Domain adaptive object detection method based on dynamic data driving and consistency constraint

Abstract

The application relates to a domain adaptive target detection method based on dynamic data driving and consistency constraint, and belongs to the technical field of target detection. The method comprises the following steps: a target detection model containing a teacher-student network is constructed, and labeled source domain images and unlabeled target domain images are input; a semantic prediction consistency loss composed of semantic context consistency loss and prediction distribution consistency loss is calculated; a dynamic modulation factor is calculated based on the loss, a frequency domain filter cutoff frequency threshold is adjusted, and frequency domain dynamic processing is performed on the source domain images; the processed source domain images are input into the student network to calculate a supervised detection loss, and a total loss function is constructed in combination with the semantic prediction consistency loss; and the model is iteratively optimized based on the total loss function, the student network parameters are updated through back propagation, and the teacher network parameters are updated through exponential moving average. While improving the robustness and accuracy of cross-domain detection, no additional inference cost is required, and the engineering application value is high.

Description

Technical Field

[0001] This application relates to the field of target detection technology, and in particular to a domain adaptive target detection method based on dynamic data-driven and consistency constraints. Background Technology

[0002] Object detection technology has made significant progress in the field of computer vision. However, in practical applications, due to the diversity of acquisition scenarios (such as changes in lighting, cross-camera conditions, and differences in weather conditions), there are significant distributional differences between the source domain data used for training and the target domain data in actual applications, leading to a substantial decrease in the detection performance of the model in the target domain. To solve this problem, domain-adaptive object detection technology has emerged, with the core objective of effectively transferring knowledge from the source domain to the unlabeled target domain.

[0003] Existing mainstream domain adaptation methods are mainly divided into two categories: feature alignment and self-training. Both essentially pursue the "statistical invariance" of the feature distributions of the source and target domains, bringing features closer together by minimizing the distribution distance between the two domains. However, these methods adopt a "brutal" alignment strategy, only looking for the intersection of features between the source and target domains. This intersection contains both the essential features that determine the object category and non-semantic environmental features shared by both domains. Indiscriminate alignment can lead the model to learn an incorrect "shortcut"—relying on easily aligned background environmental features to identify objects, rather than focusing on the objects themselves, thus limiting the upper limit of the model's detection performance.

[0004] Furthermore, some existing technologies attempt to decompose images using frequency domain transformation to filter out environmental feature interference, but most of them use fixed filter bandwidth or static frequency thresholds, ignoring the dynamic changes in feature sensitivity during model training. Fixed thresholds in the early stages of training may incorrectly filter out key target features, and when the model tends to stabilize, it cannot effectively suppress residual environmental noise. It is difficult to achieve a dynamic balance between preserving target features and removing environmental interference, which further limits the effectiveness of domain adaptation.

[0005] Therefore, there is an urgent need in related technologies for a way to improve the robustness and accuracy of cross-domain target detection models. Summary of the Invention

[0006] Therefore, it is necessary to provide a domain adaptive target detection method based on dynamic data driving and consistency constraints that can improve the robustness and accuracy of cross-domain target detection models, addressing the aforementioned technical problems.

[0007] Firstly, this application provides a domain-adaptive target detection method based on dynamic data-driven principles and consistency constraints. The method includes: Construct a target detection model that includes a teacher network and a student network, and input labeled source domain images and unlabeled target domain images into the target detection model; Calculate the semantic prediction consistency loss, which includes the semantic context consistency loss and the prediction distribution consistency loss; The dynamic modulation factor is calculated based on the semantic prediction consistency loss. The cutoff frequency threshold of the frequency domain filter is adjusted using the dynamic modulation factor. The adjusted cutoff frequency threshold is then used to perform dynamic frequency domain processing on the source domain image. The source domain image after dynamic processing in the frequency domain is input into the student network, the supervised detection loss is calculated, and the total loss function is constructed by combining the supervised detection loss and the semantic prediction consistency loss. The target detection model is iteratively optimized and trained based on the total loss function, and the parameters of the student network and the teacher network are updated.

[0008] Optionally, in one embodiment of this application, both the teacher network and the student network use a feature pyramid network as their backbone network and both include a region proposal network and a region of interest detection head.

[0009] Optionally, in one embodiment of this application, the calculation of the semantic prediction consistency loss, which includes semantic context consistency loss and prediction distribution consistency loss, includes: Perform instance normalization on the intermediate layer feature maps of the feature pyramid to remove style statistics from the image; The global semantic similarity of features between the student network and the teacher network after instance normalization is calculated. The semantic context consistency loss is obtained by calculating the cosine similarity, combined with the preset edge and hierarchical weights.

[0010] Optionally, in one embodiment of this application, the calculation of the semantic prediction consistency loss, which includes semantic context consistency loss and prediction distribution consistency loss, further includes: Using the output of the teacher network as pseudo-labels, a general anchor box sampling strategy is adopted to reuse the positive and negative sample anchor box indexes and region of interest suggestions screened by the teacher network. Based on the anchor box index and region of interest suggestions, the KL divergence of the classification probability distributions of the student network and the teacher network, the L1 loss of the regression offset, and the summation of these results in the prediction distribution consistency loss.

[0011] Optionally, in one embodiment of this application, calculating the dynamic modulation factor based on the semantic prediction consistency loss includes: The semantic prediction consistency loss is smoothed by exponential moving average, and the deviation rate of the semantic prediction consistency loss relative to the historical mean is calculated. The deviation rate is then mapped to a dynamic modulation factor using a hyperbolic tangent function.

[0012] Optionally, in one embodiment of this application, the cutoff frequency threshold of the frequency domain filter is the upper and lower thresholds of the statically held frequency band. The updated real-time cutoff frequency threshold is obtained by exponentially calculating the initial upper and lower thresholds of the statically held frequency band using the dynamic modulation factor.

[0013] Optionally, in one embodiment of this application, performing frequency domain dynamic processing on the source domain image using the adjusted cutoff frequency threshold includes: The source domain image is transformed to the frequency domain space through discrete cosine transform, and the frequency domain space is divided into statically maintained frequency bands and perturbation frequency bands using the updated cutoff frequency threshold. Based on the cutoff frequency threshold, a static hold-up band mask and a perturbation band mask are generated. Gaussian noise is injected into the perturbation band, and the source domain image is restored by combining the features of the two bands through inverse discrete cosine transform to obtain the processed source domain image.

[0014] Optionally, in one embodiment of this application, during training, the parameters of the student network are updated based on the total loss function through backpropagation, and the parameters of the teacher network are updated through the exponential moving average of the student network parameters, without participating in gradient backpropagation.

[0015] Secondly, this application also provides a computer device. The computer device includes a memory and a processor, the memory storing a computer program, and the processor executing the steps of the methods described in the various embodiments above.

[0016] Thirdly, this application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program thereon, which, when executed by a processor, implements the steps of the methods described in the various embodiments above.

[0017] The aforementioned domain-adaptive target detection method based on dynamic data-driven and consistency constraints has the following significant technical advantages compared with existing technologies: First, it solves the technical problem that static frequency domain processing cannot adapt to the model training state: This invention establishes a closed-loop feedback mechanism of "training state perception - data augmentation adjustment", using semantic prediction consistency loss as a real-time feedback signal to dynamically adjust the bandwidth of the frequency domain filter. In the early stage of training, more frequency bands are automatically retained to ensure basic feature learning. In the later stage of training, the perturbation intensity of environmental feature frequency bands is increased, avoiding the tediousness of manually searching for parameters, preventing model training from crashing or prematurely converging to suboptimal solutions, and ensuring that the model is on the optimal learning path throughout the training cycle.

[0018] Second, it effectively decouples the essential features of the target from non-semantic environmental features, improving generalization ability: by using instance normalization and attention mechanisms to remove non-semantic environmental features from the feature map, and combined with dynamic noise injection in the frequency domain, it cuts off the model's "shortcut learning" path, forcing the model to focus on the essential features of the target, such as structure and contour. In cross-domain testing from urban street scene to foggy scene, this method achieves an average precision (mAP) of 65.6% at the highest difficulty with a fog concentration of 0.02, significantly outperforming existing mainstream methods and demonstrating strong detection robustness under complex environmental changes.

[0019] Third, it achieves high-precision detection without additional inference costs: The dynamic data-driven module and consistency constraint module introduced in this invention only play a role in the training phase. The model deployment and inference phases only need to use the trained teacher network, without loading the student network or additional frequency domain transformation modules. While significantly improving cross-domain detection accuracy, it does not increase the computational burden and number of parameters in the inference phase, making it more suitable for deployment on devices with high real-time requirements such as autonomous driving and edge computing, and has high engineering application value. Attached Figure Description

[0020] Figure 1 This is a flowchart illustrating a domain adaptive target detection method based on dynamic data-driven and consistency constraints in one embodiment. Figure 2 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0021] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0022] In one embodiment, a domain-adaptive target detection method based on dynamic data-driven and consistency constraints is provided, such as... Figure 1 The diagram shown is a schematic of the overall process. It integrates the overall model architecture, the Semantic Prediction Consistency (SPC) module, the Difference-Guided Dynamic Data-Driven Refinement (DGCR) module, and the inference process. The descriptions of each part are as follows: Overall model architecture: A dual-stream design of teacher and student is adopted. The top left is the target domain data stream, where unlabeled target domain images are weakly enhanced and then enter the teacher network backbone to generate pseudo-labels and guidance features. The bottom left is the source domain data stream, where labeled source domain images are first processed by the DGCR module and then strongly enhanced before being input into the student network backbone. The middle SPC module receives the intermediate layer feature maps of the teacher and student network feature pyramids and the output of the detection head, calculates the difference signal, and the red dashed arrows indicate the closed-loop feedback path, which transmits the difference signal from the SPC module to the DGCR module for real-time adjustment of the processing intensity of the source domain data.

[0023] The SPC module performs instance normalization on the feature maps of the teacher and student networks, removing style statistics; it enhances student features through a cross-attention mechanism and calculates the semantic context consistency loss by calculating the similarity between student and teacher features. Simultaneously, using teacher network prediction as the anchor point, the difference between the classification probability distribution and the regression offset is calculated to obtain the prediction distribution consistency loss. The weighted sum of the two yields the total semantic prediction consistency loss. It is then output as a feedback signal.

[0024] DGCR module: Receives the consistency loss from the SPC module and maps it to a frequency cutoff threshold using an algorithm. and The source domain image is transformed to the frequency domain using DCT, and the frequency domain is divided into a static band (MS) and a perturbation band (Mp) using a threshold. Gaussian noise is injected into the perturbation band, and the image is restored to the processed source domain image by IDCT in combination with the static band features.

[0025] Parameter update and inference: Student network parameters are updated through backpropagation of the total loss function, and teacher network parameters are updated through the EMA of student network parameters; during the inference phase, only the trained teacher network is called to achieve detection without additional cost.

[0026] The Chinese explanations of the English and Chinese abbreviations in the attached diagram are as follows: RPN stands for Region Proposal Network; ROI stands for Region of Interest Detector; DCT / IDCT stands for Discrete Cosine Transform / Inverse Discrete Cosine Transform. For semantic context consistency loss; To predict the distribution consistency loss; For semantic prediction consistency loss; Losses due to unsupervised monitoring.

[0027] In one embodiment, firstly, a target detection model based on the Mean Teacher architecture is constructed. The student network and teacher network have completely identical structures, both using a Feature Pyramid Network (FPN) as the backbone network and including a Region Proposal Network (RPN) and a Region of Interest (ROI) detection head to adapt to multi-scale target detection requirements. The input data is divided into manually labeled source domain images (such as images of driving scenes in clear weather) and unlabeled target domain images (such as images of driving scenes in foggy weather).

[0028] In one embodiment, the semantic prediction consistency loss comprises two parts: semantic context consistency loss and prediction distribution consistency loss. This loss value is used both to optimize the network and as a real-time difference signal to quantify the degree of cognitive bias of the model in the target domain.

[0029] In one embodiment of this application, for the intermediate layer feature maps of the feature pyramid, an instance normalization operation is first introduced. This normalization is performed by calculating the mean and variance of the feature maps to remove style statistics (i.e., non-semantic context features) from the image. Specifically, for the feature layers of the feature pyramid... First, calculate the mean of the features in the spatial dimension. and standard deviation :

[0030]

[0031] in, This represents the value of the c-th channel at position (h, w). To prevent the use of tiny constants that divide by zero, the feature map is then normalized using the aforementioned statistics to obtain a destylated feature map.

[0032] Subsequently, an attention mechanism is used to calculate the global semantic similarity between student network features and teacher network features, forcing the student network to learn the more essential semantic features of the teacher network after removing style interference. Specifically, the style-de-stylized student features are mapped to queries. Teacher features are mapped as and .

[0033] Calculate attention-enhanced features :

[0034] in The projection dimension is used. To prevent feature degradation, residual connections are introduced to fuse the enhanced features with the original student features, resulting in the final student feature order. :

[0035] in It is a serialized representation of student network features. It is the residual scaling factor. Using global average pooling g( Extract feature vectors and calculate cosine similarity. Finally, the cosine similarity with edges is calculated as the semantic context consistency loss. :

[0036]

[0037] in, For teacher characteristic sequences, For preset edges, For hierarchical weights, It is the set of levels of the feature pyramid.

[0038] In one embodiment of this application, the prediction results output by the teacher network are used as pseudo-labels, and a general anchor box sampling strategy is adopted, allowing the student network to directly reuse the positive and negative sample anchor box indices and region of interest suggestions selected by the teacher network. To address the localization error problem, this paper calculates the divergence (KL divergence) between the student and teacher networks in the classification probability distribution, as well as the L1 loss of the foreground region regression offset. Specifically, it calculates the prediction distribution consistency loss. It covers two dimensions: classification and regression.

[0039] in Let KL divergence be the KL divergence. Foreground indicator function, and For teacher and student models in public The probability distribution calculated above, and This represents the predicted regression offset.

[0040] The weighted sum of the semantic context consistency loss and the prediction distribution consistency loss yields:

[0041] in, , These are weighting coefficients, which can be adjusted according to the actual testing scenario.

[0042] In one embodiment, semantic prediction consistency loss is used as a feedback signal to dynamically adjust the processing intensity of source domain data, thereby achieving dynamic frequency domain processing of the source domain image.

[0043] In one embodiment of this application, firstly, a dynamic modulation factor is calculated. Specifically, an exponential moving average smoothing process is performed on the semantic prediction consistency loss to obtain the historical mean of the loss at time t. :

[0044] in, This is the smoothing coefficient.

[0045] Calculate the deviation rate of the loss at time t relative to the historical mean. :

[0046] in, To prevent division by zero of small constants; The deviation rate is mapped to a dynamic modulation factor using the hyperbolic tangent function:

[0047] in, For amplitude factor, This is the sensitivity coefficient. If the consistency loss is large, the modulation factor increases; if the loss is small, the modulation factor decreases.

[0048] In one embodiment of this application, the initial cutoff threshold of the statically maintained frequency band is defined as... Utilizing dynamic modulation factors Update the real-time threshold at time t:

[0049] In one embodiment of this application, the source domain image is... The frequency domain is transformed using Discrete Cosine Transform (DCT); a statically preserved band mask is generated based on the updated threshold. and perturbation band mask Inject Gaussian noise into the disturbance frequency band The processed frequency domain features are then restored to a pixel-space image using Inverse Discrete Cosine Transform (IDCT), yielding the processed source domain image.

[0050] in, For DCT transformation, For IDCT transformation, ⊙ represents element-wise multiplication; when the modulation factor increases, the width of the static hold band is expanded to protect the model's basic learning ability; when the modulation factor decreases, the static hold band is compressed, the perturbation band is expanded and noise is injected to suppress the model's dependence on environmental background features.

[0051] In one embodiment, the source domain image processed by the DGCR module is input into the student network, and the supervised detection loss is calculated. The loss is the standard Faster R-CNN loss, which includes classification loss and regression loss. The total loss function is constructed by combining supervised detection loss and semantic prediction consistency loss. The model is iteratively optimized and trained based on the total loss function. The parameters of the student network are updated through backpropagation, while the parameters of the teacher network do not participate in gradient backpropagation. At the same time, the parameters of the teacher network are updated through exponential moving average (EMA) to ensure the stability of the teacher network output. The above steps are repeated until the model converges and the training is completed.

[0052] During the model inference stage, only the target domain image to be detected is input into the trained teacher network to achieve high-precision cross-domain target detection without loading additional modules, thus ensuring real-time inference.

[0053] It should be understood that although the steps in the flowcharts of the embodiments described above are shown sequentially according to the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts of the embodiments described above may include multiple steps or multiple stages. These steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the steps or stages of other steps.

[0054] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 2As shown, the computer device includes a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, mobile cellular networks, NFC (Near Field Communication), or other technologies. When executed by the processor, the computer program implements a domain-adaptive target detection method based on dynamic data-driven and consistency constraints. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0055] Those skilled in the art will understand that Figure 2 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0056] In one embodiment, a computer device is provided, including a memory and a processor, wherein the memory stores a computer program, and the processor executes the computer program to implement the steps in the above-described method embodiments.

[0057] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps in the above method embodiments.

[0058] In one embodiment, a computer program product is provided, including a computer program that, when executed by a processor, implements the steps in the above method embodiments.

[0059] It should be noted that the user information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are all information and data authorized by the user or fully authorized by all parties.

[0060] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium. When executed, the computer program can include the processes of the embodiments described above. Any references to memory, databases, or other media used in the embodiments provided in this application can include at least one of non-volatile and volatile memory. Non-volatile memory can include read-only memory (ROM), magnetic tape, floppy disk, flash memory, optical memory, high-density embedded non-volatile memory, resistive random access memory (ReRAM), magnetic random access memory (MRAM), ferroelectric random access memory (FRAM), phase change memory (PCM), graphene memory, etc. Volatile memory can include random access memory (RAM) or external cache memory, etc. By way of illustration and not limitation, RAM can take many forms, such as Static Random Access Memory (SRAM) or Dynamic Random Access Memory (DRAM). The databases involved in the embodiments provided in this application may include at least one type of relational database and non-relational database. Non-relational databases may include, but are not limited to, blockchain-based distributed databases. The processors involved in the embodiments provided in this application may be general-purpose processors, central processing units, graphics processing units, digital signal processors, programmable logic devices, quantum computing-based data processing logic devices, etc., and are not limited to these.

[0061] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0062] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this application should be determined by the appended claims.

Claims

1. A domain-adaptive target detection method based on dynamic data-driven and consistency constraints, characterized in that, The method includes: Construct a target detection model that includes a teacher network and a student network, and input labeled source domain images and unlabeled target domain images into the target detection model; Calculate the semantic prediction consistency loss, which includes the semantic context consistency loss and the prediction distribution consistency loss; The dynamic modulation factor is calculated based on the semantic prediction consistency loss. The cutoff frequency threshold of the frequency domain filter is adjusted using the dynamic modulation factor. The adjusted cutoff frequency threshold is then used to perform dynamic frequency domain processing on the source domain image. The source domain image after dynamic processing in the frequency domain is input into the student network, the supervised detection loss is calculated, and the total loss function is constructed by combining the supervised detection loss and the semantic prediction consistency loss. The target detection model is iteratively optimized and trained based on the total loss function, and the parameters of the student network and the teacher network are updated.

2. The domain adaptive target detection method based on dynamic data-driven and consistency constraints according to claim 1, characterized in that, Both the teacher network and the student network use a feature pyramid network as their backbone network and include a region proposal network and a region of interest detection head.

3. The domain adaptive target detection method based on dynamic data-driven and consistency constraints according to claim 1, characterized in that, The calculation includes semantic context consistency loss and prediction distribution consistency loss. The semantic prediction consistency loss includes: Perform instance normalization on the intermediate layer feature maps of the feature pyramid to remove style statistics from the image; The global semantic similarity of features between the student network and the teacher network after instance normalization is calculated. The semantic context consistency loss is obtained by calculating the cosine similarity, combined with the preset edge and hierarchical weights.

4. The domain adaptive target detection method based on dynamic data-driven and consistency constraints according to claim 1, characterized in that, The calculation includes semantic context consistency loss and prediction distribution consistency loss. The semantic prediction consistency loss also includes: Using the output of the teacher network as pseudo-labels, a general anchor box sampling strategy is adopted to reuse the positive and negative sample anchor box indexes and region of interest suggestions screened by the teacher network. Based on the anchor box index and region of interest suggestions, the KL divergence of the classification probability distributions of the student network and the teacher network, the L1 loss of the regression offset, and the summation of these results in the prediction distribution consistency loss.

5. The domain adaptive target detection method based on dynamic data-driven and consistency constraints according to claim 1, characterized in that, The calculation of the dynamic modulation factor based on the semantic prediction consistency loss includes: The semantic prediction consistency loss is smoothed by exponential moving average, and the deviation rate of the semantic prediction consistency loss relative to the historical mean is calculated. The deviation rate is then mapped to a dynamic modulation factor using a hyperbolic tangent function.

6. The domain adaptive target detection method based on dynamic data-driven and consistency constraints according to claim 1, characterized in that, The cutoff frequency threshold of the frequency domain filter is the upper and lower thresholds of the statically held frequency band. The updated real-time cutoff frequency threshold is obtained by exponentially calculating the initial upper and lower thresholds of the statically held frequency band using the dynamic modulation factor.

7. The domain adaptive target detection method based on dynamic data-driven and consistency constraints according to claim 1, characterized in that, The step of performing frequency domain dynamic processing on the source domain image using the adjusted cutoff frequency threshold includes: The source domain image is transformed to the frequency domain space through discrete cosine transform, and the frequency domain space is divided into statically maintained frequency bands and perturbation frequency bands using the updated cutoff frequency threshold. Based on the cutoff frequency threshold, a static hold-up band mask and a perturbation band mask are generated. Gaussian noise is injected into the perturbation band, and the source domain image is restored by combining the features of the two bands through inverse discrete cosine transform to obtain the processed source domain image.

8. The domain adaptive target detection method based on dynamic data-driven and consistency constraints according to claim 1, characterized in that, During training, the parameters of the student network are updated based on the total loss function through backpropagation, while the parameters of the teacher network are updated through the exponential moving average of the student network parameters and do not participate in gradient backpropagation.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1 to 8.

10. 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 steps of the method according to any one of claims 1 to 8.

Images