DETECTION OF DEFECT CENTERS FOR LITHOGRAPHY MASKS

Abstract

The present disclosure relates to the localization of defects in images, e.g., aerial images, representing a lithography mask. A defect center probability map (230) is predicted based on measured image data (205) of the lithography mask and reference image data (210) of the lithography mask, wherein each pixel of the defect center probability map (230) includes a probability score representing a probability of being a defect center. A method includes identifying pixels with the highest probability values in a local neighborhood as defect coordinates (250) and in the defect center probability map, and returning the one or more defect coordinates, optionally together with their corresponding probability values.

Description

FIELD OF EXPERTISE

[0001] The present disclosure relates generally to the field of sample inspection and, more particularly, to techniques for automatic defect detection in lithography mask inspection workflows. BACKGROUND

[0002] In semiconductor manufacturing, a lithography mask (or simply "mask") inspection workflow enables the evaluation and assurance of the quality of masks used to fabricate semiconductor structures. The process begins with high-resolution imaging—e.g., using a charged particle microscope or light microscopy—to capture images of the lithography mask, for example, each part of the lithography mask used to expose a specific chip. These images are then analyzed with automated software that uses comparison algorithms to detect one or more physical defects such as cracks, scratches, or contamination. This automated inspection facilitates the identification of anomalies by comparing each mask or section of a mask to a reference.

[0003] The current state of the art relies on various approaches to detect one or more defects.

[0004] One category of methods uses "non-data-driven" techniques that have been in use for some time. These methods typically involve registering two images, calculating their difference, thresholding the result, and returning local clusters as potential defect candidates.

[0005] With the advent of deep learning, segmentation-based techniques have gained popularity. These methods use semantic segmentation, provided by a machine learning segmentation module, to assign each pixel of the difference image to either a defect or non-defect class. The resulting workflow involves predicting a semantic segmentation probability map from the difference image, connecting neighboring defect pixels into components, calculating the center of each component, and returning the center coordinates along with an average defect score for each component.

[0006] Detection-based techniques, often adopted from other application areas such as natural vision, predict bounding boxes for defect instances from the difference image, suppress nearly identical predictions by non-maximum suppression, and return the coordinates of the center along with a detection score for each detected instance.

[0007] The above approaches have several disadvantages. Non-data-driven methods often struggle with accuracy issues, leading to high fault rates. Fault rates refer to the frequency or ratio, or the relative or absolute number of false or misleading results, such as false alarms or misclassified defects, which can negatively impact the performance and effectiveness of a defect detection system. Segmentation-based approaches require complex post-processing steps that consume significant computational resources and energy. Detection-based approaches include non-maximum suppression steps, which can be computationally intensive; they can also pose problems for very small defect instances. SUMMARY

[0008] Therefore, it is an object of the present disclosure to realize a defect detection method for mask inspection that enables fast, reliable and precise defect detection.

[0009] This need is met by the features of the independent claims. The features of the dependent claims define embodiments.

[0010] The present disclosure relates to a method of defect detection for photolithography mask inspection workflows. The photolithography mask may have an aspect ratio between 1:1 and 1:4, preferably between 1:1 and 1:2, most preferably 1:1 or 1:2. The photolithography mask may have a nearly rectangular shape. The photolithography mask may preferably be 12.7 cm (5 inches) to 17.8 cm (7 inches) long and wide, most preferably 15.2 cm (6 inches) long and wide. Alternatively, the photolithography mask may be 12.7 cm (5 inches) to 17.8 cm (7 inches) long and 25.4 cm (10 inches) to 35.6 cm (14 inches) wide, preferably 15.2 cm (6 inches) long and 30.5 cm (12 inches) wide. A concept disclosed here is based on modeling defects as objects completely described by their center coordinate, thus simplifying the detection process and adapting it to the specific requirements of the application.Another concept disclosed here is based on a combination of semantic segmentation and detection models.

[0011] A method is disclosed. The method is used to detect one or more defects in a lithography mask. The method comprises predicting a defect center probability map based on (i) measured image data of the lithography mask and further based on (ii) reference image data of the lithography mask. Each pixel of the defect center probability map includes a probability value representing the probability of being the defect center. For example, a specific channel of the defect center probability map may contain a pixel value encoding this probability value.

[0012] The method further comprises identifying pixels of the defect center probability map that exceed a predefined minimum probability threshold and / or have the highest probability values in a local neighborhood as defect coordinates. Various distance metrics can be used to determine the local neighborhood.

[0013] The method further comprises returning said one or more defect coordinates, optionally together with their corresponding probability values.

[0014] Optionally, one or more defect coordinates can be transferred from a coordinate system associated with the defect center probability map to a coordinate system associated with the measured image data and / or the reference image data.

[0015] This method has several advantages over conventional approaches. By focusing exclusively on the detection of defect centers, for example, computational effort and energy consumption are reduced, making it particularly suitable for large-area mask inspections. Furthermore, the approach is more reliable and precise because it eliminates complex post-processing steps and non-maximum suppression.

[0016] In general, the pixels of the defect center probability map can have different sizes and / or different pixel grid positions compared to the pixels of the measured image data and the reference image data. This means that an input pixel size of the model that determines the defect center probability map can differ from an output pixel size. Accordingly, the size of the defect center probability map can differ from the size of the measured image data and / or the reference image data. Therefore, it is usually not necessary to make a defect probability prediction for each pixel of the measured image data. For example, the number of pixels of the defect center probability map can be less than 30% or even less than 10% of the number of pixels of the measured image data.For example, each pixel of the defect center probability map can be associated with a total of four pixels of the measured image data along each direction. Then, the number of pixels of the defect center probability map is 1 / 4 * 1 / 4 = 1 / 16 of the number of pixels of the measured image data. Therefore, if a specific pixel of the defect center probability map is identified as a defect coordinate, this defect coordinate can be transformed into a defect coordinate in the measured image data.

[0017] A defect in a lithography mask can be an intrusion or a protrusion, i.e., a height deviation of the lithography mask with respect to a reference. A defect can be a contamination, e.g., dust, adder, etc. Problems can arise during the manufacturing process that result in one or more defects on the lithography mask. For example, the mask material may be faulty due to inclusions or unevenness, resulting in variations in the surface topography of the mask. In addition, errors can occur during the writing of the mask pattern, such as lines that are too thick or too thin, which can lead to deviations from the intended design. Line edge roughness (LER), in which the edges of the patterns on the mask are not smooth and even, can also occur.In some cases, individual areas may have been over-processed, with too much of the top layer being removed, or under-processed, with insufficient removal of the top layer.

[0018] A machine learning model can be used to predict the defect center probability map. Various scenarios are conceivable for implementing the machine learning model. For example, classic options such as a convolutional neural network, e.g., using an encoder-decoder-like UNet architecture, can be used, but other common options such as Vision Transformer (ViT) models or standard encoder / decoder structures are also possible. Furthermore, the exact layout of the model is not critical, so variations in the number of layers, the number of filters, the choice of normalization, and nonlinearities, among other things, are possible to achieve the desired defect detection results.

[0019] According to various examples, the machine learning model used to predict the defect center probability map includes an image encoder. The machine learning model provides, for each of a predefined number of classes and at each pixel grid position, a probability that this position is the center of an object of this class. The machine learning model can additionally predict, for each pixel grid position, an offset for each class and an object size for each class.

[0020] In one embodiment, the method may further comprise simplifying object detection from a CenterNet architecture by removing certain components, such as predicting multiple classes of one or more defects or surrounding bounding boxes.

[0021] For example, a corresponding CenterNet architecture is used in WO 2023 195036 A1. See also, for example, US20240095927 A1. The defect center probability can be predicted based on a machine learning model based on a CenterNet architecture.

[0022] In one embodiment, the method may further comprise eliminating the refinement of center coordinates via offsets, thereby reducing computational complexity and energy consumption. Accordingly, the machine learning model may not include a module for refining center coordinates, e.g., via offsets. The machine learning model may not include a module for predicting bounding boxes.

[0023] In one embodiment, the method may further comprise omitting non-maximum suppression techniques to filter near-duplicates of the predictions since only one or more defect centers are predicted without considering surface instances.

[0024] The method can be part of a die-to-die, die-to-reference, or die-to-database workflow. This impacts the selection of reference image data. For example, the reference image data can correspond to images measured for different parts of the lithography mask associated with different dies (die-to-die; D2D). In one embodiment, the defect detection method could be applied to other workflows outside of D2D, for example, die-to-reference (D2R) or die-to-database (D2DB). In these scenarios, a "good" reference image is used instead of a second acquired image. D2R: The reference image data can be obtained from another imaging process, e.g., at an earlier time before the mask was deployed, i.e., at a different aging stage of the mask. D2DB: For example, mask design data can be used to generate synthesized reference image data.The reference image data can be predefined in a database. Another example is a single-die workflow: Here, a single image is acquired and divided into the measured image data and the reference image data, which form the basis for the difference image.

[0025] The defect center probability map indicates the probability that pixels in the measured image data are the center of a defect. In this context, the center of a defect can be defined by one or more of the following features: the centroid of an arbitrarily shaped object, the visual center / centroid of an arbitrarily shaped object, the center of a surrounding geometric boundary, such as a bounding box or convex hull, or even a human-provided annotation. While several options are possible, a preferred implementation defines the defect center as the center of the surrounding bounding box of that defect, which has been shown to be more reliable than alternatives such as the centroid of the polygon describing the boundary due to issues with polygon definitions.For each identified defect coordinate, a corresponding bounding box can be determined.

[0026] As disclosed above, the defect center probability map can be predicted based on the machine learning model receiving one or more inputs. Details of the inputs are explained below. In general, the model can process a single input or various combinations of inputs. These potential inputs can be one or more of the following: a difference image determined based on the measured image data and the reference image data, and / or the original measured and reference image data. This flexibility in input options allows for adaptation to various workflow scenarios.

[0027] The method may comprise registering the measured image data and the reference image data and determining the difference image from the measured image data and the reference image data taking the registration into account.

[0028] For example, if multiple chips are available, rather than just two, the inputs can be expanded to include multiple measured image data. Furthermore, other inputs are also possible depending on the overall workflow. In a single-chip scenario, for example, only a single image would be required as input, without the need for a difference image. On the other hand, in workflows such as D2DB and D2R, the second image could be either a synthesized image or a reference image acquired at a different aging stage of the lithography mask.

[0029] While using multiple inputs in the machine learning model can be beneficial because it provides the model with more information, it also carries the risk of the model ignoring certain inputs, such as the difference image. However, this approach can lead to more robust predictions, especially when the registration between the measured images and the reference images was not perfect.

[0030] Simulation-based methods can be used to obtain accurate reference image data for predicting the defect center probability map. By obtaining the reference image data from a simulation, it is possible to generate high-quality images free of noise and artifacts, enabling more accurate detection of one or more defects in the lithography mask. This approach can be particularly useful when actual measured reference image data is unavailable or is of poor quality.

[0031] After predicting the defect center probability map and identifying one or more defect coordinates (where each of the one or more defect coordinates marks a corresponding predicted defect), control data for a mask repair tool can be determined based on these one or more defect coordinates and / or based on their corresponding probability values and / or based on the extents of the one or more defects, e.g., bounding box sizes. In other words, the determined control data can consider not only the position of the one or more defects, but also their probability of being actual defects, as quantified by the probability values.Alternatively or additionally, it would be possible for the determined control data to take into account not only the position of one or more defects, but also one or more properties of the one or more defects, e.g. their size / dimensions and / or type of the one or more defects, e.g. propagation or penetration of a mask defect. This enables more precise and targeted mask repair, where, for example, the repair process can be optimized to first repair the most likely defects and / or the most severe defects and / or defects of a specific type. Alternatively or additionally, this enables more precise and targeted mask repair, where, for example, the determination of repair shapes in the repair process can be optimized based on the provided position and / or the provided extent of the defect in the provided control data.For example, the precise shape for repairing a predicted defect can be determined more quickly if the approximate extent is already known. A computing device is disclosed. The computing device comprises a processor and a memory. The processor is configured to load program code from the memory and execute the program code. When executing the program code, the processor performs a method as disclosed above.

[0032] A system comprises such a computing device and an imaging device for acquiring one or more image data of the lithography mask. For example, the imaging device can be implemented by an optical microscope, e.g., an aerial imaging system such as an actinic optical microscope, or a particle microscopy device such as a scanning electron microscope.

[0033] It is understood that the above-mentioned features and the features to be explained below can be used not only in the specified combinations, but also in other combinations or in isolation, without departing from the scope of the invention. BRIEF DESCRIPTION OF THE DRAWINGS Fig. 1 is a flowchart of a method according to various examples. Fig. 2 schematically illustrates a data processing pipeline according to reference implementations. DETAILED DESCRIPTION OF EMBODIMENTS

[0034] Some examples of the present disclosure generally provide for a variety of circuits or other electrical devices. All references to the circuits and other electrical devices and the functionality provided by them are not intended to be limited to include only what is illustrated and described herein. Although specific labels may be assigned to the various disclosed circuits or other electrical devices, such labels are not intended to limit the functional range of the circuits and other electrical devices. Such circuits and other electrical devices may be combined with one another and / or separated in any manner based on the particular type of electrical implementation desired.It is recognized that any circuit or other electrical device disclosed herein may include any number of microcontrollers, a graphics processing unit (GPU), integrated circuits, memory devices (e.g., FLASH, random access memory (RAM), read-only memory (ROM), electrically programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or other suitable variations thereof), and software that cooperate with one another to perform the operations disclosed herein. Additionally, one or more of the electrical devices may be configured to execute program code embodied in a non-transitory, computer-readable medium programmed to perform any number of the disclosed functions.

[0035] The following discloses a technique for detecting one or more defects in lithography masks. A machine learning model can be used to predict a defect center probability map.

[0036] Fig. 1 is a flowchart of a method for detecting one or more defects in the lithography mask. For example, the method of Fig. 1 by a processor device comprising a processor configured to load program code from a memory and execute the program code. In the scenario of Fig. 1, one or more defects are detected using a machine learning model that predicts (i.e., infers) a defect center probability map.

[0037] In box 105, it is optionally possible to control a microscope to acquire image data of the lithography mask. For example, a light microscope can be used, e.g., an aerial image measurement system such as an actinic microscope.

[0038] The image data can be acquired in various ways, e.g., by a charged particle beam system such as a scanning electron microscope (SEM) or a focused ion beam microscope (FIB), or by an atomic force microscope (AFM), or by an aerial imaging system equipped with, for example, a staring array sensor, a line sensor, or a time-delay integration (TDI) sensor.

[0039] An aerial image creates the radiation intensity distribution at the substrate level. In this way, the aerial image enables analysis of the semiconductor structures printed onto the substrate during the printing process. An aerial image refers to a type of image created by the reflection or refraction of light. This can include images captured with a camera or viewed through a microscope, telescope, or other optical instrument. An aerial image is typically captured by a lens that focuses the light onto a medium such as film or a digital sensor. Aerial images can also be created by other means such as holography and interferometry.

[0040] In box 110, it is optionally possible to determine measured image data and / or reference image data based on the image data acquired in box 105.

[0041] In other examples, the measured image data and the reference image data may be provided preliminarily (e.g., stored in a memory) and / or obtained from a simulation (e.g., a digital twin); accordingly, box 105 and box 110 are optional.

[0042] In box 115 it is optionally possible to register the measured image data and the reference image data.

[0043] In box 120, a difference image can optionally be determined based on the registration from box 115. The difference image is determined for the measured image data and the reference image data. Determining the difference image is not required in all scenarios. In some examples, the machine learning model can be operated to predict a defect center probability map directly based on the measured image data and the reference image data determined in box 110 or otherwise specified.

[0044] In box 125, the defect center probability map is predicted. The machine learning model is used for this purpose. The defect center probability map is predicted based on the measured image data and the reference image data. For example, the machine learning model can receive one or more of the following elements as input: the measured image data; the reference image data; and the difference image.

[0045] For example, a convolutional neural network and / or a vision transformer neural network can be used. For example, a U-Net architecture can be used. A CenterNet architecture can be used.

[0046] In box 130, one or more defect coordinates are identified in the defect center probability map. This is achieved by identifying pixels of the defect center probability map that have pixel values that exceed a minimum probability threshold. This is achieved by using thresholding. Furthermore, pixels that have the highest probability values in a local neighborhood are identified. Various metrics for determining a local neighborhood can be implemented, e.g., based on the pixel values in the defect center probability map.

[0047] In box 135, it is possible to return one or more defect coordinates. For example, a human-machine interface can be controlled to output the one or more defect coordinates. In box 135, it is optionally possible to additionally output probability values associated with each of the one or more defects.

[0048] Post-processing can optionally be performed in box 140. For example, additional detection models and / or classification models can be executed based on the output data obtained in box 135. Interference can be suppressed. Additional information such as the shape and / or size of a defect can be determined. Such techniques are based on the realization that defects typically occur sparsely, so post-processing in box 140 is limited to a few cases. A review of one or more localized defects can be implemented. This can enable validation of one or more defects.

[0049] The post-processing in Box 140 could be implemented by completely separate models or by additional decoding branches of the machine learning model previously used in Box 125. End-to-end training of multiple machine learning models could be possible.

[0050] In box 145, it would be possible to use the output from box 135 or box 140. For example, control data for a mask repair tool can be generated. Such tools can include micromanipulators and / or charged particle beams for micromanipulating the lithography mask. The control data can be generated based on the outputs of box 135, i.e., based on the one or more defect coordinates and / or probability values.

[0051] Focused ion beam milling (FIB) tools are an example of a mask repair tool that uses a focused ion beam to remove material from the lithography mask, enabling precise repair of one or more defects such as pinholes or excess material. Electron beam lithography (EBL) systems can also be used to repair one or more defects by adding or removing material using electron beam-induced chemistry. Nanoindentation tools are another type of mask repair tool that use a sharp probe to indent and remove material from the mask surface, allowing the repair of small defects such as scratches or pits.Additionally, laser-based repair tools, such as laser ablation or laser-induced chemical etching, can be used to remove one or more defects from the mask surface. Atomic force microscopy (AFM)-based repair tools are another option, which utilizes an AFM probe to detect and repair one or more nanoscale defects, such as scratches or contaminants. Finally, gas cluster ion beam (GCIB) tools use a beam of gas clusters to remove material from the mask surface, enabling precise repair of one or more defects.

[0052] Fig.2 schematically illustrates a data processing pipeline according to a reference implementation. A measured image 205 (as an example of measured image data) and a reference image 210 (as an example of reference image data) are combined, taking the respective registration into account. This results in the difference image 220.

[0053] A machine learning model 225 is then executed. This results in a defect center probability map 230. Furthermore, a map 235 may be obtained indicating offsets of the one or more defect centers in the pixel grid of the measured image 205 relative to the pixel grid of the defect center probability map 230. A map 240 may be obtained indicating a dimension of each of the one or more defects, e.g., using bounding boxes. It may be possible not to generate the map 235 and / or the map 240.

[0054] A module 245 identifies pixels of the defect center probability map that exceed a minimum probability threshold and have the highest probability values in the local neighborhood, thereby obtaining one or more defect coordinates 250. Such output data in the form of the one or more defect coordinates may indicate one or more defects by providing bounding boxes as output.

[0055] A concrete example of the training and inference of such a machine learning model 225 is given below: Data structure: The machine learning model is trained on a dataset consisting of 100,000 image sections (512 x 512 pixels) with various patterns such as lines and spaces, pinpoints, and logic. There are also 20,000 validation images.

[0056] Model structure: The machine learning model takes a single-channel input (the difference image) and outputs a single-channel probability map representing the probability of one or more defect centers. The output has a single pixel for every 16 pixels of the input (rescaled by a factor of four in both the x and y directions). The model architecture is an encoder-decoder layout with 5 layers in the encoder and 3 layers in the decoder, initially using 16 filters and batch normalization. This is based on the CenterNet architecture.

[0057] Training setup: During training, a random sample of 448 x 448 pixels is applied as a data augmentation. The Adam W optimizer is used with a learning rate of 3e-4, which decays by 0.93 over 10 epochs, and a batch size of 32. The loss function consists of binary cross-entropy for centers and L2 for offsets. Initialization is done from scratch using standard Glorot.

[0058] Local maximum value calculation: To identify one or more defect coordinates, a local aggregation of the confidence values is performed using 3x3 max-pooling with stride 1. Locations with a defect probability less than 0.1 are filtered out, and the remaining local maxima or their corresponding positions in the image data are returned as one or more defect coordinates.

[0059] In summary, techniques for defect detection in lithography masks have been disclosed. The disclosed techniques offer several advantages over other approaches: (1) Compared to non-data-driven methods, the disclosed techniques achieve better detection rates and lower noise rates; (2) compared to segmentation-based methods, the disclosed techniques are significantly faster because no complex post-processing is required; and (3) compared to detection-based methods, the disclosed techniques are faster because fewer unnecessary computations are performed.

[0060] Although the invention has been shown and described with reference to certain preferred embodiments, equivalents and modifications will occur to those skilled in the art after reading and understanding the specification. The present invention includes all such equivalents and modifications and is limited only by the scope of the appended claims.

[0061] To illustrate, the machine learning model's output could be extended with additional offset vectors to refine the center location predictions (see Map 235). This would change the model output from a single channel representing the center probability to two or three channels, including center probability, x-offset, and / or y-offset. The advantage of this approach is that the center location predictions can be more precise, even if the pixel grid of the defect center probability map has a lower resolution compared to the measured image data. Furthermore, with this modification, it would be possible to reduce the depth of the decoder and, if necessary, achieve the same or better accuracy compared to the full decoder.

[0062] To further illustrate this, the machine learning model could be extended to predict multiple defect types, going beyond a simple yes-or-no classification. This could be achieved by using state-of-the-art object detection algorithms that predict not just one channel for center probability, but one channel per defect class. This would enable discrimination between multiple defect types, which can be advantageous in certain inspection workflows.

[0063] For further illustration, techniques related to detecting one or more defects in lithography masks were disclosed above. Similar techniques can be readily applied to detecting one or more defects in other types of devices and samples, e.g., detecting one or more defects in semiconductor structures fabricated on wafers, or detecting one or more defects in macroscopic samples, e.g., turbine blades, using invasive or non-invasive imaging, etc. REFERENCE NUMBERS 105 Box 110 boxes 115 Box 120 boxes 125 Box 130 boxes 135 Box 140 boxes 205 Input image 210 Input image 220 Input image 225 Machine Learning Model 230 Defect center probability map 235 Map 240 card 245 Module 250 output data QUOTES CONTAINED IN THE DESCRIPTION

[0000] This list of documents submitted by the applicant was generated automatically and is included solely for the convenience of the reader. This list is not part of the German patent or utility model application. The DPMA assumes no liability for any errors or omissions. Cited patent literature

[0000] WO 2023 195036 A1

[0021] US 20240095927 A1

[0021]

Claims

[1] A method for detecting one or more defects in a lithography mask, the method comprising: - predicting a defect center probability map (230) based on measured image data (205) of the lithography mask and reference image data (210) of the lithography mask, wherein each pixel of the defect center probability map (230) comprises a probability score representing a probability of being a defect center, - identifying pixels having the highest probability values in a local neighborhood as one or more defect coordinates (251) and in the defect center probability map, and - Returning the one or more defect coordinates (251), optionally together with their corresponding probability values. [2] The method of claim 1, wherein the defect center probability map is predicted based on a machine learning model, wherein the machine learning model is optionally based on a CenterNet architecture. [3] The method of claim 2, wherein the machine learning model does not include a center coordinate refinement module. [4] The method of claim 2 or 3, wherein the machine learning model does not include a bounding box prediction module. [5] A method according to any one of the preceding claims, further comprising: - Performing a registration of the measured image data and the reference image data, - based on the registration, determining a difference image (220) from the measured image data and the reference image data, wherein the defect center probability map is determined based on the difference image (220). [6] A method according to any preceding claim, wherein each defect coordinate is associated with a center of a bounding box. [7] Method according to one of the preceding claims, wherein the measured image data and the reference image data are obtained for different parts of the lithography mask, optionally in different aging states of the lithography mask. [8] Method according to one of the preceding claims, wherein the reference image data are predefined in a database. [9] Method according to one of the preceding claims, wherein the reference image data comprises aerial image data. [10] A method according to any one of the preceding claims, further comprising: - Obtaining the reference image data from a simulation. [11] A method according to any one of the preceding claims, further comprising: - Controlling an aerial image measuring system to acquire aerial image data of the lithography mask, and - Determine at least the measured image data or the reference image data based on the aerial imaging data. [12] Method according to one of the preceding claims, wherein pixels are identified as defect coordinates (250) when pixel values in the defect center probability map (230) exceed a predefined minimum probability threshold. [13] A method according to any one of the preceding claims, further comprising: - determining (145) control data for a mask repair tool based on the one or more defect coordinates and optionally based on the probability values. [14] Computing device comprising a processor and a memory, wherein the processor is arranged to load program code from the memory and to execute the program code, wherein the processor carries out the method according to one of the preceding claims when executing the program code.

Images