Digital core super-resolution training set generation method and device, equipment and medium

Abstract

Embodiments of the present application provide a digital core super-resolution training set generation method, device, equipment and medium. The present application relates to the field of digital core, and the method comprises: obtaining real digital core image samples, processing the resolution of the real digital core image samples based on a plurality of image interpolation algorithms to obtain synthetic digital core image samples; calculating the synthesis quality evaluation value of each image interpolation algorithm according to a plurality of real digital core image samples and a plurality of synthetic digital core image samples; determining a target image interpolation algorithm according to the synthesis quality evaluation value of each image interpolation algorithm; and processing the resolution of each real digital core image based on the target image interpolation algorithm according to a plurality of real digital core images to obtain a digital core super-resolution training set. The method is used to achieve the technical effect of improving the accuracy of generating a digital core super-resolution training set.

Description

Technical Field

[0001] This application relates to the field of digital core technology, and in particular to a method, apparatus, equipment and medium for generating digital core super-resolution training sets. Background Technology

[0002] In the training of deep learning-based digital core image super-resolution models, constructing a high-quality low-resolution-high-resolution paired image dataset is a key prerequisite.

[0003] In existing technologies, traditional image processing algorithms such as bicubic interpolation are typically used to manually downsample real high-resolution core images to synthesize corresponding low-resolution training samples. However, existing technologies are inaccurate in processing high-resolution images, leading to performance degradation and insufficient generalization ability of the trained super-resolution models in practical applications. Therefore, existing technologies suffer from the technical problem of poor accuracy in generating digital core super-resolution training sets. Summary of the Invention

[0004] This application provides a method, apparatus, device, and medium for generating digital core super-resolution training sets, which aims to improve the accuracy of generating digital core super-resolution training sets.

[0005] In a first aspect, embodiments of this application provide a method for generating a digital core super-resolution training set, including:

[0006] Real digital core image samples were obtained, and the resolution of the real digital core image samples was processed based on multiple image interpolation algorithms to obtain synthetic digital core image samples.

[0007] Based on multiple real digital core image samples and multiple synthetic digital core image samples, the synthesis quality evaluation value of each image interpolation algorithm is calculated;

[0008] The target image interpolation algorithm is determined based on the synthesis quality evaluation value of each image interpolation algorithm;

[0009] Based on multiple real digital core images, the resolution of each real digital core image is processed according to the target image interpolation algorithm to obtain a digital core super-resolution training set.

[0010] In one possible implementation, real digital core image samples are acquired, and the resolution of the real digital core image samples is processed based on multiple image interpolation algorithms to obtain synthetic digital core image samples, including:

[0011] Obtain the resolution type of real digital core image samples; among which, the resolution type includes high resolution and low resolution, and low resolution includes 2x resolution, 4x resolution, 8x resolution and 16x resolution;

[0012] If the resolution of the real digital core image sample is high, then based on each image interpolation algorithm, the resolution of the real digital core image sample is upsampled to obtain the synthetic digital core image sample; wherein, the synthetic digital core image sample is a low-resolution synthetic digital core image sample.

[0013] If the resolution type of the real digital core image sample is low resolution, then based on each image interpolation algorithm, the resolution of the real digital core image sample is downsampled to obtain the synthetic digital core image sample; wherein, the synthetic digital core image sample is a high-resolution synthetic digital core image sample.

[0014] In one possible implementation, based on multiple real digital core image samples and multiple synthetic digital core image samples, a synthetic quality evaluation value for each image interpolation algorithm is calculated, including:

[0015] Based on multiple real digital core image samples and multiple synthetic digital core image samples, a performance evaluation value for each image interpolation algorithm is obtained; the performance evaluation value is used to characterize the slope information of the performance of each image interpolation algorithm as a function of resolution.

[0016] Based on multiple real digital core image samples and multiple synthetic digital core image samples, a fidelity evaluation value is obtained for each image interpolation algorithm; the fidelity evaluation value is used to characterize the frequency domain energy distribution between the synthetic digital core image sample generated by each image interpolation algorithm and the corresponding real digital core image sample.

[0017] Based on the performance evaluation value and the fidelity evaluation value, the synthesis quality evaluation value of each image interpolation algorithm is obtained.

[0018] In one possible implementation, based on multiple real digital core image samples and multiple synthetic digital core image samples, a performance evaluation value for each image interpolation algorithm is obtained, including:

[0019] The peak signal-to-noise ratio and structural similarity index between each real digital core image sample and the corresponding synthetic digital core image sample were obtained.

[0020] Based on multiple peak signal-to-noise ratios and multiple structural similarity indices, the curves of peak signal-to-noise ratio and structural similarity index as a function of resolution for each image interpolation algorithm are obtained;

[0021] Based on the curves of peak signal-to-noise ratio and structural similarity index as a function of resolution, the slope information of the performance of each image interpolation algorithm as a function of resolution is obtained.

[0022] Based on the mapping relationship and slope information, the performance evaluation value of each image interpolation algorithm is obtained; where the mapping relationship is the correspondence between slope information and performance evaluation value.

[0023] In one possible implementation, based on multiple real digital core image samples and multiple synthetic digital core image samples, a fidelity evaluation value for each image interpolation algorithm is obtained, including:

[0024] Based on Fourier transform, each real digital core image sample and the corresponding synthetic digital core image sample are converted into the frequency domain to obtain the real frequency domain energy distribution map corresponding to each real digital core image sample and the synthetic frequency domain energy distribution map corresponding to each synthetic digital core image sample.

[0025] Similarity information is obtained based on each real frequency domain energy distribution map and its corresponding synthetic frequency domain energy distribution map;

[0026] Based on multiple similarity information, the fidelity evaluation value of each image interpolation algorithm is obtained.

[0027] In one possible implementation, after obtaining the digital core super-resolution training set, the method further includes:

[0028] Based on multiple real digital core images, the resolution of each real digital core image is processed according to the comparison image interpolation algorithm to obtain a digital core super-resolution comparison training set.

[0029] The digital core super-resolution training set and the digital core super-resolution control training set are respectively input into the super-resolution network to obtain the reconstructed high-resolution image set and the reconstructed high-resolution image control set.

[0030] Calculate the first similarity information between the reconstructed high-resolution image set and the original real low-resolution image set;

[0031] Calculate the second similarity information between the reconstructed high-resolution image reference set and the original real low-resolution image set;

[0032] Calculate the difference between the first similarity information and the second similarity information;

[0033] Determine whether the difference meets the preset threshold. If so, then determine the digital core super-resolution training set as the target digital core super-resolution training set.

[0034] If not, a new target image interpolation algorithm is determined until the new difference meets the preset threshold, and the generated new digital core super-resolution training set is determined as the target digital core super-resolution training set.

[0035] Secondly, embodiments of this application provide a digital core super-resolution training set generation device, comprising:

[0036] The first processing module is used to acquire real digital core image samples, and based on multiple image interpolation algorithms, process the resolution of the real digital core image samples to obtain synthetic digital core image samples.

[0037] The second processing module is used to calculate the synthesis quality evaluation value of each image interpolation algorithm based on multiple real digital core image samples and multiple synthetic digital core image samples.

[0038] The third processing module is used to determine the target image interpolation algorithm based on the synthesis quality evaluation value of each image interpolation algorithm;

[0039] The fourth processing module is used to process the resolution of each real digital core image based on multiple real digital core images according to the target image interpolation algorithm, so as to obtain a digital core super-resolution training set.

[0040] In one possible implementation, the first processing module is further configured to:

[0041] Obtain the resolution type of real digital core image samples; among which, the resolution type includes high resolution and low resolution, and low resolution includes 2x resolution, 4x resolution, 8x resolution and 16x resolution;

[0042] If the resolution of the real digital core image sample is high, then based on each image interpolation algorithm, the resolution of the real digital core image sample is upsampled to obtain the synthetic digital core image sample; wherein, the synthetic digital core image sample is a low-resolution synthetic digital core image sample.

[0043] If the resolution type of the real digital core image sample is low resolution, then based on each image interpolation algorithm, the resolution of the real digital core image sample is downsampled to obtain the synthetic digital core image sample; wherein, the synthetic digital core image sample is a high-resolution synthetic digital core image sample.

[0044] In one possible implementation, the second processing module is further configured to:

[0045] Based on multiple real digital core image samples and multiple synthetic digital core image samples, a performance evaluation value for each image interpolation algorithm is obtained; the performance evaluation value is used to characterize the slope information of the performance of each image interpolation algorithm as a function of resolution.

[0046] Based on multiple real digital core image samples and multiple synthetic digital core image samples, a fidelity evaluation value is obtained for each image interpolation algorithm; the fidelity evaluation value is used to characterize the frequency domain energy distribution between the synthetic digital core image sample generated by each image interpolation algorithm and the corresponding real digital core image sample.

[0047] Based on the performance evaluation value and the fidelity evaluation value, the synthesis quality evaluation value of each image interpolation algorithm is obtained.

[0048] In one possible implementation, the second processing module is further configured to:

[0049] The peak signal-to-noise ratio and structural similarity index between each real digital core image sample and the corresponding synthetic digital core image sample were obtained.

[0050] Based on multiple peak signal-to-noise ratios and multiple structural similarity indices, the curves of peak signal-to-noise ratio and structural similarity index as a function of resolution for each image interpolation algorithm are obtained;

[0051] Based on the curves of peak signal-to-noise ratio and structural similarity index as a function of resolution, the slope information of the performance of each image interpolation algorithm as a function of resolution is obtained.

[0052] Based on the mapping relationship and slope information, the performance evaluation value of each image interpolation algorithm is obtained; where the mapping relationship is the correspondence between slope information and performance evaluation value.

[0053] In one possible implementation, the second processing module is further configured to:

[0054] Based on Fourier transform, each real digital core image sample and the corresponding synthetic digital core image sample are converted into the frequency domain to obtain the real frequency domain energy distribution map corresponding to each real digital core image sample and the synthetic frequency domain energy distribution map corresponding to each synthetic digital core image sample.

[0055] Similarity information is obtained based on each real frequency domain energy distribution map and its corresponding synthetic frequency domain energy distribution map;

[0056] Based on multiple similarity information, the fidelity evaluation value of each image interpolation algorithm is obtained.

[0057] In one possible implementation, the fourth processing module is further configured to:

[0058] Based on multiple real digital core images, the resolution of each real digital core image is processed according to the comparison image interpolation algorithm to obtain a digital core super-resolution comparison training set.

[0059] The digital core super-resolution training set and the digital core super-resolution control training set are respectively input into the super-resolution network to obtain the reconstructed high-resolution image set and the reconstructed high-resolution image control set.

[0060] Calculate the first similarity information between the reconstructed high-resolution image set and the original real low-resolution image set;

[0061] Calculate the second similarity information between the reconstructed high-resolution image reference set and the original real low-resolution image set;

[0062] Calculate the difference between the first similarity information and the second similarity information;

[0063] Determine whether the difference meets the preset threshold. If so, then determine the digital core super-resolution training set as the target digital core super-resolution training set.

[0064] If not, a new target image interpolation algorithm is determined until the new difference meets the preset threshold, and the generated new digital core super-resolution training set is determined as the target digital core super-resolution training set.

[0065] Thirdly, embodiments of this application provide an electronic device, including: a memory and a processor;

[0066] The memory stores computer-executed instructions;

[0067] The processor executes computer execution instructions stored in the memory, causing the processor to perform the first aspect and / or various possible implementations of the first aspect as described above.

[0068] Fourthly, embodiments of this application provide a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, are used to implement the first aspect and / or various possible implementations of the first aspect.

[0069] Fifthly, embodiments of this application provide a computer program product, including a computer program that, when executed by a processor, implements the first aspect and / or various possible implementations of the first aspect.

[0070] The digital core super-resolution training set generation method, apparatus, device, and medium provided in this application acquire real digital core image samples and use various image interpolation algorithms to process these samples to generate corresponding synthetic digital core image samples. Subsequently, by comparing and analyzing the real digital core image samples with the synthetic digital core image samples generated by each algorithm, a synthetic quality evaluation value corresponding to each image interpolation algorithm is calculated. Based on these evaluation values, the optimal target image interpolation algorithm can be objectively selected. Finally, the target algorithm is used to uniformly process multiple real digital core images, thereby constructing a high-quality digital core super-resolution training set. This overcomes the systematic bias that may be introduced by a single fixed interpolation algorithm in traditional methods, significantly improves the consistency between synthetic low-resolution images and the real imaging degradation process, and the constructed training set more closely resembles the data distribution of low-resolution images in actual geological scanning scenarios, enabling the super-resolution model trained based on this training set to more accurately learn the mapping relationship from low resolution to high resolution. Attached Figure Description

[0071] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0072] Figure 1 A flowchart illustrating the digital core super-resolution training set generation method provided in this application. Figure 1 ;

[0073] Figure 2 A flowchart illustrating the digital core super-resolution training set generation method provided in this application. Figure 2 ;

[0074] Figure 3 A comparison diagram of synthetic digital core image samples and real digital core image samples for the digital core super-resolution training set generation method provided in this application. Figure 1 ;

[0075] Figure 4 A comparison diagram of synthetic digital core image samples and real digital core image samples for the digital core super-resolution training set generation method provided in this application. Figure 2 ;

[0076] Figure 5 A schematic diagram of the digital core super-resolution training set generation device provided in this application;

[0077] Figure 6 A hardware schematic diagram of the digital core super-resolution training set generation device provided in this application.

[0078] The accompanying drawings illustrate specific embodiments of this application, which will be described in more detail below. These drawings and descriptions are not intended to limit the scope of the concept in any way, but rather to illustrate the concept of this application to those skilled in the art through reference to particular embodiments. Detailed Implementation

[0079] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application. Rather, they are merely examples of methods and approaches consistent with some aspects of this application as detailed in the appended claims.

[0080] In the field of digital core image super-resolution technology based on deep learning, constructing a low-resolution-high-resolution paired training set that accurately reflects the real imaging degradation process is a crucial prerequisite for ensuring model performance. Current mainstream methods typically employ pre-defined traditional image processing algorithms, particularly bicubic interpolation, to downsample real high-resolution core images by a fixed factor, thereby artificially synthesizing low-resolution training samples. While this single-algorithm application mode is simple and easy to implement, it inherently implies an oversimplification of the complex imaging degradation process.

[0081] The core problem with existing technologies lies in their inability to accurately simulate the multi-layered, nonlinear degradation mechanisms present in real core imaging systems. Core imaging equipment such as micro-CT scanners and focused ion beam scanning electron microscopes are subject to a combination of factors in actual operation, including inherent hardware characteristics, scanning parameter settings, differences in sample preparation, and environmental noise. Their degradation process involves complex characteristics such as spatial blurring caused by the device's specific point spread function, composite noise distribution formed by the superposition of electronic and quantum noise, and edge response variations due to sample heterogeneity. The idealized linear degradation model constructed by traditional interpolation algorithms deviates significantly from this actual physical process, resulting in systematic differences between synthesized low-resolution images and real low-resolution images in terms of texture detail preservation, edge sharpness maintenance, and noise statistical characteristics.

[0082] This mismatch between training data and real data distribution directly leads to a series of technical problems. First, the super-resolution model built on such training sets establishes an incorrect degradation-reconstruction mapping relationship during the learning process, resulting in significant performance degradation when processing real low-resolution core images. This manifests as excessive smoothing of reconstructed images, morphological distortion of micropore structures, and artifact interference at mineral boundaries. More seriously, this data bias fundamentally restricts the model's generalization ability. When applied to core images acquired from different geological strata, different imaging devices, or different scanning parameters, the model's reconstruction quality fluctuates significantly, and it may even fail completely. This limitation not only affects the reliable application of super-resolution technology in geological research but may also lead to misleading conclusions in subsequent key geological engineering applications such as quantitative analysis of pore structure and seepage simulation prediction due to the propagation of reconstruction errors. This weakens the scientific value and engineering guidance significance of digital core technology in actual oil and gas exploration and development.

[0083] This invention aims to address the technical problem in existing technologies where the use of a single fixed interpolation algorithm (such as bicubic interpolation) to generate digital core super-resolution training sets leads to a mismatch between the synthesized data and the actual imaging degradation process, resulting in decreased performance and insufficient generalization ability of the super-resolution model in practical applications. The solution proposed in this application is to construct an intelligent training set generation framework that features dynamic optimization, multi-dimensional evaluation, and iterative verification. Its core is to use a systematic algorithm evaluation and screening mechanism to match the image interpolation algorithm that best matches the actual degradation characteristics of specific core image data, thereby generating training data with higher fidelity and stronger physical meaning.

[0084] Specifically, this method first breaks through the limitations of traditional single algorithms by introducing multiple image interpolation algorithms to form a candidate algorithm pool, and performs bidirectional processing on real digital core image samples: for high-resolution samples, upsampling is performed using each algorithm to simulate low-resolution images; for low-resolution samples, downsampling is performed to construct corresponding high-resolution references. This bidirectional processing mechanism ensures the comprehensiveness of algorithm evaluation, and can simultaneously test the consistency of algorithm behavior in both resolution increases and decreases. Subsequently, this application proposes a comprehensive evaluation system for synthetic quality, which not only focuses on conventional spatial domain pixel-level similarity, but also reveals in depth the law of algorithm performance change with resolution and its energy preservation ability in the frequency domain. By calculating the slope of the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) of each algorithm as a function of resolution, the smoothness of its performance degradation is evaluated; the smaller the slope, the more stable the algorithm is at different scales. At the same time, Fourier transform is used to compare the frequency domain energy distribution similarity between real and synthetic images, evaluating the algorithm's ability to preserve the image texture structure and detail spectrum from the frequency domain perspective. By combining these two evaluation values, we can objectively select the target image interpolation algorithm that performs best in both multi-scale stability and frequency domain feature fidelity.

[0085] To ensure the effectiveness of the training set generated by the selected algorithm, this application further designs an iterative verification and optimization mechanism. After generating an initial training set using the target algorithm, it is not used directly. Instead, it and a control training set generated by a traditional algorithm (as a control) are input into the same super-resolution network for training and reconstruction. By comparing the similarity difference between the reconstruction results of the two algorithms and the original real images, and determining whether the difference meets a preset threshold, the superiority of the target algorithm is verified to be reflected in the actual model performance. If it does not meet the threshold, an iterative optimization process is initiated to re-evaluate and determine a new target algorithm until a training set that can significantly improve the model's reconstruction fidelity is generated. This closed-loop verification process elevates the quality evaluation of the training set from "image similarity at the algorithm level" to "reconstruction performance gain at the model level," ensuring the practical value of the generated training set.

[0086] In summary, this application's solution constructs a systematic, rigorous, and adaptive training set generation methodology through multiple algorithm candidates, bidirectional processing verification, multi-dimensional quality assessment in the spatial and frequency domains, and iterative optimization based on model reconstruction results. It no longer relies on a single, fixed prior degradation assumption, but instead uses a data-driven approach to find and confirm the most suitable synthesis algorithm for a specific digital core image dataset. This generates a super-resolution training set that highly matches the actual imaging degradation process, fundamentally alleviating the model performance bottleneck caused by training data distortion and improving the accuracy and reliability of super-resolution technology in actual geological exploration and core analysis.

[0087] The technical solution of this application and how the technical solution of this application solves the above-mentioned technical problems are described in detail below with specific embodiments. These specific embodiments can be combined with each other, and the same or similar concepts or processes may not be described again in some embodiments. The embodiments of this application will now be described with reference to the accompanying drawings.

[0088] Figure 1 A flowchart illustrating the digital core super-resolution training set generation method provided in this application. Figure 1 ,like Figure 1 As shown, the method includes:

[0089] S101. Obtain real digital core image samples. Based on multiple image interpolation algorithms, process the resolution of the real digital core image samples to obtain synthetic digital core image samples.

[0090] This embodiment is a crucial step in constructing a foundation for differentiated training data. In practice, it first requires acquiring geologically representative real digital core image samples from core imaging equipment such as micro-CT scanners and focused ion beam electron microscopes. These samples contain pore structure information at different resolutions, ranging from nanometer to micrometer. Subsequently, various typical image interpolation algorithms (including but not limited to nearest neighbor interpolation, bilinear interpolation, bicubic interpolation, and Lanczos resampling) are used to perform resolution transformation on the same set of real samples, thereby obtaining synthetic digital core image samples. This process breaks the limitations of traditional methods that rely solely on a single interpolation algorithm. By generating multiple sets of synthetic images with different degradation characteristics in parallel, it provides rich comparative samples for subsequent algorithm evaluation. It also simulates various resolution degradation modes that may exist in real imaging systems, laying the foundation for constructing training data that better reflects physical reality.

[0091] S102. Based on multiple real digital core image samples and multiple synthetic digital core image samples, calculate the synthetic quality evaluation value of each image interpolation algorithm.

[0092] In this embodiment, a quantitative comparative analysis is performed on each set of real samples and their corresponding synthesized samples from various algorithms to calculate the synthesis quality evaluation value of each image interpolation algorithm. The significance of this step lies in going beyond simple image similarity comparison, comprehensively evaluating the interpolation algorithms from two deeper dimensions: multi-scale stability and frequency domain feature preservation. This provides an objective quantitative basis for algorithm selection that encompasses both surface quality and deep features, ensuring that the evaluation results truly reflect the applicability of the algorithm to core images, a data with complex texture structures.

[0093] S103. Determine the target image interpolation algorithm based on the synthesis quality evaluation value of each image interpolation algorithm.

[0094] In this embodiment, after obtaining the multi-dimensional synthesis quality evaluation values ​​of each interpolation algorithm, a comprehensive evaluation function is established. This function comprehensively considers the algorithm's performance in terms of spatial similarity, multi-scale stability, and frequency domain fidelity, and calculates the comprehensive score of each algorithm through methods such as weighted fusion or hierarchical analysis. Based on a preset selection strategy (such as the highest total score or the optimal algorithm for a specific geological application), the target image interpolation algorithm most suitable for the characteristics of the current core image dataset is automatically determined. This step transforms the algorithm selection process from traditional subjective experience-based judgment to objective decision-making based on quantitative data, ensuring that the selected algorithm has optimal degradation simulation capabilities under specific geological image features and resolution requirements, thereby providing the optimal algorithmic tool for generating a high-quality training set.

[0095] S104. Based on multiple real digital core images, the resolution of each real digital core image is processed according to the target image interpolation algorithm to obtain a digital core super-resolution training set.

[0096] In this embodiment, after determining the target interpolation algorithm, it is applied to process a large-scale database of real digital core images: for high-resolution original images, the target algorithm is used to systematically downsample and generate corresponding low-resolution samples; for existing low-resolution images, the same algorithm is used to upsample and generate reference high-resolution samples, ultimately constructing a large-scale, high-quality, paired super-resolution training dataset. This step transforms the research results of previous algorithm evaluation and selection into practically usable data assets. The generated training set, due to the use of an interpolation algorithm optimized for the characteristics of core images, can more accurately simulate the degradation process of real imaging systems, thereby providing training samples with better preservation of geological features and stronger physical meaning for subsequent super-resolution model training, fundamentally improving the performance and generalization ability of deep learning models in actual core image super-resolution tasks.

[0097] The digital core super-resolution training set generation method provided in this application first acquires real digital core image samples and then uses various image interpolation algorithms to process these real samples to generate corresponding synthetic digital core image samples. Subsequently, by comparing the real digital core image samples with the synthetic samples generated by each algorithm, a synthetic quality evaluation value reflecting the fidelity of each algorithm is calculated. Based on these quantitative evaluation values, the optimal target image interpolation algorithm is objectively determined. Finally, this optimized algorithm is used to process a batch of real digital core images to generate a high-quality, highly consistent digital core super-resolution training set. This method, by establishing a multi-algorithm evaluation and screening mechanism, effectively avoids the simulation bias of the degradation process that may exist in traditional single interpolation algorithms. It provides a fundamental optimization path at the data level for digital core super-resolution technology, and fundamentally improves the generalization performance and practical effect of the model by enhancing the quality of the training set. This method plays an important supporting role in promoting high-precision digital core analysis, microscopic pore structure characterization, and digital research on oil and gas reservoirs.

[0098] Figure 2 A flowchart illustrating the digital core super-resolution training set generation method provided in this application. Figure 2 ,like Figure 2 As shown, in this embodiment... Figure 1 Based on the embodiments, the method for generating a digital core super-resolution training set is described in detail. Specifically, according to step S201, synthetic digital core image samples are obtained; according to steps S202 to S204, the synthetic quality evaluation value of each image interpolation algorithm is obtained; and according to steps S205 to S207, a digital core super-resolution training set is obtained. This method includes:

[0099] S201. Obtain the resolution type of the real digital core image sample; if the resolution type of the real digital core image sample is high resolution, then based on each image interpolation algorithm, the resolution of the real digital core image sample is upsampled to obtain a synthetic digital core image sample; if the resolution type of the real digital core image sample is low resolution, then based on each image interpolation algorithm, the resolution of the real digital core image sample is downsampled to obtain a synthetic digital core image sample.

[0100] In this embodiment, traditional training set construction typically involves downsampling low-resolution samples from high-resolution images. This one-way approach implicitly assumes that the information loss during downsampling is equivalent to the degradation process of a real imaging system. However, real core imaging often contains image data of varying resolutions (e.g., both high-resolution micro-CT scans and rapidly acquired low-resolution overview images). This step actively identifies the resolution type of the input samples and adopts different synthesis directions accordingly. High-resolution samples are upsampled to test the algorithm's performance in "information generation," while low-resolution samples are downsampled to test the algorithm's fidelity in "information compression," thus achieving a more comprehensive and rigorous two-way evaluation of the interpolation algorithm.

[0101] In practice, the resolution type is first automatically determined by parsing image metadata or based on features such as image sharpness and texture detail (e.g., pixels smaller than 1 micrometer are classified as high resolution, and those larger than 5 micrometers as low resolution). For samples classified as high resolution, each candidate interpolation algorithm is used for upsampling. For samples classified as low resolution, downsampling is performed to directly test their ability to simulate resolution degradation. Through this two-way testing, an excellent interpolation algorithm should exhibit high self-consistency and fidelity in both directions.

[0102] For example, image interpolation algorithms may include nearest neighbor interpolation, box interpolation, bilinear interpolation, bicubic interpolation, and Lanczos2 and Lanczos3 algorithms. By obtaining the resolution type of each real digital core image sample, which can be divided into high resolution and low resolution (including 2x, 4x, 8x, and 16x resolution), when the resolution type of the real digital core image sample is high resolution, the resolution of the real digital core image sample is upsampled based on each image interpolation algorithm to obtain 2x, 4x, 8x, and 16x resolution synthetic digital core image samples.

[0103] When the resolution type of the real digital core image sample is low resolution, the resolution of the real digital core image sample is downsampled based on each image interpolation algorithm to obtain a synthetic digital core image sample with one times the resolution of the real digital core image sample.

[0104] Figure 3 A comparison diagram of synthetic digital core image samples and real digital core image samples for the digital core super-resolution training set generation method provided in this application. Figure 1 ; Figure 4 A comparison diagram of synthetic digital core image samples and real digital core image samples for the digital core super-resolution training set generation method provided in this application. Figure 2 ;like Figure 3 As shown, real digital core image samples at one-fold resolution are processed using real interpolation, bilinear interpolation, bicubic interpolation, box interpolation, nearest neighbor interpolation, Lanczos2 interpolation, and Lanczos3 interpolation to obtain corresponding synthetic digital core image samples at two-fold, four-fold, eight-fold, and sixteen-fold resolutions. Each image interpolation algorithm corresponds to a combined digital core image sample. At the same time, during horizontal comparison, there are multiple sets of synthetic digital core image samples at each resolution, and each synthetic digital core image sample corresponds to an image interpolation algorithm.

[0105] like Figure 4 As shown, bilinear interpolation and bicubic interpolation algorithms are used to reconstruct the image. By comparison, it can be clearly seen that the bilinear reconstructed image is closer to the real image in terms of pore edges, crack continuity, and texture clarity, while the bicubic reconstructed image may have blurred edges, lost details, or false textures.

[0106] S202. Obtain the peak signal-to-noise ratio (PSNR) and structural similarity index between each real digital core image sample and the corresponding synthetic digital core image sample; based on multiple PSNR and structural similarity indices, obtain the PSNR and structural similarity index variation curves of each image interpolation algorithm with resolution; based on the PSNR and structural similarity index variation curves with resolution, obtain the slope information of the performance of each image interpolation algorithm as a function of resolution; based on the mapping relationship and slope information, obtain the performance evaluation value of each image interpolation algorithm.

[0107] In this embodiment, in the field of digital core imaging, the geological information density and feature scale carried by images at different resolution levels vary significantly. The stability of interpolation algorithms at different resolution transformation factors directly determines their applicability in complex multi-scale geological modeling. This step systematically calculates the peak signal-to-noise ratio and structural similarity index of each interpolation algorithm at multiple resolution transformation levels (e.g., 2x, 4x, 8x, 16x), and analyzes the slope characteristics of these indicators as a function of resolution, thereby quantitatively evaluating the degradation law of algorithm performance with scale changes. This evaluation method not only focuses on the absolute performance of the algorithm at a specific scale, but also emphasizes its robustness in cross-scale applications. The flatter the slope, the less sensitive the algorithm is to scale changes, and the more practical it is in multi-scale analysis scenarios.

[0108] Specifically, for each candidate interpolation algorithm, the peak signal-to-noise ratio (PSNR) and structural similarity index (SSIM) of its synthesized digital core image samples and real digital core image samples are first calculated at different resolution transformation factors, forming two sets of data sequences that vary with the resolution factor. For example, when evaluating the bicubic interpolation algorithm, the PSNR / SSIM values ​​of the images obtained after downsampling the same high-resolution core sample by 2x, 4x, and 8x are calculated, along with the original image. Subsequently, by performing linear or nonlinear fitting on these discrete data points, PSNR-resolution curves and SSIM-resolution curves are obtained, and the average slope or slope variation characteristics of each curve in a specific interval are calculated. These slope values ​​reveal the rate at which the algorithm's quality degrades as the resolution decreases. A steeper negative slope indicates that the algorithm is more sensitive to high-magnification downsampling, and the quality degrades rapidly; a gentler slope indicates that the algorithm maintains better quality consistency across different scales. Finally, these slope information are converted into standardized performance evaluation values ​​through a pre-defined mapping relationship, providing key input for subsequent comprehensive evaluation.

[0109] In one possible implementation, the PSNR-resolution curve and SSIM-resolution curve corresponding to each candidate interpolation algorithm for each digital core image sample are obtained, and the PSNR slope information and SSIM slope information are obtained. The average of multiple PSNR slope information and multiple SSIM slope information is calculated to obtain the PSNR slope information and SSIM slope information of each candidate interpolation algorithm. Furthermore, through a pre-defined mapping relationship, the numerical values ​​corresponding to the PSNR slope information and the SSIM slope information are obtained. The two values ​​are then subjected to a weighted fusion algorithm to obtain the final performance evaluation value of each image interpolation algorithm.

[0110] This step establishes a multi-scale quantitative basis for algorithm selection. By introducing resolution variation curve analysis, this method can effectively identify algorithms that perform well in conventional evaluations but experience a sharp decline in performance under extreme scale transformations, thus avoiding the selection of "scale-sensitive" interpolation methods. This is particularly important for digital core analysis, as practical research often requires information fusion and scale transformation between images of different resolutions, and the cross-scale stability of the algorithm directly affects the accuracy of multi-scale pore network modeling. Furthermore, by simultaneously incorporating the variation patterns of the two complementary indices, PSNR and SSIM, both pixel-level error control and structural fidelity perceived by human vision are considered, resulting in a more comprehensive performance profile of the algorithm. This dynamic evaluation mechanism based on curve slope shifts algorithm selection from the traditional "single-point optimality" to "overall robustness," ensuring that the selected algorithm not only performs well under ideal conditions but also adapts to the complex multi-scale processing requirements of practical applications, providing a more scientific selection criterion for generating high-quality training data.

[0111] S203. Based on Fourier transform, convert each real digital core image sample and its corresponding synthetic digital core image sample into the frequency domain to obtain the real frequency domain energy distribution map corresponding to each real digital core image sample and the synthetic frequency domain energy distribution map corresponding to each synthetic digital core image sample; obtain similarity information based on each real frequency domain energy distribution map and its corresponding synthetic frequency domain energy distribution map; obtain the fidelity evaluation value of each image interpolation algorithm based on multiple similarity information.

[0112] In this embodiment, key geological features of digital core images, such as pore structure and mineral boundaries, are not only reflected in the grayscale changes of spatial pixels but also contained in their frequency domain energy distribution characteristics. Sharp edges correspond to high-frequency components, uniform matrix regions correspond to low-frequency components, and the periodic arrangement of pores at different scales reflects the energy concentration in specific frequency bands. Traditional evaluation metrics based on spatial pixels (such as PSNR) are difficult to fully capture these structural feature changes. Fourier transform converts the image from the spatial domain to the frequency domain, decomposing the image into harmonic components of different frequencies and directions, thus providing a mathematical basis for evaluating the fidelity of the image's spectral features.

[0113] Specifically, a two-dimensional Fourier transform is first performed on each set of real digital core image samples and the synthetic samples generated by a certain interpolation algorithm. The transformed samples are obtained as complex frequency domain representations. By calculating the square of the modulus (i.e., the power spectrum), the "real frequency domain energy distribution map" of the real sample and the "synthetic frequency domain energy distribution map" of the synthetic sample are obtained. These two maps visually demonstrate the distribution of image energy at different frequencies and directions. To quantify the similarity between the two, this step uses a frequency domain extension method based on normalized cross-power spectrum correlation or structural similarity. Taking the normalized cross-correlation method as an example: the two energy distribution maps are multiplied point-by-point in the frequency domain and summed, then divided by the geometric mean of their respective energy sums to obtain a correlation coefficient ranging from [-1, 1]. The closer this coefficient is to 1, the more similar the synthetic image is to the real image in terms of frequency domain energy distribution; that is, the algorithm better preserves the spectral characteristics of the original image.

[0114] After obtaining the frequency domain similarity information for each pair of images (real and synthetic), a comprehensive evaluation of the algorithm's performance on all test samples is needed to obtain its "fidelity evaluation value." In practice, the frequency domain similarity values ​​calculated by the algorithm at all resolution transformation levels (e.g., 2x, 4x, 8x downsampling) and on all test core samples are first collected to form a similarity dataset. Then, these data are summarized using statistical methods (e.g., calculating the mean, median, or, more robustly, a weighted average after removing extreme values) to finally generate a scalar evaluation value between 0 and 1. For example, Algorithm A obtained 40 frequency domain similarity values ​​on 10 different core samples and 4 different downsampling factors, with a mean of 0.92; while Algorithm B, under the same conditions, had a mean of 0.78. This directly indicates that Algorithm A is significantly better than Algorithm B in terms of frequency domain feature preservation. This fidelity evaluation value quantifies the algorithm's ability to preserve the essential texture and structural information of the core image from a frequency domain perspective. It complements the performance evaluation value based on spatial multi-scale stability obtained in step S202, together forming a comprehensive and in-depth evaluation system for the interpolation algorithm.

[0115] This step provides an image quality assessment dimension that transcends pixel-level errors, sensitively detecting interpolation defects that perform well on spatial domain metrics but may obscure microscopic textures or introduce spurious periodic structures. This is particularly crucial for core studies relying on fine pore structures for seepage analysis. Secondly, frequency domain analysis has a stronger ability to characterize the overall structural features and directional texture of images, enabling the assessment of whether the algorithm preserves the anisotropic characteristics of different minerals in the core image. Finally, by combining frequency domain fidelity evaluation values ​​with spatial domain performance evaluation values, this method can select interpolation algorithms that are not only stable in pixel accuracy but also excellent in preserving deep structural features. This ensures that the constructed training set can retain the essential geological information of the original core image to the greatest extent, laying the foundation for training a super-resolution model capable of reconstructing geologically realistic and structurally faithful images.

[0116] S204. Based on the performance evaluation value and the fidelity evaluation value, obtain the synthesis quality evaluation value of each image interpolation algorithm.

[0117] In this embodiment, a quantitative evaluation system that comprehensively reflects the overall performance of interpolation algorithms is constructed through the systematic integration of multi-dimensional indicators. In steps S202 and S203, each candidate algorithm has been evaluated from two orthogonal and complementary dimensions: the performance evaluation value focuses on the stability and consistency of the quantification algorithm under different resolution scale transformations, revealing the algorithm's robustness in handling multi-scale processing tasks; while the fidelity evaluation value measures the algorithm's realism in feature preservation and structure reconstruction from the perspective of frequency domain energy distribution. However, these two types of evaluation values ​​have different physical meanings and numerical ranges, and direct comparison or simple superposition will lose its scientific validity and comparability. Therefore, step S204 needs to establish a scientific fusion mechanism to organically integrate these two heterogeneous evaluation dimensions into a unified synthetic quality evaluation value, thereby providing a unique and clear decision-making basis for algorithm selection.

[0118] Specifically, the two types of evaluation values ​​first need to be standardized and preprocessed to normalize them to the same numerical range (e.g., between 0 and 1) to eliminate differences in dimensions and orders of magnitude. Standardization methods can be selected based on the distribution characteristics of the evaluation values, such as linear scaling, percentile mapping, or transformations based on cumulative distribution. After standardization, a weighted fusion approach is used to calculate the final synthetic quality evaluation value. The core of this approach lies in rationally determining the weight allocation between the two types of evaluation values. The weights are not fixed but need to be dynamically adjusted based on the specific requirements of the digital core super-resolution mission or pre-calibrated based on domain knowledge.

[0119] For example, if the application scenario focuses more on the reliability of the reconstructed image in the interpretation of geological structures (such as pore network extraction), then the ability to preserve structural features reflected by frequency domain fidelity should be given higher weight; if the scenario focuses more on the consistency of quantitative parameters (such as porosity) between images of different resolutions, then the performance evaluation value corresponding to multi-scale stability may be more critical.

[0120] A typical implementation is as follows: Synthetic quality evaluation value = α × Normalized performance evaluation value + β × Normalized fidelity evaluation value, where α and β are weighting coefficients determined based on prior knowledge or experimental calibration, and satisfy α + β = 1. Through this weighted fusion, each algorithm ultimately obtains a single synthetic quality evaluation value, which integrates its performance in both spatial scale stability and frequency domain feature fidelity.

[0121] This step elevates the evaluation dimensions from discrete to unified, guiding algorithm selection towards the final application goal. First, it solves the common challenge of "how to comprehensively evaluate" in multi-dimensional assessments, avoiding biased selection due to the advantage of a single metric. An algorithm might perform moderately well in multi-scale stability (medium performance rating) but excel in frequency domain feature preservation (high fidelity rating). Through weighted fusion, its comprehensive advantage is revealed, and vice versa. Second, through the design of weighting coefficients, this step directly transmits the upper-level application requirements to the algorithm selection stage, ensuring that the final selected target algorithm is not only balanced and excellent in technical indicators but also targeted in terms of task orientation. For example, by setting a higher β weight, algorithms with superior fidelity in reconstructing core microstructure and texture can be selected, even if their traditional metrics such as PSNR are not the highest. Ultimately, the synthetic quality evaluation value output by this fusion mechanism provides a clear and sortable quantitative basis for "determining the target image interpolation algorithm" in step S205, ensuring the logical rigor and optimality of the entire method from evaluation to decision-making, and laying a solid foundation for algorithm selection to generate a high-quality digital core super-resolution training set.

[0122] S205. Determine the target image interpolation algorithm based on the synthesis quality evaluation value of each image interpolation algorithm; based on multiple real digital core images, process the resolution of each real digital core image according to the target image interpolation algorithm to obtain a digital core super-resolution training set.

[0123] In this embodiment, this step transforms the preliminary multi-dimensional evaluation results into optimal decisions that can guide practice through a quantitative evaluation-based selection mechanism, and generates high-quality training data on a large scale. In step S204, each candidate interpolation algorithm has obtained a synthetic quality evaluation value that comprehensively reflects its multi-scale stability and frequency domain feature fidelity. This value is a scalar score that has been standardized and weighted. Step S205 first sorts all algorithms based on these evaluation values, and usually selects the algorithm with the highest score as the target image interpolation algorithm. This selection mechanism is not simply selecting the best in a single indicator, but rather a global selection based on the comprehensive performance of the algorithm in two key dimensions: spatial scale robustness and frequency domain structure fidelity. This ensures that the selected algorithm has a balanced advantage and comprehensive reliability when dealing with the complex characteristics of digital core images.

[0124] For example, suppose that after evaluation, the bicubic interpolation algorithm scores 85 points for synthetic quality, Lanczos interpolation scores 92 points, and nearest neighbor interpolation scores 70 points. The Lanczos interpolation algorithm will be automatically selected as the target image interpolation algorithm. After determining the target algorithm, it will be invoked to batch process a large-scale database of real digital core images. If the input is a high-resolution image, the algorithm will be used for downsampling to generate corresponding low-resolution samples; if the input contains low-resolution images, upsampling can also be performed to construct reference high-resolution samples, ultimately forming a set of paired low-resolution-high-resolution image pairs, i.e., the digital core super-resolution training set. This process is fully automated and scalable, efficiently utilizing the selected optimal algorithm to generate consistent and high-quality degradation simulation results for massive amounts of core image data.

[0125] Optionally, in order to effectively amplify the detail differences between images of different resolutions and make them easier for deep learning models to capture, a global thresholding method is used to binarize all paired real high-resolution grayscale images. The selected target image interpolation algorithm is then used to downsample the binarized real high-resolution images to generate corresponding synthetic low-resolution images. In this process, a secondary thresholding step is introduced, with a truncation value of 0.5. The interpolation results are then re-binarized to ensure that the pixel values ​​of all synthetic low-resolution images are strictly 0 or 1, thus maintaining consistency with the binarized real high-resolution images in the numerical domain.

[0126] This step eliminates human subjectivity and arbitrariness in algorithm selection through objective quantitative scoring and a merit-based selection mechanism, ensuring the theoretical rigor and reproducibility of the training set generation process. Secondly, the selected target algorithm represents the most adaptable and fidelity-enhancing interpolation method for specific types of core images under the current evaluation system. The training set generated by this algorithm can more accurately simulate resolution degradation in real imaging processes, thus providing learning samples closer to the real data distribution for subsequent super-resolution model training. Finally, the large-scale generated training set directly serves the training of the deep learning model. Its quality advantages will translate into significant improvements in model reconstruction accuracy, detail recovery ability, and generalization performance. This provides a fundamental data foundation for solving the model performance bottleneck caused by insufficient training data fidelity, and powerfully promotes the effective transformation of digital core super-resolution technology from methodological research to practical geological applications.

[0127] S206. Based on multiple real digital core images, the resolution of each real digital core image is processed according to the comparison image interpolation algorithm to obtain a digital core super-resolution comparison training set; the digital core super-resolution training set and the digital core super-resolution comparison training set are respectively input into a super-resolution network to obtain a reconstructed high-resolution image set and a reconstructed high-resolution image comparison set.

[0128] In this embodiment, after determining the target image interpolation algorithm and generating the corresponding training set in step S205, the evaluation score of the algorithm itself is insufficient to fully demonstrate its superiority in practical applications, as there may be a deviation between the evaluation metrics and the final super-resolution model performance. Therefore, this step requires setting a reference benchmark—a comparison image interpolation algorithm. This algorithm is typically a standard method currently widely used in the field (such as the classic bicubic interpolation algorithm) or a previously mainstream technical solution. Using this comparison algorithm, another set of digital core super-resolution comparison training sets is generated with identical real digital core image source data and the same processing flow (including the same resolution transformation factor and the same sample pairing method). The establishment of this comparison set provides a fair and consistent benchmark for subsequent performance comparisons, ensuring that any observed performance differences can be clearly attributed to differences in the interpolation algorithm itself, rather than the influence of data or other variables.

[0129] In practice, two training sets are processed simultaneously: one is the "digital core super-resolution training set" generated by the target image interpolation algorithm selected in step S205, and the other is the "digital core super-resolution control training set" generated by the control image interpolation algorithm. Subsequently, these two training sets are input into the same predefined deep learning super-resolution network model (such as SRCNN, ESPCN, or a more advanced Transformer-based network) with predefined architecture and initialization parameters for training. To ensure fairness in the comparison, both training processes use the exact same training configuration, including network structure, optimizer, learning rate, loss function, training epochs, and batch size. After training, the same independent validation or test set (usually composed of real low-resolution core images not used in training and their corresponding high-resolution reference images) is used to infer the two trained models, thereby obtaining the "reconstructed high-resolution image set" output by the model supported by the target algorithm training set, and the "reconstructed high-resolution image control set" output by the model supported by the control algorithm training set. These two sets of reconstructed image sets are the direct outputs for evaluating the quality of the training sets.

[0130] This step establishes a closed-loop verification system. First, by introducing a control algorithm and generating a control training set, this method establishes rigorous experimental control, ensuring that subsequent performance comparisons have the necessary control conditions for scientific experiments, leading to more reliable conclusions. Second, by training and testing the same super-resolution network separately, the impact of training sets generated by different interpolation algorithms on the final model's reconstruction capability can be directly observed. This is more practically meaningful than simply comparing the spatial or frequency domain metrics of the interpolation algorithms themselves, as the ultimate goal is to improve the performance of the super-resolution model. Finally, this step provides the necessary input for the quantitative comparison in step S207. Two comparable sets of reconstructed images allow for the quantification of the actual performance gain of the target algorithm relative to traditional methods by calculating the similarity difference between them and real high-resolution images. This provides the final and most direct decision-making basis for whether to adopt the target algorithm and its corresponding training set. This design ensures that the entire method not only stays at the algorithm selection stage but also extends to the final stage of verifying its practical application value.

[0131] S207. Calculate the first similarity information between the reconstructed high-resolution image set and the original real low-resolution image set; calculate the second similarity information between the reconstructed high-resolution image set and the original real low-resolution image set; calculate the difference information between the first similarity information and the second similarity information; determine whether the difference meets the preset threshold. If yes, determine the digital core super-resolution training set as the target digital core super-resolution training set; if not, redetermine a new target image interpolation algorithm until the new difference meets the preset threshold, and determine the generated new digital core super-resolution training set as the target digital core super-resolution training set.

[0132] In this embodiment, after obtaining the super-resolution models trained on the training sets generated by the target algorithm and the control algorithm respectively in step S206, and their reconstruction results, subjective observation or the output of a single model is insufficient to scientifically determine the quality of the training sets. Therefore, this step introduces a quantitative comparison based on objective image quality indicators: calculating the first similarity information between the "reconstructed high-resolution image set" and the "original real low-resolution image set," and the second similarity information between the "reconstructed high-resolution image control set" and the same set of "original real low-resolution images." The key here is that the similarity calculation does not directly compare the reconstructed image with the real high-resolution image (which is often unavailable in real-world scenarios), but cleverly downsamples the reconstructed high-resolution image back to its original low-resolution size and then compares it with the original, real low-resolution image. The core logic of this design is that the high-resolution image reconstructed by an excellent super-resolution model, when downsampled to its original low resolution, should be highly consistent with the original real low-resolution image—this conforms to the theoretical consistency constraint of image degradation and reconstruction processes.

[0133] In practice, for each high-resolution reconstructed image in the two reconstructed image sets (models trained from the target algorithm's training set and the control algorithm's training set, respectively), the same downsampling method (usually the selected target algorithm or standard algorithm) used when constructing the training set is applied to downsample it to the low-resolution size of the original input. Then, the similarity between these downsampled images and their corresponding original ground truth low-resolution images is calculated, using metrics such as PSNR, SSIM, or other perceptual quality indicators. The similarity of all image pairs is then averaged or weighted to obtain the first similarity information (corresponding to the target algorithm) and the second similarity information (corresponding to the control algorithm) representing the overall reconstruction quality.

[0134] Next, the difference between the two sets of similarity information is calculated, for example, by using a relative improvement percentage: Difference = (First Similarity - Second Similarity) / Second Similarity × 100%. Finally, this difference is compared with a preset performance improvement threshold set according to actual application requirements. If the difference is greater than or equal to the threshold, it proves that the model trained on the training set generated by the target algorithm has a significant and sufficient improvement in reconstruction fidelity compared to the control baseline. Therefore, this training set is determined as the final adopted "target digital core super-resolution training set". If the threshold is not met, it means that the performance gain brought by the currently selected target algorithm has not met expectations, and an iterative optimization process will be triggered: based on the synthetic quality evaluation value in step S204, the algorithm with the second-best score is selected as the new "target image interpolation algorithm", and the process from S205 to S207 is repeated until an algorithm that can produce a difference that meets the threshold requirement and its generated training set are found.

[0135] This step implements a closed-loop, self-optimizing training set generation and selection process oriented towards final model performance. First, it cleverly bypasses the absolute dependence on true high-resolution ground truth through an indirect evaluation strategy of "reconstruction-downsampling-comparison," better aligning with real-world applications where only low-resolution images are available, thus enhancing the method's practicality. Second, by setting a performance improvement threshold, this method ultimately anchors the evaluation criteria for training set quality from algorithmic metrics (such as S202 to S204 evaluation values) to the performance of the actual super-resolution task, ensuring that the selected training set can genuinely improve the model's capabilities and aligning the evaluation criteria with the final application goal. Finally, the introduced iterative optimization mechanism endows the method with self-improvement and fault tolerance: even if the initially selected algorithm fails to meet expectations during actual model training, it can automatically try alternative solutions, approaching the optimal solution through multiple iterations, thereby significantly improving the robustness of the entire method to different data characteristics and the reliability of the final output results.

[0136] Figure 5 This is a schematic diagram of the digital core super-resolution training set generation device provided in this application, as shown below. Figure 5 As shown, the digital core super-resolution training set generation device 50 provided in this embodiment includes:

[0137] The first processing module 501 is used to acquire real digital core image samples, and process the resolution of the real digital core image samples based on multiple image interpolation algorithms to obtain synthetic digital core image samples.

[0138] The second processing module 502 is used to calculate the synthesis quality evaluation value of each image interpolation algorithm based on multiple real digital core image samples and multiple synthetic digital core image samples.

[0139] The third processing module 503 is used to determine the target image interpolation algorithm based on the synthesis quality evaluation value of each image interpolation algorithm;

[0140] The fourth processing module 504 is used to process the resolution of each real digital core image based on multiple real digital core images according to the target image interpolation algorithm to obtain a digital core super-resolution training set.

[0141] In one possible implementation, the first processing module 501 is further configured to:

[0142] Obtain the resolution type of real digital core image samples; among which, the resolution type includes high resolution and low resolution, and low resolution includes 2x resolution, 4x resolution, 8x resolution and 16x resolution;

[0143] If the resolution of the real digital core image sample is high, then based on each image interpolation algorithm, the resolution of the real digital core image sample is upsampled to obtain the synthetic digital core image sample; wherein, the synthetic digital core image sample is a low-resolution synthetic digital core image sample.

[0144] If the resolution type of the real digital core image sample is low resolution, then based on each image interpolation algorithm, the resolution of the real digital core image sample is downsampled to obtain the synthetic digital core image sample; wherein, the synthetic digital core image sample is a high-resolution synthetic digital core image sample.

[0145] In one possible implementation, the second processing module 502 is further configured to:

[0146] Based on multiple real digital core image samples and multiple synthetic digital core image samples, a performance evaluation value for each image interpolation algorithm is obtained; the performance evaluation value is used to characterize the slope information of the performance of each image interpolation algorithm as a function of resolution.

[0147] Based on multiple real digital core image samples and multiple synthetic digital core image samples, a fidelity evaluation value is obtained for each image interpolation algorithm; the fidelity evaluation value is used to characterize the frequency domain energy distribution between the synthetic digital core image sample generated by each image interpolation algorithm and the corresponding real digital core image sample.

[0148] Based on the performance evaluation value and the fidelity evaluation value, the synthesis quality evaluation value of each image interpolation algorithm is obtained.

[0149] In one possible implementation, the second processing module 502 is further configured to:

[0150] The peak signal-to-noise ratio and structural similarity index between each real digital core image sample and the corresponding synthetic digital core image sample were obtained.

[0151] Based on multiple peak signal-to-noise ratios and multiple structural similarity indices, the curves of peak signal-to-noise ratio and structural similarity index as a function of resolution for each image interpolation algorithm are obtained;

[0152] Based on the curves of peak signal-to-noise ratio and structural similarity index as a function of resolution, the slope information of the performance of each image interpolation algorithm as a function of resolution is obtained.

[0153] Based on the mapping relationship and slope information, the performance evaluation value of each image interpolation algorithm is obtained; where the mapping relationship is the correspondence between slope information and performance evaluation value.

[0154] In one possible implementation, the second processing module 502 is further configured to:

[0155] Based on Fourier transform, each real digital core image sample and the corresponding synthetic digital core image sample are converted into the frequency domain to obtain the real frequency domain energy distribution map corresponding to each real digital core image sample and the synthetic frequency domain energy distribution map corresponding to each synthetic digital core image sample.

[0156] Similarity information is obtained based on each real frequency domain energy distribution map and its corresponding synthetic frequency domain energy distribution map;

[0157] Based on multiple similarity information, the fidelity evaluation value of each image interpolation algorithm is obtained.

[0158] In one possible implementation, the fourth processing module 504 is further configured to:

[0159] Based on multiple real digital core images, the resolution of each real digital core image is processed according to the comparison image interpolation algorithm to obtain a digital core super-resolution comparison training set.

[0160] The digital core super-resolution training set and the digital core super-resolution control training set are respectively input into the super-resolution network to obtain the reconstructed high-resolution image set and the reconstructed high-resolution image control set.

[0161] Calculate the first similarity information between the reconstructed high-resolution image set and the original real low-resolution image set;

[0162] Calculate the second similarity information between the reconstructed high-resolution image reference set and the original real low-resolution image set;

[0163] Calculate the difference between the first similarity information and the second similarity information;

[0164] Determine whether the difference meets the preset threshold. If so, then determine the digital core super-resolution training set as the target digital core super-resolution training set.

[0165] If not, a new target image interpolation algorithm is determined until the new difference meets the preset threshold, and the generated new digital core super-resolution training set is determined as the target digital core super-resolution training set.

[0166] The digital core super-resolution training set generation device provided in this embodiment can execute the method provided in the above method embodiment. Its implementation principle and technical effect are similar, and will not be described in detail here.

[0167] Figure 6 A hardware schematic diagram of the digital core super-resolution training set generation device provided in this application. Figure 6 As shown, the electronic device 60 provided in this embodiment includes at least one processor 601 and a memory 602. Optionally, the device 60 further includes a communication component 603. The processor 601, memory 602, and communication component 603 are connected via a bus 604.

[0168] In a specific implementation, at least one processor 601 executes computer execution instructions stored in memory 602, causing at least one processor 601 to perform the above-described method.

[0169] The specific implementation process of processor 601 can be found in the above method embodiments, and its implementation principle and technical effect are similar. It will not be repeated here.

[0170] In the above embodiments, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly implemented by a hardware processor, or implemented by a combination of hardware and software modules within the processor.

[0171] The memory may include random access memory (RAM) and may also include non-volatile memory (NVM), such as at least one disk storage device.

[0172] The bus can be an Industry Standard Architecture (ISA) bus, a Peripheral Component Interconnect (PCI) bus, or an Extended Industry Standard Architecture (EISA) bus, etc. Buses can be categorized as address buses, data buses, control buses, etc. For ease of illustration, the buses shown in the accompanying drawings are not limited to a single bus or a single type of bus.

[0173] This application also provides a computer program product, including a computer program that, when executed by a processor, implements the above-described method.

[0174] This application also provides a computer-readable storage medium storing computer-executable instructions, which, when executed by a processor, implement the above-described method.

[0175] The aforementioned readable storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk. The readable storage medium can be any available medium accessible to a general-purpose or special-purpose computer.

[0176] An exemplary readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application Specific Integrated Circuit (ASIC). Alternatively, the processor and the readable storage medium can exist as discrete components in the device.

[0177] The division of units is merely a logical functional division; in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be indirect coupling or communication connection through some interfaces, methods, or units, and may be electrical, mechanical, or other forms.

[0178] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment according to actual needs.

[0179] In addition, the functional units in the various embodiments of the present invention can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0180] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0181] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0182] Finally, it should be noted that other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the art not disclosed herein, and is not limited to the precise structures described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

Claims

1. A digital core super-resolution training set generation method, characterized in that, include: Real digital core image samples are obtained, and the resolution of the real digital core image samples is processed based on multiple image interpolation algorithms to obtain synthetic digital core image samples. Based on multiple real digital core image samples and multiple synthetic digital core image samples, the synthesis quality evaluation value of each image interpolation algorithm is calculated; The target image interpolation algorithm is determined based on the synthesis quality evaluation value of each of the image interpolation algorithms; Based on multiple real digital core images, the resolution of each real digital core image is processed according to the target image interpolation algorithm to obtain a digital core super-resolution training set.

2. The method according to claim 1, characterized in that, The process of obtaining real digital core image samples, based on multiple image interpolation algorithms, involves processing the resolution of the real digital core image samples to obtain synthetic digital core image samples, including: Obtain the resolution type of the real digital core image sample; wherein, the resolution type includes high resolution and low resolution, and the low resolution includes double resolution, quadruple resolution, eight-fold resolution and sixteen-fold resolution; If the resolution type of the real digital core image sample is high, then based on each of the image interpolation algorithms, the resolution of the real digital core image sample is upsampled to obtain the synthetic digital core image sample; wherein, the synthetic digital core image sample is a low-resolution synthetic digital core image sample. If the resolution type of the real digital core image sample is low resolution, then based on each of the image interpolation algorithms, the resolution of the real digital core image sample is downsampled to obtain the synthetic digital core image sample; wherein, the synthetic digital core image sample is a high-resolution synthetic digital core image sample.

3. The method according to claim 2, characterized in that, The step of calculating the synthesis quality evaluation value of each image interpolation algorithm based on multiple real digital core image samples and multiple synthesized digital core image samples includes: Based on multiple real digital core image samples and multiple synthetic digital core image samples, a performance evaluation value for each image interpolation algorithm is obtained; wherein, the performance evaluation value is used to characterize the slope information of the performance of each image interpolation algorithm as a function of resolution; Based on multiple real digital core image samples and multiple synthetic digital core image samples, a fidelity evaluation value is obtained for each image interpolation algorithm; wherein, the fidelity evaluation value is used to characterize the frequency domain energy distribution between the synthetic digital core image sample generated by each image interpolation algorithm and the corresponding real digital core image sample; Based on the performance evaluation value and the fidelity evaluation value, the synthesis quality evaluation value of each image interpolation algorithm is obtained.

4. The method according to claim 3, characterized in that, The step of obtaining a performance evaluation value for each image interpolation algorithm based on multiple real digital core image samples and multiple synthetic digital core image samples includes: The peak signal-to-noise ratio and structural similarity index between each real digital core image sample and the corresponding synthetic digital core image sample are obtained; Based on multiple peak signal-to-noise ratios and multiple structural similarity indices, the curves of the peak signal-to-noise ratio and structural similarity index of each image interpolation algorithm as a function of resolution are obtained; Based on the curves showing the change of peak signal-to-noise ratio and structural similarity index with resolution, the slope information of the performance of each image interpolation algorithm as a function of resolution is obtained. Based on the mapping relationship and the slope information, a performance evaluation value for each image interpolation algorithm is obtained; wherein, the mapping relationship is the correspondence between the slope information and the performance evaluation value.

5. The method according to claim 3, characterized in that, The step of obtaining a fidelity evaluation value for each image interpolation algorithm based on multiple real digital core image samples and multiple synthetic digital core image samples includes: Based on Fourier transform, each real digital core image sample and the corresponding synthetic digital core image sample are converted into the frequency domain to obtain the real frequency domain energy distribution map corresponding to each real digital core image sample and the synthetic frequency domain energy distribution map corresponding to each synthetic digital core image sample. Similarity information is obtained based on each of the real frequency domain energy distribution maps and the corresponding synthetic frequency domain energy distribution maps; Based on multiple similarity information, the fidelity evaluation value of each image interpolation algorithm is obtained.

6. The method according to any one of claims 1-5, characterized in that, After obtaining the digital core super-resolution training set, the following is also included: Based on multiple real digital core images, the resolution of each real digital core image is processed according to the comparison image interpolation algorithm to obtain a digital core super-resolution comparison training set. The digital core super-resolution training set and the digital core super-resolution control training set are respectively input into a super-resolution network to obtain a reconstructed high-resolution image set and a reconstructed high-resolution image control set. Calculate the first similarity information between the reconstructed high-resolution image set and the original real low-resolution image set; Calculate the second similarity information between the reconstructed high-resolution image reference set and the original real low-resolution image set; Calculate the difference information between the first similarity information and the second similarity information; Determine whether the difference meets a preset threshold. If so, then determine the digital core super-resolution training set as the target digital core super-resolution training set. If not, a new target image interpolation algorithm is determined until the new difference meets the preset threshold, and the generated new digital core super-resolution training set is determined as the target digital core super-resolution training set.

7. A digital core super-resolution training set generation device, characterized in that, include: The first processing module is used to acquire real digital core image samples, and process the resolution of the real digital core image samples based on multiple image interpolation algorithms to obtain synthetic digital core image samples. The second processing module is used to calculate the synthesis quality evaluation value of each of the image interpolation algorithms based on the multiple real digital core image samples and the multiple synthetic digital core image samples. The third processing module is used to determine the target image interpolation algorithm based on the synthesis quality evaluation value of each of the image interpolation algorithms; The fourth processing module is used to process the resolution of each real digital core image based on multiple real digital core images according to the target image interpolation algorithm to obtain a digital core super-resolution training set.

8. An electronic device, characterized in that, include: Memory, processor; The memory stores computer-executed instructions; The processor executes computer execution instructions stored in the memory, causing the processor to perform the method as described in any one of claims 1-6.

9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer-executable instructions, which, when executed by a processor, are used to implement the method as described in any one of claims 1-6.

10. A computer program product, characterized in that, Includes a computer program that, when executed by a processor, implements the method described in any one of claims 1-6.

Images