Ice detection model training method, device, electronic device and storage medium

Abstract

The present invention discloses an ice detection model training method, device, electronic device and storage medium, including: inputting a one-dimensional sequence of image samples of a transmission line into a backbone layer for processing to fuse the features of the one-dimensional sequence at multiple different data partitioning scales to obtain a target feature sequence; inputting the target feature sequence into an ice recognition module to obtain a second ice region and a second ice type; training the backbone layer and the ice recognition module with the known first ice region and the first ice type, and the predicted second ice region and the second ice type as supervision; when the training of the backbone layer and the ice recognition module is completed, locking the parameters in the backbone layer and the ice recognition module to generate an ice detection model according to the pixel partitioning module, the backbone layer and the ice recognition module. Thereby improving the accuracy of ice detection by the ice detection model. Different data partitioning scales can enhance the correlation between each subsequence and pay attention to local features that are easily overlooked.

Description

Technical Field

[0001] The present invention relates to the technical field of power grid line maintenance, and in particular to an icing detection model training method, device, electronic equipment and storage medium. Background Art

[0002] When ice covers the transmission lines, it may cause power transmission safety problems. In order to ensure the stable operation and reliable power supply of the power system, it is necessary to detect ice on the transmission lines. Microwaves have penetrating power and can penetrate ice, rain, snow and other media better than other waves. Therefore, it is more appropriate and reliable to use microwave sensing technology in ice detection. In addition, the use of automated algorithms such as deep learning to automatically screen and detect ice from the data extracted by microwave sensing technology can significantly reduce the time consumption and human resources of the process.

[0003] Alexey Dosovitskiy et al. proposed the Vision Transformer model. The Vision Transformer proposed that attention does not need to rely on convolutional neural networks. Therefore, the pure Transformer applied to the line image patch sequence in the ice-covered scene can well complete the computer vision task. The encoder of the Vision Transformer can compress the transmission line imaging data collected by microwave detection into a one-dimensional representation sequence, which is mapped into another continuous sequence information in the model, that is, the output information of the decoder, to predict whether the transmission line is covered with ice. The disadvantage of this method is that the correlation between sequences is low when performing self-attention calculations, and local features cannot be effectively focused on. Summary of the invention

[0004] The present invention provides an icing detection model training method to solve the problem of icing detection model training.

[0005] In a first aspect, the present invention provides an ice detection model training method, comprising:

[0006] Acquire an image sample of the power transmission line, and acquire a first ice-covered area and a first ice-covered type in the image sample;

[0007] Inputting the image samples into a pixel partitioning module for pixel partitioning to obtain a one-dimensional sequence of the image samples;

[0008] Inputting the one-dimensional sequence into the backbone layer for processing to fuse the features of the one-dimensional sequence at multiple different data division scales to obtain a target feature sequence;

[0009] Inputting the target feature sequence into an ice covering recognition module to obtain a second ice covering area and a second ice covering type;

[0010] Training the backbone layer and the ice recognition module by aligning the first ice-covered area and the first ice-covered type, and the second ice-covered area and the second ice-covered type as supervision;

[0011] When the training of the backbone layer and the ice recognition module is completed, the parameters in the backbone layer and the ice recognition module are locked to generate an ice detection model according to the pixel partitioning module, the backbone layer and the ice recognition module.

[0012] In a second aspect, the present invention provides an ice detection model training device, comprising:

[0013] A sample acquisition module, used to acquire an image sample of the transmission line, and acquire a first ice-covered area and a first ice-covered type in the image sample;

[0014] A pixel partitioning module, used for inputting the image samples into the pixel partitioning module for pixel partitioning to obtain a one-dimensional sequence of the image samples;

[0015] A feature extraction and fusion module is used to input the one-dimensional sequence into the backbone layer for processing, so as to fuse the features of the one-dimensional sequence at multiple different data division scales to obtain a target feature sequence;

[0016] An ice-covering recognition module, used for inputting the target feature sequence into the ice-covering recognition module to obtain a second ice-covering area and a second ice-covering type;

[0017] A training module, configured to train the backbone layer and the ice recognition module by aligning the first ice-covered area and the first ice-covered type, and the second ice-covered area and the second ice-covered type as supervision;

[0018] A model determination module is used to lock the parameters in the backbone layer and the ice recognition module when the training of the backbone layer and the ice recognition module is completed, so as to generate an ice detection model according to the pixel partitioning module, the backbone layer and the ice recognition module.

[0019] In a third aspect, the present invention provides an electronic device, the electronic device comprising:

[0020] at least one processor; and

[0021] a memory communicatively connected to the at least one processor; wherein,

[0022] The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the ice detection model training method described in the first aspect of the present invention.

[0023] In a fourth aspect, the present invention provides a computer-readable storage medium storing computer instructions, wherein the computer instructions are used to enable a processor to implement the ice detection model training method described in the first aspect of the present invention when executed.

[0024] An embodiment of the present invention provides an ice detection model training method, which inputs a one-dimensional sequence of image samples of a transmission line into a backbone layer for processing to fuse the features of the one-dimensional sequence at multiple different data partitioning scales to obtain a target feature sequence; the target feature sequence is input into an ice recognition module to obtain a second ice area and a second ice type; the backbone layer and the ice recognition module are trained with the known first ice area and the first ice type, and the predicted second ice area and the second ice type as supervision; when the training of the backbone layer and the ice recognition module is completed, the parameters in the backbone layer and the ice recognition module are locked to generate an ice detection model according to the pixel partitioning module, the backbone layer and the ice recognition module.

[0025] First, the one-dimensional sequence of image samples is divided at multiple different data division scales to obtain different image features, and the ice-covered area and ice-covered type are identified based on the image features. This allows the ice-covered detection model to learn to reasonably divide the one-dimensional sequence data and extract sequence features with more ice-covered features from the divided sequence data, thereby improving the accuracy of the ice-covered detection model in ice-covered detection.

[0026] Second, through multiple data division scales, the image division method is enriched, making the model's feature extraction of one-dimensional sequences more refined, avoiding the use of a fixed division method that causes the original sequence with ice-covered features to be segmented, making it difficult to detect the ice-covered features of the image. Therefore, the multiple data division scales of this scheme can obtain a relatively complete sequence with ice-covered features, enhance the correlation between each subsequence, and then pay attention to local features that are easy to ignore.

[0027] It should be understood that the contents described in this section are not intended to identify the key or important features of the embodiments of the present invention, nor are they intended to limit the scope of the present invention. Other features of the present invention will become easily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0028] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following briefly introduces the drawings required for use in the description of the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without creative work.

[0029] Figure 1is a flow chart of an ice detection model training method provided in Embodiment 1 of the present invention;

[0030] Figure 2 It is a structural schematic diagram of an ice detection model provided in the first embodiment of the present invention;

[0031] Figure 3 It is an image obtained after 3*3 data partitioning scale division provided by the first embodiment of the present invention;

[0032] Figure 4 is a flow chart of an ice detection model training method provided by Embodiment 2 of the present invention;

[0033] Figure 5 is a structural schematic diagram of an ice detection model provided in Embodiment 2 of the present invention;

[0034] Figure 6 It is a structural schematic diagram of an ice detection model training device provided in Embodiment 3 of the present invention;

[0035] Figure 7 It is a schematic diagram of the structure of an electronic device provided in Embodiment 4 of the present invention. DETAILED DESCRIPTION

[0036] In order to enable those skilled in the art to better understand the scheme of the present invention, the technical scheme in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work should fall within the scope of protection of the present invention.

[0037] Embodiment 1

[0038] Figure 1 This is a flow chart of an ice detection model training method provided in the first embodiment of the present invention. This embodiment is applicable to the case of ice detection model training. The method can be executed by an ice detection model training device. The ice detection model training device can be implemented in the form of hardware and / or software. The ice detection model training device can be configured in an electronic device. Figure 1 As shown, the ice detection model training method includes:

[0039] S101. Acquire an image sample of a power transmission line, and acquire a first ice-covered area and a first ice-covered type in the image sample.

[0040] Among them, the image samples of the transmission line are obtained by microwave sensing technology, and the image samples are usually three-dimensional images. Microwaves are a type of electromagnetic wave, which is in the high frequency band of the electromagnetic spectrum. Its wavelength is usually between radio frequency and infrared, and the frequency range ranges from hundreds of megahertz (MHz) to hundreds of gigahertz (GHz). The microwave sensor is connected to the substation or transmission tower in the low temperature area. The transmitting antenna of the microwave sensor is loaded with the transmission microwave detection signal to the transmission line. The microwave detection signal will be reflected and scattered after reaching the interface. The collection module can receive the reflected detection data through each antenna, collect the transmission line images corresponding to each microwave sensor of the transmission line to construct the original data set, and the image sample is the picture in the original data set.

[0041] Label each image sample, that is, label the first ice-covered area and the first ice-covered type of the image sample. Labeling can be done manually. After labeling is completed, the first ice-covered area and the first ice-covered type in the image sample can be obtained. Specifically, after collecting the data set, each data is labeled in turn, and the coordinates and categories are labeled in turn for subsequent model training and testing. The label corresponding to each transmission line image uses the VOC labeling format, and its content is defined as follows:

[0042]

[0043] Example: Assume that there is an ice phenomenon at a certain point in the transmission line image. The coordinates of the upper left corner and lower right corner of the ice are two points (rectangular box), P1 = (x1, y1) and P2 = (x2, y2). Then the content of the annotation file corresponding to the image is:

[0044]

[0045] The annotation file is stored in XML format, and the annotation format here corresponds to the coordinates and classification information of each target box detected by the model. <annotation>The large tag for the entire annotation file indicates that the XML file is a annotation file, and all objects in an image are within the scope of the tag; <filename>It is the relative path of the image corresponding to the tag, and corresponds one-to-one with the image; <size>The width of the label is passed inside the label <width>,high <height>Capture the corresponding image, and only images within this size will be detected; <object>The label corresponds to an object in the image. <name>is the category to which each target belongs. The example icing in the figure means that the target category is ice-covered; <bndbox>The label describes the coordinate frame of each target, where xmin, ymin, xmax, and ymax are the minimum coordinate of the target's horizontal axis, the minimum coordinate of the target's vertical axis, the maximum coordinate of the target's horizontal axis, and the maximum coordinate of the target's vertical axis. Each image will correspond to a labeling text file. At the same time, each image may have multiple ice-covered areas, so each labeling text file will correspond to multiple target labeling information.

[0046] In an optional embodiment, after obtaining the image samples of the power transmission line, the method further includes: performing contrast normalization processing on the image samples.

[0047] The scene of ice detection on transmission lines is usually at a low temperature (usually below zero), and ice and snow are prone to appear in the environment where the transmission lines are located. When the transmission lines are covered with ice and snow, the color channels of the ice and snow in the environment and the transmission lines are likely to be similar in the collected image samples, which will increase the noise and difficulty of ice detection. Therefore, the image samples can be contrast normalized to improve the contrast of the image samples and increase the difference in color channels between the transmission line area covered with ice and snow in the picture and the ice and snow area in the environment, so as to facilitate ice detection.

[0048] Specifically, for an image sample of a transmission line, the corresponding image matrix is ​​I, and the normalized image matrix is ​​I norm , the specific steps are as follows:

[0049] Assuming that the number of rows and columns of the image are M and N respectively, calculate the average brightness of the image matrix

[0050] Among them, M and N are the number of rows and columns of the image respectively, and i and j represent different rows and columns.

[0051] Get the average brightness of the image matrix Then, calculate the standard deviation σ of the image I ;

[0052]

[0053] Finally, the original matrix I and standard deviation σ of the image are I and average brightness Calculate and get the contrast normalized image matrix I norm :

[0054]

[0055] The above formula is applied to each pixel in the image sample, subtracting the mean brightness and then dividing by the standard deviation so that each pixel value has the same range across the entire image sample.

[0056] S102: Input the image samples into a pixel partitioning module for pixel partitioning to obtain a one-dimensional sequence of image samples.

[0057] First, we can build an initial ice detection model. Figure 2 is a structural diagram of an ice detection model. Figure 2 As shown, the ice detection model 100 includes a pixel partitioning module 1, a backbone layer 2 and an ice recognition module 3.

[0058] The pixel partitioning module is used to process the image samples of the initial input model and transform the image samples of the transmission line into a one-dimensional sequence suitable for model processing.

[0059] Assume that the input image of an ice-covered power transmission line has a height of H, a width of W, and three RGB channels. According to the defined partition block structure size patch_size*patch_size, the image can be cut into several partitions with pixels as the minimum unit. Here, it is assumed that the partition block structure is 4*4 pixels, with a total of 16 pixels. The segmented image is channel-adjusted and a linear embedding vector layer is used to map the original value feature to an arbitrary dimension.

[0060] Specifically, for an RGB image of size H×W, it is assumed here that a two-dimensional convolutional layer is used, and the step size and convolution kernel size are both the block structure size of the partition patch_size, and the feature dimension of each partition is:

[0061] P×P×C=4×4×3=48;

[0062] Among them, P is the patch_size;

[0063] The total number of partitions N is:

[0064]

[0065] Therefore, after the RGB image is patch segmented, the image dimension size becomes (H / 4)×(W / 4)×48, that is, (H / 4)×(W / 4)×48 pixels are obtained. Finally, the linear embedding vector layer is used to convolve the channel dimension of each pixel to map the original transmission line image into a one-dimensional sequence.

[0066] S103, inputting the one-dimensional sequence into the backbone layer for processing to fuse the features of the one-dimensional sequence at multiple different data division scales to obtain a target feature sequence.

[0067] The backbone layer can be improved by any of the convolutional neural networks, residual neural networks, recurrent neural networks and deep neural networks.

[0068] The data division scale here may refer to the division method and division ratio of image segmentation. For example, the division method may be 3*3, 4*4 or 3*4, etc., and the division ratio may include a horizontal division ratio and a vertical division ratio. For example, when the division method is 3*3, the horizontal division ratio is 1:2:1, and the vertical division ratio is 1:1:1.

[0069] The backbone layer is equipped with multiple interconnected modules, and different modules can divide the one-dimensional sequence into different data partitioning scales. For example, 3 is an image obtained after 3*3 data partitioning scale division. Figure 3 As shown, the sizes of the 9 regions obtained are not the same. Of course, since the data division scales are multiple and different, it is conceivable that the one-dimensional sequence of image samples can also be divided into multiple regions of the same size.

[0070] By obtaining the detection frames under various division methods, the image features under different data division methods can be obtained and integrated. Through various data division scales, the image division methods are enriched, making the model more refined in extracting the features of the one-dimensional sequence, avoiding the use of a single division method, which causes the sequence with ice-covered features to be segmented and difficult to detect the ice-covered features of the image. Therefore, the method of multiple data division scales in this scheme can obtain a relatively complete sequence with ice-covered features, enhance the correlation between sequences, and then pay attention to local features that are easy to ignore.

[0071] S104: Input the target feature sequence into an ice-covering recognition module to obtain a second ice-covering area and a second ice-covering type.

[0072] The ice recognition module is the output module of the ice detection model. Its function is to locate and classify ice, which is used to achieve the final model monitoring. Specifically, the regression algorithm in the deep learning algorithm is a supervised learning algorithm used to predict continuous numerical outputs. It can predict one or more target variables based on input features. The ice recognition module uses the regression algorithm to generate the position information of the prediction box to achieve the fitting of the real box.

[0073] The ice recognition module can include a regression branch and a classification branch. The regression branch is used to predict the position of each detection box, including the coordinate information of the upper left and lower right points of the bounding box. The classification branch is used to predict the ice category (name) of the target in each detection box. Icing means that the target in the detection box is covered with ice, and none means that there is no ice or snow.

[0074] S105, training a backbone layer and an ice recognition module by aligning the first ice-covered area and the first ice-covered type, and the second ice-covered area and the second ice-covered type as supervision.

[0075] Aligning the ice-covered area and the ice-covered type may mean that the second ice-covered area is constantly approaching the first ice-covered area and the second ice-covered type is constantly approaching the second ice-covered type, that is, the ice-covered features output by the ice-covered recognition module are constantly approaching the ice-covered features preset in the image sample.

[0076] It should be noted that for the icing recognition module, its prediction function is affected by the input target feature sequence and its own parameters, and the target feature sequence is affected by the parameters of the backbone layer. Therefore, to obtain a better prediction effect, it is necessary to adjust the parameters of the backbone layer and the icing recognition module, rather than just adjusting the parameters of the icing recognition module.

[0077] Exemplarily, the training process of the backbone layer and the ice recognition module is as follows:

[0078] Input the one-dimensional sequence into the backbone layer for processing to fuse the features of the one-dimensional sequence at multiple different data division scales to obtain the target feature sequence, input the target feature sequence into the ice recognition module to obtain the second ice area and the second ice type; calculate the overlap degree of the first ice area and the second ice area, and calculate the similarity between the first ice type and the second ice type; determine whether the overlap degree is greater than the preset overlap degree threshold and whether the similarity is greater than the preset similarity threshold; if so, it means that the parameter adjustment of the backbone layer and the ice recognition module is completed, that is, the training of the backbone layer and the ice recognition module is completed, if not, calculate the loss value based on the overlap degree and the similarity, and adjust the parameters of the backbone layer and the ice recognition module based on the loss value, and return to execute the step of inputting the one-dimensional sequence into the backbone layer for processing to fuse the features of the one-dimensional sequence at multiple different data division scales to obtain the target feature sequence. In addition, a validation data set can be set to verify the prediction accuracy of the model, repeat this process, and continuously update the model parameter weights until the number of training iterations is reached, the model training is completed, and the trained transmission line ice detection model is obtained.

[0079] It should be noted that, in the process of training the backbone layer and the ice cover recognition module, the parameters of one part of the backbone layer and the ice cover recognition module can be adjusted in turn, or the parameters of both parts can be adjusted at the same time, or the parameters of both parts can be adjusted at the same time, and then the parameters of each part can be adjusted in turn. The present invention does not limit the training order and method of the backbone layer and the ice cover recognition module.

[0080] It can be understood that under the above supervision constraints, the ice recognition module in the ice detection model can further improve the expressiveness of ice features, enhance the individual differences of images, and improve the ability to identify ice on image samples.

[0081] S106: When the training of the backbone layer and the ice recognition module is completed, the parameters in the backbone layer and the ice recognition module are locked to generate an ice detection model according to the pixel partitioning module, the backbone layer and the ice recognition module.

[0082] When the training of the backbone layer and the ice recognition module is completed, the parameters in the backbone layer and the ice recognition module are locked, and then the pixel partitioning module, the backbone layer and the ice recognition module are spliced ​​to obtain the ice detection model.

[0083] The embodiment of the present invention provides an ice detection model training method, which inputs a one-dimensional sequence of image samples of a transmission line into a backbone layer for processing to fuse the features of the one-dimensional sequence at multiple different data division scales to obtain a target feature sequence; inputs the target feature sequence into an ice recognition module to obtain a second ice region and a second ice type; trains the backbone layer and the ice recognition module with the known first ice region and the first ice type, and the predicted second ice region and the second ice type as supervision; when the training of the backbone layer and the ice recognition module is completed, the parameters in the backbone layer and the ice recognition module are locked to generate an ice detection model according to the pixel partition module, the backbone layer and the ice recognition module. First, the one-dimensional sequence of image samples is divided at multiple different data division scales to obtain different image features, and the ice region and ice type are identified based on the image features, so that the ice detection model can learn to reasonably divide the one-dimensional sequence data and extract sequence features with more ice features from the divided sequence data, thereby improving the accuracy of the ice detection model for ice detection. First, the one-dimensional sequence of image samples is divided at multiple different data division scales to obtain different image features, and the ice-covered area and ice-covered type are identified based on the image features. This allows the ice-covered detection model to learn to reasonably divide the one-dimensional sequence data and extract sequence features with more ice-covered features from the divided sequence data, thereby improving the accuracy of ice-covered detection by the ice-covered detection model. Second, through multiple data division scales, the image division method is enriched, making the model's feature extraction of the one-dimensional sequence more refined, avoiding the use of a fixed division method that causes the original sequence with ice-covered features to be segmented, and making it difficult to detect the ice-covered features of the image. Therefore, the multiple data division scales of this scheme can obtain a relatively complete sequence with ice-covered features, enhance the correlation between each subsequence, and then be able to focus on local features that are easy to ignore.

[0084] Embodiment 2

[0085] Figure 4 This is a flow chart of an ice detection model training method provided in the second embodiment of the present invention. The embodiment of the present invention is optimized on the basis of the above-mentioned first embodiment. Figure 4 As shown, the ice detection model training method includes:

[0086] S401: Acquire an image sample of a power transmission line, and acquire a first ice-covered area and a first ice-covered type in the image sample.

[0087] S402: Input the image samples into a pixel partitioning module for pixel partitioning to obtain a one-dimensional sequence of image samples.

[0088] S401-S402 are similar to S101-S102 in the first embodiment, and the details can be referred to the relevant description.

[0089] S403, inputting the one-dimensional sequence into a backbone layer, wherein the backbone layer includes a plurality of first division modules, a second division module, a feature enhancement module and a fusion module.

[0090] like Figure 2 As shown, the ice detection model 100 includes a pixel partitioning module 1, a backbone layer 2 and an ice recognition module 3.

[0091] Figure 5 This is a schematic diagram of the structure of an ice detection model provided in this embodiment. Figure 5 As shown, the backbone layer 2 includes multiple first division modules 21, second division modules 22, feature enhancement modules 23 and fusion modules 24, wherein the first division modules 21 and the second division modules 22 are respectively connected to the feature enhancement module 23, and the feature enhancement module 23 is connected to the fusion module 24.

[0092] It should be noted that Figure 5 The connection mode and placement order of each module in the backbone layer 2 are examples, and the setting order and setting quantity of the first division module 21 and the second division module 22 can be set based on actual needs. In other words, the second division module 22 can also be used as the first module in the backbone layer 2, and the number of the first division module 21 and the second division module 22 can also be unequal.

[0093] Generally speaking, in the backbone layer, the number of the first division modules and the second division modules is 7-10. Of course, the number of each division module can also be set according to actual needs, and the present invention is not limited to this.

[0094] S404: Use the one-dimensional sequence as an initial input sequence.

[0095] That is, the initial input sequence of the entire backbone layer, that is, the input sequence of the first partitioning module (not necessarily the first partitioning module) in the backbone layer. In the present invention, except for the pixel partitioning module at the head and the ice recognition module at the end, other modules are connected to each other in a head-to-tail manner. Therefore, the output data of the previous module is the input sequence of the next module. For the convenience of description, the data received by each module at its input interface is called the input sequence.

[0096] S405. In the first division module, the input sequence is randomly divided into a plurality of subsequences as the input sequence of the next module.

[0097] The data division scale may refer to the division method and division ratio of image segmentation. The data division scale in this step is random, that is, the division method and the division ratio are both random. It should be noted that even if the division method is determined to be the same, there may be different division ratios. Therefore, the size of the subsequence obtained by the one-dimensional sequence of image samples in the random division method is also random, and generally speaking, the size of the subsequence obtained is different.

[0098] S406: In the second division module, the input sequence is divided into a plurality of subsequences of the same size as the input sequence of the next module.

[0099] It should be noted that the data division scale in this step is not fixed, but the size of each subsequence after division is the same, that is, the values ​​in the division ratio are all 1. For example, when the division method is 3*3, the horizontal division ratio and the vertical division ratio are both 1:1:1, and when the division method is 3*4, the horizontal division ratio is 1:1:1, and the vertical division ratio is 1:1:1:1.

[0100] If a single data partitioning method is used, for example, the input sequence is fixedly divided into the same n subsequences, then there may be a situation where the ice cover feature is divided into different subsequences and cannot be detected, especially when the ice cover feature corresponds only to a small local area. If the local area is divided into different subsequences, it will be more difficult to detect, making the local ice cover feature easy to be ignored, which may cause the model to fail to identify the local ice cover feature and cause incorrect prediction results.

[0101] It can be known that, in the present embodiment, in the first division module and the second division module, subsequences of various division methods and division ratios can be obtained, and the ice cover features covered by the subsequences of different division methods and division ratios are different. Through a variety of data division scales, the image division method is enriched, and it is avoided that the sequence originally having ice cover features is segmented due to the use of a fixed division method, and it is difficult to detect the ice cover features of the image. Therefore, the method of multiple data division scales in this scheme can obtain a relatively complete sequence with ice cover features, enhance the correlation between each subsequence, and then pay attention to local features that are easy to ignore.

[0102] S407: In the feature enhancement module, a subsequence with ice-covering features in the input sequence is identified, and features of the subsequence with ice-covering features are enhanced, and the processed input sequence is used as an input sequence of the next module.

[0103] In the feature enhancement module, the subsequences with ice-covering features in the input sequence are identified, and the subsequences with ice-covering features are feature enhanced, and the processed input sequence is used as the input sequence of the next module, including: determining the input sequence of the feature enhancement module to obtain the target input sequence; identifying the subsequences with ice-covering features in the target input sequence and marking them as target subsequences. For each subsequence in the target input sequence, feature enhancement is performed based on the attention mechanism and the attention weight is updated to obtain a feature enhancement sequence. In the target input sequence, the attention weight of the target subsequence is higher than that of other subsequences; and the processed target input sequence is used as the input sequence of the next module.

[0104] The corresponding attention weight Attention is calculated within the region where each subsequence is located:

[0105]

[0106] Where Q is the query matrix, K is the key matrix, V is the value matrix, and d k is the dimension of the query or key, usually equal to the number of columns of the query or key matrix, Is the calculation weight. If there is ice-covered feature on a part of the transmission line, the calculation weight value of the subsequence corresponding to this part will be correspondingly larger, so that the model can learn to pay more attention to the ice-covered area, thereby improving the detection accuracy of the model.

[0107] When the feature enhancement module obtains the input sequence, a higher degree of attention is paid to the subsequence of the ice-covered area of ​​the transmission line in the sequence. At the same time, since only the subsequence of the cut part of the image (ice-covered area) is paid attention to instead of the entire image, the calculation parameters of the model are greatly reduced, and the speed and efficiency of the transmission line ice detection process are improved.

[0108] Optionally, before identifying a subsequence with ice-covering characteristics in the target input sequence and marking it as a target subsequence, the method further includes: for each subsequence in the target input sequence, standardizing the subsequence to obtain a standardized subsequence; and scaling and translating the standardized subsequence to obtain an updated target input sequence.

[0109] Assume that is the input sequence, z1,z2,...,z n For the subsequences, the designed expressions for normalization and offset are as follows:

[0110] For each input sequence Calculate the standardized Layer Normalization (LN) sequence:

[0111]

[0112] Among them, ε is a small constant used to ensure that the expression of the standardized sequence will not have a divisor of 0 when σ=0, μ is the mean, and σ 2 After the sequence is standardized, in order to increase the expressive power of the model, make the model more flexible to adapt to different input data, and avoid certain features from having too much impact on model training, it is necessary to standardize the sequence again. Scaling and translation are performed to obtain the sequence y i :

[0113]

[0114] Among them, γ and β are learnable parameters, which are used to scale and translate the normalized results respectively.

[0115] In order to obtain the sequence y i After that, the sequence y can be i Perform feature enhancement attention.

[0116] In an optional embodiment, before using the processed target input sequence as the input sequence of the next module, it also includes: for each subsequence in the target input sequence, performing residual superposition on the subsequence and the corresponding feature enhancement sequence to obtain an output sequence of each subsequence.

[0117] In order to make it easier for the network to learn the identity mapping and avoid the gradient vanishing problem, the original input sequence The residual superposition is performed on the sequence after feature enhancement to obtain the sequence

[0118] The overall expression of the input sequence processing is:

[0119]

[0120] The role of the LFE function is to enhance the features of the sequence and calculate the attention according to different weights in each subsequence. The role of the LN function is standardization.

[0121] If the scaling and translation steps are taken into account, the overall expression for the input sequence processing is:

[0122]

[0123] For the input sequence Normalized sequence, sequence y i for Zoom and translate the resulting sequence.

[0124] S408: In the fusion module, the input sequence is fused to obtain a fused feature sequence as the input sequence of the next module.

[0125] Fusing the input sequences means concatenating the features obtained after each feature enhancement operation.

[0126] In the process of interactive feature extraction, although a feature enhancement module has been used to reduce the number of parameters of the model of the present invention, the computational complexity of the model is still large. Therefore, the present invention designs a fusion module, which is used to embed the tail of the feature enhancement module to reduce the number of parameters of the model of the present invention and improve the efficiency of the model.

[0127] First, assume that the input image feature sequence is X, and its size is H×W×C, where H is the height, W is the width, the values ​​of H and W are both 4, and C is the number of channels. Since it is a sequence input, its value is 1. ij Each block of the input sequence X is 2×2 in size. The same positions of each block are spliced ​​to obtain 4 side-by-side features, and the regional fusion operation is performed on the feature:

[0128] Y ij =PartitionFusion(X ij );

[0129] Among them, Y ij Represents the feature map after PartitionFusion, PartitionFusion(X ij ) means Concat(X ij ), that is, the splicing operation is performed in the depth dimension.

[0130] For the obtained Y ij Perform LN operation to obtain the standardized sequence Z:

[0131] Z = LayerNorm(Y);

[0132] Perform a fully connected linear transformation operation on the standardized feature sequence Z to obtain

[0133] Finally, the height and width of the image feature sequence X after passing through the fusion module will be reduced by half, and the depth will also become half of the original. Therefore, the down sampling of the features is achieved in the fusion module, which improves the detection efficiency of the model.

[0134] S409: Use the fused feature sequence output by the last fusion module as the target feature sequence.

[0135] The fused feature sequence output by the last fusion module is the output data of the entire backbone layer, which is also the input data of the ice cover recognition module.

[0136] S410: Input the target feature sequence into an ice-covering recognition module to obtain a second ice-covering area and a second ice-covering type.

[0137] S411, training a backbone layer and an ice recognition module by aligning the first ice-covered area and the first ice-covered type, and the second ice-covered area and the second ice-covered type as supervision.

[0138] S410-S411 are similar to S105-S106 in the first embodiment, and the details may refer to the relevant description in the first embodiment.

[0139] S412: When the training of the backbone layer and the ice recognition module is completed, the parameters in the backbone layer and the ice recognition module are locked to generate an ice detection model according to the pixel partitioning module, the backbone layer and the ice recognition module.

[0140] Optionally, after the ice detection model is trained, a picture of the transmission line to be tested may be obtained, and the picture to be tested is input into the ice detection model to obtain the ice area and ice type.

[0141] After the transmission line icing model is trained, the model is used to process the image data extracted by microwave sensing technology to complete the icing detection on the transmission line. In practical applications, microwave sensors embedded in transmission line towers or substations collect transmission line data containing icing conditions after microwave data imaging. Through the preprocessing of transmission line data and the icing detection model, the icing conditions of the transmission line are output to facilitate subsequent further processing.

[0142] In this embodiment, on the one hand, in the first division module and the second division module, subsequences of various division methods and division ratios can be obtained. The ice features covered by the subsequences of different division methods and division ratios are different. Through multiple data division scales, the image division method is enriched, avoiding the use of a fixed division method to cause the sequence originally having ice features to be segmented, and it is difficult to detect the ice features of the image. Therefore, the method of multiple data division scales in this scheme can obtain a relatively complete sequence with ice features, enhance the correlation between each subsequence, and then pay attention to the local features that are easy to ignore. On the other hand, when the feature enhancement module obtains the input sequence, a higher degree of attention is paid to the subsequence of the ice-covered area of ​​the transmission line in the sequence. At the same time, since only the subsequence of the cut part of the image (ice-covered area) is paid attention to instead of the entire image, the calculation parameter amount of the model is greatly reduced, and the speed and efficiency of the transmission line ice detection process are improved.

[0143] In order to explain the training process of the ice detection model in the present invention simply and clearly, the following example is used for illustration: This example uses a data set of 2,000 transmission lines with ice conditions scanned by microwave sensors on transmission line towers to train the transmission line ice detection model.

[0144] Step 1: 2000 transmission line data with ice coverage are preprocessed for contrast normalization, with the slope alpha set to 1.5 and the offset beta adjusted to 30. The higher the alpha and beta, the greater the contrast enhancement.

[0145] Step 2: Divide the transmission line ice data set into 1200 images out of 2000 as training sets, 400 as validation sets, and 400 as test sets; set the number of training times for data preprocessing T = 200, the number of pixel partitions b = 9, the number of layers of the feature partition module (the first partition module and 10 second partition modules) n = 10, and the weight of the adaptive feature fusion module α = 0.4.

[0146] Step 3: Build the model

[0147] 1) Input the three-channel RGB data of the transmission line with ice coverage into the pixel partitioning module and map it into one-dimensional sequence data through the pixel partitioning module;

[0148] 2) The one-dimensional sequence format data of the transmission line with ice coverage is input into the backbone layer, and finally the target feature sequence is obtained.

[0149] 3) Input the target features into the icing recognition module, record the coordinates of each icing area, and classify the medium on the transmission line in the area into icing (indicating that there is icing on the transmission line) and 1 (indicating that there is no icing on the transmission line);

[0150] Step 4: Train the transmission line ice detection model and adjust the parameters in the model according to the loss value in each iteration.

[0151] After adjusting relevant parameters, the accuracy of the transmission line icing detection model finally trained tends to be stable, mainly fluctuating between 97.0% and 97.5%.

[0152] Embodiment 3

[0153] Figure 6 This is a schematic diagram of the structure of an ice detection model training device provided in Embodiment 3 of the present invention. Figure 6 As shown, the ice detection model training device includes:

[0154] A sample acquisition module 601 is used to acquire an image sample of a power transmission line, and acquire a first ice-covered area and a first ice-covered type in the image sample;

[0155] A pixel partitioning module 602, configured to input the image samples into a pixel partitioning module for pixel partitioning to obtain a one-dimensional sequence of the image samples;

[0156] The feature extraction and fusion module 603 is used to input the one-dimensional sequence into the backbone layer for processing, so as to fuse the features of the one-dimensional sequence at multiple different data division scales to obtain a target feature sequence;

[0157] An ice-covering recognition module 604 is used to input the target feature sequence into an ice-covering recognition module to obtain a second ice-covering area and a second ice-covering type;

[0158] A training module 605 is used to train the backbone layer and the ice recognition module by aligning the first ice-covered area and the first ice-covered type, and the second ice-covered area and the second ice-covered type as supervision;

[0159] The model determination module 606 is used to lock the parameters in the backbone layer and the ice recognition module when the training of the backbone layer and the ice recognition module is completed, so as to generate an ice detection model according to the pixel partitioning module, the backbone layer and the ice recognition module.

[0160] Optionally, the icing detection model training device further includes:

[0161] The preprocessing module is used to perform contrast normalization processing on the image samples.

[0162] Optionally, the feature extraction and fusion module 603 includes:

[0163] A one-dimensional sequence input submodule, used to input the one-dimensional sequence into a backbone layer, wherein the backbone layer includes a plurality of first division modules, a second division module, a feature enhancement module and a fusion module, wherein the first division module and the second division module are respectively connected to the feature enhancement module, and the feature enhancement module is connected to the fusion module;

[0164] An input sequence determination submodule, used to take the one-dimensional sequence as an initial input sequence;

[0165] A first division submodule, used for randomly dividing the input sequence into a plurality of subsequences in the first division module as the input sequence of the next module;

[0166] A second division submodule, configured to divide the input sequence into a plurality of subsequences of the same size in the second division module as the input sequence of the next module;

[0167] A feature enhancement submodule, used for identifying a subsequence with ice-covering features in the input sequence in the feature enhancement module, and performing feature enhancement on the subsequence with ice-covering features, and using the processed input sequence as an input sequence of the next module;

[0168] An input sequence fusion submodule is used to fuse the input sequence in the fusion module to obtain a fusion feature sequence as the input sequence of the next module;

[0169] The target feature sequence output submodule is used to use the fused feature sequence output by the last fusion module as the target feature sequence.

[0170] Optionally, the feature enhancement submodule includes:

[0171] A target input sequence determining unit, used to determine the input sequence of the feature enhancement module to obtain a target input sequence;

[0172] a target subsequence determination unit, configured to identify a subsequence having ice-covering characteristics in the target input sequence and mark it as a target subsequence;

[0173] A feature enhancement unit, configured to perform feature enhancement and update the attention weight for each subsequence in the target input sequence based on an attention mechanism to obtain a feature enhanced sequence, wherein the attention weight of the target subsequence is higher than that of other subsequences in the target input sequence;

[0174] The input sequence determination unit is used to use the processed target input sequence as the input sequence of the next module.

[0175] Optionally, the feature enhancement submodule further includes:

[0176] A standardization unit, used for standardizing each subsequence in the target input sequence to obtain a standardized subsequence;

[0177] The offset unit is used to scale and translate the standardized subsequence to obtain the updated target input sequence.

[0178] Optionally, the feature enhancement submodule further includes:

[0179] The superposition unit is used to perform residual superposition on each subsequence in the target input sequence and the corresponding feature enhancement sequence to obtain an output sequence of each subsequence.

[0180] Optionally, the icing detection model training device further includes:

[0181] A test picture acquisition module, used for acquiring the test picture of the power transmission line after the ice detection model training is completed;

[0182] The ice detection module is used to input the image to be tested into the ice detection model to obtain the ice area and ice type.

[0183] The icing detection model training device provided in the embodiment of the present invention can execute the icing detection model training method provided in any embodiment of the present invention, and has the corresponding functional modules and beneficial effects of the execution method.

[0184] Embodiment 4

[0185] Figure 7 A schematic diagram of an electronic device 40 that can be used to implement an embodiment of the present invention is shown. The electronic device is intended to represent various forms of digital computers, such as laptop computers, desktop computers, workstations, personal digital assistants, servers, blade servers, mainframe computers, and other suitable computers. The electronic device can also represent various forms of mobile devices, such as personal digital processing, cellular phones, smart phones, and other similar computing devices. The components shown herein, their connections and relationships, and their functions are merely examples and are not intended to limit the implementation of the present invention described and / or required herein.

[0186] like Figure 7 As shown, the electronic device 40 includes at least one processor 41, and a memory connected to the at least one processor 41, such as a read-only memory (ROM) 42, a random access memory (RAM) 43, etc., wherein the memory stores a computer program that can be executed by at least one processor, and the processor 41 can perform various appropriate actions and processes according to the computer program stored in the read-only memory (ROM) 42 or the computer program loaded from the storage unit 48 to the random access memory (RAM) 43. In the RAM 43, various programs and data required for the operation of the electronic device 40 can also be stored. The processor 41, the ROM 42, and the RAM 43 are connected to each other through a bus 44. An input / output (I / O) interface 45 is also connected to the bus 44.

[0187] A number of components in the electronic device 40 are connected to the I / O interface 45, including: an input unit 46, such as a keyboard, a mouse, etc.; an output unit 47, such as various types of displays, speakers, etc.; a storage unit 48, such as a disk, an optical disk, etc.; and a communication unit 49, such as a network card, a modem, a wireless communication transceiver, etc. The communication unit 49 allows the electronic device 40 to exchange information / data with other devices through a computer network such as the Internet and / or various telecommunication networks.

[0188] The processor 41 may be a variety of general and / or special processing components with processing and computing capabilities. Some examples of the processor 41 include, but are not limited to, a central processing unit (CPU), a graphics processing unit (GPU), various special artificial intelligence (AI) computing chips, various processors running machine learning model algorithms, digital signal processors (DSPs), and any appropriate processors, controllers, microcontrollers, etc. The processor 41 executes the various methods and processes described above, such as the ice detection model training method.

[0189] In some embodiments, the ice detection model training method can be implemented as a computer program, which is tangibly contained in a computer-readable storage medium, such as a storage unit 48. In some embodiments, part or all of the computer program can be loaded and / or installed on the electronic device 40 via the ROM 42 and / or the communication unit 49. When the computer program is loaded into the RAM 43 and executed by the processor 41, one or more steps of the ice detection model training method described above can be performed. Alternatively, in other embodiments, the processor 41 can be configured to perform the ice detection model training method in any other appropriate manner (for example, by means of firmware).

[0190] Various implementations of the systems and techniques described above herein can be implemented in digital electronic circuit systems, integrated circuit systems, field programmable gate arrays (FPGAs), application specific integrated circuits (ASICs), application specific standard products (ASSPs), systems on chips (SOCs), complex programmable logic devices (CPLDs), computer hardware, firmware, software, and / or combinations thereof. These various implementations can include: being implemented in one or more computer programs that can be executed and / or interpreted on a programmable system including at least one programmable processor, which can be a special purpose or general purpose programmable processor that can receive data and instructions from a storage system, at least one input device, and at least one output device, and transmit data and instructions to the storage system, the at least one input device, and the at least one output device.

[0191] Computer programs for implementing the methods of the present invention may be written in any combination of one or more programming languages. These computer programs may be provided to a processor of a general-purpose computer, a special-purpose computer, or other programmable data processing device, so that when the computer program is executed by the processor, the functions / operations specified in the flow chart and / or block diagram are implemented. The computer program may be executed entirely on the machine, partially on the machine, partially on the machine and partially on a remote machine as a stand-alone software package, or entirely on a remote machine or server.

[0192] In the context of the present invention, a computer-readable storage medium may be a tangible medium that may contain or store a computer program for use by or in combination with an instruction execution system, device or equipment. A computer-readable storage medium may include, but is not limited to, electronic, magnetic, optical, electromagnetic, infrared, or semiconductor systems, devices or equipment, or any suitable combination of the foregoing. Alternatively, a computer-readable storage medium may be a machine-readable signal medium. A more specific example of a machine-readable storage medium may include an electrical connection based on one or more lines, a portable computer disk, a hard disk, a random access memory (RAM), a read-only memory (ROM), an erasable programmable read-only memory (EPROM or flash memory), an optical fiber, a portable compact disk read-only memory (CD-ROM), an optical storage device, a magnetic storage device, or any suitable combination of the foregoing.

[0193] To provide interaction with a user, the systems and techniques described herein may be implemented on an electronic device having: a display device (e.g., a CRT (cathode ray tube) or LCD (liquid crystal display) monitor) for displaying information to the user; and a keyboard and a pointing device (e.g., a mouse or trackball) through which the user can provide input to the electronic device. Other types of devices may also be used to provide interaction with the user; for example, the feedback provided to the user may be any form of sensory feedback (e.g., visual feedback, auditory feedback, or tactile feedback); and input from the user may be received in any form (including acoustic input, voice input, or tactile input).

[0194] The systems and techniques described herein may be implemented in a computing system that includes backend components (e.g., as a data server), or a computing system that includes middleware components (e.g., an application server), or a computing system that includes frontend components (e.g., a user computer with a graphical user interface or a web browser through which a user can interact with implementations of the systems and techniques described herein), or a computing system that includes any combination of such backend components, middleware components, or frontend components. The components of the system may be interconnected by any form or medium of digital data communication (e.g., a communication network). Examples of communication networks include: a local area network (LAN), a wide area network (WAN), a blockchain network, and the Internet.

[0195] A computing system may include a client and a server. The client and the server are generally remote from each other and usually interact through a communication network. The client and server relationship is generated by computer programs running on the corresponding computers and having a client-server relationship with each other. The server may be a cloud server, also known as a cloud computing server or cloud host, which is a host product in the cloud computing service system to solve the defects of difficult management and weak business scalability in traditional physical hosts and VPS services.

[0196] It should be understood that the various forms of processes shown above can be used to reorder, add or delete steps. For example, the steps described in the present invention can be executed in parallel, sequentially or in different orders, as long as the desired results of the technical solution of the present invention can be achieved, and this document does not limit this.

[0197] The above specific implementations do not constitute a limitation on the protection scope of the present invention. It should be understood by those skilled in the art that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modification, equivalent substitution and improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.< / bndbox> < / name> < / object> < / height> < / width> < / size> < / filename> < / annotation>

Claims

1. A method for training an ice detection model, characterized in that: include: Acquire an image sample of the power transmission line, and acquire a first ice-covered area and a first ice-covered type in the image sample; Inputting the image samples into a pixel partitioning module for pixel partitioning to obtain a one-dimensional sequence of the image samples; Inputting the one-dimensional sequence into the backbone layer for processing to fuse the features of the one-dimensional sequence at multiple different data division scales to obtain a target feature sequence; Inputting the target feature sequence into an ice covering recognition module to obtain a second ice covering area and a second ice covering type; Training the backbone layer and the ice recognition module by aligning the first ice-covered area and the first ice-covered type, and the second ice-covered area and the second ice-covered type as supervision; When the training of the backbone layer and the ice recognition module is completed, the parameters in the backbone layer and the ice recognition module are locked to generate an ice detection model according to the pixel partitioning module, the backbone layer and the ice recognition module.

2. The ice detection model training method according to claim 1, characterized in that: After acquiring the image sample of the power transmission line, the method further includes: Performing contrast normalization processing on the image samples.

3. The ice detection model training method according to claim 1, characterized in that: The one-dimensional sequence is input into the backbone layer for processing to fuse the features of the one-dimensional sequence at multiple different data division scales to obtain a target feature sequence, including: Inputting the one-dimensional sequence into a backbone layer, wherein the backbone layer includes a plurality of first division modules, a second division module, a feature enhancement module and a fusion module, wherein the first division module and the second division module are respectively connected to the feature enhancement module, and the feature enhancement module is connected to the fusion module; Taking the one-dimensional sequence as an initial input sequence; In the first division module, the input sequence is randomly divided into a plurality of subsequences as the input sequence of the next module; In the second division module, the input sequence is divided into a plurality of subsequences of the same size as the input sequence of the next module; In the feature enhancement module, a subsequence with ice-covering features in the input sequence is identified, and features of the subsequence with ice-covering features are enhanced, and the processed input sequence is used as the input sequence of the next module; In the fusion module, the input sequence is fused to obtain a fusion feature sequence as the input sequence of the next module; The fused feature sequence output by the last fusion module is used as the target feature sequence.

4. The ice detection model training method according to claim 3, characterized in that: In the feature enhancement module, a subsequence with ice-covering features in the input sequence is identified, and features of the subsequence with ice-covering features are enhanced, and the processed input sequence is used as an input sequence of the next module, including: Determine the input sequence of the feature enhancement module to obtain a target input sequence; Identifying a subsequence having ice-covering characteristics in the target input sequence and marking it as a target subsequence; For each subsequence in the target input sequence, feature enhancement is performed based on an attention mechanism and attention weights are updated to obtain a feature enhanced sequence, wherein in the target input sequence, the attention weight of the target subsequence is higher than that of other subsequences; The processed target input sequence is used as the input sequence of the next module.

5. The ice detection model training method according to claim 4, characterized in that: Before identifying a subsequence having ice-covering characteristics in the target input sequence and marking it as a target subsequence, the method further includes: For each subsequence in the target input sequence, standardize the subsequence to obtain a standardized subsequence; The standardized subsequence is scaled and translated to obtain the updated target input sequence.

6. The ice detection model training method according to claim 4, characterized in that: Before using the processed target input sequence as the input sequence of the next module, the method further includes: For each subsequence in the target input sequence, residual superposition is performed on the subsequence and the corresponding feature enhancement sequence to obtain an output sequence of each subsequence.

7. The ice detection model training method according to any one of claims 1 to 6, characterized in that: Also includes: After the ice detection model training is completed, obtaining a picture of the transmission line to be tested; The image to be tested is input into the ice coverage detection model to obtain the ice coverage area and ice coverage type.

8. An ice detection model training device, characterized in that: include: A sample acquisition module, used to acquire an image sample of the transmission line, and acquire a first ice-covered area and a first ice-covered type in the image sample; A pixel partitioning module, used for inputting the image samples into the pixel partitioning module for pixel partitioning to obtain a one-dimensional sequence of the image samples; A feature extraction and fusion module is used to input the one-dimensional sequence into the backbone layer for processing, so as to fuse the features of the one-dimensional sequence at multiple different data division scales to obtain a target feature sequence; An ice-covering recognition module, used for inputting the target feature sequence into the ice-covering recognition module to obtain a second ice-covering area and a second ice-covering type; A training module, configured to train the backbone layer and the ice recognition module by aligning the first ice-covered area and the first ice-covered type, and the second ice-covered area and the second ice-covered type as supervision; A model determination module is used to lock the parameters in the backbone layer and the ice recognition module when the training of the backbone layer and the ice recognition module is completed, so as to generate an ice detection model according to the pixel partitioning module, the backbone layer and the ice recognition module.

9. An electronic device, characterized in that: The electronic device comprises: at least one processor; and a memory communicatively connected to the at least one processor; wherein, The memory stores a computer program executable by the at least one processor, and the computer program is executed by the at least one processor so that the at least one processor can execute the ice detection model training method according to any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores computer instructions, and the computer instructions are used to enable a processor to implement the ice detection model training method according to any one of claims 1 to 7 when executed.

Images