Cross-view-angle image matching method for artillery reconnaissance aircraft

Abstract

The invention discloses a cross-view-angle image matching method for an artillery reconnaissance plane, and the method comprises the steps: carrying out the feature extraction of images after considering the large difference between a ground view angle image and a view angle image of an unmanned plane, and reducing the difference between the images, so as to enable the classes shot at the same place to be classified into the same class. Firstly, after a ConvNeXt network is used for carrying out feature extraction, the extracted features are input into Transformer to further search for the local dependency relationship of images, and then a result with the highest matching degree is obtained according to the similarity of the features between a matched image and an image to be matched. The method can be used for unmanned aerial vehicle navigation under a GPS denial condition.

Description

Technical Field

[0001] This invention belongs to the field of image processing technology, specifically relating to a cross-view image matching method for artillery reconnaissance aircraft. Background Technology

[0002] Unmanned aerial vehicles (UAVs) have gradually become a major source of remote sensing information due to their ease of image acquisition and low cost. However, their positioning primarily relies on high-precision sensors and the Global Positioning System (GPS). In recent years, in response to the requirements of next-generation weapon intelligence, the diversification and integration of information platforms, and integrated ground-air warfare, UAVs have gradually become an important pillar of next-generation weapons. Positioning based on reconnaissance aircraft images can be used indirectly in sparsely populated areas, areas with poor signal base station coverage, or when interfered with by enemy positioning information. Modern weapons are equipped with a large amount of image and video acquisition information, resulting in a vast amount of ground images. However, these images may lack geographic location information when geolocation fails. With the continuous development of remote sensing satellites, a large number of satellite images containing geographic data tags are being collected. Naturally, matching cross-view ground perspective images with airspace satellite images is used to determine the location of weapons and equipment. However, due to the significant differences in features between cross-view images, traditional methods cannot effectively solve this problem. In recent years, due to the significant achievements of deep learning in image processing, it is feasible to use deep learning technology to match cross-view ground perspective images with satellite perspective images.

[0003] Before the introduction of deep learning, cross-view image matching techniques mostly relied on manually designed features. The SIFT algorithm was used to extract features from relevant locations within the image, followed by KNN or RANSAC algorithms to remove errors and achieve accurate matching. However, due to the significant differences between ground-view and drone-view images, manual feature matching was not feasible. Some methods attempted to locate images using visual cues, such as the sun's position, shadows, or weather conditions. Deep learning methods model cross-view image localization as an image retrieval and matching problem, for example, matching ground-view images with satellite image databases containing location information. Existing methods transform the localization problem into a multi-classification problem, treating satellite images with the same geographic landmarks as the same class. This typically involves designing a Siamese convolutional neural network to extract features from both views, fusing the features, and then inputting them into a feature fusion module for classification. In summary, image-view localization has evolved from previous ground-to-ground or high-altitude-to-high-altitude methods to the more challenging and broader ground-to-high-altitude cross-view matching.

[0004] Previous methods mostly employed convolutional neural networks, using L2 loss functions or cosine distance to calculate the loss between features. However, using two-branch convolutional neural networks, due to the significant differences between cross-view images, meant that the convolutional neural networks could not explicitly define the positional information of each view. Some proposed using cGANs to synthesize satellite view images corresponding to terrestrial view images, fusing features from both terrestrial and synthetic satellite images as input features; however, the synthesized images could not guarantee granularity and detail.

[0005] In summary, by transforming the problem of determining the geographical location of captured images into matching them with satellite images containing geographical location information, a solution can be provided for assisted positioning problems in autonomous driving, augmented reality, and GPS-denied conditions. Summary of the Invention

[0006] To overcome the shortcomings of existing technologies, this invention provides a cross-view image matching method for artillery reconnaissance aircraft. Considering the significant differences between ground-view images and UAV-view images, the method extracts image features to reduce these differences, allowing images taken at the same location to be classified into the same category. First, a ConvNeXt network is used for feature extraction. The extracted features are then input into a Transformer to further identify local dependencies between the images. Finally, based on the similarity of features between the matching image and the image to be matched, the result with the highest matching degree is obtained. This invention can be used for UAV navigation under GPS denial conditions.

[0007] The technical solution adopted by this invention to solve its technical problem is as follows:

[0008] Step 1: Set up and train the feature extraction network model;

[0009] The feature extraction network model consists of two parts: the first part is a preliminary feature extraction based on the ConvNeXt network; the second part is a multi-head attention network based on ViT for learning global features; and a feature fusion layer is connected after the feature extraction network model to obtain the final features.

[0010] The training process of the feature extraction network model is as follows:

[0011] The images used in the training dataset were from both ground-based and drone-based perspectives; a GPS-based sampling strategy was used to initialize negative samples before training, i.e., negative samples were selected based on GPS location information.

[0012] The samples from the ground view are preprocessed using polar coordinate transformation;

[0013] After the network has been trained a certain number of times, the method of mining negative samples based on GPS sampling will be replaced with a method based on image cosine similarity, that is, the negative sample with the highest similarity will be put into each training batch.

[0014] Step 2: After the training in Step 1 is completed, the ground images and satellite images that need to be matched are sequentially input into the network for feature extraction to obtain the feature description vector of each image;

[0015] Based on the feature description vector, the feature phase velocity of the ground image to be matched and the satellite image in the satellite image library to be matched is calculated, and the image with the highest feature similarity is taken as the matching result of the ground image to be matched this time.

[0016] Preferably, the training process of the feature extraction network model is as follows:

[0017] Step 1-1: Perform polar coordinate transformation on the ground view image;

[0018] The relationship between terrestrial and satellite imagery can be expressed using the formula for converting from Cartesian to polar coordinates:

[0019]

[0020] Steps 1-2: Construct a preliminary feature extraction network based on the ConvNeXt network;

[0021] CovNeXt-T was used as the initial feature extraction network, and the original classification layer and the last convolutional block were removed, retaining only the feature extraction backbone network. Furthermore, the input Siamese network did not share network parameters. Therefore, the backbone network contained 42 convolutional layers, divided into four parts. The first part was a single-layer convolution without pooling, using a 4x4 kernel. The second part was a three-layer ConvNeXt block with kernel sizes of 7x7, 1x1, and 1x1, respectively. The 7x7 kernel used a depth-scalable... The first part is a three-layer ConvNeXt block, with three convolutional kernels of 7x7, 1x1, and 1x1. The output feature size is 56x56x96, where the last 96 represents the number of channels. The second part is also a three-layer ConvNeXt block, with three convolutional kernels of 7x7, 1x1, and 1x1. The output feature size is 28x28x192, and this part is repeated three times. The third part uses a three-layer ConvNeXt block, with convolutional kernels of 7x7, 1x1, and 1x1. This block is repeated nine times, and the output feature size is 14x14x384, resulting in a 14x14 block.

[0022] Steps 1-3: Transformer-based self-attention feature module;

[0023] The 14x14 blocks from the previous step are treated as a feature sequence with a dimension of 384, i.e., the input is {x1, x2, ... x... p The sequence} is projected by the self-attention mechanism onto k and v respectively;

[0024] Z′ l =LayerNorm(X l )

[0025] Q = W q Z′ l K = W k Z′ l V = W v Z′ l

[0026]

[0027] Among them W q W k W v These are the weights of query, key, and value, respectively, and H represents the number of multi-heads;

[0028] Steps 1-4: Feature fusion module;

[0029] The features obtained in the previous step are passed through two fully connected layers. The first fully connected layer increases the dimension to four times the original, and the second fully connected layer restores the dimension. Then, they are projected onto a K-dimensional space.

[0030] F′=W2σ(W1(X T ))

[0031] F K =W3F′

[0032]

[0033] The negative sample loss is calculated using the feature vectors obtained from the two branches in the previous step, and the loss function is continuously optimized through backpropagation to improve the network's accurate discrimination ability. The negative sample loss uses the InfoNce loss, and the specific calculation formula is shown below:

[0034]

[0035] Where q represents the encoding of the ground-view image, γ is the encoded danger-view image, and only γ... + For a correct match, the InfoNCE loss uses a dot product to calculate the similarity between the query and the reference image, resulting in a lower similarity when the query and the positive match are similar, and a lower similarity when the negative match is positive. i The value is higher when it differs from q; τ is a parameter.

[0036] Preferably, step 2 specifically comprises:

[0037] Step 2-1: Feature extraction;

[0038] Based on the feature extraction model obtained from the training phase, the images to be matched and images from the subsequent image library are input into the feature extraction network model. After convolution to extract preliminary features, these features are then input into the self-attention feature module, and finally through the feature fusion module of the fully connected layer.

[0039] Step 2-2: Feature similarity calculation;

[0040] Based on the similarity of image features obtained in the previous step, image matching is performed.

[0041] Preferably, the similarity is measured by cosine distance, and its calculation formula is as follows:

[0042]

[0043] Where A and B are the features of the image to be matched and the features of the target image, and ||A|| is the size of the feature.

[0044] Preferably, the ConvNeXt network is ConvNeXt-T.

[0045] Preferably, the loss function is an information-noise contrast estimation loss function.

[0046] A computer program that causes a computer to perform the above-described cross-view image matching method.

[0047] An electronic device includes: a processor and a memory; the memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to cause the electronic device to perform the above-described cross-view image matching method.

[0048] A computer-readable storage medium having a computer program stored thereon that, when executed by a processor, implements the above-described cross-view image matching method.

[0049] A chip includes a processor for retrieving and running a computer program from a memory, causing a device on which the chip is mounted to perform the aforementioned cross-view image matching method.

[0050] A computer program product includes a computer storage medium storing a computer program, the computer program including instructions executable by at least one processor, which, when executed by the at least one processor, implement the above-described cross-view image matching method.

[0051] The beneficial effects of this invention are as follows:

[0052] This invention utilizes the geographical location information contained in satellite-view images to match ground-view images with satellite-view images, thereby obtaining the location of the ground image. It employs a Transformer and hard sample mining strategy, with the former learning global dependencies and the latter improving the model's generalization ability. This method can be used for image-assisted positioning and navigation in autonomous driving under GPS-denied environments. Attached Figure Description

[0053] Figure 1 This is a schematic diagram illustrating the matching process between ground-view and satellite-view images according to the present invention.

[0054] Figure 2 This is a schematic diagram of the training network model of the present invention.

[0055] Figure 3 This is a schematic diagram of the self-attention module of the present invention. Detailed Implementation

[0056] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0057] The purpose of this invention is to: match ground-view images with satellite images (drone images) containing geographic location information, addressing the inability of traditional convolutional neural networks, which rely on locality assumptions, to obtain the dependencies between global contexts. By introducing a Transformer structure, the visual ambiguity between images from different viewpoints is resolved, achieving more accurate cross-view image matching and thus having wider applicability.

[0058] The method for matching ground-view images with satellite-view images according to the present invention includes the following steps:

[0059] Step 1: Set up and train the network model for feature extraction;

[0060] Feature extraction consists of two parts: the first part is preliminary feature extraction based on the ConvNeXt network, and the second part is a multi-head attention network based on ViT to learn global features. A feature fusion layer is connected after the feature extraction network to obtain the final features.

[0061] The training of the feature extraction network for cross-view images is as follows:

[0062] The images in the dataset used for model training are from both ground-based and drone-based perspectives. To address the problem of not being able to select difficult samples before training, a simple GPS-based sampling strategy is used to initialize difficult negative samples before training, i.e., selecting negative samples based on GPS location information.

[0063] For samples viewed from the ground, polar coordinate transformation preprocessing is performed. This method utilizes the geometric relationship between ground-view images and satellite-view images to significantly reduce the domain differences between the images.

[0064] After the network has been trained for a certain number of generations, the method of mining negative samples based on GPS sampling is replaced with a method based on image cosine similarity, that is, the negative sample with the highest similarity is put into each training batch.

[0065] Step 2: After the training in Step 1 is completed, the ground images and satellite images that need to be matched are sequentially input into the network for feature extraction to obtain the feature description vector of each image;

[0066] Based on the feature description vector, the feature phase velocity of the ground image to be matched and the satellite image in the satellite image library to be matched is calculated, and the image with the highest feature similarity is taken as the matching result of the ground image to be matched this time.

[0067] Example:

[0068] To address the limitations of cross-view matching methods based on convolutional neural networks in visual location, which cannot learn global dependencies and hinder performance in complex scenes, performance may degrade in the presence of obstacles or dynamic objects (cars or pedestrians). Referring to the human visual system, which not only utilizes local image information but also makes more accurate predictions based on global context when visual information is unclear or incomplete, Transformer's positional encoding allows the network to flexibly learn location-related feature representations. Transformer divides the feature map into equal-sized blocks and can model the relationship between any two blocks. This invention proposes an image matching method for ground-view image localization and location navigation based on satellite images with geographic location information. Addressing the significant differences between ground-view images and UAV-view images, techniques such as polar coordinate transformation and hard sample mining are used to calculate negative sample loss, narrowing the feature spaces of correctly matched images and widening the feature spaces of incorrectly matched images. Once the model is trained, it can perform bidirectional image matching. That is, it can use a ground-view image as a request to match a satellite-view image, or use a satellite-view image as a request to match a ground-view image. By comparing the feature similarity of the target database, the most similar images are considered to belong to the same geographical location.

[0069] refer to Figure 1 The cross-view image matching and localization technology of the present invention mainly includes two steps: first, training a feature extraction network; and finally, using the extracted features to extract features of the image to be matched and the target image, and finding the most matching image features.

[0070] The training of the network model can be divided into four steps.

[0071] 1. The first step is to perform a polar coordinate transformation on the ground-view image. Because the ground image and satellite image have a certain geometric relationship, pixels on the same horizontal line in the ground image have the same depth; that is, connecting the horizontal lines in the ground image corresponds to a concentric circle in the satellite-view image. Secondly, on the vertical lines of the ground image, the depth increases with the increase of the ordinate, which corresponds to the radial rays in the satellite-view image. The relationship between the two can be expressed using the formula for converting from Cartesian coordinates to polar coordinates:

[0072]

[0073] This target transformation can significantly reduce domain differences in cross-view images.

[0074] 2. The second step involves constructing a preliminary feature extraction network based on the ConvNeXt network. In this invention, CovNeXt-T is used as the preliminary feature extraction network, and the original classification layer and the last convolutional block are removed, retaining only the backbone network for feature extraction. Furthermore, the input Siamese network does not share network parameters. Therefore, the backbone network contains 42 convolutional layers, which are divided into four parts. The first part is a single-layer convolution without pooling, using a 4x4 kernel. The second part is a three-layer ConvNeXt block with kernel sizes of 7x7, 1x1, and 1x1, respectively. The 7x7 kernel uses depthwise separable convolution. This module is repeated three times, and the output feature size is 56x56x96, where the last 96 represents the number of channels. The third part is also a three-layer ConvNeXt block, with the convolutional kernels still being 7x7, 1x1, 1x1, and the output feature size is 28x28x192. This part is repeated three times. The fourth part uses a three-layer ConvNeXt block, with the convolutional kernel size being 7x7, 1x1, 1x1. This module is repeated 9 times, and the output feature size is 14x14x384, resulting in a 14x14 small block. This feature will be input into the multi-head self-attention network.

[0075] 3. The third step is the Transformer-based self-attention feature module. The 14x14 blocks from the previous step are treated as a feature sequence with a dimension of 384, i.e., the input is {x1, x2, ... x...}. p The sequence} is projected into k and v by the self-attention mechanism.

[0076] Z′ l =LayerNorm(X l )

[0077] Q = W q Z′ l K = W k Z′ l V = W v Z′ l

[0078]

[0079] Among them W q W k W v These represent the weights of the query, key, and value, respectively, and H represents the number of multi-heads.

[0080] 4. The fourth step is the feature fusion module. The features obtained in the previous step are passed through two fully connected layers. The first fully connected layer increases the dimensionality to four times the original, and the second fully connected layer restores the dimensionality, which is then projected onto a K-dimensional space. Here, instead of global or average pooling to obtain the final descriptor, fully connected layers are used, allowing the network to freely learn how to aggregate feature information.

[0081] F′=W2σ(W1(X T ))

[0082] F K =W3F′

[0083]

[0084] The negative sample loss is calculated using the feature vectors obtained from the two branches in the previous step, and the loss function is continuously optimized through backpropagation to improve the network's accurate discrimination ability. The negative sample loss uses the InfoNce loss method, and the specific calculation formula is shown below:

[0085]

[0086] Where q represents the encoding of the ground-view image, and γ is the encoded danger-view image, where only γ... + For a correct match, the InfoNCE loss uses a dot product to calculate the similarity between the query and the reference image, resulting in a lower similarity when the query and the positive match are similar, and a lower similarity when the match is negative. i The value is higher when it differs from q. Here, τ is a parameter.

[0087] The image matching process can be divided into two steps, specifically:

[0088] The first step is feature extraction. Based on the feature extraction model obtained from the training part, the images to be matched and the images from the subsequent image library are input into the feature extraction network. After convolution, the preliminary features are extracted and then input into the autonomous one module. Then, the feature fusion module of the fully connected layer is used. In short, the trained model is used to extract the features of the two input images.

[0089] The second step involves calculating the similarity of features. Based on the similarity of image features obtained in the previous step, image matching is performed. In this specific implementation method, the similarity measure is cosine distance, and its calculation formula is as follows:

[0090]

[0091] Where A and B are the features of the image to be matched and the features of the target image, and ||A|| is the size of the feature.

[0092] The cross-view image matching method of the present invention is applied to the CVUSA & CVACT dataset to achieve cross-view matching between ground view images and satellite view images. The method implemented by the present invention has the advantages of strong robustness, wide applicability, and high matching accuracy. At the same time, it takes into account the global relationship between images and uses InfoNCE loss to regress images from different viewpoints belonging to the same geographical location to the same feature space.

[0093] In this embodiment, the selected datasets are the CVUSA dataset and the CVACT dataset. These are terrestrial and satellite images from the United States and Canberra. The training sets of both datasets consist of 35,532 pairs of training images, and the test sets each consist of 8,884 pairs of images. It is assumed that the center of the satellite image is strongly aligned with the corresponding terrestrial image.

[0094] (1) Polar coordinate transformation of ground view image data and sampling of difficult negative samples based on GPS location.

[0095] Polar coordinate transformation is applied to the ground-view images in the training set to reduce the feature differences between the ground-view images and the satellite-view images. Then, based on GPS information, the geographically closest negative samples are selected as hard negative samples to participate in training. Generally speaking, negative samples that are far away do not contribute much to the loss function.

[0096] (2) Construct a deep learning network model.

[0097] The model consists of three parts. The first part is a ConvNeXt network with the last convolutional and classification layers removed. The network is then fed into a self-attention module to learn dependencies. Finally, the features are fused using a fully connected layer to learn the relationships between different channels.

[0098] (3) Loss function regression training.

[0099] Using the features extracted in the previous step, we calculate the loss function between features to bring the distance between two correctly matched feature vectors closer to the same feature space, while keeping the feature spaces of incorrectly matched feature vectors far apart. After training for a period of time, we replace the sampling of difficult negative samples based on GPS location information with mining of difficult negative samples based on feature similarity, that is, we add the negative sample with the closest cosine distance to the training.

[0100] (4) Feature extraction and matching.

[0101] Once the network training is complete, the model can be used to perform matching between ground viewpoints and satellite viewpoints, or between satellite viewpoints and drone viewpoints. Specifically, the image to be matched and the image to be matched are input into the feature map extraction network, which calculates the true similarity of all images in the matching library, sorts them, and uses the image with the highest similarity as the final matching result.

Claims

1. A method for cross-view image matching for artillery reconnaissance aircraft, characterized in that, Includes the following steps: Step 1: Set up and train the feature extraction network model; The feature extraction network model consists of two parts: the first part is the preliminary feature extraction based on the ConvNeXt network; the second part is the multi-head attention network based on ViT for learning global features. A feature fusion layer is connected after the feature extraction network model to obtain the final features; The training process of the feature extraction network model is as follows: The images used in the training dataset were from both ground-based and drone-based perspectives; a GPS-based sampling strategy was used to initialize negative samples before training, i.e., negative samples were selected based on GPS location information. The samples from the ground view are preprocessed using polar coordinate transformation; After the network has been trained a certain number of times, the method of mining negative samples based on GPS sampling will be replaced with a method based on image cosine similarity, that is, the negative sample with the highest similarity will be put into each training batch. Step 2: After the training in Step 1 is completed, the ground images and satellite images that need to be matched are sequentially input into the network for feature extraction to obtain the feature description vector of each image; Based on the feature description vector, the feature phase velocity of the ground image to be matched and the satellite image in the satellite image library to be matched is calculated, and the image with the highest feature similarity is taken as the matching result of the ground image to be matched this time.

2. The method for cross-view image matching for artillery reconnaissance aircraft according to claim 1, characterized in that, The training process of the feature extraction network model is as follows: Step 1-1: Perform polar coordinate transformation on the ground view image; The relationship between terrestrial and satellite imagery can be expressed using the formula for converting from Cartesian to polar coordinates: Steps 1-2: Construct a preliminary feature extraction network based on the ConvNeXt network; CovNeXt-T was used as the initial feature extraction network, and the original classification layer and the last convolutional block were removed, retaining only the feature extraction backbone network. Furthermore, the input Siamese network did not share network parameters. Therefore, the backbone network contained 42 convolutional layers, divided into four parts. The first part was a single-layer convolution without pooling, using a 4x4 kernel. The second part was a three-layer ConvNeXt block with kernel sizes of 7x7, 1x1, and 1x1, respectively. The 7x7 kernel used a depth-scalable... The first part is a three-layer ConvNeXt block, with three convolutional kernels of 7x7, 1x1, and 1x1. The output feature size is 56x56x96, where the last 96 represents the number of channels. The second part is also a three-layer ConvNeXt block, with three convolutional kernels of 7x7, 1x1, and 1x1. The output feature size is 28x28x192, and this part is repeated three times. The third part uses a three-layer ConvNeXt block, with convolutional kernels of 7x7, 1x1, and 1x1. This block is repeated nine times, and the output feature size is 14x14x384, resulting in a 14x14 block. Steps 1-3: Transformer-based self-attention feature module; The 14x14 blocks from the previous step are treated as a feature sequence with a dimension of 384, i.e., the input is {x1, x2, ... x... p The sequence} is projected by the self-attention mechanism onto k and v respectively; Z′ l =LayerNorm(X l ) Q=W q Z′ l ,K=W k Z′ l ,V=W v Z′ l Among them W q W k W v These are the weights of query, key, and value, respectively, and H represents the number of multi-heads; Steps 1-4: Feature fusion module; The features obtained in the previous step are passed through two fully connected layers. The first fully connected layer increases the dimension to four times the original, and the second fully connected layer restores the dimension. Then, they are projected onto a K-dimensional space. F′=W2σ(W1(X T )) F K =W3F′ The negative sample loss is calculated using the feature vectors obtained from the two branches in the previous step, and the loss function is continuously optimized through backpropagation to improve the network's accurate discrimination ability. The negative sample loss uses InfoNceloss, and the specific calculation formula is shown below: Where q represents the encoding of the ground-view image, γ is the encoded danger-view image, and only γ... + For a correct match, the InfoNCE loss uses a dot product to calculate the similarity between the query and the reference image, resulting in a lower similarity when the query and the positive match are similar, and a lower similarity when the negative match is positive. i The value is higher when it differs from q; τ is a parameter.

3. The method for cross-view image matching for artillery reconnaissance aircraft according to claim 2, characterized in that, Step 2 specifically involves: Step 2-1: Feature extraction; Based on the feature extraction model obtained from the training phase, the images to be matched and images from the subsequent image library are input into the feature extraction network model. After convolution to extract preliminary features, these features are then input into the self-attention feature module, and finally through the feature fusion module of the fully connected layer. Step 2-2: Feature similarity calculation; Based on the similarity of image features obtained in the previous step, image matching is performed.

4. The method for cross-view image matching for artillery reconnaissance aircraft according to claim 3, characterized in that, The similarity is measured by cosine distance, and its calculation formula is as follows: Where A and B are the features of the image to be matched and the features of the target image, and ||A|| is the size of the feature.

5. A method for cross-view image matching for artillery reconnaissance aircraft according to claim 4, characterized in that, The ConvNeXt network is ConvNeXt-T.

6. The method for cross-view image matching for artillery reconnaissance aircraft according to claim 5, characterized in that, The loss function is the information-noise contrast estimation loss function.

7. A computer program, characterized in that, The computer program causes the computer to perform the method as described in any one of claims 1 to 6.

8. An electronic device, characterized in that, include: Processor and memory; The memory is used to store a computer program, and the processor is used to execute the computer program stored in the memory to cause the electronic device to perform the method as described in any one of claims 1 to 6.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the method as described in any one of claims 1 to 6.

10. A chip, characterized in that, include: A processor for retrieving and running a computer program from memory, causing a device on which the chip is mounted to perform the method as described in any one of claims 1 to 6.

Images