Scene text image super-resolution method and system based on semantic perception interaction

Abstract

The invention belongs to the technical field of computer image processing, and discloses a scene text image super-resolution method and system based on semantic perception interaction, and the method comprises the steps: collecting an image; the method comprises the following steps: performing segmentation operation on an image to obtain text regions, performing feature extraction on the text regions to obtain text region features, performing mapping operation on the text regions, screening the mapped text regions, performing operation on the screened text regions to obtain a text probability sequence, and obtaining text probability features based on the text probability sequence; extracting image features, performing information exchange on the image features and the text region features, adding the two features after information exchange to obtain features after interactive fusion, and obtaining a super-resolution image based on the features after interactive fusion and the text probability features; defining a loss function, and optimizing model parameters. The method effectively solves the problem of inaccurate text positioning and processing under the condition of high noise, and can better adapt to continuously changing resolution images.

Description

Technical Field

[0001] The present invention relates to the technical field of computer image processing, and in particular to a scene text image super-resolution method and system based on semantic perception interaction. Background Art

[0002] The statements in this section merely provide background information related to the present invention and do not necessarily constitute prior art.

[0003] With the rapid development of digital image processing, scene text image super-resolution technology has become a crucial research direction. It not only improves the clarity and readability of scene text and ensures the accurate transmission of information, but also significantly enhances the ability to display details of the image. Scene text image super-resolution can help solve the challenges of text recognition in different environments, such as illumination changes, text blur, and background complexity, and promote the innovation and application of computer vision technology. More importantly, this technology is of great significance for improving the practicality of intelligent transportation, automatic translation, augmented reality and other fields. It not only promotes the advancement of information extraction technology, but also promotes the improvement of data analysis and processing capabilities, thereby supporting the comprehensive development of smart cities and smart life. By continuously optimizing scene text image super-resolution technology, we can achieve more efficient and accurate text recognition, and promote scientific and technological progress and social development.

[0004] In the current field of scene text image processing, improving the resolution of images to enhance the readability and prominence of text is a key task. Existing super-resolution methods often have difficulty accurately locating and extracting text areas in environments with severe noise interference, which limits their application effects in specific scenarios. At present, many scene text image super-resolution models use text priors to guide the generation of super-resolution images. This method is often unable to cope with high-noise images. It is often difficult for the model to distinguish and locate the text position in high-noise images. This limitation is more obvious in extremely low-resolution images. Therefore, the text priors obtained are often inaccurate and will seriously affect the performance of the model. Summary of the invention

[0005] In order to solve the above problems, the present invention proposes a scene text image super-resolution method and system based on semantic perception interaction, which utilizes a pre-trained text region segmentation network to extract and mark text region information. By guiding the model at the semantic level, it effectively solves the problem of inaccurate text positioning and processing under high noise conditions, and can better adapt to the ever-changing resolution images.

[0006] In order to achieve the above object, the present invention adopts the following technical solution: In a first aspect, the present invention provides a scene text image super-resolution method based on semantic perception interaction, comprising the following steps: Collect text images in different scenarios and preprocess the images; Perform segmentation operation on the image to obtain text area, perform feature extraction on the text area to obtain text area features, perform mapping operation on the text area, filter the mapped text area, operate on the filtered text area to obtain text probability sequence, and obtain text probability features based on the text probability sequence; Extract features from the image to obtain image features, exchange information between the image features and the text region features, add the two features after information exchange to obtain interactively fused features, and obtain a super-resolution image based on the interactively fused features and the text probability features; Define the loss function and optimize the model parameters.

[0007] As an optional implementation, the image is preprocessed, specifically: The image is cropped, the text area is annotated on the cropped image, and the data set is divided into a training set and a test set according to the set ratio.

[0008] As an optional implementation, a text region segmentor based on a feature pyramid network is used to perform a segmentation operation on the image, and the text region segmentor is a pre-trained model.

[0009] As an optional implementation, a cross-attention mechanism is used to exchange information between image features and text area features, and the two features after the information exchange are added to obtain interactively fused features.

[0010] As an optional implementation, the interactively fused features are refined to enhance the feature expression capability, and the final interactively fused features are obtained. The final interactively fused features and text probability features are used to obtain a super-resolution image.

[0011] As an alternative implementation, the loss function is:

[0012] in, SR i is the generated super-resolution image, HR i It is a real high-resolution image. N is the total number of pixels in the image.

[0013] In a second aspect, the present invention provides a scene text image super-resolution system based on semantic perception interaction, comprising: The data collection and preprocessing module is configured to: collect text images in different scenes and preprocess the images; The text probability generation module is configured to: segment the image to obtain the text area, extract features from the text area to obtain text area features, map the text area, filter the mapped text area, operate the filtered text area to obtain a text probability sequence, and obtain text probability features based on the text probability sequence; The super-resolution module is configured to: extract features from the image to obtain image features, exchange information between the image features and text region features, add the two features after the information exchange to obtain interactively fused features, and obtain a super-resolution image based on the interactively fused features and the text probability features; The model training module is configured to: define the loss function and optimize the model parameters.

[0014] In a third aspect, the present invention provides an electronic device comprising a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method described in the first aspect is performed.

[0015] In a fourth aspect, the present invention provides a computer-readable storage medium for storing computer instructions, wherein when the computer instructions are executed by a processor, the method described in the first aspect is performed.

[0016] In a fifth aspect, the present invention provides a computer program product, comprising a computer program, which, when executed by a processor, implements the method described in the first aspect.

[0017] Compared with the prior art, the present invention has the following beneficial effects: The present invention discloses a scene text image super-resolution method and system based on semantic perception interaction, which uses a pre-trained text region segmentation network to extract and mark text region information. By guiding the model at the semantic level, the problem of inaccurate text positioning and processing under high noise conditions is effectively solved. A multimodal semantic information interaction module is created, which is committed to semantically interacting and fusing text region features with image features to achieve more accurate and efficient information integration. By introducing semantic segmentation and multimodal information interaction, the clarity and recognizability of text regions in images are significantly improved, and it can better adapt to changing resolution images. Not only can it accurately extract and enhance text regions under high noise and complex backgrounds, but it can also maintain excellent generalization ability and stable performance in a variety of scenarios. It not only promotes the development of scene text image processing technology, but also provides core technical support for intelligent document processing, augmented reality applications and other fields that require high-quality text recognition, helping to improve user experience and application accuracy.

[0018] Advantages of additional aspects of the present invention will be given in part in the following description, and in part will become obvious from the following description, or will be learned through practice of the present invention. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] The accompanying drawings in the specification, which constitute a part of the present invention, are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute improper limitations on the present invention.

[0020] Figure 1 A flowchart of a scene text image super-resolution method based on semantic perception interaction provided in Example 1 of the present invention; Figure 2 A framework diagram of a scene text image super-resolution method based on semantic perception interaction provided in Example 1 of the present invention; Figure 3 A framework diagram of a super-resolution module provided in Embodiment 1 of the present invention; Figure 4 A framework diagram for information exchange between image features and text region features provided in Example 1 of the present invention; Figure 5 This is a framework diagram of the text region segmenter provided in Example 1 of the present invention. DETAILED DESCRIPTION

[0021] The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

[0022] It should be noted that the following detailed descriptions are exemplary and are intended to provide further explanation of the present invention. Unless otherwise specified, all technical and scientific terms used herein have the same meanings as those commonly understood by those skilled in the art to which the present invention belongs.

[0023] It should be noted that the terms used herein are only for describing specific embodiments, and are not intended to limit exemplary embodiments according to the present invention. As used herein, unless the context clearly indicates otherwise, the singular form is also intended to include the plural form. In addition, it should be understood that the terms "include" and "have" and any of their variations are intended to cover non-exclusive inclusions, for example, a process, method, system, product or device comprising a series of steps or units is not necessarily limited to those steps or units clearly listed, but may include other steps or units that are not clearly listed or inherent to these processes, methods, products or devices.

[0024] In the absence of conflict, the embodiments of the present invention and the features of the embodiments may be combined with each other.

[0025] Example 1 like Figure 1-2As shown, this embodiment provides a scene text image super-resolution method based on semantic perception interaction, comprising the following steps: Step 1: Collect text images in different scenes and preprocess the images; Step 2: segment the image to obtain text regions, extract features from the text regions to obtain text region features, map the text regions, filter the mapped text regions, operate the filtered text regions to obtain text probability sequences, and obtain text probability features based on the text probability sequences; Step 3: Extract features from the image to obtain image features, exchange information between the image features and the text region features, add the two features after information exchange to obtain interactively fused features, and obtain a super-resolution image based on the interactively fused features and the text probability features; Step 4: Define the loss function and optimize the model parameters.

[0026] Step 1: Collect text images in different scenarios and perform preprocessing on the images. The specific implementation process is as follows: In order to fully evaluate the performance of the model in various scenarios, we collected datasets from different scenarios and application areas. The dataset includes a variety of complex text images, covering a variety of scenarios such as street signs, billboards, store signs, document scans, etc. Specifically, these images involve various environmental conditions, such as different lighting, text orientation, font style, and background complexity, to ensure the diversity and representativeness of the dataset.

[0027] The dataset is cropped and other operations are performed, the text areas are annotated on the processed images, and then the dataset is divided into training set and test set according to the set ratio.

[0028] Step 2: segment the image to obtain text areas, extract features from the text areas to obtain text area features, map the text areas, filter the mapped text areas, operate on the filtered text areas to obtain text probability sequences, and obtain text probability features based on the text probability sequences.

[0029] First, in the text information branch, a text region segmentation extractor based on a feature pyramid network is built to segment the text region location to serve as a priori for the subsequent extraction of text features. The specific implementation process is as follows: like Figure 5 As shown in the figure, the text region segmentation extractor is a priori branch before text feature extraction. Its function is to segment the text region position in the low-resolution image to provide prior knowledge for subsequent text feature extraction, making the subsequent text feature extraction more accurate. Among them, this text region segmentor is a pre-trained model. We willH×W×C The low-resolution image is scaled and fed into the text region segmentation extractor. This step ensures the consistency of the size of the input low-resolution image, making it more suitable for the model.

[0030] Specifically, the low-resolution image LR is processed by the feature pyramid network to process the features of different scales of the image. The feature pyramid network outputs feature maps at multiple scales, which capture different semantic information from shallow to deep. Then the deep, low-resolution feature maps are upsampled in a top-down manner and laterally connected and fused with adjacent high-resolution feature maps. Finally, the features extracted from each layer are concatenated and fused, and then the obtained P l After further processing by the text region feature extraction module, the final text region features are obtained. This module uses a progressive scale expansion algorithm ( PSE ), this algorithm combines multi-dimensional feature maps and starts from the smallest scale, gradually merging feature maps upwards to finally form a complete text region. The specific process is shown in formula (1) (2): (1) (2) in, Ci Indicates i The feature map of the layer, U ( ) indicates upsampling operation, Conv represents the convolution operation, P l is the final feature map, PSE Represents the operation of the progressive scale expansion algorithm.

[0031] Secondly, a text probability generation module is built after the text region segmentation extractor, which is dedicated to extracting text probability features based on the segmentation results of the text region segmentation extractor as a priori. The specific implementation process is as follows: like Figure 2 As shown in Figure 1, the text probability generation module is another module in the text information branch. The main function of this module is to extract text probability features. First, this module extracts the text region R segmented by the text region feature extraction module. Ta As a priori, R Ta Map to image R H×W×C Then the mapped image R Ta H×W×CApplying clustering algorithms to fine-tune the mapped images, the similarities of the features of these local regions can be analyzed and divided into different categories, thereby helping the model to better identify these local features. Through this clustering strategy, the model can apply different processing methods to different regions, and adaptively process them according to the feature distribution of each region, thereby improving the overall super-resolution effect. This can not only alleviate the information inconsistency problem caused by the direct fusion of different modal information, but also more effectively focus on important areas, improve the positioning accuracy of text areas and the detail restoration ability of super-resolution images. The filtered images are then added to the text probability generation module to generate text probability features. Specifically, the input filtered image R Ta H×W×C First, it is resized by the bicubic interpolator to match the input of the text probability generation module, and then passed to the text probability generation module to generate a text probability matrix with a width of L. Each position of the text probability is a vector of size |A|, where |A| is the number of categories of the letter A used in the CRNN. ​​The specific formula is shown in formula (3)(4)(5): (3) (4) (5) in N ε ( x i ) represents a data point x i of ε -neighborhood, that is, all x i Distance not exceeding ε The set of points, ∣ N ε ( x i )∣ represents the number of points in the neighborhood, MinPts is the preset minimum neighborhood point threshold, which is used to define whether a point is a core point. Ta It is the text region output by the text region feature extraction module, with a size of H×W×C. TP_Generation represents the operation of the text probability generation module.

[0032] Generated text probability sequence T DIts size may not match the image feature map in the SR branch. To solve this problem, we use TP Transformer to convert the text probability sequence into a feature map. TP Transformer mainly consists of four Deconv modules, and all deconvolution kernels are 3×3. After passing through these four modules, TD will match the dimension of the image feature. Its specific formula is shown in formula (6): (6) in, Deconv refers to the deconvolution module, T F Represents the obtained text probability features.

[0033] Step 3: Extract features from the image to obtain image features, exchange information between the image features and the text area features, add the two features after the information exchange to obtain interactively fused features, and obtain a super-resolution image based on the interactively fused features and the text probability features.

[0034] First, an image feature extraction module is installed in the image feature branch to extract image features from low-resolution images for subsequent guidance of super-resolution image generation. The specific implementation process is as follows: like Figure 2 As shown in the figure, in the image feature branch, the low-resolution image will first pass through the STN network to correct the irregularities in the data. Its function is to enable the neural network to have the ability to perform geometric transformations on the input data, thereby improving the model's robustness to image deformation, perspective changes, etc. Subsequently, the corrected image is passed through the image feature extraction part to extract the low-resolution image features. F The specific formula is shown in formula (7): (7) in, IFE Refers to the image feature extraction part, U refers to the input image, Localization ( ) refers to the positioning network, Grid ( ) refers to mapping the output image coordinates to the input image coordinates, Sampler Refers to sampling the input image according to the mapped coordinates.

[0035] Secondly, a multimodal semantic information interaction module is created to integrate text region features with image features through semantic interaction to achieve more accurate and efficient information integration. The specific implementation process is as follows: like Figure 4As shown in Figure 1, the module consists of two branches, which input text region features and image features respectively. First, we use the cross-attention mechanism to perform global information exchange between the two branches. Then, we add the features of the two branches after the information exchange. Finally, in order to enhance the expressiveness of the features and prevent the interference of noise generated during the information exchange process, we perform a refinement operation after adding the two features.

[0036] Specifically, we first transform the size of R H×W×C The input features are flattened, where C is the number of feature channels, and the flattened value is R N×C , where C is H×W. Then we will interact information between the two flattened branch features. The cross attention mechanism for information interaction is based on the traditional self-attention mechanism. The self-attention mechanism first encodes the input vector into Query (Q), Key (K) and Value (V). Multiply the transpose of Q and K to get the global attention map. However, we use the global context vector G=K T V. The attention result is calculated by QG, ​​which can reduce the amount of calculation. In MSIIM, we embed the flattened features into K and V, and then get the vector G. After getting G, we multiply G with the flattened vector of another branch to get the output, that is, to exchange information. Then the result vector U is concatenated with the flattened feature X to generate a vector of size R N×2C Then it goes through a channel layer to resize it to R N×C Next, a reshape operation is performed to resize it to R H×W×C , and then add it to the features before flattening. Through such steps, the information of each branch interacts to obtain the final feature representation. I , F IR Then we add and fuse these two features. Then, in order to enhance the feature expression capability and prevent the noise interference generated by information interaction, we refine the fused features and finally obtain the interactive fused features. Through this module, we can strengthen the fusion between these two types of features to achieve more accurate and efficient information integration. The above process can be expressed by formula (8) (9): (8) (9) in, X IR represents the flattened text region features, X I represents the flattened image features, U I represents the image features after cross attention, UIR Represents the text region features after cross attention. Here F I , F IR Represents the final image features and text area features after information interaction. Ch represents the channel layer, N Represents a reshape operation.

[0037] Finally, we build a super-resolution module, which is dedicated to using text probability features and image features to guide the super-resolution module to perform super-resolution on images. The specific implementation process is as follows: like Figure 3 As shown, after obtaining the image feature I F and text probability feature T F After that, it is added to the super-resolution module. First, the image features are simply denoised, and then the image features and text probability features are spliced ​​and fused. Then a residual link is introduced to prevent the degradation problem and gradient disappearance problem in deep network training. Then it is processed by a recurrent neural network to obtain the final features. After obtaining the final features, it is upsampled to obtain a high-resolution image. The above process can be expressed by formula (10)(11): (10) (11) in, Y represents the final feature, GRU represents a recurrent neural network, BN represents batch normalization, PReLU represents the activation function layer, Conv Represents a convolution operation, PixelShuffle represents an upsampling operation, S Represents the upsampling multiple, and SR represents high-resolution image.

[0038] Step 4: Define the loss function and optimize the model parameters.

[0039] After the model is built, the model is trained using the training set and super-resolution image loss. The optimal weight after training is selected and the model is tested using the test set. The specific implementation process is as follows: In the super-resolution task, the difference between the generated high-resolution image and the target high-resolution image is one of the core indicators to measure the performance of the model. To this end, the model needs to evaluate the difference between the generated image and the real image at the pixel level through an error function, and use the error backpropagation to optimize the parameters of the model, thereby continuously improving the quality of the model-generated images. Specifically, the super-resolution image SR generated by the model is first compared pixel by pixel with the target high-resolution image HR. These loss functions quantify the closeness of the generated image to the real image by calculating the difference between the two images in pixels. The above process can be expressed as formula (12): (12) in, SR i is the generated high-resolution image, HR i It is a real high-resolution image. N is the total number of pixels in the image.

[0040] Example 2 This embodiment provides a scene text image super-resolution system based on semantic perception interaction, including: The data collection and preprocessing module is configured to: collect text images in different scenes and preprocess the images; The text probability generation module is configured to: segment the image to obtain the text area, extract features from the text area to obtain text area features, map the text area, filter the mapped text area, operate the filtered text area to obtain a text probability sequence, and obtain text probability features based on the text probability sequence; The super-resolution module is configured to: extract features from the image to obtain image features, exchange information between the image features and text region features, add the two features after the information exchange to obtain interactively fused features, and obtain a super-resolution image based on the interactively fused features and the text probability features; The model training module is configured to: define the loss function and optimize the model parameters.

[0041] It should be noted that the above modules correspond to the steps described in Example 1, and the examples and application scenarios implemented by the above modules and the corresponding steps are the same, but are not limited to the contents disclosed in the above Example 1. It should be noted that the above modules, as part of the system, can be executed in a computer system such as a set of computer executable instructions.

[0042] In further embodiments, there is also provided: An electronic device includes a memory and a processor, and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method described in Embodiment 1 is performed. For the sake of brevity, it will not be described in detail here.

[0043] It should be understood that in this embodiment, the processor may be a central processing unit CPU, and the processor may also be other general-purpose processors, digital signal processors DSP, application-specific integrated circuits ASIC, off-the-shelf programmable gate arrays FPGA or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor may be a microprocessor or the processor may also be any conventional processor, etc.

[0044] The memory may include a read-only memory and a random access memory, and provide instructions and data to the processor. A portion of the memory may also include a non-volatile random access memory. For example, the memory may also store information about the device type.

[0045] A computer-readable storage medium is used to store computer instructions, and when the computer instructions are executed by a processor, the method described in Example 1 is completed.

[0046] The method in Example 1 can be directly embodied as a hardware processor, or a combination of hardware and software modules in the processor. The software module can be located in a mature storage medium in the field such as a random access memory, a flash memory, a read-only memory, a programmable read-only memory, or an electrically erasable programmable memory, a register, etc. The storage medium is located in the memory, and the processor reads the information in the memory and completes the steps of the above method in combination with its hardware. To avoid repetition, it is not described in detail here.

[0047] A computer program product includes a computer program, and when the computer program is executed by a processor, the method described in embodiment 1 is implemented.

[0048] The present invention also provides at least one computer program product tangibly stored on a non-transitory computer-readable storage medium. The computer program product includes computer executable instructions, such as instructions included in a program module, which are executed in a device on a real or virtual processor of the target to perform the process / method as described above. Typically, program modules include routines, programs, libraries, objects, classes, components, data structures, etc. that perform specific tasks or implement specific abstract data types. In various embodiments, the functions of program modules can be combined or divided between program modules as needed. Machine executable instructions for program modules can be executed in local or distributed devices. In distributed devices, program modules can be located in local and remote storage media.

[0049] The computer program code for implementing the method of the present invention can be written in one or more programming languages. These computer program codes can be provided to the processor of a general-purpose computer, a special-purpose computer or other programmable data processing device, so that the program code, when executed by the computer or other programmable data processing device, causes the function / operation specified in the flow chart and / or block diagram to be implemented. The program code can be executed completely on a computer, partially on a computer, as an independent software package, partially on a computer and partially on a remote computer or completely on a remote computer or server.

[0050] In the context of the present invention, computer program codes or related data may be carried by any appropriate carrier to enable a device, apparatus or processor to perform the various processes and operations described above. Examples of carriers include signals, computer readable media, and the like. Examples of signals may include electrical, optical, radio, acoustic or other forms of propagation signals, such as carrier waves, infrared signals, and the like.

[0051] Those of ordinary skill in the art will appreciate that the units and algorithm steps of each example described in conjunction with this embodiment can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered to be beyond the scope of this application.

[0052] Although the above describes the specific implementation mode of the present invention in conjunction with the accompanying drawings, it is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art on the basis of the technical solution of the present invention without creative work are still within the scope of protection of the present invention.

Claims

1. A scene text image super-resolution method based on semantic perception interaction, characterized in that: The following steps are involved: Collect text images in different scenarios and preprocess the images; Perform segmentation operation on the image to obtain text area, perform feature extraction on the text area to obtain text area features, perform mapping operation on the text area, filter the mapped text area, operate on the filtered text area to obtain text probability sequence, and obtain text probability features based on the text probability sequence; Extract features from the image to obtain image features, exchange information between the image features and the text region features, add the two features after information exchange to obtain interactively fused features, and obtain a super-resolution image based on the interactively fused features and the text probability features; Define the loss function and optimize the model parameters.

2. The scene text image super-resolution method based on semantic perception interaction according to claim 1, characterized in that: Preprocess the image, specifically: The image is cropped, the text area is annotated on the cropped image, and the data set is divided into a training set and a test set according to the set ratio.

3. The scene text image super-resolution method based on semantic perception interaction according to claim 1, characterized in that: A text region segmentor based on a feature pyramid network is used to perform a segmentation operation on the image, and the text region segmentor is a pre-trained model.

4. The scene text image super-resolution method based on semantic perception interaction according to claim 1, characterized in that: The cross-attention mechanism is used to exchange information between image features and text area features, and the two features after information exchange are added to obtain the interactively fused features.

5. The scene text image super-resolution method based on semantic perception interaction according to claim 4, characterized in that: The interactively fused features are refined to enhance the feature expression capability and obtain the final interactively fused features. The final interactively fused features and text probability features are used to obtain a super-resolution image.

6. The scene text image super-resolution method based on semantic perception interaction according to claim 1, characterized in that: The loss function is: in, SR i is the generated super-resolution image, HR i It is a real high-resolution image. N is the total number of pixels in the image.

7. A scene text image super-resolution system based on semantic perception interaction, characterized in that: include: The data collection and preprocessing module is configured to: collect text images in different scenes and preprocess the images; The text probability generation module is configured to: segment the image to obtain the text area, extract features from the text area to obtain text area features, map the text area, filter the mapped text area, operate the filtered text area to obtain a text probability sequence, and obtain text probability features based on the text probability sequence; The super-resolution module is configured to: extract features from the image to obtain image features, exchange information between the image features and text region features, add the two features after the information exchange to obtain interactively fused features, and obtain a super-resolution image based on the interactively fused features and text probability features; The model training module is configured to: define the loss function and optimize the model parameters.

8. An electronic device, characterized in that: The method comprises a memory and a processor and computer instructions stored in the memory and executed on the processor, wherein when the computer instructions are executed by the processor, the method according to any one of claims 1 to 6 is completed.

9. A computer-readable storage medium, characterized in that: Used to store computer instructions, which, when executed by a processor, complete the method described in any one of claims 1 to 6.

10. A computer program product, characterized in that The invention comprises a computer program, which, when executed by a processor, implements the method according to any one of claims 1 to 6.

Images