A small sample image recognition method and system

Abstract

The application belongs to the field of image recognition, and specifically discloses a small sample image recognition method and system. Through the application, when training a small sample image recognition network, a query set containing multiple to-be-tested images and a support set containing a small amount of labeled samples are input, a channel weight calculation module is used to construct a difference matrix of each query set sample and the support set, the inconsistency of each dimension of the matrix from different categories of data is captured, and the importance of the current channel or dimension for the current query set sample is judged as a judgment standard, so that the distance modulation module pays more attention to the channels that are more discriminative and distinguishable for the current task. Most current channel weighting algorithms (such as SENet, Attention module) are based on the interaction and connection between feature channels to generate weights, and the application is completely different from most current channel weighting algorithms, and generates weights for each channel based on the inconsistency of different class data in each channel.

Description

TECHNICAL FIELD

[0001] The present application belongs to the field of image recognition, and more particularly, to a small sample image recognition method and system. BACKGROUND

[0002] In recent years, with the rapid development of theoretical modeling, technical innovation, software and hardware, deep learning has achieved excellent results in many computer vision tasks, such as image classification, object detection, etc. Despite this, current deep learning models all need a large amount of annotated data to support training to develop various intelligent services. However, in some fields such as health care, rare disasters, etc., due to privacy or high cost of annotation, it is often difficult to obtain labeled samples. To solve this problem, small sample learning has emerged, and small sample image classification has become one of the important research directions in the field of deep learning. Small sample image classification refers to given an unknown class image, only a small amount of labeled image can be used to correctly identify the image to be identified. Compared with traditional deep learning image classification which needs a large number of labeled samples to drive, small sample image classification only needs a small amount of labeled image data to identify new unknown images, which is more close to the application scenario in the real world.

[0003] The method based on distance metric learning has become a prominent method for small sample image classification, which usually uses a distance function to measure the difference between query samples and support samples to identify the class of query samples. For example, in “Prototypical Networks for Few-shot Learning, Proceedings of the Neural Information Processing Systems (NeurIPS), 2017.”, it takes the average vector of the labeled samples from the same class as the class prototype, and takes the Euclidean distance as the distance metric classifier.

[0004] However, when calculating the class score, these distance classifiers only treat each channel difference between query and support features equally, ignoring the importance of different channels for different tasks. Since different channels in the learned features often represent different patterns, these distance metrics do not take into account the different importance of different channels for small sample image classification, so they cannot accurately measure the similarity between samples, resulting in poor classification recognition effect in actual application scenarios. SUMMARY

[0005] In view of the defects of the prior art, the purpose of the present application is to provide a small sample image recognition method and system, aiming to solve the problems of poor classification recognition effect and poor generalization caused by the equal treatment of each channel when calculating the class score by the existing method.

[0006] To achieve the above-mentioned purpose, in a first aspect, the present application provides a small sample image recognition method, comprising:

[0007] obtaining a pre-trained small sample image recognition network;

[0008] inputting the query set and the support set into the pre-trained small sample image recognition network to obtain a class score vector of each test image in the query set, and then judging the class;

[0009] The small sample image recognition network comprises a feature extraction module, a class prototype determination module, a channel weight calculation module and a distance modulation module.

[0010] The feature extraction module has a first parameter and is used to extract a feature vector of an input image, wherein the input image comprises each labeled image in the support set and each query image in the query set.

[0011] The class prototype determination module is used to obtain the feature vectors of the labeled images of the same class in the support set, and then determine the class prototype of each class.

[0012] The channel weight calculation module is used to obtain the feature vectors of a single query image and all labeled images in the support set, construct a difference matrix related to the feature vector of the query image, and obtain a channel weight matrix related to the feature vector of the query image through a calculation function with a second parameter. The weight matrix is used to modulate the distance between the query sample and other sample features in the current classification task.

[0013] The distance modulation module is used to calculate the distance between the feature vector of a single query image and each class prototype, and modulate the distance through the channel weight matrix to obtain the class score vector of the query image.

[0014] The first parameter and the second parameter are optimized through the training stage.

[0015] Preferably, the class prototype determination module determines the class prototype of each class in the nearest neighbor judgment mode, wherein,

[0016] Class prototype is expressed as:

[0017]

[0018] wherein, n∈[1, N] represents a category, N represents the total number of categories of the support set, l represents the lth channel, and d represents the total number of channels of the feature, represents a category prototype c n , K represents the number of sample images contained in each category of the support set, and z j represents the jth sample feature contained in the category n.

[0019] Preferably, the difference matrix τ q is represented as:

[0020]

[0021]

[0022] wherein, and respectively represent the lth channel of the qth query sample feature z q and the lth channel of the sth support sample feature z s , represents the distribution of the difference of the lth channel, N represents the total number of categories of the support set, K represents the number of sample images contained in each category of the support set, and d represents the total number of channels of the feature, represents the distance between the lth channel of the qth query sample feature z q and the sth support sample feature z s , l=1, 2, …, d, and s=1, 2, …, NK.

[0023] Preferably, the channel weight matrix W q is represented as:

[0024]

[0025] wherein, g φ (·) represents a calculation function with a second parameter φ, represents the distribution of the difference of the lth channel, l=1, 2, …, d, and d represents the total number of channels of the feature.

[0026] Preferably, the calculation function with the second parameter is a double-layer multilayer perceptron, which includes two fully connected layers fc1(·) and fc2(·), and ReLU is used as the activation function between layers.

[0027] Preferably, the distance between the query sample feature z q and the category prototype c n after modulation is represented as:

[0028] ​

[0029] wherein W q denotes a channel weight matrix, denotes the distance between the l-th channel of the q-th query sample feature z q and the n-th class prototype c n , l = 1, 2, …, d, n ∈ [1, N], N denotes the total number of classes of the support set.

[0030] Preferably, the training phase updates the first parameter and the second parameter until convergence:

[0031]

[0032] wherein θ denotes the first parameter, φ denotes the second parameter, and η denotes the update step, denotes the partial derivative of the loss function with respect to the parameters θ and φ, which is used to measure the performance of the model on the current small sample task, denotes the classification loss function of the network prediction sampling generation query set sample label and the real label.

[0033] Preferably, in the training phase, given a training set D train and a non-overlapping test set D test , first train on a series of tasks randomly sampled from D train , and then test on a series of tasks randomly sampled from D test Each small sample task contains two disjoint sets, the support set and the query set follows the "N-way K-shot" setting, the support set is composed of N classes, each containing K labeled images; the query set shares the same label space as .

[0034] To achieve the above purpose, in a second aspect, the present application provides a small sample image recognition system, comprising:

[0035] at least one memory for storing programs;

[0036] at least one processor for executing the programs stored in the memory, when the programs stored in the memory are executed, the processor is used to execute the method as described in the first aspect.

[0037] To achieve the above purpose, in a third aspect, the present application provides a computer readable storage medium, the computer readable storage medium stores a computer program, when the computer program runs on the processor, so that the processor executes the method as described in the first aspect.

[0038] It can be understood that the beneficial effects of the above-mentioned second aspect to the third aspect can be referred to the related description in the above-mentioned first aspect, which will not be repeated here.

[0039] Overall, compared with the prior art, the above technical solutions conceived by the present application have the following beneficial effects:

[0040] (1) The present application provides a small sample image recognition method and system. When training a small sample image recognition network, a query set containing multiple to-be-tested images and a support set containing a small number of labeled samples are input. A channel weight calculation module is used to construct a difference matrix of each query set sample and the support set. By capturing the inconsistency of each dimension from different categories of data in the matrix, the importance of the current channel or dimension for the current query set sample is judged as a criterion, so as to make the distance modulation module pay more attention to the channels that are more discriminative and distinguishable for the current task. Most current channel weighting algorithms (such as SENet, Attention module) generate weights based on the interaction and connection between feature channels. The present application is completely different from the current most channel weighting algorithms, and generates weights for each channel based on the inconsistency of different class data in each channel.

[0041] (2) The present application provides a small sample image recognition method and system. The distance modulation module provided only operates on distance measurement, so it does not directly interfere with the feature extraction process. Therefore, it can be inserted into any backbone-based distance metric learning small sample image recognition method.

[0042] (3) The present application provides a small sample image recognition method and system. The parameter amount of the distance modulation module provided is only related to the number of support set samples, and is independent of the depth of the backbone network. The number of added learnable parameters is (NK+1) 2 Compared with the current channel weighting method which adds tens of thousands of parameters, the distance modulation module is a lightweight network that almost does not consume computing resources. BRIEF DESCRIPTION OF DRAWINGS

[0043] Figure 1 is a small sample image recognition method provided by an embodiment of the present application.

[0044] Figure 2 is a channel weight calculation module structure diagram provided by an embodiment of the present application.

[0045] Figure 3 is a structure diagram of an electronic device provided by an embodiment of the present application.

[0046] In all the drawings, the same reference signs are used to represent the same elements or structures, wherein:

[0047] 1-feature extraction module; 2-category prototype determination module; 3-channel weight calculation module; 4-distance modulation module; 810-processor; 820-communication interface; 830-memory; 840-communication bus. DETAILED DESCRIPTION

[0048] In order to make the purpose, technical scheme and advantages of the present application clearer, the present application will be further described in detail below in combination with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present application and not to limit the present application.

[0049] The term "and / or" herein is a description of the association relationship of the associated objects, which means that there can be three relationships, for example, A and / or B can mean that there are three cases of A alone, A and B together, and B alone. The symbol " / " in this paper represents the relationship of or, for example, A / B represents A or B.

[0050] The terms "first" and "second" and the like in the description and claims herein are used to distinguish different objects, not to describe a specific order of the objects. For example, the first response message and the second response message are used to distinguish different response messages, not to describe the specific order of the response messages.

[0051] In the embodiments of the present application, the words "exemplary" or "for example" are used to mean serving as an example, instance, or illustration. Any embodiment or design scheme described as "exemplary" or "for example" in the embodiments of the present application should not be construed as being more preferred or having more advantages than other embodiments or design schemes. Rather, the use of "exemplary" or "for example" is intended to present relevant concepts in a concrete manner.

[0052] In the description of the embodiments of the present application, unless otherwise specified, "a plurality of" means two or more, for example, a plurality of processing units means two or more processing units, and the like; a plurality of elements means two or more elements, and the like.

[0053] Next, the technical scheme provided in the embodiments of the present application will be introduced.

[0054] As shown in the figure, the present application provides a small sample image recognition method, comprising: Figure 1

[0055] Obtaining a pre-trained small sample image recognition network;

[0056] The query set and the support set are jointly input into the pre-trained small sample image recognition network to obtain a category score vector of each to-be-tested image in the query set, and then the category is judged.

[0057] ​The small sample image recognition network comprises a feature extraction module, a category prototype determination module, a channel weight calculation module and a distance modulation module.

[0058] The feature extraction module 1 is provided with a first parameter and is used to extract a feature vector of an input image, wherein the input image comprises each labeled image in a support set and each query image in a query set.

[0059] The category prototype determination module 2 is used to obtain the feature vectors of the labeled images of the same category in the support set, and then determine the category prototype of each category.

[0060] The channel weight calculation module 3 is used to obtain the feature vectors of a single query image and all labeled images in the support set, construct a difference matrix related to the feature vector of the query image, and obtain a channel weight matrix related to the feature vector of the query image through a calculation function provided with a second parameter, wherein the weight matrix is used for modulation when the query sample calculates the distance with other sample features in the current classification task.

[0061] The distance modulation module 4 is used to calculate the distance between the feature vector of a single query image and each category prototype, and modulate the distance through the channel weight matrix to obtain a category score vector of the query image.

[0062] The first parameter and the second parameter are optimized through a training stage.

[0063] Preferably, the feature extraction module remaps the pixel-level features into high-dimensional features in an embedding space through a series of operations such as convolution.

[0064] Suppose the feature z of the sample x is wherein z l is the lth channel or dimension of the feature z, and d is the number of dimensions, and the feature of the query sample is The feature of the support sample is

[0065] Preferably, the category prototype determination module determines the category prototype of each category in a nearest neighbor judgment mode, wherein

[0066] The category prototype is expressed as:

[0067]

[0068] wherein category n∈[1, N], N represents the total number of categories of the support set, l represents the lth channel, d represents the total number of channels of the feature, the lth channel of the category prototype c n is expressed as: jrepresents the jth sample feature included in the category n.

[0069] As shown in Figure 2 , preferably, the difference matrix τ q is represented as:

[0070]

[0071]

[0072] wherein, and respectively represent the lth channel of the qth query sample feature z q and the lth channel of the sth support sample feature z s . represents the distribution of the difference of the lth channel, N represents the total number of categories of the support set, K represents the number of sample images included in each category of the support set, and d represents the total number of channels of the feature. represents the distance between the lth channel of the qth query sample feature z q and the sth support sample feature z s , l = 1, 2, …, d, and s = 1, 2, …, NK.

[0073] Each column of the difference matrix can reflect the inconsistency of data from different categories in the same dimension. According to the idea of the variance threshold method in feature selection, the more inconsistent the data from different categories in the same dimension, the more important the dimension, and a higher weight should be allocated to the dimension. On this basis, a meta-learner is used to learn a judgment rule for evaluating the importance of the channel.

[0074] As shown in Figure 2 , represents the distribution of the difference of the lth channel, which can be generated by a distance modulation module g with φ as a parameter for the lth channel of the qth query sample feature z q .

[0075]

[0076] Preferably, the channel weight matrix W q is represented as:

[0077]

[0078] wherein g φ (·) represents a calculation function with a second parameter φ, represents the distribution of the difference of the lth channel, l = 1, 2, …, d, and d represents the total number of channels of the feature.

[0079] Preferably, a meta-learner is used to compute a weighted weight for each dimension of the distance classifier based on the inconsistency of different classes within each channel.

[0080] Preferably, the second parameterized function is a two-layer multi-layer perceptron, including two fully connected layers fc1(·) and fc2(·), and ReLU as the activation function between layers.

[0081] The generation process of the weight can be represented as:

[0082]

[0083] The distance between the feature of the query sample and the feature of the support sample can be represented as:

[0084]

[0085] Preferably, the Euclidean distance or the cosine similarity is used as the distance metric function to measure the distance between the feature of the query sample and the feature of the support sample. The greater the similarity or the smaller the distance, the more likely the image belongs to the category.

[0086] It is noted that when calculating the total distance , the distance in each channel has equal weight, and for l = 1, …, d, the distance metric can be adjusted by combining the learnable channel weight . The above formula can be rewritten as:

[0087]

[0088]

[0089] where the column vector

[0090] The category prototype can be represented as:

[0091]

[0092] The distance between the query sample feature z q and the category prototype c n after modulation can be represented as:

[0093]

[0094] Preferably, the query sample feature z q and the category prototype cn the distance after modulation is expressed as:

[0095]

[0096] wherein W q denotes a channel weight matrix, denotes the distance between the qth query sample feature z q and the lth channel of the nth class prototype c n , l = 1, 2, …, d, n ∈ [1, N], N represents the total number of classes of the support set.

[0097] Preferably, the training phase updates the first parameter and the second parameter until convergence:

[0098]

[0099] wherein θ represents the first parameter, φ represents the second parameter, and η represents the update step, denotes the partial derivative of the loss function with respect to the parameters θ and φ, which is used to measure the performance of the model on the current small sample task, denotes the classification loss function of the network prediction sampling generating query set sample label and real label. The above process is repeated T times from the training set D train , and f θ and g φ are updated until convergence.

[0100] Preferably, in the training phase, given the training set D train and the non-overlapping test set D test , first train on a series of tasks randomly sampled from D train , and then test on a series of tasks randomly sampled from D test ; each small sample task T contains two disjoint sets, a support set S and a query set Q follows the "N-way K-shot" setting, the support set is composed of N classes, each containing K labeled images; the query set shares the same label space as .

[0101] When training the recognition network, K labeled training samples and M training samples to be predicted during the training process are determined for each class of image, and the training samples corresponding to the N classes of images are input into the recognition network for training.

[0102] In a real application scenario, the query set image collection to be identified and the support image collection are input into the pre-trained identification network to classify the to-be-identified image, and an identification result is output.

[0103] It can be understood that the detailed function implementation of each unit / module described above can refer to the description in the foregoing method embodiments, which will not be repeated here.

[0104] The pseudo code of the entire algorithm is shown in Table 1, which will not be repeated here.

[0105] Table 1

[0106]

[0107] Embodiment

[0108] Due to privacy and annotation cost, etc., a small amount of labeled samples can be obtained in most cases, resulting in only a small amount of labeled instances for each class. The present embodiment provides a small sample medical image recognition system, wherein the feature extraction module adopts ResNet-12, and the pre-training weight of minimizing the cross-entropy loss on the training set is used as the initialization, d=640. The prototype network is used to determine the class of a rare case, and in the case of only a small amount of labeled cases in each class, the prototypes of all classes to be identified (the mean of the features of only a small amount of labeled samples in each class) are calculated, a difference matrix of the features of the to-be-identified case and the labeled case is constructed, and the distance modulation module trained based on each column of the difference matrix generates a weighted weight for each channel in turn, which is used to modulate the distance between the features of the to-be-identified case and the prototype of each class, and the one with the largest similarity is the class to which it belongs. The final prediction class score of the query sample is obtained, and the cross-entropy loss of the predicted class score and the true label is calculated, and the gradient descent is used to optimize the overall network parameters.

[0109] It can be understood that the detailed function implementation of each unit / module described above can refer to the description in the foregoing method embodiments, which will not be repeated here.

[0110] It should be understood that the above device is used to execute the method in the above embodiment, and the corresponding program module in the device, the implementation principle and technical effect of which are similar to the description in the above method, and the working process of the device can refer to the corresponding process in the above method, which will not be repeated here.

[0111] Based on the method in the above embodiment, as Figure 3As shown, the electronic device provided by the embodiments of the present application can include a processor 810, a communications interface 820, a memory 830 and a communications bus 840, wherein the processor 810, the communications interface 820 and the memory 830 can complete mutual communication through the communications bus 840. The processor 810 can invoke the logical instructions in the memory 830 to execute the method in the above-mentioned embodiments.

[0112] In addition, the logical instructions in the memory 830 described above can be implemented in the form of a software functional unit and sold or used as an independent product, and can be stored in a computer readable storage medium. Based on such understanding, the technical solutions of the present application can be embodied in the form of a software product in essence or in the form of a part of the technical solutions that make contributions to the prior art, or the technical solutions of the present application. The computer software product is stored in a storage medium and includes a plurality of instructions for causing a computer device (which can be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the method described in the embodiments of the present application.

[0113] Based on the method in the above-mentioned embodiments, the embodiments of the present application provide a computer readable storage medium, which stores a computer program. When the computer program runs on a processor, the processor executes the method in the above-mentioned embodiments.

[0114] Based on the method in the above-mentioned embodiments, the embodiments of the present application provide a computer program product, which runs on a processor. When the computer program product runs on the processor, the processor executes the method in the above-mentioned embodiments.

[0115] It can be understood that the processor in the embodiments of the present application can be a central processing unit (CPU), and can also be other general-purpose processors, digital signal processors (DSPs), application specific integrated circuits (ASICs), field programmable gate arrays (FPGAs) or other programmable logic devices, transistor logic devices, hardware components or any combination thereof. The general-purpose processor can be a microprocessor or any conventional processor.

[0116] The method steps in the embodiments of the present application can be implemented by means of hardware, or by means of a processor executing software instructions. The software instructions can be composed of corresponding software modules, which can be stored in a random access memory (RAM), a flash memory, a read-only memory (ROM), a programmable read-only memory (PROM), an erasable PROM (EPROM), an electrically EPROM (EEPROM), a register, a hard disk, a mobile hard disk, a CD-ROM, or any other form of storage medium well known in the art. An exemplary storage medium is coupled to a processor, so that the processor can read information from the storage medium and write information to the storage medium. Of course, the storage medium can also be an integral part of the processor. The processor and the storage medium can be located in an ASIC.

[0117] In the above embodiments, all or part of the embodiments can be implemented by means of software, hardware, firmware, or any combination thereof. When implemented by means of software, all or part of the embodiments can be implemented in the form of a computer program product. The computer program product includes one or more computer instructions. When loaded and executed by a computer, all or part of the computer program instructions generate the processes or functions described in the embodiments of the present application. The computer can be a general-purpose computer, a special-purpose computer, a computer network, or other programmable devices. The computer instructions can be stored in a computer-readable storage medium or transmitted by means of the computer-readable storage medium. The computer instructions can be transmitted from one website, computer, server, or data center to another website, computer, server, or data center by means of a wire (such as a coaxial cable, an optical fiber, a digital subscriber line (DSL)) or a wireless (such as infrared, wireless, microwave, etc.). The computer-readable storage medium can be any available medium accessible by a computer or a data storage device such as a server, data center, etc. that includes one or more available media sets. The available media can be a magnetic medium (such as a floppy disk, a hard disk, a magnetic tape), an optical medium (such as a DVD), or a semiconductor medium (such as a solid state disk (SSD)), etc.

[0118] It can be understood that the various numerical numbers involved in the embodiments of the present application are only used for differentiation for convenience of description, and do not limit the scope of the embodiments of the present application.

[0119] Those skilled in the art can easily understand that the above description is only the preferred embodiment of the present application, and is not used to limit the present application, and any modification, equivalent replacement and improvement made within the spirit and principle of the present application should be included in the protection scope of the present application.

Claims

1. A small sample image recognition method, characterized in that, The method comprises the following steps: obtaining a pre-trained small sample image recognition network; inputting a query set and a support set into the pre-trained small sample image recognition network to obtain a class score vector of each image to be tested in the query set, and then judging the class; the small sample image recognition network comprises a feature extraction module, a class prototype determination module, a channel weight calculation module and a distance modulation module; the feature extraction module is configured to extract a feature vector of an input image using a feature calculation function with a first parameter, wherein the input image comprises each labeled image in the support set and each query image in the query set; the class prototype determination module is configured to obtain the feature vectors of the labeled images of the same class in the support set, and then determine the class prototype of each class; the channel weight calculation module is configured to obtain the feature vectors of a single query image and all labeled images in the support set, construct a difference matrix related to the feature vector of the query image, and obtain a channel weight matrix related to the feature vector of the query image by passing the difference matrix through a weight calculation function with a second parameter, wherein the weight matrix is used to modulate the distance calculation between the query sample and other samples in the current classification task; the distance modulation module is configured to calculate the distance between the feature vector of a single query image and the class prototype of each class, modulate the distance by the channel weight matrix, and obtain the class score vector of the query image; the first parameter and the second parameter jointly constitute network parameters, which are optimized in a training phase.

2. The method of claim 1, wherein, The class prototype determination module determines the class prototype of each class in a nearest neighbor judgment manner, wherein Class Prototype is represented as: wherein the class , represents the total number of classes of the support set, represents the i th channel, represents the total number of channels of the feature, represents the i th channel of the class prototype , represents the number of sample images contained in each class of the support set, represents the i th sample feature contained in the class .

3. The method of claim 1, wherein, the difference matrix is represented as: wherein, and respectively represent the first channel of the first query sample feature and the first channel of the first support sample feature , represents the distribution of the difference of the first channel, represents the total number of categories of the support set, represents the number of sample images contained by each category of the support set, represents the total number of channels of the feature, represents the distance of the first channel of the first query sample feature and the first channel of the first support sample feature , .

4. The method of claim 3, wherein, The channel weight matrix is represented as: wherein, represents a computing function with second parameters , represents a distribution of differences of the first channel, , represents a total number of channels of the feature.

5. The method of claim 1, wherein, The second parameterized calculation function is a two-layer multi-layer perceptron, including two fully connected layers and , ReLU as the activation function between layers.

6. The method of claim 1, wherein, Query sample features With category prototypes Distance after modulation Is represented as: wherein, denotes a channel weight matrix, denotes the distance of the i-th channel to the i-th query sample feature and the i-th class prototype , , denotes the total number of classes of the support set.​​​ 7. The method of claim 1, wherein, the first parameter and the second parameter are updated in the training phase until convergence: wherein, denotes a first parameter, denotes a second parameter, denotes an update step size, denotes a partial derivative of a loss function with respect to a parameter and for measuring the performance of the model on the current small-sample task, denotes a classification loss function of the network prediction sampling generating query set sample labels and real labels.

8. The method of claim 1, wherein, In the training phase, given a training set and a disjoint test set , the model is first trained on a series of tasks sampled randomly from the training set and then tested on a series of tasks sampled randomly from the test set; each small-sample task contains two disjoint sets, the support set and the query set ; following the "N-way K-shot" setting, the support set is composed of classes, each containing labeled images; the query set shares the same label space as .

9. A small sample image recognition system, characterized in that, The method comprises the following steps: at least one memory for storing a program; at least one processor for executing the program stored in the memory, wherein when the program stored in the memory is executed, the processor is configured to execute the method according to any one of claims 1 to 8.

10. A computer-readable storage medium storing a computer program, the computer program comprising instructions that, when executed by a computer, cause the computer to perform the method of any one of claims 1 to 9. When the computer program runs on the processor, the processor is caused to execute the method according to any one of claims 1 to 8.

Images