A sea surface small target detection method based on multi-domain multi-dimensional feature combination

Abstract

The application discloses a sea surface small target detection method based on multi-domain multi-dimensional feature combination, and adaptively extracts deep features of sea clutter and target echo signals by using a stack sparse auto-encoder, so that the feature dimension is improved to ensure the feature feasibility. Since the feature extraction is a complete adaptive process, the complexity of the model is reduced. Meanwhile, aiming at the problem that the feature distinction degree of the sea clutter and target echo data in a single domain is low, a time-frequency domain feature combination method is provided to improve the feature difference and ensure the stable and efficient detection performance of the detector. Through an adaptive genetic algorithm, the convergence process of the super parameter group optimization is accelerated, and the local optimization of the super parameter is prevented to a certain extent, so that the final detection probability is effectively improved. The experimental results show that the detector provided by the application has better detection effect on high sea state data, the detection probability is improved by 27.6%, and the sea surface small target detection under high sea state can be coped with.

Description

A method for detecting small targets on the sea surface based on the combination of multi-domain and multi-dimensional features Technical Field

[0001] This invention relates to a method for detecting small targets on the sea surface based on the combination of multi-domain and multi-dimensional features, belonging to the field of radar signal processing. Background Technology

[0002] Small targets on the sea surface are affected by factors such as their small radar cross section (RCS) and complex sea conditions, resulting in very weak signals that are easily submerged in the background of sea clutter, making them difficult to detect and leading to low detection rates. Therefore, finding a suitable difference between the sea clutter and the target signal is an important research topic.

[0003] Currently, methods for detecting small targets against sea clutter backgrounds mainly fall into two categories. The first category analyzes sea clutter characteristics from the perspective of statistical models. The most widely used statistical model in sea clutter backgrounds is the K-distribution model, which considers the correlation of radar echo data and can fit the true distribution well. However, the sea surface is affected by many factors such as wind and tides, exhibiting strong non-stationarity. Statistical models of sea clutter characteristics are often based on stable sea surface conditions; therefore, in complex sea conditions and with low signal-to-clutter ratios, the detection performance of statistical models is poor. The second category is based on nonlinear methods, directly analyzing the properties of sea clutter from a spatiotemporal perspective. In recent years, neural network learning methods have also made some progress in sea clutter target detection. Haykin treats sea clutter as a chaotic system and proposes a method for detecting weak targets in sea clutter based on a BP neural network model; Guo Zixun et al. proposed an improved K-nearest neighbor method for detecting small targets on the sea surface with controllable false alarms using feature detection; Shi Sainan proposed a method for detecting small targets on the sea surface based on multi-domain and multi-dimensional feature fusion, which can extract multi-dimensional features from multiple domains and has good detection performance on IPIX data.

[0004] Current methods for detecting small targets on the sea surface based on multi-domain and multi-dimensional feature fusion suffer from problems such as complex models and uncontrollable false alarms. Summary of the Invention

[0005] This invention provides a method for detecting small targets on the sea surface based on the combination of multi-domain and multi-dimensional features, which solves the problems disclosed in the background art.

[0006] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows:

[0007] A method for detecting small targets on the sea surface based on the combination of multi-domain and multi-dimensional features:

[0008] The sea clutter and target echo signals are acquired and divided into balanced training samples and detection samples.

[0009] The detection sample is input into a pre-built detector based on a stacked sparse autoencoder and the GA-XGBoost algorithm to obtain the predicted value ρ of the detection sample.

[0010] The detector update decision threshold γ based on stacked sparse autoencoder and GA-XGBoost algorithm;

[0011] The predicted value ρ of the detected sample is compared with the decision threshold γ. If the predicted value ρ of the detected sample is greater than the decision threshold γ, the echo signal contains the target. If the predicted value ρ of the detected sample is less than the decision threshold γ, the echo signal does not contain the target.

[0012] Furthermore, the construction process of the detector based on the stacked sparse autoencoder and GA-XGBoost algorithm, which is pre-built from the input of the detection sample, is as follows:

[0013] A stacked sparse autoencoder is constructed to extract features that can characterize the signal from the training samples. By adding constraints to the sparse representation, deep features of the echo are extracted.

[0014] An XGBoost model is constructed to evaluate and classify the extracted deep echo features. The hyperparameter set in the XGBoost model is extracted, and the hyperparameter set is optimized using an adaptive genetic optimization algorithm to obtain the optimal hyperparameter set.

[0015] By combining the optimal hyperparameter set with a stacked sparse autoencoder, a detector based on a stacked sparse autoencoder and the GA-XGBoost algorithm is obtained.

[0016] Furthermore, a stacked sparse autoencoder is constructed to extract features that characterize the signal from the training samples. By adding constraints to the sparse representation, the process of extracting deep-level features of the echo is as follows:

[0017] The sea clutter and target echo signals are labeled separately, with 1 indicating the presence of a target and 0 indicating the absence of a target.

[0018] Randomly generate weights W1 and W2 and biases B1 and B2 for each layer of the autoencoder network to construct an initial autoencoder network; the network formula is as follows:

[0019] h = σ(W1·x + B1)

[0020] y = σ(W²·h + B²)

[0021] Where x is the input layer, h is the hidden layer, y is the output layer, and the loss function during training is:

[0022]

[0023] Where m is the number of units in the input layer, n l s represents the total number of layers in the stacked autoencoder network. l Let λ be the number of neurons in the first layer, and λ be the weight decay parameter. The first term of the loss function is the sum of the minimum mean squared errors between the actual values ​​and the network output, where ||·|| is the L2 norm. The second term of the loss function is a regularization term, which reduces the weight magnitude to prevent overfitting; specifically, it is the sum of the squares of the coefficients of each term in the encoding and decoding weight matrices. A sparsity penalty term, KL(||), is introduced into the loss function:

[0024]

[0025] Where ρ is the sparsity constant, and ρi represents the average activation of the i-th neuron. Two constructed initialization stacked sparse autoencoder networks are used to extract features from the signal to be detected in the time domain and frequency domain, respectively. In each stacked sparse autoencoder, the hidden layer of the previous layer is used as the input layer of the next layer, and finally outputs 64-dimensional features.

[0026] The output 64-dimensional time-domain features are concatenated with the 64-dimensional frequency-domain features to form a 128-dimensional deep feature.

[0027] Furthermore, the hyperparameter set is extracted from the XGBoost model, and the optimal hyperparameter set is obtained by using an adaptive genetic optimization algorithm:

[0028] S31: Set the range of hyperparameters in the XGBoost model and perform binary encoding on the hyperparameter set;

[0029] S32: Calculate the fitness value of the hyperparameter set. The fitness value is selected as the detection probability of the signal to be detected by the XGBoost model.

[0030] S33: Use selection operators, crossover operators, and mutation operators to iteratively update hyperparameters using a genetic algorithm;

[0031] Among them, the selection operator uses a roulette wheel method to calculate the fitness value of each individual and form a roulette wheel ratio for random selection;

[0032] The crossover operator sets a threshold k and performs crossover at the k-th position of the gene.

[0033] The mutation operator selects the gene mutation location based on the mutation probability and performs a 0-1 transition to prevent local optimization.

[0034] S34: Update the values ​​of crossover probability Pc and mutation probability Pm according to the following formula to speed up the iteration:

[0035]

[0036] Where f is the fitness value, favg and fmax are the average and maximum fitness values, respectively, Pcmax, Pcmin, Pmmax and Pmmin are the upper and lower limits of the crossover probability and mutation probability, respectively, and A is the weight update parameter.

[0037] S35: Determine if the fitness value has converged. If the fitness value has converged, terminate the current iteration and output the optimal hyperparameter set; otherwise, return to steps S33 and S34 and continue iterating.

[0038] Furthermore, the method for allocating training and detection samples for sample balancing is as follows: sea clutter data and target echo data of equal length are selected respectively, and overlapping data selection is performed, with a cumulative pulse number of N = 1024 and an overlap of P = 924.

[0039] Furthermore, the method for updating the decision threshold γ of the detector based on the stacked sparse autoencoder and the GA-XGBoost algorithm is as follows:

[0040] The detector evaluation classification actually obtains n sets of predicted values ​​from the sea clutter data, arranged in descending order, denoted as ρ1, ρ2, ..., ρ3. n ;

[0041] Selected false alarm rate P fa The calculation formula is as follows:

[0042]

[0043] The decision threshold γ is calculated using the following formula:

[0044]

[0045] Accordingly, a computer-readable storage medium storing one or more programs: the one or more programs include instructions that, when executed by a computing device, cause the computing device to perform any of the methods described above.

[0046] Accordingly, a computing device includes:

[0047] One or more processors, one or more memories, and one or more programs, wherein the one or more programs are stored in the one or more memories and configured to be executed by the one or more processors, and the one or more programs include instructions for performing any of the methods described above.

[0048] The beneficial effects achieved by this invention are as follows: Compared with existing technologies, this invention uses a stacked sparse autoencoder to adaptively extract deep features of sea clutter and target echo signals, improving feature dimensionality and ensuring feature feasibility. Since feature extraction is a fully adaptive process, the complexity of the model is reduced. Simultaneously, addressing the problem of low feature discrimination between sea clutter and target echo data in a single domain, a method combining time-frequency domain features is proposed to improve feature discriminability, ensuring stable and efficient detection performance of the detector. Through an adaptive genetic algorithm, the convergence process of hyperparameter optimization is accelerated, while preventing local hyperparameter optimization to a certain extent, effectively improving the final detection probability. Experimental results show that the detector proposed in this invention has better detection performance for high sea state data, increasing the detection probability by 27.6%, and can handle the detection of small targets on the sea surface under high sea states. Attached Figure Description

[0049] Figure 1 is a schematic diagram of the detection process of the present invention;

[0050] Figure 2 is a schematic diagram of the decision threshold update process;

[0051] Figure 3 shows a comparison of the detection performance of three different domain features.

[0052] Figure 4 shows the detection performance of the detector proposed in this invention under four polarizations for 10 sets of data.

[0053] Figure 5 shows a comparison of the performance of four detectors under HH polarization with 10 sets of data. Detailed Implementation

[0054] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0055] As shown in Figure 1, the sea surface small target detection method based on multi-domain and multi-dimensional feature combination of the present invention includes the following steps:

[0056] S1: Acquire the signal to be detected and perform noise reduction preprocessing on the signal to be detected. The signal to be detected includes sea clutter signal and sea clutter and target echo signal. Label the sea clutter signal and the target echo signal with “0” and “1” respectively.

[0057] S2: Construct a two-layer stacked sparse autoencoder to extract a series of features that characterize the signal from the preprocessed signal to be detected. By adding constraints to the sparse representation, deep-level features of the echo are extracted;

[0058] S3: Construct an XGBoost model, evaluate and classify the extracted features, extract the hyperparameter set in the XGBoost network, and use an adaptive genetic optimization algorithm to optimize the hyperparameter set to quickly obtain the optimal hyperparameter set.

[0059] S4: Based on the selected sea clutter and target echo data, the samples are divided into balanced training samples and detection samples. The training samples are processed through the above steps to obtain the optimal hyperparameter set of the trained stacked sparse autoencoder network and XGBoost network, and a detector based on the stacked sparse autoencoder and GA-XGBoost algorithm is obtained.

[0060] S5: Combine the predicted values ​​of a series of sea clutter data obtained from the detector with the false alarm rate calculation, update the decision threshold γ, and realize the controllable false alarm characteristics of the detector.

[0061] S6: Input the detection sample into the detector to obtain the predicted value ρ of the detection sample. By comparing it with the decision threshold γ, determine whether there is a target in the echo signal.

[0062] In step S1, the detection problem is categorized as a binary hypothesis test:

[0063]

[0064] Where c represents the sea clutter signal, s represents the target echo signal, H0 assumption means that the echo signal contains only sea clutter signal, and H1 assumption means that the echo signal contains the target echo.

[0065] In step S2, the process of constructing a deep stacked sparse autoencoder and extracting deep features of the target signal includes the following steps:

[0066] S21: Label the sea clutter and target echo signals in the signal to be tested, with 1 for the presence of a target and 0 for the absence of a target;

[0067] S22: Randomly generate the weights and biases W1, W2, B1, B2 for each layer of the autoencoder network to construct the initial autoencoder network. The network and formulas are as follows:

[0068] h = σ(W1·x + B1)

[0069] y = σ(W²·h + B²)

[0070] S23: The loss function during training is

[0071]

[0072] Introduce a sparse penalty term KL(||) into the loss function:

[0073]

[0074] S24: Two constructed initial stacked sparse autoencoder networks are used to extract features from the signal to be detected in the time domain and frequency domain respectively. In each stacked sparse autoencoder, the hidden layer of the previous layer is used as the input layer of the next layer, and finally 64-dimensional features are output respectively.

[0075] S25: Connect the output 64-dimensional time-domain features with the 64-dimensional frequency-domain features to form a 128-dimensional deep feature.

[0076] Step S3, the process of extracting the hyperparameter set from the XGBoost network and optimizing the hyperparameters using an adaptive genetic optimization algorithm, includes the following steps:

[0077] S31: Set the range of hyperparameters in the XGBoost network and perform binary encoding on the hyperparameter group;

[0078] S32: Calculate the fitness value of the hyperparameter set. The fitness value is selected as the detection probability of the signal to be detected through the XGBoost network.

[0079] S33: Use selection, crossover, and mutation operators to iteratively update hyperparameters using a genetic algorithm. The three operators are as follows:

[0080] (1) Selection Operator. Using a roulette wheel method, the fitness value of each individual is calculated to form a roulette wheel ratio for random selection;

[0081] (2) Crossover operator. A threshold k is set, and crossover is performed at the k-th position of the gene;

[0082] (3) Mutation operator. Select the gene mutation location based on the mutation probability, perform a 0-1 transition, and prevent local optimization;

[0083] S34: Update the values ​​of crossover probability a and mutation probability b according to the following formula to speed up the iteration:

[0084]

[0085] S35: Determine if the fitness value has converged. If the fitness value has converged, terminate the current iteration and output the optimal hyperparameter set; otherwise, return to steps S43 and S44 and continue iterating.

[0086] In step S4, the method for balancing sea clutter data and echo data is as follows: select sea clutter data and target echo data of equal length respectively, and perform overlapping data selection. The cumulative pulse number is N = 1024, and the overlap is P = 924.

[0087] In step S5, the process of realizing the controllable false alarm characteristic includes the following steps:

[0088] S51: Obtain n sets of predicted values ​​for the actual sea clutter data used for detector evaluation and classification, and arrange them in descending order, denoted as ρ1, ρ2, ..., ρ n ;

[0089] S52: Select the false alarm rate P fa Its calculation formula is

[0090]

[0091] S53: Calculate the decision threshold γ, the formula is as follows:

[0092]

[0093] S54: Update the decision threshold γ to meet the false alarm rate condition, thereby achieving controllable false alarms.

[0094] In step S6, the judgment criteria are as follows: when ρ>γ, it is judged that there is a target in the detection signal, which belongs to the H1 hypothesis; when ρ<γ, it is judged that there is no target in the detection signal, which belongs to the H0 hypothesis.

[0095] The data used in this paper comes from the IPIX radar target database, collected by Professor Haykin of McMaster University on the east coast of Canada. Ten sets of data were used in the experiment, each consisting of 14 adjacent range cells, each containing 131,072 pulses, with a range resolution of 30m. The target was a polystyrene foam ball wrapped in wire mesh, approximately 1m in diameter. Four polarization schemes were obtained based on the different data transmission and reception methods: HH, HV, VH, and VV.

[0096] Figure 2 illustrates the decision threshold update process. There are a total of 1134 samples with an actual label of "0". Initializing the decision threshold γ = 0.5 results in 9 false alarms. Increasing the decision threshold γ = 0.9688 results in only one false alarm. The decision threshold is updated in real-time during training to ensure controllable false alarms.

[0097] Figure 3 compares the detection performance of three different domain features. The detection probability of time-domain features is 56.60%, that of frequency-domain features is 69.62%, and that of combined time-frequency domain features is 94.30%. It is evident that the detection performance of high-dimensional features extracted by the time-frequency domain feature combination method is significantly higher than that of high-dimensional features extracted from a single domain.

[0098] Figure 4 shows the detection performance of the detector proposed in this invention under four polarizations with 10 sets of data. The parameters are as follows: P fa =10 -3 N = 1024, P = 924. It can be seen that the detection performance under HH, HV, and VH polarization modes is good, and the detection probability under HH polarization mode is the best.

[0099] Figure 5 shows a performance comparison of four detectors under HH polarization for 10 sets of data. It can be seen that the detector proposed in this invention generally outperforms the other three detectors. The detection probabilities for sea state #17 and #280 data are significantly improved compared to the other detectors, increasing by 27.6% and 57.7% respectively. This demonstrates that the detection method proposed in this invention has better detection performance for sea state data.

[0100] In summary, this invention solves the problem of complex design of high-dimensional feature classifiers, effectively addresses the poor detection performance of traditional detectors on high sea state data, and effectively improves the detection probability of small floating targets on the sea surface.

[0101] The above description is only a preferred embodiment of the present invention. It should be noted that for those skilled in the art, several improvements and modifications can be made without departing from the technical principles of the present invention, and these improvements and modifications should also be considered within the scope of protection of the present invention.

[0102] A computer-readable storage medium storing one or more programs, the one or more programs including instructions that, when executed by a computing device, cause the computing device to perform a method for detecting small targets on the sea surface based on a combination of multi-domain and multi-dimensional features.

[0103] A computing device includes one or more processors, one or more memories, and one or more programs, wherein the one or more programs are stored in the one or more memories and configured to be executed by the one or more processors, and the one or more programs include instructions for executing a method for detecting small targets on the sea surface based on the combination of multi-domain and multi-dimensional features.

[0104] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0105] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create means for implementing the functions specified in one or more blocks of the flowchart illustrations and / or one or more blocks of the block diagrams.

[0106] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means that implement the functions specified in one or more flowcharts and / or one or more block diagrams.

[0107] These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, such that the instructions, which execute on the computer or other programmable apparatus, provide steps for implementing the functions specified in one or more flowcharts and / or one or more block diagrams.

[0108] The above are merely embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention are included within the scope of the claims of the present invention pending approval.

Claims

1. A method for detecting small targets on the sea surface based on the combination of multi-domain and multi-dimensional features, characterized in that: Sea clutter and target echo signals are acquired and divided into balanced training and detection samples. The detection samples are then input into a pre-built detector based on a stacked sparse autoencoder and the GA-XGBoost algorithm to obtain the predicted values ​​of the detection samples. The process of constructing the detector based on the stacked sparse autoencoder and GA-XGBoost algorithm, which is input to the detection sample, is as follows: A stacked sparse autoencoder is constructed to extract features that characterize the signal from the training samples. By adding sparse representation constraints, deep echo features are extracted. An XGBoost model is constructed to evaluate and classify the extracted deep echo features. Hyperparameter sets are extracted from the XGBoost model, and the fitness value of the hyperparameter set is selected as the detection probability of the signal to be detected by the XGBoost model. An adaptive genetic optimization algorithm is used to optimize the hyperparameter set to obtain the optimal hyperparameter set. The optimal hyperparameter set is combined with the stacked sparse autoencoder to obtain the detector based on the stacked sparse autoencoder and GA-XGBoost algorithm. Two constructed initialized stacked sparse autoencoder networks are used to extract features from the signal to be detected in the time and frequency domains respectively. In each stacked sparse autoencoder, the hidden layer of the previous layer is used as the input layer of the next layer, and finally, 64-dimensional features are output respectively. The output 64-dimensional time-domain features are concatenated with the 64-dimensional frequency-domain features to form a 128-dimensional deep feature; the detector based on the stacked sparse autoencoder and GA-XGBoost algorithm updates the decision threshold. The detector update decision threshold based on stacked sparse autoencoder and GA-XGBoost algorithm The method is as follows: Obtain n sets of predicted values ​​for the actual sea clutter data used for detector evaluation and classification, and arrange them in descending order, denoted as . ， ，…， Select the false alarm rate The calculation formula is as follows: ; Calculate the decision threshold The calculation formula is as follows: ； Compare the predicted values ​​of the test samples With the judgment threshold If the predicted value of the detected sample Greater than the judgment threshold The echo signal contains the target; if the sample prediction value is measured... Less than the judgment threshold Then there is no target in the echo signal.

2. The method for detecting small targets on the sea surface based on the combination of multi-domain and multi-dimensional features according to claim 1, characterized in that: A stacked sparse autoencoder is constructed to extract features that characterize the signal from the training samples. By incorporating sparse representation constraints, the process of extracting deep-level echo features is as follows: Sea clutter and target echo signals are labeled separately, with 1 for the presence of a target and 0 for the absence of a target; the weights W1 and W2 and biases B1 and B2 of each layer in the autoencoder network are randomly generated to construct an initial autoencoder network; the network formula is as follows: ； ;in, For the input layer, For hidden layers, For the output layer, the loss function during training is: ;in, The number of units in the input layer. The total number of layers in the stacked autoencoder network. This represents the number of neural units in the first layer. Here, represents the weight decay parameter; the first term of the loss function is the sum of the minimum mean square error between the actual value and the network output, and ||·|| is... The paradigm; the second term of the loss function is a regularization term, which reduces the magnitude of the weights to prevent overfitting. Specifically, it is represented as the sum of the squares of the coefficients of each term in the encoding and decoding weight matrices; a sparsity penalty term is introduced into the loss function. ： ;in, It is a sparsity constant. Indicates the first Average activation level of each neuron.

3. The method for detecting small targets on the sea surface based on the combination of multi-domain and multi-dimensional features according to claim 1, characterized in that: The process of extracting the hyperparameter set from the XGBoost model and optimizing it using an adaptive genetic optimization algorithm to obtain the optimal hyperparameter set is as follows: S31: Set the range for the hyperparameters in the XGBoost model and perform binary encoding on the hyperparameter set; S32: Calculate the fitness value of the hyperparameter set; S33: Use selection, crossover, and mutation operators to iteratively update the hyperparameters using a genetic algorithm; Specifically, the selection operator uses a roulette wheel method to calculate the fitness value of each individual and randomly selects individuals; the crossover operator sets a threshold... In the first gene Crossover occurs at each position; the mutation operator selects the gene mutation position based on the mutation probability, performing a 0-1 transition to prevent local optimization; S34: the crossover probability is updated according to the following formula. With mutation probability The value of [value] speeds up the iteration process. ； ;in, For fitness value, 、 These are the average and maximum fitness values, respectively. 、 、 、 These are the upper and lower limits of the crossover probability and the mutation probability, respectively. Update the parameters for the weights; S35: Determine if the fitness value has converged. If the fitness value has converged, the current iteration terminates and the optimal hyperparameter set is output; otherwise, return to steps S33 and S34 and continue iterating.

4. The method for detecting small targets on the sea surface based on the combination of multi-domain and multi-dimensional features according to claim 1, characterized in that: The method for allocating training and detection samples for sample balancing is as follows: Sea clutter data and target echo data of equal length are selected and overlapped, with a cumulative pulse count of [number missing]. The degree of overlap is 。 5. A computer-readable storage medium for storing one or more programs, characterized in that: The one or more programs include instructions that, when executed by a computing device, cause the computing device to perform any of the methods according to claims 1 to 4.

6. A computing device, characterized in that, include: One or more processors, one or more memories, and one or more programs, wherein the one or more programs are stored in the one or more memories and configured to be executed by the one or more processors, the one or more programs including instructions for performing any of the methods according to claims 1 to 4.

Images