Generating composite image for training machine learning model
By gradually adding noise using a diffusion model to generate style-appropriate synthetic images, the problem of time-consuming and costly annotation training examples is solved, improving the training efficiency and effectiveness of machine learning models and reducing domain offset.
Patent Information
- Application Number
- CN202510629148.X
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Priority Date
- 2024-05-15
- Filing Date
- 2025-05-15
- Publication Date
- 2025-11-18
AI Technical Summary
In the training process of existing machine learning models, the process of labeling training examples is time-consuming, costly, and the difference in domain and distribution between synthetic images and physically recorded images leads to poor training results.
A diffusion model is used to train images by progressively adding noise, and then synthesize images that conform to the style by combining them with pre-given conditions. The model parameters are optimized by a cost function to ensure that the synthesized images have the same style features as the physically recorded images.
The generated synthetic images can effectively supplement the physical recording images, improve training efficiency and effectiveness, reduce domain offset, and enhance the generalization ability of machine learning models.
Smart Images
Figure CN120976031A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to generating synthetic images that can be used as training examples for machine learning models and, in particular, can alleviate the shortage of training examples “annotated” with prior knowledge. Background Technology
[0002] Machine learning models are increasingly being used to evaluate images, particularly in environmental monitoring where vehicles or robots are at least partially autonomous in factory areas or public transportation. These models possess the advantageous property that, once trained, they can generalize to images not seen during training, based on a limited number of training examples. This, in the broadest sense, simulates the learning process of a human driver who, after only a few dozen hours and less than 1000 kilometers of driving experience, has encountered a very limited selection of traffic situations. Even after such limited training, a driver can typically still grasp situations not seen during training.
[0003] Machine learning model training is typically performed using a monitoring approach. That is, training examples are "labeled" with prior knowledge in the form of a predetermined output from which the machine learning model should ideally generate the predetermined output. Training progress is then measured by how well the machine learning model's average output across all training examples matches the predetermined output.
[0004] Labeling training examples is essentially a manual process, and therefore greatly affects the training time and cost. Summary of the Invention
[0005] This invention provides a method for training a diffusion model. The diffusion model itself transforms a statistical distribution (e.g., noise with a normal distribution) into another distribution (e.g., the distribution of a realistic-looking image). Combined with pre-given conditions, such as text or semantic segmentation, the diffusion model can iteratively generate synthetic images that meet those conditions. Therefore, for example, text input can be given in advance as conditions to generate synthetic images with pre-given content. In this sense, the diffusion model can therefore be constructed to iteratively generate synthetic images that meet the pre-given conditions from noise.
[0006] Within the scope of this method, a pre-defined style (S ti1) is given for the synthesized images. A set of training images x0 is provided, which correspond to the pre-defined style to varying degrees.
[0007] Noise is progressively added to the training image x0 in a predetermined number of iterations T, thereby generating noisy versions x1, ..., x2. T From the noisy versions x1, ..., xT Sample x t The diffusion model, combined with pre-given conditions, will extract the sample x. t Processed for each of the previous noisy versions x t-1 Prediction
[0008] Evaluate the prediction using a pre-given cost function. Compared to the actual noisy version x t-1 Consistency. To achieve the goal of optimizing the parameters characterizing the behavior of the diffusion model, such that further processing of the training image x0 and the resulting samples x... t The evaluation is improved by using the aforementioned cost function.
[0009] When the sample x is extracted t When, and / or when the resulting prediction is evaluated through the cost function At that time, the more the corresponding training image x0 corresponds to the pre-given style, the more likely these samples x can still identify the style of the corresponding training image x0. t The more representative it is, the better.
[0010] Recognizing that, in this way
[0011] • Diffusion models can be trained to generate synthetically produced images that conform to a pre-given style.
[0012] • This does not require restricting the training examples to this pre-given style from the beginning.
[0013] Generating synthetic images with a specific pre-given style improves the applicability of these synthetically generated images as training examples for training machine learning models. For such training, synthetically generated images are typically not used alone; instead, they are usually used to supplement a finite number of existing training examples of physical records. For optimal training, the synthetically generated training examples should belong to the same domain and / or distribution as the training examples of physical records. The characteristics of the training examples of physical records typically lie in the specific properties of the image records.
[0014] For example, if images are recorded using a camera mounted on a vehicle, the images may not be as perfect as those recorded, for instance, with a professional camera used for feature film shooting, due to the camera's limited size. Now, synthetically generated images may, in this sense, be "too perfect," meaning they are far superior in quality to what could be achieved with a vehicle-mounted camera. Such synthetically generated images therefore do not belong to the domain and / or distribution of the physically recorded images, but rather create a domain shift. In contrast, the method proposed herein can generate images that are significantly more similar to existing physically recorded images.
[0015] The same applies when synthetic images have already been generated from other sources and the inventory should be meaningfully replenished. Methods for synthetically generating images can also impart their own style to the images, for example, in the form of characteristic artifacts.
[0016] In principle, it's possible to force the generation of images to conform to a specific style by restricting the training examples to those corresponding to a pre-defined style from the outset. This would sacrifice a significant portion of the total available training examples. However, it's recognized that as noise is progressively added to the training images, style-related information becomes unrecognizable faster than content-related information. While it's still possible to identify what should be displayed in the image over a relatively long period with increasing noise, it becomes relatively quick to lose the ability to identify which camera captured the image.
[0017] Therefore, for example, for a training image x0 that does not correspond to a pre-given style, iterative x... t Sampling is performed, and the noise-adding process continues, making it impossible to definitively reconstruct the style from it. This allows training the diffusion model to reconstruct the basic ability to reconstruct content with greater variability. However, iterative styles x can be definitively reconstructed from it. t Therefore, only the training image x0 corresponding to the pre-given style can be sampled. Thus, whenever the diffusion model reconstructs a style, this always occurs only for the training image x0 that matches the style.
[0018] Alternatively or in combination with this, the sample x, which can still be clearly identified as having an "incorrect" style, can also be suppressed in the cost function. t The impact on the training results of the diffusion model. Whether modifying the cost function or modifying the sampling is easier to implement depends on the specific application.
[0019] In a particularly advantageous design, the set of training images x0 is divided into a correct subset of training images x0 corresponding to a pre-given style and an incorrect subset of training images x0 not corresponding to the pre-given style. When the sample x is drawn... t When, and / or when the resulting prediction is evaluated through the cost function Even at that time, it can still identify samples x with the style of the corresponding training image x0. t It is only considered if it originates from the training image x0 in the correct subset. As mentioned earlier, in this way, the information content of the training image x0 in the incorrect subset can also be optimally utilized.
[0020] For this purpose, a threshold S can be specified, for example, up to which samples x have t ≤ S. tIt can still identify the style of the corresponding training image x0. A threshold S can be quickly found; samples x with t > S are considered valid if their values are above this threshold in any case. t All style information is lost. Within the scope of this method, it is not a problem to set the threshold S too high. This will only exclude some contributions from the training image x0 from the incorrect subset, but will not change the style of the generated image to correspond to the pre-given style as desired.
[0021] When a training image x0 is noise-added, for example, in T = 1000 iterations, a threshold S = 200 iterations can be specified. Below this threshold, samples x0 with t ≤ S are considered valid. t It can still identify the style of the corresponding training image x0.
[0022] To optimize the threshold S, in a further particularly advantageous design, multiple candidate thresholds S are considered. * Test whether it is still possible to obtain samples The style of the corresponding training image x0 is explicitly determined. For this test, a classifier can be used, for example, which is constructed based on one or more styles as samples. Assigning category scores. For example, when comparable category scores are assigned to multiple different styles, the judgment of a particular style becomes unclear.
[0023] The pre-given style can, in particular, represent a transfer function that translates the semantic content of the image into the image itself. Therefore, it can involve a process for generating the corresponding image, and specifically, include, for example, the traces left by this process in the training image x0. Synthetic images that appear to have been obtained using the same process as the training image x0 can then be generated particularly well using this method.
[0024] This also applies to another particularly advantageous design, in which a pre-defined style characterizes a device for recording images and / or an algorithm for synthetically generating image x0. For example, the style could characterize a camera used for image recording or roughly describe a method for synthetically generating images.
[0025] This style definition differs from common usage in machine learning, which essentially distinguishes between semantic content and style. According to this usage, factors such as the object's color or material, lighting conditions, time of day, and season also fall under the category of style. However, more precisely, these are elements of "semantic style," related to the characteristics of a specific object rather than the entire imaging process. In the context of the proposed method, the primary focus is on preserving the generative style of the training image x0, regardless of whether this generation is achieved through a physical imaging system (e.g., a camera) or an algorithm.
[0026] Therefore, the pre-given style, in particular, includes, for example,
[0027] • Image distortion, and / or
[0028] • Blurry focus, and / or
[0029] • Coloring and / or color deviation, and / or
[0030] • One or more textures, and / or
[0031] • One or more artifacts that appear when the training image x0 is generated.
[0032] Among a large number of training images x0 that mix multiple styles, only a relatively small number of training images x0 correspond to a pre-given style. Therefore, for most training images x0, a noisy version x can be expected. t This is restricted to iterative indices t where the style becomes truly unrecognizable due to noise. This can lead to a situation where, in samples x drawn from all training images x0... t Of the total number of iterations, the lower iteration indices t belonging to the weaker noisy versions are underrepresented. To counteract this trend, in a further particularly advantageous design scheme, [the following is omitted as the text is incomplete and requires further context].
[0033] • Used to extract samples x that can still be identified in the style of the corresponding training image x0 t frequency, and / or
[0034] • Used to extract samples x that trace back to a training image x0 with a pre-given style t The frequency of the sample is adapted such that the total iterative index t of the sampled samples is distributed according to a pre-given distribution. This pre-given distribution can be, in particular, a uniform distribution or a normal distribution.
[0035] In another particularly advantageous design, the pre-given conditions include
[0036] • The training image x0 consists of objects, and / or
[0037] • The edges of training image x0, and / or
[0038] • Additional information about the layout of the training image x0.
[0039] In this way, variants of the training image x0 can be selectively generated, possessing the same spatial layout and / or semantic content but displaying that content in different ways. Here, the synthesized images always belong to the same domain and / or distribution as those generated in the same way as the original training image x0. Therefore, the synthesized images are particularly well-suited as training examples for machine learning models. Especially during monitored training of such models, the labels of the training image x0 can be reused in the form of nominal outputs that the machine learning model should generate from the training image x0.
[0040] Once the diffusion model has completed training, in another particularly advantageous design, noise samples are drawn from the noise distribution and fed into the trained diffusion model, along with pre-given conditions. This produces a synthetically generated image. Following the method suggested here, the synthetically generated image corresponds to a pre-given style.
[0041] As mentioned earlier, these synthetically generated images are particularly well-suited as training examples for machine learning models. Therefore, in another particularly advantageous design, synthetically generated images are used as training examples to train the machine learning model. Specifically, the synthetically generated images better integrate into the domain and / or distribution of the existing training examples. In this way, the synthetically generated training examples are a genuine aid to the ongoing training, rather than a distraction that leads the training in a different direction than intended due to domain shifts. Machine learning models are typically trained for a specific task and are therefore also called task models.
[0042] In another particularly advantageous design, an input image recorded by at least one sensor is fed to a machine learning model that has been trained in this way. The output subsequently provided by the machine learning model forms a control signal. This control signal is used to control vehicles, driver assistance systems, robots, systems for quality control, systems for area monitoring, and / or systems for medical imaging. Based on this improved training, the likelihood that the controlled system will respond appropriately to the control signal in the context embodied in the input image is increased.
[0043] This method can be implemented, in particular, wholly or partially, by a computer. Therefore, the invention also relates to a computer program having machine-readable instructions that, when executed on one or more computers and / or computing instances, cause those machine-readable instructions to perform the described method. In this sense, control devices for vehicles and embedded systems for technical devices, also capable of executing machine-readable instructions, should also be considered computers. Computing instances can be, for example, virtual machines, containers, or serverless execution environments, which can be provided, in particular, in the cloud.
[0044] Similarly, the present invention also relates to machine-readable data carriers and / or downloadable products having the aforementioned computer program. A downloadable product is a digital product that can be transmitted via a data network, i.e., downloaded by a user of the data network, and can be sold instantly for download, for example, in an online store.
[0045] In addition, one or more computers and / or computing instances may be equipped with the computer program, the machine-readable data carrier, or the download product. Attached Figure Description
[0046] Other improvements to the invention are shown in more detail below with reference to the accompanying drawings and a description of preferred embodiments thereof.
[0047] Figure 1 An embodiment of a method 100 for training diffusion model 1 is shown;
[0048] Figure 2 Examples illustrating the impact of increased noise on image style recognizability;
[0049] Figure 3 This demonstrates weakly noisy iterations x only for training images x0 that conform to a pre-given style 5. t A schematic diagram. Detailed Implementation
[0050] Figure 1 This is a schematic flowchart of an embodiment of a method 100 for training a diffusion model 1. The diffusion model 1 can be used to iteratively generate a synthetic image 4 from noise 2, in conjunction with pre-given conditions 3.
[0051] In step 110, a style 5 is pre-given, which is the style that the image 4 synthesized by the completed diffusion model 1 should have.
[0052] According to box 111, the pre-given style 5 can characterize the transfer function that transforms the semantic content of the image into the image.
[0053] According to box 112, the pre-given style 5 can characterize the device used to record the image and / or the algorithm used to synthesize the image.
[0054] According to box 113, the pre-given style 5 can include
[0055] • Image distortion, and / or
[0056] • Blurry focus, and / or
[0057] • Coloring and / or color deviation, and / or
[0058] • One or more textures, and / or
[0059] • One or more artifacts that appear when the training image x0 is generated.
[0060] In step 120, a training image set x0 is provided, which corresponds to a pre-given style 5 to varying degrees.
[0061] According to box 121, the training image set x0 can be divided into a correct subset R of those training images x0 that correspond to a pre-given style 5 and an incorrect subset F of those training images x0 that do not correspond to a pre-given style 5.
[0062] In step 130, the training image x0 is progressively subjected to noise 2 in a predetermined number of iterations T, thereby forming noisy versions x1, ..., x2 respectively. T .
[0063] In step 140, from the noisy versions x1, ..., x T Sample x t .
[0064] In step 150, the extracted sample x is processed by diffusion model 1 in accordance with the pre-given condition 3. t Process into pairs of previous noisy versions x t-1 Prediction
[0065] According to box 151, the pre-given condition 3 includes
[0066] • The training image x0 consists of objects, and / or
[0067] • The edges of training image x0, and / or
[0068] • Additional information about the layout of the training image x0.
[0069] According to box 152, the pre-given condition 3 may include features of the training image x0, which should be determined by the machine learning model 8 to be trained, and prior knowledge exists regarding these features for the monitored training of the machine learning model 8. In this way, enhanced versions of the training image x0 can be generated, for which the label of the training image x0 can continue to be used.
[0070] In step 160, these predictions are evaluated using a pre-given cost function 7. Compared with the actual noise-added version x t-1 Consistency. Formation of assessment 7a.
[0071] In step 170, to achieve the objective, the parameter 1a representing the behavior of diffusion model 1 is optimized, such that in further processing of the training image x0 and the resulting sample x t The cost function is used to improve the evaluation of 7a. The state of optimization of parameter 1a is marked with the attached figure 1a. * This indicates and determines the completed training state 1 of diffusion model 1. * .
[0072] When 140 samples are drawn x t When, and / or when the resulting forecast is evaluated through cost function 7, 160 At that time, the more the corresponding training image x0 corresponds to the pre-given style 5, the more likely these samples x can still identify the style of the corresponding training image x0. t The more representative it is, the better.
[0073] Specifically, for example, according to boxes 141 or 161, this could mean that when 140 samples x are drawn... t When, and / or when the resulting forecast is evaluated through cost function 7, 160 Even at that time, it can still identify samples x with the style of the corresponding training image x0. t It will only be considered if it originates from the training image x0 in the correct subset R formed according to box 121.
[0074] According to boxes 142 or 162, a threshold S can be specified up to which samples x have t ≤ S. t The style of the corresponding training image x0 can still be identified. To specify a threshold, specifically, for example, it is possible to...
[0075] • Based on box 142a or 162a, for multiple candidate thresholds S * Test, is it still possible to obtain samples? Clearly define the style of the corresponding training image x0, and
[0076] • Select the candidate threshold S that has been proven to be no longer possible, based on box 142b or 162b.* As the threshold S.
[0077] According to box 143 or 163, we can...
[0078] • Used to extract samples x whose style can still be identified from the corresponding training image x0 t frequency, and / or
[0079] • Used to extract samples x that trace back to a training image x0 with a pre-given style t The frequency of is adapted such that the iterative index t of the total number of samples drawn is distributed according to a pre-given distribution.
[0080] exist Figure 1 In the example shown, in step 180, samples of noise 2 from the noise distribution are fed to the trained diffusion model 1 along with pre-given conditions 3. The resulting synthesized image 4 is then formed.
[0081] In step 190, the synthesized image 4 is used as a training example to train a machine learning model 8 constructed to solve a pre-given task. The completed training state of the machine learning model is labeled with figure 8. * express.
[0082] In step 200, the trained machine learning model 8 is fed... * The input image 9, recorded by at least one sensor, is transmitted. This then produces the output 8a.
[0083] In step 210, control signals 210a are formed from these outputs 8a. In step 220, the control signals 210a are used to control the vehicle 50, the driver assistance system 51, the robot 60, the quality control system 70, the area monitoring system 80, and / or the medical imaging system 90.
[0084] Figure 2 Five examples (a) through (e) of the training image x0 are shown. These training images x0 clearly differ not only in their individual content but also in their generation style. Thus, for example, in example (a), it can be noticed that the image is noticeably distorted due to the fisheye effect of the camera used. The traffic conditions on the highway shown in example (b) appear "too good" at first glance for a photograph, and the lane textures exhibit artifacts typical of image synthesis. In example (c), image areas with light intensities below a certain value are completely black. Example (d) shows weakened color contrast. Example (e) shows focus blur, and image areas with light intensities above a certain value are completely white.
[0085] For each of the training images x0 Figure 2 The image shows the versions x with added noise. 50 x 100 and x 150 These are generated after 50, 100, or 150 progressively noisy iterations. Here, even after 150 noisy iterations, the basic semantic content of the training image x0 remains recognizable. However, the differences in generated styles are significantly diminished. For example, the fisheye effect in example (a) is almost imperceptible, the texture artifacts in example (b) are no longer visible, and the focus blur in example (e) is also masked by noise. Therefore, the more heavily noisy iterations x0... t It can be used on all training images x0 without "contaminating" the training of Model 1 with an "incorrect" style. Iterative x with weaker noise addition. t It should only be used on the training image x0 corresponding to the pre-given style 5.
[0086] Figure 3 This is illustrated in more detail in a schematic way. Figure 3 The example shown includes: three training images x0 that correspond to the pre-given style 5 and therefore belong to the correct set R formed in box 121; and two training images x0 that do not correspond to the pre-given style 5. And therefore belong to the error set F formed in box 121. All training images x0 are noisy in T iterations. A threshold S is defined such that the noisy iterations x t＞S It no longer contains any information about the generation style of the original training image x0, but the noisy iterative x... t≤S This generation style can still be identified.
[0087] Among all training images x0, consider the iterations x with stronger noise respectively. t＞S However, the weaker noisy iteration x is considered only if the corresponding training image x0 belongs to the correct set R. t≤S All extracted samples x t All samples are pooled and fed into the diffusion model 1 to be trained. Here, for the training image x0 from the correct set R, the weaker noisy sample x... t≤S Compare the noisy samples x t＞S As each is considered more individually, the total number of iterative indices t in the pool is roughly evenly distributed.
[0088] Diffusion Model 1 applies to each sample x in the pool. t For example, a previous iteration x with weaker noise. t-1 Create predictions separately In step 160 of method 100, the prediction Compared to the actual weaker noisy iteration x t-1 A comparison is made. The results of this comparison are evaluated according to the given cost function 7, and in step 170 of method 100, the feedback of parameter 1a, which characterizes the behavior of diffusion model 1, is determined. The optimized parameter 1a is then generated. * These parameters define the state 1 of the completed training of the diffusion model 1. * .
Claims
1. A method (100) for training a diffusion model (1), the method being able to generate a synthetic image (4) conforming to said pre-given conditions (3) in an iterative manner from noise (2), the method comprising the following steps: • The style (5) that the synthesized image (4) should have is given in advance (110); • Provide a set of (120) training images x0, which correspond to a pre-given style (5) to varying degrees; • Noise (2) is progressively applied to the training image x0 at a predetermined number of iterations T, thereby generating noisy versions x1, ..., x T ; • From the noisy versions x1, ..., x T Draw (140) samples x t ; • Combining the pre-given conditions (3), the extracted sample x is obtained by the diffusion model (1). t Processing (150) for each of the previous noisy versions x t-1 Prediction • Evaluate the prediction (160) using a pre-given cost function (7). Compared to the actual noisy version x t-1 Consistency; and • To achieve the goal of optimizing (170) the parameters (1a) characterizing the behavior of the diffusion model (1), such that in further processing of the training image x0 and the samples x0 generated therefrom, t The evaluation (7a) is improved by the aforementioned cost function. When the sample x described in (140) is drawn t When, and / or when the resulting prediction is evaluated by the cost function (7) (160) When the corresponding training image x0 corresponds more closely to the pre-given style (5), the samples x that can still identify the style of the corresponding training image x0 are... t The more representative it is, the better.
2. The method (100) according to claim 1, wherein The set of training images x0 is divided (121) into a correct subset (R) of training images x0 that correspond to a pre-given style (5) and an incorrect subset (F) of training images x0 that do not correspond to a pre-given style (5), and • When the sample x of (140) is drawn t When, and / or when the resulting prediction is evaluated by the cost function (7) (160) Even at that time, it can still identify samples x with the style of the corresponding training image x0. t It is only considered when it originates from the training image x0 in the correct subset (R) (141, 161).
3. The method (100) according to any one of claims 1 to 2, wherein a threshold S is defined (142, 162) until said threshold, for samples x having t ≤ S. t It can still identify the style of the corresponding training image x0.
4. The method (100) according to claim 3, wherein • For multiple candidate thresholds S * Tests (142a, 162a) are still available from the samples. Clearly define the style of the corresponding training image x0, and • Select the candidate threshold S* that is proven to be no longer possible (142b, 162b) as the threshold S.
5. The method (100) according to any one of claims 1 to 4, wherein the pre-given style (5) characterizes (111) a transfer function that converts the semantic content of an image into an image.
6. The method (100) according to any one of claims 1 to 5, wherein the pre-given style (5) characterizes (122) the device used to record the image and / or the algorithm used to synthesize the image.
7. The method (100) according to any one of claims 1 to 6, wherein the pre-given style (5) includes (113). • Image distortion, and / or • Blurry focus, and / or • Coloring and / or color deviation, and / or • One or more textures, and / or • One or more artifacts that appear when the training image x0 is generated.
8. The method (100) according to any one of claims 1 to 7, wherein... • Used to extract samples x that can still be identified in the style of the corresponding training image x0 t frequency, and / or • Used to extract samples x that trace back to a training image x0 with a pre-given style t frequency, The fit (143, 163) is such that the iterative index t of the total number of samples drawn is distributed according to a pre-given distribution.
9. The method (100) according to any one of claims 1 to 8, wherein the pre-given condition (3) includes (151). • The training image x0 consists of objects, and / or • The edges of training image x0, and / or • Additional information about the layout of the training image x0.
10. The method (100) according to any one of claims 1 to 9, wherein the pre-given condition (3) includes (152) characteristics of the training image x0, the characteristics being determined by the machine learning model (8) to be trained and having prior knowledge for monitored training of the machine learning model (8) for the characteristics.
11. The method (100) according to any one of claims 1 to 9, wherein samples of noise (2) from the noise distribution are fed together with pre-given conditions (3) to a trained diffusion model (1) to produce a synthetically generated image (4).
12. The method (100) of claim 11, wherein the synthetically generated image (4) is used as a training example to train (190) the machine learning model (8).
13. The method (100) according to claim 12, wherein • Input images (9) recorded by at least one sensor (10) are fed (200) into the trained machine learning model (8*); • A control signal (210a) is formed (210) from the output (8a) subsequently provided by the machine learning model (8*); and • The control signal (210a) is used to control (220) the vehicle (50), the driver assistance system (51), the robot (60), the system for quality control (70), the system for area monitoring (80), and / or the system for medical imaging (90).
14. A computer program comprising machine-readable instructions that, when executed on one or more computers and / or computing instances, cause the one or more computers and / or computing instances to perform the method (100) according to any one of claims 1 to 13.
15. A machine-readable data carrier and / or downloadable product having the computer program according to claim 14.
16. One or more computers and / or computing instances having a computer program as claimed in claim 14, and / or having a machine-readable data carrier and / or downloadable product as claimed in claim 15.