DE102021214465A1 - Analysis of the behavior of image classifiers - Google Patents

Abstract

Method (100) for examining the behavior of an image classifier (1), which assigns an input image (2) to one or more classes of a specified classification, with the steps: • the input image (2) is scanned with an encoder (3a) a trained encoder-decoder arrangement (3) onto an input representation (4) (110);• representation modifications (4') of this input representation (4) are determined (120);• these representation modifications (4') are mapped (130) to image modifications (2') of the input image (2) with the decoder (3b) of the encoder-decoder arrangement (3);• each image modification (2') is evaluated (140) with a predetermined cost function (5) from a number of contributions (141, 142, 143); • according to a predetermined criterion, which depends on the evaluation (5a) by the cost function (5), decision-relevant image modifications (2nd *) selected (150).

Description

The present invention relates to the analysis of the behavior of image classifiers that can be used, for example, for quality control in the context of series production of products.

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").

Disclosure of Invention

The invention provides a method for examining the behavior of an image classifier. The image classifier assigns an input image to one or more classes of a given classification.

As part of this method, the input image is mapped to an input representation using an encoder of a trained encoder-decoder arrangement. Representation modifications of this input representation are determined. These representation modifications are mapped to image modifications of the input image with the decoder of the encoder-decoder arrangement.

The image modifications are evaluated using a predetermined cost function. Decision-relevant image modifications are selected according to a predetermined criterion. The predetermined criterion depends on the evaluation by the cost function.

The decision-relevant image modifications contain the information sought about the behavior of the image classifier. With this, it is possible to examine which changes in the input image make the image classifier more or less certain in its decision, even without reference to concrete spatially defined image parts. For example, a statement can be obtained as to how much the uncertainty of the image classifier's decision changes in the event of global changes in color, contrast and other image properties. Such a statement then relates to the image classifier as such and is only linked to this specific image classifier, but not to a specific input image.

In a particularly advantageous embodiment, image components on which the image classifier bases its decision are evaluated from the selected, decision-relevant image modifications. For this purpose, for example, a difference between the input image on the one hand and one or more decision-relevant image modifications on the other hand can be evaluated. In concrete applications, it is often asked to what extent the image classifier uses precisely those image areas for its decision, in which the information relevant for the respective application should also be found. For example, if a face in a portrait photograph is to be classified based on certain aspects, the image classifier should ideally base its decision only on the face itself, and not on the background of the image.

A search is therefore made for those image modifications for which the evaluation by the cost function in connection with the specified criterion indicates that these modifications are relevant to the decision of the image classifier. However, this search is not sought directly in the space of image modifications, but in the space of representational modifications.

The point of this "detour" is that if the input image is changed directly, it is difficult to ensure that the input image still belongs to a desired distribution. For example, you may want the image modifications to "look realistic" just as much. This requirement can be mapped and quantified very well by training an encoder-decoder arrangement on training images from the desired distribution of, for example, "realistic" images to map the training images to representations with the encoder and then use the decoder to map the training images to representations to reconstruct the respective training image. The representations used here can be significantly reduced in their dimensionality, for example compared to the training images. The encoder-decoder arrangement then learns to accommodate the most important information for the reconstruction in the representation and at the same time to evaluate the small amount of information contained in the representation in the best possible way. If now representation modifications are formed and translated back to image modifications of the input image, the training of the decoder ensures that the image modification belongs to the same distribution as the original input image.

The cost function contains a number of contributions which can be combined with one another in any way, for example additively and/or multiplicatively, and which can be weighted among one another.

The value of the cost function becomes the better, the better the value of a given task cost function which depends on a processing result determined by the image classifier from the image modification. So if the image modification is changed from the input image in such a way that the image classifier can in some way process this image modification better than the original input image, it follows that the change is relevant to the image classifier's decision.

For this purpose, the task cost function can contain, for example, an uncertainty measured with a predefined uncertainty metric, with which the image classifier assigns the image modification to one or more classes of the predefined classification. For example, if the change in image morph leads to a better recognition of a feature of the input image that is important for the image classifier, the uncertainty with which the image classifier makes its decision decreases.

Alternatively or in combination with this, the task cost function can also contain, for example, one or more classification scores in relation to classes of the specified classification that the image classifier determines for the image modification. It can also be reflected in these classification scores how well or how poorly certain features can be recognized in the image modification.

Particularly advantageously, only those image modifications are taken into account for which a highest classification score determined by the image classifier relates to the same class as the highest classification score determined by the image classifier for the input image. Changes to the input image that do not change the input image so much that it changes to a different class are searched for. As a result, the image modifications ultimately selected as relevant to the decision become more meaningful in relation to the image classifier. It's all about examining the behavior of the image classifier in the working point defined by the input image and to determine to what extent even small changes to the input image have a noticeable effect on the output of the image classifier.

Also serving this purpose is that the smaller a distance, measured with a given distance metric, between this image modification and the input image, the better the value of the cost function. It is to be expected that with a drastic change in the input image, an equally drastic change in the behavior of the image classifier can be achieved. However, if even a small change in the input image significantly changes the output of the image classifier, this shows that the image classifier is particularly sensitive to this change.

Here, for example, the distance can include a component that directly compares the input image with the image modification and depends on the result of this comparison. Alternatively or in combination with this, the distance can contain a component, for example, which compares classification scores determined by the image classifier for the input image on the one hand and for the image modification on the other hand and depends on the result of this comparison. In particular, with the latter component of distance, the search for image modifications can be focused on those image modifications that change the uncertainty of the class assignments made by the image classifier, but leave those class assignments essentially unchanged.

Furthermore, the value of the cost function becomes better the larger the distances between the representation modification, on which this image modification is based, and further representation modifications, measured with a predetermined distance metric, are. In particular, this contribution ensures that the image modifications selected as decision-relevant cover different aspects of the input image instead of all being focused on a single such aspect. For this purpose, it is measured how well the area in the representation space in which representation modifications are searched for is covered by the representation modifications used.

This is roughly comparable to the fact that a potential buyer of a property is likely to be skeptical if he is only presented with photos from the living room of that property. The suspicion then arises that only this living room has been particularly renovated and spruced up, while the rest of the property clearly leaves a lot to be desired. It is much more confidence-inspiring if the buyer is presented with photos from all areas of the property.

The distance metric can, for example, include distances in pairs between the representation modification just examined on the one hand and the other representation modifications on the other. Alternatively or in combination with this, the distance metric can, for example, comprise mutual information (transinformation) in pairs between the representation modification just examined on the one hand and the further representation modifications on the other. In particular, a differentiable implementation or approximation of this transinformation can be used, for example.

haben. Hierin bezeichnet z* alle anderen Repräsentations-Abwandlungen als diejenige Repräsentations-Abwandlung z („Kandidat“), die aktuell bewertet werden soll. H ist ein, vorzugsweise differenzierbares, Maß für die Unsicherheit der Ausgabe des Bildklassifikators, wie hier beispielsweise eine Klasse y, der der Bildklassifikator eine aus einer Repräsentations-Abwandlung z erzeugte Bild-Abwandlung µ_θ (x | z) primär zuordnet. Hier soll die Notation µ_θ(x | z) verdeutlichen, dass die Bild-Abwandlung x durch den Decoder der trainierten Encoder-Decoder-Anordnung als Vorhersagemittel erzeugt wird unter der Voraussetzung, dass die Repräsentations-Abwandlung z ist.Starting from an input image x ₀ , the cost function L dependent on the representation modifications z can, for example, have the form $$L\left(\mathrm{e.g},{\mathrm{e.g}}^{\ast }\right)=H\left(y|{\mu }_{\theta }\left(x|\mathrm{e.g}\right)\right)+\mathrm{i.e}({\mu }_{\theta }\left(x|\mathrm{e.g}\right),x_0)-\lambda \cdot \mathrm{i.e}\text{'}\left(\mathrm{e.g},{\mathrm{e.g}}^{\ast }\right)$$

have. Here z* denotes all representation variants other than the representation variant z (“candidate”) that is currently to be evaluated. H is a preferably differentiable measure for the uncertainty of the output of the image classifier, such as a class y here, for example, to which the image classifier primarily assigns an image modification μ _θ (x | z) generated from a representation modification z. Here, the notation µ _θ (x|z) is intended to clarify that the image modification x is generated by the decoder of the trained encoder-decoder arrangement as a prediction means, provided that the representation modification is z.

d'(z, z*) misst, wie „weit verstreut“ die Kandidaten z und z* sind. Zu diesem Zweck kann beispielsweise

• • eine Summe der paarweisen Distanzen über alle Elemente von z* gebildet werden, wobei diese einzelnen Distanzen dann beispielsweise wieder mit einem der bereits erwähnten Ähnlichkeitsmaße gebildet werden können; und/oder
• • eine Summe der paarweisen gegenseitigen Information (Transinformation) über alle Elemente von z* gebildet werden, wobei dann für die gegenseitige Information idealerweise eine differenzierbare Näherung verwendet wird.

d(µ _θ (x | z), x ₀ ), d(x, x ₀ ) for short, denotes the distance between the image modification µ _θ (x | z) and the input image x ₀ . As explained before, this distance d(x, x ₀ ) can contain two contributions d x and _{d y} measured in the space of the images x, or in the space of the classification scores f( _x ): $$\mathrm{i.e}\left(x,x_0\right)={\lambda }_x{\mathrm{i.e}}_x\left(x,x_0\right)+{\lambda }_y{\mathrm{i.e}}_y(f\left(x\right),f\left(x_0\right)).$$

d'(z,z*) measures how "dispersed" the candidates z and z* are. For this purpose, for example

• • a sum of the paired distances over all elements of z* is formed, with these individual distances then being able to be formed again, for example, using one of the similarity measures already mentioned; and or
• • a sum of the mutual information in pairs (transinformation) is formed over all elements of z*, in which case a differentiable approximation is then ideally used for the mutual information.

Minimizing the cost function L(z,z*) leads to one or more optimal representation modifications z ^# , and these one or more optimal representation modifications z ^# then result in one or more optimal image modifications x ^# = µ _θ ( x | z ^# ).

In particular, the term d′(z, z*) in the cost function L(z, z*) has the effect that representation modifications z ^# that are as dissimilar as possible to one another are generated from the outset as part of the optimization. This is illustrated with a simple example. For example, if there is a similarity matrix M that tells each pair of two candidates z how similar these candidates z are, with smaller numerical values corresponding to greater similarity, d'(z, z*) can be defined as the sum of all matrix elements of M be calculated.

In a further particularly advantageous embodiment, the representation transformations are sampled from a predetermined environment around the input representation. For example, the environment can be a hypersphere with a given radius δ around the input representation. This favors the image modifications being sufficiently similar to the input image and at the same time belonging to the desired distribution to which the input image also belongs.

In a further advantageous refinement, one or more further representation modifications are determined from at least one representation modification by means of an optimization step carried out on the basis of the evaluation by the cost function. In particular, starting from a number of initial candidates for representation modifications, for example, optimizations can be carried out in each case in order to arrive at a number of optimal image modifications which relate to different aspects of the input image. In this case, these optimizations do not necessarily have to be carried out one after the other and independently of one another. Instead, the optimizations can be coupled to one another by the term d'(z,z*) in the cost function L(z,z*). A joint optimization can therefore take place in which each candidate z is not only based on its own advantages history, but also depends on the positions of all other candidates z* in the space Z of representations.

In this case, the optimization step can be performed, for example, using a gradient of the cost function after the representation modification. This can then optionally be restricted such that the optimization step is only carried out if the specified environment around the input representation would not be left. However, optimization steps can also be performed according to any other optimization method based on the evaluation by the cost function. For example, a gradient descent method can be supplemented with a step size control that increases the step size of optimization steps if successive optimization steps consistently point in the same direction.

In a further particularly advantageous embodiment, one of these representation modifications is rejected in response to the fact that two representation modifications are similar to one another according to a predefined degree of similarity. For this purpose, the degree of similarity can be compared, for example, with an absolute or relative threshold value. In this way, redundant representation variations that ultimately say the same thing are filtered out. A large number of found optimal representation modifications, and thus optimal image modifications, no longer deceives the fact that these modifications all cluster around one and the same local minimum of the cost function. Instead, the probability is increased that several different optimal representation modifications or image modifications actually relate to different local minima of the cost function.

The degree of similarity can include, for example, a distance between the two representation modifications. This distance can be measured, for example, in the form of a Euclidean distance and/or a cosine distance between the two representation variants. In particular, two representation modifications can be evaluated as similar, for example, if they are less than a predetermined distance ε away from one another.

The degree of similarity can also contain, for example, a distance between image modifications onto which the decoder of the encoder-decoder arrangement maps the two representation modifications. The similarity is then assessed directly in the room in which the optimal image modifications are ultimately also evaluated.

In a further particularly advantageous embodiment, the representation modification to be discarded is selected probabilistically. It can thus be specified, for example, that one representation modification is rejected with a probability of p and the other with a probability of 1-p. Alternatively or in combination with this, it can continue to be designed probabilistically as to whether a representation modification is discarded at all or whether both representation modifications are retained.

Furthermore, for example, a representation modification can also be discarded in response to the fact that the image modification generated therefrom does not lead to any improvement in the task cost function compared to the original input image. This, as well as the discarding of one of two representation modifications that are too similar to one another, is particularly advantageous in connection with a continuous optimization of the representation modifications. No further optimization time is then used for representation modifications from which no further gain in knowledge can be expected. A budget of optimization time that is available overall can therefore be focused on those representation modifications that together provide as complete a picture as possible of the behavior of the image classifier.

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 zur Bestimmung von Bildanteilen, auf die ein Bildklassifikator 1 seine Entscheidung stützt;
• 2 Anwendungsbeispiel des Verfahrens 100 in der Qualitätskontrolle.

It shows:

• 1 Exemplary embodiment of the method 100 for determining image parts on which an image classifier 1 bases its decision;
• 2 Application example of method 100 in quality control.

1 1 is a schematic embodiment of the method 100 for examining the behavior of an image classifier 1, which assigns an input image 2 to one or more classes of a given classification. in the in 1 In the example shown, the method 100 is used to determine image parts on which the image classifier 1 bases its decision to allocate an input image 2 to one or more classes of a specified classification.

In step 110, the input image 2 is mapped onto an input representation 4 using an encoder 3a of a trained encoder-decoder arrangement 3 .

In step 120, representation modifications 4' of this input representation 4 are determined.

In step 130, these representation modifications 4' are mapped onto image modifications 2' of the input image 2 using the decoder 3b of the encoder-decoder arrangement 3.

In step 140, each image modification 2' is evaluated using a predetermined cost function 5.

According to block 141, the better the value of a predetermined task cost function 141, which depends on a processing result determined by the image classifier 1 from the image modification 2', the better the value 5a of the cost function 5 becomes.

According to block 142, the smaller a distance between the examined image modification 2' and the input image 2, measured with a predetermined distance metric, the better the value 5a of the cost function 5.

According to block 143, the value 5a of the cost function 5 improves the larger the distances between the representation modification 4′, on which the examined image modification 2′ is based, and further representation modifications 4′ are measured with a predetermined distance metric.

In step 150, image modifications 2* relevant to the decision are selected according to a predetermined criterion. The predetermined criterion depends on the evaluation 5a by the cost function 5.

In step 160, from the selected, decision-relevant image modifications 2*, the searched-for image parts 2**, on which the image classifier (1) bases its decision, are evaluated.

According to block 121, the representation modifications 4' can be sampled from a predetermined environment around the input representation 4.

According to block 122, one or more further representation modifications 4' can be determined for at least one representation modification 4' by at least one optimization step performed by the cost function 5 on the basis of the evaluation 5a. Such an optimization step can be carried out according to block 122a, for example using a gradient of the cost function 5 after the representation modification 4',

According to block 123, in response to the fact that two representation modifications 4' are similar to one another according to a predetermined degree of similarity, one of these representation modifications 4' can be discarded.

The representation modification 4' to be discarded can be selected probabilistically according to block 123a.

According to block 123b, the measure of similarity can include a distance between the two representation modifications 4'.

According to block 123c, the degree of similarity can contain a distance between image modifications 2, onto which the decoder 3b of the encoder-decoder arrangement 3 maps the two representation modifications 4'.

According to block 124, at least one representation modification 4' may be discarded in response to the resulting image modification 2' resulting in no improvement in the task cost function over the input image 2.

According to block 141a, the task cost function can be an uncertainty measured with a predetermined uncertainty metric with which the image class sifikator 1 assigns the image modification 2' to one or more classes of the given classification.

According to block 141b, the task cost function may include one or more classification scores related to classes of the predetermined classification that the image classifier 1 determines for the image modification 2'.

According to block 144, only such image modifications 2' can be considered for which a highest classification score determined by image classifier 1 relates to the same class as the highest classification score determined by image classifier 1 for input image 2.

2 shows a simple application example from material testing. In this example, the image classifier 1 is designed to classify input images 2 into the binary classes “OK” and “not OK=NOK”.

The input image 2 shows a screw nut 10 with a thread 11 in the middle and a crack 12 that can be seen. Since the crack 12 is not yet clearly visible, the input image 2 is evaluated by the image classifier 1 used for quality control still classified in the "OK" class, but only with a comparatively weak classification score of 0.6.

In step 110 of the method 100, the input image 2 is transformed to the input representation 4 in the space Z of realistic image representations.

In step 120, a modified representation 4' is formed in this space Z from the input representation 4.

In step 130, the representation modification 4' is transformed into an image modification 2'. The crack 12 can no longer be seen in this image modification 2'. Therefore, image modification 2' is still classified as OK by image classifier 1, but with a significantly higher classification score of 0.97. The evaluation with the cost function in step 140 reveals this, and subsequently the image modification 150 is selected as the decision-relevant image modification 2*. The evaluation in step 160, for example by comparison with the input image 2, identifies the area in which the crack is missing in the image modification 2' as a decision-relevant image portion.

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 2018197074 A1 [0003]

Claims (18)

Method (100) for examining the behavior of an image classifier (1), which assigns an input image (2) to one or more classes of a specified classification, with the steps: • the input image (2) is mapped (110) to an input representation (4) using an encoder (3a) of a trained encoder-decoder arrangement (3); • Representation modifications (4') of this input representation (4) are determined (120); • these representation modifications (4') are mapped (130) to image modifications (2') of the input image (2) with the decoder (3b) of the encoder-decoder arrangement (3); • each image modification (2') is evaluated (140) with a predetermined cost function (5), the value (5a) of which becomes the better, ◯ the better the value of a predetermined task cost function (141), which depends on a processing result determined by the image classifier (1) from the image modification (2'); ◯ the smaller a distance measured with a predetermined distance metric is between this image modification (2') and the input image (2) (142); and ◯ the larger the distances between the representation modification (4') measured with a predetermined distance metric, to which this image modification (2') is based, and further representation modifications (4') are (143); • Decision-relevant image modifications (2*) are selected (150) according to a predetermined criterion, which depends on the evaluation (5a) by the cost function (5). Method (100) according to claim 1 , image components (2**) on which the image classifier (1) bases its decision being evaluated (160) from the selected, decision-relevant image modifications (2*). Method (100) according to any one of Claims 1 until 2 , wherein the task cost function includes an uncertainty measured with a predetermined uncertainty metric, with which the image classifier (1) assigns the image modification (2') to one or more classes of the predetermined classification (141a). Method (100) according to any one of Claims 1 until 3 , wherein the task cost function includes (141b) one or more classification scores relating to classes of the predetermined classification which the image classifier (1) determines for the image modification (2'). Method (100) according to any one of Claims 1 until 4 , whereby only such image modifications (2') are considered (144) for which a highest classification score determined by the image classifier (1) relates to the same class as the highest by the image classifier (1) for the input Image (2) determined classification score. Method (100) according to any one of Claims 1 until 5 , wherein the representation modifications (4') are sampled (121) from a predetermined environment around the input representation (4). Method (100) according to any one of Claims 1 until 6 , wherein one or more further representation modifications (4') are determined (122) from at least one representation modification (4') by at least one optimization step guided by the cost function (5) on the basis of the evaluation (5a). Method (100) according to claim 7 , wherein the optimization step is performed (122a) using a gradient of the cost function (5) after the representation modification (4'). Method (100) according to any one of Claims 1 until 8th , wherein in response to the fact that two representation modifications (4') are similar to one another according to a predetermined degree of similarity, one of these representation modifications (4') is discarded (123). Method (100) according to claim 9 , wherein the representation modification (4') to be discarded is selected probabilistically (123a). Method (100) according to any one of claims 9 until 10 , wherein the measure of similarity includes a distance between the two representation modifications (4') (123b). Method (100) according to any one of claims 9 until 11 , wherein the measure of similarity includes a distance between image modifications (2) onto which the decoder (3b) of the encoder-decoder arrangement (3) maps the two representation modifications (4') (123c). Method (100) according to any one of Claims 1 until 12 wherein at least one representation modification (4') is discarded (124) in response to the image modification (2') generated therefrom resulting in no improvement in the task cost function over the input image (2). Method (100) according to any one of Claims 1 until 13 , wherein the input image (2) comprises an image of a manufactured product and wherein the image classifier (1) is designed to classify the input image (2) with regard to a quality assessment of the product. Method (100) according to any one of Claims 1 until 13 , wherein the input image (2) comprises an image of a traffic situation or an image of a monitored area and wherein the image classifier (1) is designed to classify the input image (2) with regard to objects contained therein. 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 15 to execute. Machine-readable data carrier with the computer program Claim 16 . One or more computers with the computer program after Claim 16 , and/or with the machine-readable data medium Claim 17 .

Images