Binary mask 4 and interference 6 are merged in step 150: at least one variant 3* is created from intermediate product 3, in which the pixels defined by binary mask 4 are replaced by corresponding pixels of interference 6.

[0048] In this case, according to block 131, in particular, interference 6 can be formed by at least one intermediate product 3', wherein image classifier 1 processes one or more inverse images 5a-5c into said intermediate product 3'.

[0049] For example, according to block 131a, the interference 6 can be formed, for example, by forming or averaging other summary statistics about a plurality of intermediate products 3', wherein the image classifier 1 processes different inverse images 5a-5c into the plurality of intermediate products 3'.

[0050] For example, according to block 131b, at least one of the inverse images 5a-5c can be selected from a plurality of inverse images 5a-5c, wherein for the at least one inverse image, the intermediate product 3' formed by the image classifier 1 is closest to the intermediate product 3 formed by the input image 2 according to a pre-given distance metric.

[0051] According to block 132, if multiple masks 4 are used, at least one inverse image 5a-5c can be randomly selected to evaluate each mask 4 so that the final result is not too strongly constrained by the individual inverse images 5a-5c.

[0052] In step 160, the image classifier 1 maps variant 3* to a classification score 7* about a pre-given category. In step 170, based on the classification score 7*, a quality function 8 is used to determine how the binary mask 4 describes the desired decision-related share 2a of the input image 2 using a metric 8a. Therefore, the metric 8a is the value of the quality function 8 for a specific binary mask 4.

[0053] According to block 161, the pre-given category for determining the classification score 7* according to variant 3* can be, for example, the target category. Thus, according to block 171, the quality function 8 can include a comparison of the classification score 7* with the classification score 7 determined for the intermediate product 3.

[0054] According to block 141, multiple binary masks 4 can be provided. The desired decision-related share 2a of the input image 2 can then be determined in step 180 by the totality of the masks 4 and the corresponding evaluation 8a performed by the quality function 8.

[0055] As long as the decision-related share has been determined in the space of the intermediate product beforehand, the decision-related share can be transformed into the desired decision-related share 2a of the input image 2 in step 190 by interpolation or other upsampling.

[0056] 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 200 with shares 2b of the input image, which have been determined to be relevant to the quality evaluation of the product by means of other imaging modalities of the same product. The quality evaluation 1a for image classifier 1 can then be determined in step 210 based on the result 200a of this comparison 200.

[0057] 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 220 with known shares 2b of the input image that are relevant to the assessment of traffic conditions. Then, in step 230, the quality rating 1a for image classifier 1 can be determined based on the result 220a of this comparison 220.

[0058] Figure 2 This study elucidates the relationship between relevant image features for product quality assessment and the activation of features in the intermediate product 3 obtained from input image 2 in the latent space.

[0059] Sub-image a) is input image 2, which shows a WLAN router 15 for home use. During a visual inspection of the router 15, the casing 10, three antennas 11a-11c, LED 12, and a tag 13 with a password for initial operation are typically visible. In sub-image a), all these components are also present, such that the image classifier 1 used for quality inspection should classify the input image 2 into the category "OK". Sub-image b) shows the intermediate product 3 obtained during the evaluation process in the latent space. Activations 50, 51a-51c, 52, and 53 correspond to image features 10, 11a-11c, 12, or 13.

[0060] Sub-image c) is input image 2, which shows a router 15 that is defective for two reasons. On the one hand, the casing 10 has a large crack 14. On the other hand, it lacks a label 13, which makes initial commissioning impossible and renders the entire router 15 useless. Each of these defects is a sufficient reason in itself for the router 15 to no longer be marketable. Therefore, this input image 2 is classified by image classifier 1 into the category "NOK". Sub-image d) is an intermediate product 3 in the latent space formed during this evaluation process. Here, compared to sub-image b), activation 54 is added, corresponding to the crack 14 in the casing 10. At the same time, activation 53 corresponding to the missing label 13 is also added.

[0061] This means that in the intermediate product 3 according to sub-image b), setting the activation 53 corresponding to label 13 to zero does not merely mean non-critically "resetting" the information to the "no information" standard state. Instead, it adds the information that router 15 has a serious defect that hinders its operation.

[0062] This effect is even more pronounced in the inverse situation, where the training of image classifier 1 converges such that zero activation 53 indicates the presence of label 13, while non-zero activation 53 (e.g., 1) indicates the absence of label 13. If we now examine the input image 2 of router 15 with label 13—that is, why router 15 has been classified as "OK"—then simply setting the activations to zero does not reveal any connection between label 13 and the quality of router 15. Activation 53 is already zero, meaning the corresponding interference 6 does not lead to any results.

[0063] The proposed method here determines the interference 6 based on the inverse images 5a-5c. Therefore, whether the interference 6 alters the intermediate product 3 no longer depends on the encoding of individual features in the intermediate product 3. Only by utilizing the interference 6, which alters the intermediate product 3, can the decision-related share 2a of the input image 2 be determined.

[0064] In sub-images b) and d), an exemplary mask 4 is drawn, which defines possible decision-related regions in the intermediate product 3. The interference 6 used within the scope of method 100 may, for example, include copying regions from the intermediate products 3 belonging to the inverse images 5a-5c, as marked by mask 4, into the variant 3* formed by the current intermediate product 3.

Claims

1. A method (100) for measuring the 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 method comprising the steps of: • The input image (2) is processed by the image classifier (1) through one or more convolutional layers (110) to form a first intermediate product (3); • The first intermediate product (3) is mapped (120) to a first classification score (7) about at least one target category by the image classifier (1); • The interference (6) in the space of the first intermediate product (3) is determined (130) by one or more inverse images (5a-5c), wherein the image classifier (1) preferentially assigns the inverse image to at least one other category compared to the target category. The interference (6) is formed (131) by at least one second intermediate product (3') as follows, wherein the image classifier (1) processes one or more inverse images (5a-5c) into the second intermediate product (3'). Among a plurality of inverse images, at least one inverse image is selected, wherein for the at least one inverse image, a second intermediate product (3') formed by the image classifier is closest to a first intermediate product (3) formed by the input image according to a pre-given distance metric, wherein the cosine distance between vectors containing the pixel values ​​of the corresponding intermediate product is used as the distance metric; • Provide (140) at least one binary mask (4) having the same number of pixels as the first intermediate product (3); • Create at least one variant (3) from the first intermediate product (3) (150) In the aforementioned variant, the pixel defined by the binary mask (4) is replaced by the corresponding pixel of the interference (6). The decision-related share of the first intermediate product (3) determined according to one or more binary masks (4) is transformed (190) into the desired decision-related share (2a) of the input image (2) by upsampling, wherein the binary mask m and the interference P set according to the input image x are used. L (x) and take one of the variations (3) )x' L Written as: , where x is the input image (2), and f L (x) is the first intermediate product (3) in the latent space generated by the image classifier (1); • The variant (3) is classified by the image classifier (1). Mapping (160) to a second classification score (7) about a pre-given category. ); •Based on the second category score (7) Using the quality function (8), determine (170): how the binary mask (4) describes the desired decision-related share (2a) of the input image (2) using the metric (8a). The quality evaluation R of each binary mask m x,f (m) corresponds to the third category score f c The image classifier (1) assigns the third classification score to the variant x'. L : N binary masks m are extracted from the space of the intermediate product. i In this case, the evaluation numbers Written as , where footnote f represents the image classifier (1) and where E[m] represents the expected value of the binary mask.

2. The method (100) according to claim 1, wherein (111) a first intermediate product (3) is selected, the first intermediate product (3) being mapped by a classifier layer to at least one first classification score (7) in the image classifier (1).

3. The method (100) according to claim 1 or 2, wherein the interference (6) is formed (131a) by forming or averaging other summary statistics about a plurality of second intermediate products (3'), wherein the image classifier (1) processes different inverse images (5a-5c) into the plurality of second intermediate products.

4. The method (100) according to claim 1, wherein (141) a plurality of binary masks (4) are provided and wherein the desired decision-related share (2a) of the input image (2) is determined (180) by the totality of the binary masks (4) and the evaluation (8a) performed by the quality function (8).

5. The method (100) according to claim 4, wherein at least one inverse image (5a-5c) is randomly selected (132) for evaluating each binary mask (4).

6. The method (100) according to claim 4 or 5, wherein the decision-related share (2a) of the input image (2) is evaluated (181) based on the sum of the following binary masks (4), wherein the binary masks are weighted using the evaluation (8a) of the binary masks (4) by the quality function (8).

7. The method (100) according to claim 1 or 2, wherein according to the variant (3) And determine the second category score (7) The pre-given category is the target category (161), and the quality function (8) includes: the second classification score (7) (171) Comparison with the first classification score (7) determined for the first intermediate product (3).

8. The method (100) according to claim 1 or 2, wherein (106) 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.

9. The method (100) according to claim 8, wherein the determined share (2a) of the input image (2) on which the image classifier (1) makes its own decisions is compared (200) with the share (2b) of the input image (2) that has been determined to be relevant to the quality evaluation of the product by means of other imaging modalities, and wherein the quality evaluation (1a) of the image classifier (1) is determined (210) based on the result (200a) of the comparison (200).

10. The method (100) according to claim 1 or 2, wherein (107) 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.

11. The method (100) according to claim 10, wherein the determined share (2a) of the input image (2) on which the image classifier (1) makes its own decisions is compared (220) with the known share (2b) of the input image (2) as relevant to the assessment of the traffic conditions, and wherein a quality evaluation (1a) for the image classifier (1) is determined (230) based on the result of the comparison (220a).

12. 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 11.

13. A machine-readable data carrier and / or downloadable product, wherein the machine-readable data carrier and / or downloadable product has the computer program according to claim 12.

14. A computer, wherein the computer has a computer program according to claim 12 and / or a machine-readable data carrier and / or download product according to claim 13.

Images