Multi-scale data processing method and device, equipment, storage medium and program product

Abstract

The invention discloses a multi-scale data processing method and device, equipment, a storage medium and a program product, and the method comprises the steps: constructing a first multi-scale data set according to a pre-collected multi-scale data sample; performing frequency feature extraction on the first multi-scale data set to construct a first frequency feature set; constructing a neural network according to the first frequency feature set, and performing data fitting on the first multi-scale data set through the neural network to perform frequency adaptive optimization on the neural network to obtain a final neural network model; processing the to-be-processed multi-scale data by adopting a neural network model to obtain a data processing result; according to the method, the neural network is constructed and adaptively optimized based on the overall frequency characteristics of the data set, the expression ability of the neural network model on multi-scale data can be improved, the calculation amount is reduced, the data processing efficiency is improved, the operation pressure can be reduced, and the operation speed can be improved.

Description

Technical Field

[0001] The present invention relates to the technical field of data processing, and in particular, to a multi-scale data processing method, apparatus, device, storage medium, and program product. Background Art

[0002] Neural networks, with their powerful non-linear modeling capabilities, exhibit excellent fitting and generalization performance when dealing with high-dimensional and complex data distributions. Traditional neural networks have achieved remarkable results in multiple fields such as data fitting, pattern recognition, image classification, natural language processing, and data clustering analysis. However, when faced with certain special types of data, especially data with obvious multi-scale structures, the performance of existing neural networks is often limited.

[0003] Multi-scale data refers to data that contains information with different spatial / temporal / feature scales simultaneously. For example, it contains both slow-changing large-scale low-frequency components and rapidly changing small-scale high-frequency components. Such structures widely exist in practical problems such as image texture processing, time series, and biological signal analysis. Due to the existence of a "spectral bias" phenomenon in neural networks themselves, that is, they tend to first learn low-frequency components in the initial stage of training and thus it is difficult to effectively capture high-frequency details. Therefore, when dealing with such multi-scale data, there are often problems such as inaccurate fitting, non-convergence in training, and poor generalization ability. Currently, the existing technology mainly constructs multi-scale feature mapping layers to extract multi-scale features of each data. Although it can solve to some extent the problem that traditional neural networks are difficult to effectively capture high-frequency details, the fitting effect is not good, and multi-scale feature mapping needs to be performed for each data, resulting in a large amount of calculation and low data processing efficiency. Summary of the Invention

[0004] In view of the problems existing in the prior art, embodiments of the present invention provide a multi-scale data processing method, apparatus, device, storage medium, and program product, which can effectively improve the expression ability of neural networks on multi-scale data, improve the accuracy of data fitting and the model precision of neural networks when dealing with multi-scale data, and at the same time can effectively reduce the amount of calculation and improve the data processing efficiency.

[0005] In a first aspect, embodiments of the present invention provide a multi-scale data processing method, including: Construct a first multi-scale data set according to pre-collected multi-scale data samples; Extract frequency features of the first multi-scale data set, and construct a first frequency feature set according to the extracted frequency features of the first multi-scale data set; Construct a neural network according to the first frequency feature set; Perform data fitting on the first multi-scale dataset through the neural network to perform frequency adaptive optimization on the neural network and obtain a neural network model; Process the multi-scale data to be processed using the neural network model to obtain corresponding data processing results.

[0006] As an improvement to the above solution, the multi-scale data to be processed includes image data to be processed.

[0007] As an improvement to the above solution, the extraction of frequency features from the first multi-scale dataset includes: Perform discrete Fourier transform on the first multi-scale dataset to extract the frequency information of the first multi-scale dataset; Select the maximum frequency information from the frequency information and calculate the dynamic threshold; Select the frequency information whose absolute value of the frequency information is greater than the dynamic threshold from the frequency information, and use the frequency corresponding to the selected frequency information as the frequency feature of the first multi-scale dataset.

[0008] As an improvement to the above solution, the construction of a neural network according to the first frequency feature set includes: Construct a neural network according to the number of frequency features in the first frequency feature set; wherein, the neural network includes a first sub-network, a second sub-network, and a third sub-network; the number of the first sub-network and the second sub-network is determined according to the number of frequency features in the first frequency feature set.

[0009] As an improvement to the above solution, the performing data fitting on the first multi-scale dataset through the neural network to perform frequency adaptive optimization on the neural network and obtain a neural network model includes: Use the multi-scale data samples in the first multi-scale dataset as the input of the neural network, and the observed data of the corresponding multi-scale data samples as the output of the neural network to perform data fitting on the first multi-scale dataset; During the data fitting process, use the stochastic gradient descent algorithm to optimize the loss function of the neural network to obtain the fitting solution output in this round; According to the fitting solution output in this round, determine the regular region for solution, and uniformly sample the first multi-scale dataset on the regular region to construct a second multi-scale dataset; Extract the frequency features of the second multi-scale dataset, and update the first frequency feature set according to the frequency features of the second multi-scale dataset; According to the number of frequency features in the updated first frequency feature set, reconstruct the neural network, and use the reconstructed neural network to perform data fitting on the first multi-scale data set until the number of data fitting times after the reconstruction of the neural network reaches a preset adaptive number of times, so as to obtain the neural network model.

[0010] As an improvement of the above solution, the constructing the first multi-scale data set according to the pre-collected multi-scale data samples includes: Perform uniform distribution sampling or Latin hypercube sampling on the pre-collected multi-scale data samples; Construct the first multi-scale data set according to the sampled multi-scale data samples.

[0011] In a second aspect, an embodiment of the present invention provides a multi-scale data processing method, including: Construct a first multi-scale image data set according to the pre-collected image data samples; Extract frequency features from the first multi-scale image data set, and construct a second frequency feature set according to the extracted frequency features of the first multi-scale image data set; Construct a neural network according to the second frequency feature set; Perform data fitting on the first multi-scale image data set through the neural network to perform frequency adaptive optimization on the neural network, so as to obtain an image recognition model; Use the image recognition model to process the image data to be processed to obtain an image recognition result.

[0012] As an improvement of the above solution, the extracting frequency features from the first multi-scale image data set includes: Perform discrete Fourier transform on the first multi-scale image data set to extract the frequency information of the first multi-scale image data set; Select the maximum frequency information from the frequency information and calculate the dynamic threshold; Select the frequency information whose absolute value of the frequency information is greater than the frequency information corresponding to the dynamic threshold from the frequency information, and use the frequency corresponding to the selected frequency information as the frequency feature of the first multi-scale image data set.

[0013] As an improvement of the above solution, the constructing a neural network according to the second frequency feature set includes: Construct a neural network according to the number of frequency features in the second frequency feature set; wherein, the neural network includes a first sub-network, a second sub-network, and a third sub-network; the numbers of the first sub-network and the second sub-network are determined according to the number of frequency features in the second frequency feature set.

[0014] As an improvement of the above solution, the data fitting of the first multi-scale image dataset by the neural network to perform frequency adaptive optimization on the neural network to obtain an image recognition model includes: Taking the image data samples in the first multi-scale image dataset as the input of the neural network, and the observation data of the corresponding image data samples as the output of the neural network, and performing data fitting on the first multi-scale image dataset; During the data fitting process, the random gradient descent algorithm is used to optimize the loss function of the neural network to obtain the fitting solution output in this round; According to the fitting solution output in this round, determine the regular region to be solved, and uniformly sample the first multi-scale image dataset on the regular region to construct a second multi-scale image dataset; Extract the frequency features of the second multi-scale image dataset, and update the second frequency feature set according to the frequency features of the second multi-scale image dataset; According to the number of frequency features in the updated second frequency feature set, reconstruct the neural network, and use the reconstructed neural network to perform data fitting on the first multi-scale image dataset until the number of data fitting times after the reconstruction of the neural network reaches the preset adaptive number of times to obtain the image recognition model.

[0015] As an improvement of the above solution, the construction of the first multi-scale image dataset according to the pre-collected image data samples includes: Performing uniform distribution sampling or Latin hypercube sampling on the pre-collected image data samples; Construct a first multi-scale image dataset according to the sampled image data samples.

[0016] In a third aspect, an embodiment of the present invention provides a multi-scale data processing device, including: A first multi-scale dataset construction module, configured to construct a first multi-scale dataset according to pre-collected multi-scale data samples; A first frequency feature extraction module, configured to extract frequency features from the first multi-scale dataset, and construct a first frequency feature set according to the extracted frequency features of the first multi-scale dataset; A first neural network construction module, configured to construct a neural network according to the first frequency feature set; A first network adaptive optimization module, configured to perform data fitting on the first multi-scale dataset through the neural network to perform frequency adaptive optimization on the neural network to obtain a neural network model; A first data processing module, configured to process the multi-scale data to be processed by using the neural network model to obtain a corresponding data processing result.

[0017] Fourthly, an embodiment of the present invention provides a multi-scale data processing device, including: A second multi-scale data set construction module, configured to construct a first multi-scale image data set according to pre-collected image data samples; A second frequency feature extraction module, configured to perform frequency feature extraction on the first multi-scale image data set, and construct a second frequency feature set according to the extracted frequency features of the first multi-scale image data set; A second neural network construction module, configured to construct a neural network according to the second frequency feature set; A second network adaptive optimization module, configured to perform data fitting on the first multi-scale image data set through the neural network, so as to perform frequency adaptive optimization on the neural network to obtain an image recognition model; A second data processing module, configured to process the to-be-processed image data by using the image recognition model to obtain an image recognition result.

[0018] Fifthly, an embodiment of the present invention provides a multi-scale data processing device, including: a processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, the multi-scale data processing method described in any item of the first aspect or the multi-scale data processing method described in any item of the second aspect is implemented.

[0019] Sixthly, an embodiment of the present invention provides a computer-readable storage medium storing a computer program. When the computer program runs, the device where the computer-readable storage medium is located is controlled to execute the multi-scale data processing method described in any item of the first aspect or the multi-scale data processing method described in any item of the second aspect.

[0020] Seventhly, an embodiment of the present invention provides a computer program product, including a computer program / instructions. When the computer program / instructions are executed by a processor, the multi-scale data processing method described in any item of the first aspect or the multi-scale data processing method described in any item of the second aspect is implemented.

[0021] Compared with the prior art, a multi-scale data processing method, apparatus, device, storage medium, and program product provided by an embodiment of the present invention first constructs a first multi-scale data set based on pre-collected multi-scale data samples; then extracts frequency features from the first multi-scale data set, and constructs a first frequency feature set according to the extracted frequency features of the first multi-scale data set; then constructs a neural network according to the first frequency feature set, and performs data fitting on the first multi-scale data set through the neural network to perform frequency adaptive optimization on the neural network to obtain a final neural network model; finally, uses the neural network model to process the multi-scale data to be processed, and corresponding data processing results can be obtained. The embodiment of the present invention constructs and adaptively optimizes a neural network based on the overall frequency features of a multi-scale data set. On the one hand, it can improve the expression ability of the neural network model on multi-scale data, improve the accuracy of data fitting and the model precision when the neural network processes multi-scale data. On the other hand, it can effectively reduce the amount of calculation and improve the data processing efficiency, thereby reducing the operation pressure of the terminal processor and improving the operation speed. BRIEF DESCRIPTION OF THE DRAWINGS

[0022] In order to more clearly illustrate the technical solutions of the present invention, the accompanying drawings to be used in the embodiments will be briefly introduced below. Obviously, the accompanying drawings in the following description are only some embodiments of the present invention. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0023] Figure 1 is a flowchart of a multi-scale data processing method provided by an embodiment of the present invention; Figure 2 is a schematic diagram of the construction process of the neural network model provided by an embodiment of the present invention; Figure 3 is another flowchart of a multi-scale data processing method provided by an embodiment of the present invention; Figure 4 is a schematic diagram of the multi-scale function representation of the multi-scale data set provided by an embodiment of the present invention; Figure 5 is the first Fourier spectrum image provided by an embodiment of the present invention; Figure 6 is the first comparison schematic diagram between the multi-scale function representation and its predicted data provided by an embodiment of the present invention; Figure 7 is the second Fourier spectrum image provided by an embodiment of the present invention; Figure 8 is the second comparison schematic diagram between the multi-scale function representation and its predicted data provided by an embodiment of the present invention; Figure 9 is the third Fourier spectrum image provided by an embodiment of the present invention; Figure 10 is a structural block diagram of a multi-scale data processing device provided by an embodiment of the present invention; Figure 11 is another structural block diagram of a multi-scale data processing device provided by an embodiment of the present invention; Figure 12 is a structural block diagram of a multi-scale data processing device provided by an embodiment of the present invention. Detailed implementation manners

[0024] Next, the technical solutions in the embodiments of the present invention will be clearly and completely described with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative efforts shall fall within the protection scope of the present invention.

[0025] It can be understood that in the embodiments of the present invention, the various numerical numbers involved are only for the convenience of description and do not limit the scope of the present application. The size of the serial numbers of each process does not mean the sequence of execution, and the execution sequence of each process should be determined according to its function and internal logic. In the embodiments of the present invention, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or sequence between these entities or operations. The terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or further includes elements inherent to such process, method, article or device. Without further limitation, the elements defined by the statement "including..." do not exclude the existence of additional identical elements in the process, method, article or device including the said elements.

[0026] See Figure 1 , Figure 1 is a flowchart of a multi-scale data processing method provided by an embodiment of the present invention. The multi-scale data processing method specifically includes: S11: Construct a first multi-scale data set according to pre-collected multi-scale data samples; S12: Extract frequency features from the first multi-scale data set, and construct a first frequency feature set according to the extracted frequency features of the first multi-scale data set; S13: Construct a neural network according to the first frequency feature set; S14: Perform data fitting on the first multi-scale data set through the neural network to perform frequency adaptive optimization on the neural network and obtain a neural network model; S15: Use the neural network model to process the multi-scale data to be processed and obtain corresponding data processing results.

[0027] It should be noted that the embodiments of the present invention can be executed by intelligent terminal devices such as servers, computers, and computers.

[0028] Among them, the multi-scale data to be processed includes image data to be processed. Such as remote sensing images, medical images, images captured by imaging devices, map images, and other image data; in this case, the constructed neural network module can perform prediction processing such as image classification and recognition on the image data. Further, the multi-scale data to be processed can also be time series data. Such as environmental information (temperature, meteorology) time series, signal (such as biological signal) time series, and other time series data; in this case, the constructed neural network module can perform prediction processing on the time series data. Similarly, the multi-scale data samples can be corresponding image data samples and time series data samples.

[0029] In the embodiments of the present invention, multi-scale data samples are collected and an initial first multi-scale data set is constructed. Then, based on the initial first multi-scale data set as a whole, frequency features are extracted from the entire first multi-scale data set, and an initial first frequency feature set is constructed; using this first frequency feature set, a corresponding neural network is constructed; among them, the neural network includes multiple different sub-networks, and different sub-networks are used to capture different frequency features to capture different scale information of the data. After that, use this neural network to perform data fitting on the first multi-scale data set and its corresponding observation data set; among them, the observation data set includes the observation data corresponding to the multi-scale data samples in the first multi-scale data set; based on the data fitting results, perform frequency adaptive optimization adjustment on the neural network to construct a final neural network model. Subsequently, this neural network model can be used to process the multi-scale data to be processed to obtain corresponding data processing results; for example, this neural network model can be used to perform recognition processing on the image data to be processed to obtain corresponding image recognition results.

[0030] The embodiment of the present invention constructs and adaptively optimizes a neural network based on the frequency characteristics of the overall multi-scale data set. On the one hand, it can improve the expression ability of the neural network model on multi-scale data, and improve the accuracy of data fitting and the model precision of the neural network when processing multi-scale data. On the other hand, it can effectively reduce the amount of calculation and improve the data processing efficiency, thereby reducing the operation pressure of the terminal processor and improving the operation speed.

[0031] Further, constructing a first multi-scale data set according to the pre-collected multi-scale data samples includes: Performing uniform distribution sampling or Latin hypercube sampling on the pre-collected multi-scale data samples; Constructing a first multi-scale data set according to the sampled multi-scale data samples.

[0032] In the embodiment of the present invention, for the pre-collected multi-scale data samples, uniform distribution sampling or Latin hypercube sampling can be used to sample a small amount of sample data for constructing the first multi-scale data set , and at the same time sampling the corresponding observation data to construct an observation data set , where , represents the j-th multi-scale data sample, represents the observation data corresponding to the j-th multi-scale data sample, represents the number of samples in the first multi-scale data set, that is, the initial number of samples.

[0033] Further, extracting the frequency characteristics of the first multi-scale data set includes: Performing discrete Fourier transform on the first multi-scale data set to extract the frequency information of the first multi-scale data set; Selecting the maximum frequency information from the frequency information and calculating the dynamic threshold; Selecting the frequency information whose absolute value of the frequency information is greater than the dynamic threshold from the frequency information, and taking the frequency corresponding to the selected frequency information as the frequency characteristic of the first multi-scale data set.

[0034] Performing discrete Fourier transform on the above-constructed first multi-scale data set to extract the frequency information of the first multi-scale data set , where the frequency information has the following function expression: (1); where k represents the frequency, i represents the imaginary unit, and the frequency information can be understood as the coefficient corresponding to the frequency k (i.e., the Fourier coefficient).

[0035] Select the maximum frequency information , and order , and define the frequency coefficient , calculate the dynamic threshold ; To set a frequency filter: , select the frequency information that meets the frequency screening condition The corresponding frequency k is used as the frequency feature of the first multi-scale data set to construct the first frequency feature set ,in, , represents the nth frequency, Represents the number of frequency features of the first multiscale dataset.

[0036] The embodiment of the present invention takes into account the overall frequency characteristics of a multi-scale data set, performs a discrete Fourier transform on the entire data set to obtain the corresponding frequency, and can capture the global frequency distribution across samples and scales at one time, convert the multi-scale problem into a multi-frequency problem, and mine the intrinsic correlation of multi-scale data samples in the multi-scale data set. On the one hand, it reduces the complexity of network learning cross-scale correlations and can greatly reduce the computational complexity. On the other hand, it does not need to evaluate the multi-scale characteristics of each multi-scale data sample separately, which greatly reduces the amount of calculation, thereby reducing the computing pressure of the terminal and improving the computing speed.

[0037] In an optional embodiment, constructing a neural network according to the first frequency feature set includes: A neural network is constructed according to the number of frequency features in the first frequency feature set; wherein the neural network includes a first subnetwork, a second subnetwork, and a third subnetwork; and the number of the first subnetwork and the second subnetwork is determined according to the number of frequency features in the first frequency feature set.

[0038] In the embodiment of the present invention, the high-frequency components in the data set can be captured by setting the frequency screening conditions, and then the corresponding neural network is constructed based on the proposed frequency features. The specific process is as follows: The number of frequency features in the first frequency feature set initially established is counted, and then multiple sub-networks are constructed based on the number of frequency features to capture different frequency features of the data set. For example, a first sub-network and a second sub-network corresponding to the number of frequency features are constructed to capture different frequency features (i.e., high-frequency components) of the frequency feature set, and a first sub-network is constructed to capture the global features of the data (mainly low-frequency components). The specific examples are as follows: (3); Among them, x represents the input of the neural network (such as multi-scale data samples ), represents the predicted output of the neural network, , , respectively represent different sub - networks, namely the first sub - network, the second sub - network, and the third sub - network. , , respectively represent the trainable network parameters of the corresponding sub - networks, represents the set of network parameters of the neural network (i.e., including all the trainable parameters of the neural network, such as the above - mentioned , , ). It can be understood that the network parameters of each sub - network are independent of each other.

[0039] The embodiment of the present invention constructs a neural network based on the frequency feature set constructed from a multi - scale data set, which can solve the problem that traditional neural networks are difficult to effectively capture high - frequency details, improve the expression ability of the neural network on multi - scale data, and enhance the fitting effect and generalization ability of the network.

[0040] In an optional embodiment, the data fitting of the first multi - scale data set by the neural network to perform frequency - adaptive optimization on the neural network to obtain a neural network model includes: Taking the multi - scale data samples in the first multi - scale data set as the input of the neural network, and the observed data of the corresponding multi - scale data samples as the output of the neural network, and performing data fitting on the first multi - scale data set; During the data fitting process, using the stochastic gradient descent algorithm to optimize the loss function of the neural network to obtain the fitting solution of the current round of output; According to the fitting solution of the current round of output, determining the regular region for solution, and uniformly sampling the first multi - scale data set on the regular region to construct a second multi - scale data set; Extracting the frequency features of the second multi - scale data set, and updating the first frequency feature set according to the frequency features of the second multi - scale data set; According to the number of frequency features in the updated first frequency feature set, reconstructing the neural network, and using the reconstructed neural network to perform data fitting on the first multi - scale data set until the number of data fitting times after the reconstruction of the neural network reaches the preset adaptive number of times to obtain the neural network model.

[0041] After the initial construction of the neural network, based on the preset adaptive number of times, performing adaptive optimization and reconstruction on the neural network, as Figure 2 shown, the specific process is as follows: Use the multi-scale data samples in the first multi-scale dataset as the input of the neural network, and the observed data of the corresponding multi-scale data samples as the output of the neural network. Fit the first multi-scale dataset through the neural network constructed above. The loss function of the neural network is as follows: (4); Use the stochastic gradient descent algorithm to optimize the above loss function to obtain a preliminary fitting solution ; for example, calculate the gradient of the loss function with respect to all trainable parameters of the neural network using the differential formula, and then update the parameters by combining the gradient and the learning rate , for example , where represents the updated parameters, and is the gradient of the loss function loss with respect to all trainable parameters of the neural network .

[0042] When the number of iterations of the neural network reaches the preset number or the gradient of the loss function with respect to all trainable parameters of the neural network no longer decreases for several consecutive rounds, end the training of the neural network and obtain a preliminary fitting solution for the first multi-scale dataset , that is, the predicted output.

[0043] For the preliminary fitting solution obtained, a regular region (i.e., the solution rule region) can be determined, such as a rectangular region. If the region where the fitting solution is located is an irregular region, it is embedded in a rectangular region. Then, resample the data points (i.e., perform uniform sampling again in the fitting solution) on this regular region, and then construct a new dataset (i.e., the above-mentioned second multi-scale dataset) based on the multi-scale data samples corresponding to the sampled data points, and re-perform frequency feature extraction on this dataset to update the first frequency feature set , where , represents the nth frequency, and represents the number of frequency features of the newly constructed second multi-scale dataset. It should be noted that the frequency feature extraction process of the newly established dataset refers to the frequency feature extraction process of the above-mentioned first multi-scale dataset, which will not be repeated here.

[0044] Since the initial first multi-scale dataset may be generated by irregular sampling in a region, which may lead to a large error in the extracted frequency features. Subsequently, based on the fitting solution of the neural network, uniform sampling within the regular region can be realized to reduce the error of the frequency features extracted from the subsequent generated multi-scale datasets.

[0045] After that, according to the new frequency feature set , reconstruct the corresponding neural network. The specific example is as follows: (5); Similarly, use the multi-scale data samples in the first multi-scale data set as the input of the newly constructed neural network, and the observed data of the corresponding multi-scale data samples as the output of the newly constructed neural network. Fit the first multi-scale data set through the newly constructed neural network, and at the same time use the stochastic gradient descent algorithm to optimize the fitting of the multi-scale data set to obtain a better fitting solution .

[0046] Then, based on the newly obtained fitting solution , repeat the above process to obtain the resampled data points , and further construct the neural network and perform data fitting to obtain a better fitting solution . By analogy, the resampled data points and fitting solutions can be obtained until the number of times of constructing the neural network reaches the preset adaptive number of times. For example, if the adaptive number of times is set to 3, it means that the frequency features need to be extracted three times, and the neural network performs data fitting four times. One time is to construct the neural network and perform data fitting based on the frequency features of the initial data set to obtain a preliminary fitting solution. Three times are to resample the sample data based on the fitting solution obtained in the previous time, and reconstruct the neural network and perform data fitting based on the frequency features of the resampled data set, so as to realize the adaptive optimization of the neural network based on the frequency features of the multi-scale data set and obtain the final neural network model. For example ; Since the frequency features extracted adaptively each time can be used to reconstruct the neural network, the fitting effect of the neural network on the data set can be improved.

[0047] The neural network model can be used for prediction processing corresponding to the multi-scale data types to be fitted. For example, in medical image recognition, sample the neural network model (such as a convolutional neural network model) constructed based on the frequency adaptive mechanism above, and use this neural network model to perform hierarchical modeling on the low-frequency contour information and high-frequency texture details of the medical image to be recognized, and identify the lesion area in the medical image to improve the accuracy of image recognition; Another example is in remote sensing image analysis, sample the neural network model (such as a convolutional neural network model) constructed based on the frequency adaptive mechanism above, and use this neural network model to recognize the remote sensing image to be recognized, and identify the ground object boundaries and texture features of the remote sensing image at different scales, enhancing the robustness and resolution of target extraction.

[0048] Furthermore, a frequency selective filter or an attention mechanism can also be embedded in the finally constructed neural network model to achieve an adaptive response to the frequency components of the input multi-scale data to be recognized, effectively dealing with problems such as large differences in the spectral distribution of image content and multi-scale aliasing. Among them, introducing a frequency selective filter or an attention mechanism into a neural network belongs to the prior art and will not be elaborated here.

[0049] Exemplarily, as Figure 3 shown, taking the multi-scale data as image data as an example, a multi-scale data processing method provided by an embodiment of the present invention includes: S21: Construct a first multi-scale image data set according to the pre-collected image data samples; S22: Extract frequency features from the first multi-scale image data set, and construct a second frequency feature set according to the extracted frequency features of the first multi-scale image data set; S23: Construct a neural network according to the second frequency feature set; S24: Perform data fitting on the first multi-scale image data set through the neural network to perform frequency adaptive optimization on the neural network to obtain an image recognition model; S25: Process the image data to be processed by using the image recognition model to obtain an image recognition result.

[0050] In an alternative embodiment, the extracting frequency features from the first multi-scale image data set includes: Perform a discrete Fourier transform on the first multi-scale image data set to extract the frequency information of the first multi-scale image data set; Select the maximum frequency information from the frequency information and calculate a dynamic threshold; Select the frequency information whose absolute value of the frequency information is greater than the dynamic threshold from the frequency information, and use the frequency corresponding to the selected frequency information as the frequency feature of the first multi-scale image data set.

[0051] In an alternative embodiment, the constructing a neural network according to the second frequency feature set includes: Construct a neural network according to the number of frequency features in the second frequency feature set; wherein, the neural network includes a first sub-network, a second sub-network, and a third sub-network; the number of the first sub-network and the second sub-network is determined according to the number of frequency features in the second frequency feature set.

[0052] In an alternative embodiment, the performing data fitting on the first multi-scale image data set through the neural network to perform frequency adaptive optimization on the neural network to obtain an image recognition model includes: Use the image data samples in the first multi-scale image dataset as the input of the neural network, and the observed data of the corresponding image data samples as the output of the neural network to perform data fitting on the first multi-scale image dataset; During the data fitting process, use the stochastic gradient descent algorithm to optimize the loss function of the neural network to obtain the fitting solution output in this round; According to the fitting solution output in this round, determine the regular region to be solved, and uniformly sample the first multi-scale image dataset on the regular region to construct a second multi-scale image dataset; Extract the frequency features of the second multi-scale image dataset, and update the second frequency feature set according to the frequency features of the second multi-scale image dataset; According to the number of frequency features in the updated second frequency feature set, reconstruct the neural network, and use the reconstructed neural network to perform data fitting on the first multi-scale image dataset until the number of data fitting times after the reconstruction of the neural network reaches the preset adaptive number of times to obtain the image recognition model.

[0053] In an optional embodiment, the constructing the first multi-scale image dataset according to the pre-collected image data samples includes: Perform uniform distribution sampling or Latin hypercube sampling on the pre-collected image data samples; Construct the first multi-scale image dataset according to the sampled image data samples.

[0054] Exemplarily, as Figure 4 shown, abstract the first multi-scale image dataset constructed based on the image data samples into a multi-scale function: (6); Wherein, represents the coarse-scale feature of the first multi-scale image dataset, that is, the low-frequency component, such as the overall contour of the image data sample, represents the fine-scale feature of the first multi-scale image dataset, that is, the high-frequency component, such as the detailed texture of the image data sample. At this time, can be understood as the image data sample in the first multi-scale image dataset, represents the result after the first multi-scale image dataset is mapped by the multi-scale function.

[0055] Set the adaptive number of times to 3. For the initial first multi-scale image dataset, if it is constructed by uniformly sampling 100 and 300 image data samples from the interval and the interval respectively, that is, the first multi-scale image dataset and the corresponding observed data set Among them, 399. Since the data sampling at this time is obtained by irregular sampling in the irregular area 、 and is not uniform sampling in the regular area , when using the discrete Fourier transform for the first frequency feature extraction, it is impossible to accurately capture individual high-frequency information in the data set (such as ). As shown in the Fourier spectrum image Figure 5 , it represents the frequency feature image obtained by performing the discrete Fourier transform on the multi-scale function characterization of the first multi-scale image data set shown Figure 4 . After the discrete Fourier transform, the four frequency features can be extracted. After that, the network structure of the initial neural network is constructed as: (7); After that, use the neural network with the above structure to fit the initial first multi-scale image data set and the corresponding observed data set , and the corresponding loss function is: (8); Use the stochastic gradient descent algorithm to optimize the above loss function to obtain a preliminary fitting solution , as shown Figure 6 . Due to inaccurate frequency feature extraction caused by irregular sampling and the small number of sample data in , the fitting effect of the prediction solution at is poor, that is, the prediction effect is poor. At this time, based on the preliminary fitting solution uniformly sample in the regular area to construct a new multi-scale data set (i.e., the second multi-scale image data set), and perform frequency feature extraction again. As shown Figure 7 , it gives the frequency features of the data set constructed after uniformly sampling in the regular area at the preliminary fitting solution . For example, the frequency features extracted at this time are , which are consistent with the frequency features shown Figure 5 . At this time, the neural network can be constructed as: (9); Among them, . After that, use the neural network with this structure to fit the initial first multi-scale image data set and the corresponding observed data set again to obtain a better fitting solution , as Figure 8 shown. Figure 8 It is shown that the predicted output of the neural network fits well with the multi-scale function characterization of the first multi-scale dataset, and the data fitting effect is better than that of the neural network in the previous round , indicating the effectiveness of the embodiments of the present invention.

[0056] Repeat the above process, and continue to perform two similar frequency adaptations based on the fitting solution . The frequency features extracted from the newly established dataset are as Figure 9 shown, and the extracted frequency features are . The results of frequency feature extraction do not change much. Reconstruct the neural network and perform data fitting to construct the final neural network model.

[0057] Subsequently, the image to be recognized can be input into the finally established neural network model for processing to obtain the recognition result of the image to be recognized.

[0058] Compared with the prior art, the beneficial effects of the embodiments of the present invention are as follows: Construct and adaptively optimize the neural network based on the frequency features of the multi-scale dataset as a whole. On the one hand, it enables the network to dynamically focus on the modeling ability of different frequency components in different training stages, which can not only improve the expression ability of the neural network model on multi-scale data, improve the accuracy of data fitting and the model precision of the neural network when processing multi-scale data, but also has a positive significance for improving the convergence speed of the model and the final precision. On the other hand, each time the frequency features of the whole dataset are calculated, the calculation cost is relatively low, which can effectively reduce the calculation amount, improve the data processing efficiency, thereby reducing the operation pressure of the terminal processor and improving the operation speed.

[0059] Reconstruct the neural network and perform data fitting based on the frequency features of the resampled dataset, which can realize the adaptive optimization of the neural network based on the frequency features of the multi-scale dataset. Since the frequency features extracted each time for adaptation can be used to reconstruct the neural network, it can solve the problem that traditional neural networks are difficult to effectively capture high-frequency details, improve the expression ability of the neural network on multi-scale data, and improve the fitting effect and generalization ability of the network.

[0060] After obtaining the initial fitting solution, perform uniform sampling in the regular region based on the fitting solution, and then further perform frequency feature extraction, which overcomes the error of the discrete Fourier transform caused by the original data from irregular sampling.

[0061] See Figure 10 , Figure 10The following is a structural block diagram of a multi-scale data processing device provided by an embodiment of the present invention. The multi-scale data processing device includes: A first multi-scale data set construction module 11, configured to construct a first multi-scale data set according to pre-collected multi-scale data samples; A first frequency feature extraction module 12, configured to perform frequency feature extraction on the first multi-scale data set, and construct a first frequency feature set according to the extracted frequency features of the first multi-scale data set; A first neural network construction module 13, configured to construct a neural network according to the first frequency feature set; A first network adaptive optimization module 14, configured to perform data fitting on the first multi-scale data set through the neural network to perform frequency adaptive optimization on the neural network and obtain a neural network model; A first data processing module 15, configured to process the multi-scale data to be processed by using the neural network model and obtain a corresponding data processing result.

[0062] In an optional embodiment, the first frequency feature extraction module 12 includes: A first discrete Fourier transform unit, configured to perform a discrete Fourier transform on the first multi-scale data set and extract the frequency information of the first multi-scale data set; A first threshold calculation unit, configured to select the maximum frequency information from the frequency information and calculate a dynamic threshold; A first frequency screening unit, configured to select the frequency information whose absolute value of the frequency information is greater than the frequency information corresponding to the dynamic threshold from the frequency information, and use the frequency corresponding to the selected frequency information as the frequency feature of the first multi-scale data set.

[0063] In an optional embodiment, the first neural network construction module 13 includes: A first neural network construction unit, configured to construct a neural network according to the number of frequency features in the first frequency feature set; wherein, the neural network includes a first sub-network, a second sub-network, and a third sub-network; the numbers of the first sub-network and the second sub-network are determined according to the number of frequency features in the first frequency feature set.

[0064] In an optional embodiment, the first network adaptive optimization module 14 includes: A first data fitting unit, configured to use the multi-scale data samples in the first multi-scale data set as the input of the neural network and the observed data of the corresponding multi-scale data samples as the output of the neural network to perform data fitting on the first multi-scale data set; The first loss optimization unit is used to optimize the loss function of the neural network by using the stochastic gradient descent algorithm during the data fitting process, and obtain the fitting solution output in this round. The first data sampling unit is used to determine the regular area to be solved according to the fitting solution output in this round, and uniformly sample the first multi-scale data set on the regular area to construct a second multi-scale data set. The first frequency feature extraction unit is used to extract the frequency features of the second multi-scale data set, and update the first frequency feature set according to the frequency features of the second multi-scale data set. The second neural network construction unit is used to reconstruct the neural network according to the number of frequency features in the updated first frequency feature set, and perform data fitting on the first multi-scale data set by using the reconstructed neural network until the number of data fitting times after the reconstruction of the neural network reaches the preset adaptive number of times, so as to obtain the neural network model.

[0065] In an optional embodiment, the first multi-scale data set construction module 11 includes: The second data sampling unit is used to perform uniform distribution sampling or Latin hypercube sampling on the pre-collected multi-scale data samples. The first data set construction unit is used to construct a first multi-scale data set according to the sampled multi-scale data samples.

[0066] It should be noted that the working processes of the various modules in the multi-scale data processing device described in the embodiments of the present invention can refer to the working processes of the multi-scale data processing method described in the above embodiments, and the technical effects achieved by them are also the same as those of the multi-scale data processing method described in the above embodiments, which will not be elaborated here.

[0067] See Figure 11 , Figure 11 FIG. is another structural block diagram of a multi-scale data processing device provided by an embodiment of the present invention. The multi-scale data processing device includes: The second multi-scale data set construction module 21 is used to construct a first multi-scale image data set according to the pre-collected image data samples. The second frequency feature extraction module 22 is used to extract the frequency features of the first multi-scale image data set, and construct a second frequency feature set according to the extracted frequency features of the first multi-scale image data set. The second neural network construction module 23 is used to construct a neural network according to the second frequency feature set. The second network adaptive optimization module 24 is used to perform data fitting on the first multi-scale image data set through the neural network to perform frequency adaptive optimization on the neural network, so as to obtain an image recognition model. The second data processing module 25 is configured to process the image data to be processed by using the image recognition model to obtain an image recognition result.

[0068] In an optional embodiment, the second frequency feature extraction module 22 includes: A second discrete Fourier transform unit, configured to perform a discrete Fourier transform on the first multi-scale image data set to extract frequency information of the first multi-scale image data set; A second threshold calculation unit, configured to select the maximum frequency information from the frequency information and calculate a dynamic threshold; A second frequency screening unit, configured to select frequency information whose absolute value is greater than the frequency information corresponding to the dynamic threshold from the frequency information, and use the frequency corresponding to the selected frequency information as the frequency feature of the first multi-scale data set.

[0069] In an optional embodiment, the second neural network construction module 23 includes: A third neural network construction unit, configured to construct a neural network according to the number of frequency features in the second frequency feature set; wherein, the neural network includes a first sub-network, a second sub-network, and a third sub-network; the number of the first sub-network and the second sub-network is determined according to the number of frequency features in the second frequency feature set.

[0070] In an optional embodiment, the second network adaptive optimization module 24 includes: A second data fitting unit, configured to use the image data samples in the first multi-scale image data set as the input of the neural network, and the observed data of the corresponding image data samples as the output of the neural network to perform data fitting on the first multi-scale image data set; A second loss optimization unit, configured to optimize the loss function of the neural network by using a stochastic gradient descent algorithm during the data fitting process to obtain a fitting solution output in this round; A third data sampling unit, configured to determine a solution rule region according to the fitting solution output in this round, and uniformly sample the first multi-scale image data set on the rule region to construct a second multi-scale image data set; A second frequency feature extraction unit, configured to extract the frequency features of the second multi-scale image data set, and update the second frequency feature set according to the frequency features of the second multi-scale image data set; A fourth neural network construction unit, configured to reconstruct a neural network according to the number of frequency features in the updated second frequency feature set, and perform data fitting on the first multi-scale image data set by using the reconstructed neural network until the number of data fitting times after the reconstruction of the neural network reaches a preset adaptive number of times, so as to obtain the image recognition model.

[0071] In an optional embodiment, the second multi-scale data set construction module 21 includes: A fourth data sampling unit, configured to perform uniform distribution sampling or Latin hypercube sampling on pre-collected image data samples; A second data set construction unit, configured to construct a first multi-scale image data set according to the sampled image data samples.

[0072] It should be noted that the working processes of the various modules in the multi-scale data processing device described in the embodiments of the present invention may refer to the working processes of the multi-scale data processing method described in the above embodiments, and the technical effects achieved by them are also the same as those of the multi-scale data processing method described in the above embodiments, which will not be elaborated here.

[0073] See Figure 12 , Figure 12 is a structural block diagram of a multi-scale data processing device provided by an embodiment of the present invention. The multi-scale data processing device includes a processor 31, a memory 32, and a computer program stored in the memory 32 and executable on the processor 31. When the processor 31 executes the computer program, the steps in the above various embodiments of the multi-scale data processing method are implemented, such as steps S11 to S15 or steps S21 to S25.

[0074] Exemplarily, the computer program may be divided into one or more modules / units, and the one or more modules / units are stored in the memory 32 and executed by the processor 31 to complete the present invention. The one or more modules / units may be a series of computer program instruction segments capable of performing specific functions, and the instruction segments are used to describe the execution process of the computer program in the multi-scale data processing device.

[0075] The multi-scale data processing device may include, but is not limited to, a processor 31 and a memory 32. Those skilled in the art can understand that the schematic diagram is only an example of the multi-scale data processing device, and does not constitute a limitation on the multi-scale data processing device. It may include more or fewer components than shown in the figure, or combine some components, or different components. For example, the multi-scale data processing device may further include input / output devices, network access devices, buses, etc.

[0076] The processor 31 may be a Central Processing Unit (CPU), or may also be other general-purpose processors, Digital Signal Processors (DSPs), Application Specific Integrated Circuits (ASICs), Field-Programmable Gate Arrays (FPGAs), 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. The processor 31 is the control center of the multi-scale data processing device, and connects various parts of the entire multi-scale data processing device through various interfaces and lines.

[0077] The memory 32 can be used to store the computer programs and / or modules. The processor 31 realizes various functions of the multi-scale data processing device by running or executing the computer programs and / or modules stored in the memory 32, and by calling the data stored in the memory 32. The memory 32 may mainly include a program storage area and a data storage area. Among them, the program storage area may store an operating system, application programs required for at least one function (such as a sound playback function, an image playback function, etc.); the data storage area may store data created according to the use of the mobile phone (such as audio data, phone book, etc.). In addition, the memory 32 may include high-speed random access memory, and may also include non-volatile memory, such as a hard disk, memory, plug-in hard disk, Smart Media Card (SMC), Secure Digital (SD) card, Flash Card, at least one magnetic disk storage device, flash device, or other volatile solid-state storage devices.

[0078] Among them, if the modules / units integrated in the multi-scale data processing device are implemented in the form of software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on such an understanding, to implement all or part of the processes in the above-described embodiment methods of the present invention, it can also be completed by instructing relevant hardware through a computer program. The computer program can be stored in a computer-readable storage medium. When the computer program is executed by the processor 31, the steps of the above-described method embodiments can be implemented. Among them, the computer program includes computer program code, and the computer program code can be in the form of source code, object code, executable file, or some intermediate form, etc. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording medium, USB flash drive, mobile hard disk, magnetic disk, optical disc, computer memory, read-only memory (ROM, Read-Only Memory), random access memory (RAM, Random Access Memory), electrical carrier signal, telecommunication signal, and software distribution medium, etc.

[0079] It should be noted that the device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separated. The components shown as units may or may not be physical units, that is, they may be located in one place or distributed to multiple network units. Some or all of the modules can be selected according to actual needs to achieve the purpose of the solution of this embodiment. In addition, in the attached drawings of the device embodiments provided by the present invention, the connection relationship between the modules indicates that there is a communication connection between them, which can be specifically implemented as one or more communication buses or signal lines. Those of ordinary skill in the art can understand and implement it without creative efforts.

[0080] The above is the preferred implementation manner of the present invention. It should be pointed out that for those of ordinary skill in the art in this technical field, without departing from the principle of the present invention, various improvements and refinements can be made, and these improvements and refinements are also regarded as the protection scope of the present invention.

Claims

1. A multi-scale data processing method, characterized in that, including: Construct a first multi-scale data set according to pre-collected multi-scale data samples; Extract frequency features from the first multi-scale data set, and construct a first frequency feature set according to the extracted frequency features of the first multi-scale data set; Construct a neural network according to the first frequency feature set; Perform data fitting on the first multi-scale data set through the neural network to perform frequency adaptive optimization on the neural network and obtain a neural network model; Process the multi-scale data to be processed using the neural network model to obtain corresponding data processing results.

2. The multi-scale data processing method according to claim 1, characterized in that, The extracting frequency features from the first multi-scale data set includes: Perform discrete Fourier transform on the first multi-scale data set to extract the frequency information of the first multi-scale data set; Select the maximum frequency information from the frequency information and calculate the dynamic threshold; Select the frequency information whose absolute value of the frequency information is greater than the dynamic threshold from the frequency information, and use the frequency corresponding to the selected frequency information as the frequency feature of the first multi-scale data set.

3. The multi-scale data processing method according to claim 1, wherein The constructing a neural network according to the first frequency feature set includes: Construct a neural network according to the number of frequency features in the first frequency feature set; wherein, the neural network includes a first sub-network, a second sub-network, and a third sub-network; the number of the first sub-network and the second sub-network is determined according to the number of frequency features in the first frequency feature set.

4. The multi-scale data processing method according to claim 3, wherein The performing data fitting on the first multi-scale data set through the neural network to perform frequency adaptive optimization on the neural network and obtain a neural network model includes: Use the multi-scale data samples in the first multi-scale data set as the input of the neural network, and the observed data of the corresponding multi-scale data samples as the output of the neural network to perform data fitting on the first multi-scale data set; During the data fitting process, use the stochastic gradient descent algorithm to optimize the loss function of the neural network to obtain the fitting solution output in this round; Determine the regular region to be solved according to the fitting solution output in this round, and uniformly sample the first multi-scale data set on the regular region to construct a second multi-scale data set; Extract the frequency features of the second multi-scale data set, and update the first frequency feature set according to the frequency features of the second multi-scale data set; Re-construct the neural network according to the number of frequency features in the updated first frequency feature set, and use the re-constructed neural network to perform data fitting on the first multi-scale data set until the number of data fitting times after the reconstruction of the neural network reaches the preset adaptive number of times to obtain the neural network model.

5. The multi-scale data processing method according to claim 1, wherein The constructing a first multi-scale data set according to pre-collected multi-scale data samples includes: Perform uniform distribution sampling or Latin hypercube sampling on the pre-collected multi-scale data samples; Construct a first multi-scale data set according to the sampled multi-scale data samples.

6. A multi-scale data processing method, characterized in that, including: Construct a first multi-scale image data set according to pre-collected image data samples; Extract frequency features from the first multi-scale image dataset, and construct a second frequency feature set according to the extracted frequency features of the first multi-scale image dataset; Construct a neural network according to the second frequency feature set; Perform data fitting on the first multi-scale image dataset through the neural network to perform frequency adaptive optimization on the neural network to obtain an image recognition model; Process the image data to be processed using the image recognition model to obtain an image recognition result.

7. The multi-scale data processing method according to claim 6, wherein The extracting frequency features from the first multi-scale image dataset includes: Perform discrete Fourier transform on the first multi-scale image dataset to extract the frequency information of the first multi-scale image dataset; Select the maximum frequency information from the frequency information and calculate the dynamic threshold; Select the frequency information whose absolute value of the frequency information is greater than the frequency information corresponding to the dynamic threshold from the frequency information, and use the frequency corresponding to the selected frequency information as the frequency feature of the first multi-scale image dataset.

8. The multi-scale data processing method according to claim 6, characterized in that The constructing a neural network according to the second frequency feature set includes: Construct a neural network according to the number of frequency features in the second frequency feature set; wherein, the neural network includes a first sub-network, a second sub-network, and a third sub-network; the number of the first sub-network and the second sub-network is determined according to the number of frequency features in the second frequency feature set.

9. The multi-scale data processing method according to claim 8, characterized in that, The performing data fitting on the first multi-scale image dataset through the neural network to perform frequency adaptive optimization on the neural network to obtain an image recognition model includes: Use the image data samples in the first multi-scale image dataset as the input of the neural network, and the observed data of the corresponding image data samples as the output of the neural network to perform data fitting on the first multi-scale image dataset; During the data fitting process, use the stochastic gradient descent algorithm to optimize the loss function of the neural network to obtain the fitting solution output in this round; Determine the rule region to be solved according to the fitting solution output in this round, and uniformly sample the first multi-scale image dataset on the rule region to construct a second multi-scale image dataset; Extract the frequency features of the second multi-scale image dataset, and update the second frequency feature set according to the frequency features of the second multi-scale image dataset; Reconstruct the neural network according to the number of frequency features in the updated second frequency feature set, and use the reconstructed neural network to perform data fitting on the first multi-scale image dataset until the number of data fitting times after the reconstruction of the neural network reaches the preset adaptive number of times to obtain the image recognition model.

10. The multi-scale data processing method according to claim 6, wherein The constructing a first multi-scale image dataset according to the pre-collected image data samples includes: Perform uniform distribution sampling or Latin hypercube sampling on the pre-collected image data samples; Construct a first multi-scale image dataset according to the sampled image data samples.

11. A multi-scale data processing device, characterized in that, Includes: A first multi-scale dataset construction module, configured to construct a first multi-scale dataset according to pre-collected multi-scale data samples; A first frequency feature extraction module, configured to extract frequency features from the first multi-scale data set, and construct a first frequency feature set according to the extracted frequency features of the first multi-scale data set; A first neural network construction module, configured to construct a neural network according to the first frequency feature set; A first network adaptive optimization module, configured to perform data fitting on the first multi-scale data set through the neural network to perform frequency adaptive optimization on the neural network, and obtain a neural network model; A first data processing module, configured to process the multi-scale data to be processed by using the neural network model, and obtain a corresponding data processing result.

12. A multi-scale data processing device, characterized in that, Including: A second multi-scale data set construction module, configured to construct a first multi-scale image data set according to pre-collected image data samples; A second frequency feature extraction module, configured to extract frequency features from the first multi-scale image data set, and construct a second frequency feature set according to the extracted frequency features of the first multi-scale image data set; A second neural network construction module, configured to construct a neural network according to the second frequency feature set; A second network adaptive optimization module, configured to perform data fitting on the first multi-scale image data set through the neural network to perform frequency adaptive optimization on the neural network, and obtain an image recognition model; A second data processing module, configured to process the image data to be processed by using the image recognition model, and obtain an image recognition result.

13. A multi-scale data processing device, characterized in that, Including: A processor, a memory, and a computer program stored in the memory and configured to be executed by the processor. When the processor executes the computer program, the multi-scale data processing method according to any one of claims 1 to 5 or the multi-scale data processing method according to any one of claims 6 to 10 is implemented.

14. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, wherein when the computer program runs, it controls the device where the computer-readable storage medium is located to execute the multi-scale data processing method according to any one of claims 1 to 5 or the multi-scale data processing method according to any one of claims 6 to 10.

15. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, the multi-scale data processing method according to any one of claims 1 to 5 or the multi-scale data processing method according to any one of claims 6 to 10 is implemented.

Images