Coal digital core sequence picture multi-scale fusion and super-resolution reconstruction method

Abstract

This invention discloses a method for multi-scale fusion and super-resolution reconstruction of digital core images of coal and rock, comprising the following steps: acquiring a dataset, including low-resolution and high-resolution images; data preprocessing, performing feature point alignment and fusion of multi-scale images to create a time-series dataset; inputting the time-series dataset into a sequence image super-resolution reconstruction (EDVR) model for training, obtaining the mapping relationship between the fused high-resolution image and the original low-resolution image; based on this mapping relationship, reconstructing the unfused low-resolution image into a high-resolution image using the EDVR model, obtaining corresponding structural similarity and peak signal-to-noise ratio indices, and testing the model; using Avico software to perform three-dimensional reconstruction of two-dimensional slices, obtaining a three-dimensional plunger-shaped coal and rock core filled with pore and fracture details. This invention is based on deep learning, enabling rapid and accurate acquisition of high-resolution images displaying a large field of view, reducing costs and increasing efficiency, and eliminating the need for extensive manual operation, thus reducing human error.

Description

Technical Field

[0001] This invention belongs to the field of coal and rock image processing technology, and in particular relates to a method for multi-scale fusion and super-resolution reconstruction of digital core sequence images of coal and rock. Background Technology

[0002] In core image analysis, low-resolution core images can display a large field of view and have good global representativeness, but they cannot accurately represent small-scale information. High-resolution core images can accurately represent small-scale information of the core, but they can usually only display a small field of view.

[0003] Currently, the existing methods for obtaining core images include the following two:

[0004] 1. Computed tomography (CT) scans are used to acquire high-resolution images of the rock core. However, the insufficient resolution of CT images makes it impossible to fully characterize micron-sized pores. To obtain clear images of micron-sized and smaller pore structures, the rock must be cut into samples of a few millimeters or even smaller, which to some extent leads to a lack of representativeness of the samples.

[0005] 2. High-resolution two-dimensional core images are obtained through optical microscopy imaging of thin sections of cores and scanning electron microscopy imaging of rock samples. Existing methods for acquiring high-resolution images with large fields of view are costly, have low imaging efficiency and slow imaging speed, and require highly skilled operators, making them unsuitable for widespread application. Summary of the Invention

[0006] The problem this invention aims to solve is to provide a method for multi-scale fusion and super-resolution reconstruction of digital core sequence images of coal and rock, especially a method based on deep learning that can quickly and accurately obtain high-resolution images with a large field of view, without requiring a lot of manual operation and reducing human error, to reconstruct low-resolution images of two-dimensional slices of plunger-shaped coal and rock cores into high-resolution images.

[0007] To solve the above-mentioned technical problems, the technical solution adopted in this invention is: a method for multi-scale fusion and super-resolution reconstruction of digital core sequence images of coal and rock, comprising the following steps:

[0008] S1: Obtain the dataset, including low-resolution and high-resolution images;

[0009] S2: Data preprocessing, performing feature point alignment and fusion of multi-scale images, and creating time series datasets;

[0010] S3: Input the time series dataset into the training of the sequence image super-resolution reconstruction EDVR model to obtain the mapping relationship between the fused high-resolution image and the original low-resolution image;

[0011] S4: Based on the mapping relationship, the unfused low-resolution image is reconstructed into a high-resolution image using the EDVR model to obtain the corresponding structural similarity and peak signal-to-noise ratio indices, and the model is tested.

[0012] S5: Use Avico software to perform three-dimensional reconstruction on the two-dimensional slices to obtain a three-dimensional plunger-shaped coal core filled with detailed information about pores and fractures.

[0013] Furthermore, S1 includes the following steps:

[0014] S11: Acquire low-resolution images with a pixel diameter of 9µm and high-resolution images with a pixel diameter of 0.3µm through CT scanning;

[0015] S12: The ratio of the number of high-resolution two-dimensional slice images to the number of low-resolution two-dimensional slice images is 1:3.

[0016] Furthermore, step S2 includes the following steps:

[0017] S21: Upsample the low-resolution image at the optimal registration ratio while preserving the original low-resolution image information;

[0018] S22: Align the upsampled low-resolution image with the high-resolution image using SIFT feature points;

[0019] S23: The aligned high-resolution image information is fused and reconstructed into a label set in the low-resolution image. A time series dataset is created based on the preserved original low-resolution image samples.

[0020] Furthermore, step S3 includes the following steps:

[0021] S31: Create a time series dataset using a sliding window and input the time series dataset into the model;

[0022] S32: Frame alignment via pyramid-cascaded deformable convolutional layers;

[0023] S33: Utilize the temporal relationship between frames and the spatial relationship within frames to perform TSA fusion and assign pixel-level aggregation weights to each frame.

[0024] S34: The feature block F after spatial and temporal attention fusion is fed into the SR network for reconstruction, and a high-resolution two-dimensional core slice support frame image is reconstructed.

[0025] S35: The reconstructed 2D core slice image is passed through an upsampling layer, and the size of the feature block is increased by using subpixel convolution to generate an HR-level image. The predicted image is then concatenated with the 2D core slice reference frame image obtained from the input end through upsampling, and its super-resolution image is output.

[0026] Furthermore, S32 includes the following steps:

[0027] S321: The two-dimensional core slice sequence images under the same time window are input and their respective feature blocks are output through convolution as the feature information of the first layer; the downsampled feature information of this layer is obtained by using a convolution operation with a stride of 2, and their image size is attenuated by a factor of 2;

[0028] S322: Starting from the top layer L3, the feature blocks of the two-dimensional core slice reference frame image of this layer are stitched and fused to output the offset of this layer. Then, the offset and the feature blocks of the two-dimensional core slice support frame image of this layer are deformably convolved to output the feature blocks of the two-dimensional core slice reference frame image after alignment of this layer.

[0029] S323: The offset generated by the third layer and the feature map of the two-dimensional core slice reference frame after alignment with the third layer are upsampled and fed to the second layer using bilinear interpolation. The offset of the second layer is obtained not only from the two input feature blocks of the layer, but also from the offset of the third layer. In addition, the aligned feature blocks output by the second layer are not only from the output of the deformable convolution of the layer, but also depend on the aligned feature blocks of the third layer, and so on, until the feature blocks of the frame image of the aligned two-dimensional core slice are output.

[0030] S324: Outside the pyramid structure, the feature blocks of the first-layer two-dimensional core slice reference frame image and the feature information of the first layer are spliced ​​and fused to output an offset. Then, this offset and the feature information of the first layer are used to perform deformable convolution to output an aligned version of the two-dimensional core slice support frame feature blocks.

[0031] Furthermore, S33 includes the following steps:

[0032] S331: Perform convolutional learning on the input two-dimensional core slice frame image to obtain an embedding space, then calculate the similarity of feature information between the input two-dimensional core slice support frame image and the two-dimensional core slice reference frame image in the embedding space, and obtain the feature block after fusion with temporal attention.

[0033] S332: The feature blocks that incorporate temporal attention are placed into a pyramid-shaped spatial attention architecture. First, alignment is performed using two convolutional downsampling operations. Then, starting from the top layer, the feature blocks that combine spatial attention are output from top to bottom through upsampling, addition, and dot product operations.

[0034] Furthermore, the reconstruction method in S34 is bicubic interpolation.

[0035] Furthermore, the present invention also provides an apparatus for performing the above-described data processing method.

[0036] Furthermore, the present invention also provides an apparatus including a memory, a processor, and an algorithm stored in the memory and executable on the processor, wherein the processor implements the above-described data processing method when executing the computer program.

[0037] Furthermore, the present invention also provides a computer-readable storage medium storing a computer algorithm, which, when executed by a processor, implements the above-described data processing method.

[0038] The advantages and positive effects of this invention are:

[0039] This invention is based on deep learning and can quickly and accurately obtain high-resolution images with a large field of view. Compared with traditional methods, this invention has the advantages of cost reduction and efficiency improvement, and does not require a lot of manual operation, thus reducing human error. Attached Figure Description

[0040] Figure 1 This is an overall flowchart of an embodiment of the present invention.

[0041] Figure 2 This is a schematic diagram of coal core sampling according to an embodiment of the present invention.

[0042] Figure 3 This is a network structure diagram of the EDVR model according to an embodiment of the present invention.

[0043] Figure 4 This is a PCD network structure diagram according to an embodiment of the present invention. Detailed Implementation

[0044] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0045] In the description of this invention, it should be noted that the terms "center," "upper," "lower," "left," "right," "vertical," "horizontal," "inner," and "outer," etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are used only for the convenience of describing the invention and for simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on the invention. Furthermore, the terms "first," "second," and "third" are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0046] In the description of this invention, it should be noted that, unless otherwise explicitly specified and limited, the terms "installation", "connection" and "linking" should be interpreted broadly, and those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0047] The embodiments of the present invention will be further described below with reference to the accompanying drawings:

[0048] like Figure 1 As shown, the method for multi-scale fusion and super-resolution reconstruction of digital core images of coal and rock includes the following steps:

[0049] S1: Obtain the dataset, including low-resolution and high-resolution images. Specifically, S1 includes the following steps.

[0050] S11: Acquire low-resolution images with a pixel diameter of 9µm and high-resolution images with a pixel diameter of 0.3µm through CT scanning.

[0051] S12: The ratio of the number of high-resolution two-dimensional slice images to the number of low-resolution two-dimensional slice images is 1:3.

[0052] Specifically, in this embodiment, CT scanning is used on the core of the plunger-shaped coal and rock sample to obtain core sequence slices, generating several low-resolution two-dimensional core slice images with a pixel diameter of 9 μm. Subsamples are taken from inside the coal and rock sample core and subjected to CT scanning to generate several high-resolution two-dimensional slice images with a pixel diameter of 0.3 μm. The number of slices from the subsamples inside the core should be less than the number of slices from the plunger-shaped coal and rock sample. The ratio of the number of subsamples inside the core to the number of plunger-shaped coal and rock samples is N:3N, where N is a natural number. Specifically, in this embodiment, the ratio of the number of subsamples inside the core to the number of plunger-shaped coal and rock samples is 1:3. A schematic diagram of the coal and rock core sampling is shown below. Figure 2 As shown, A represents a sample taken from the core of the coal and rock sample, and B represents a plunger-shaped coal and rock sample.

[0053] When aligning digital core images, direct image alignment cannot accurately determine the position of the subsample within the plunger sample due to the different resolutions of the plunger sample and subsamples. To address this problem, a method must be found that can accurately align images even with different resolutions and precisely determine the position of the subsample within the plunger sample. Since the SIFT algorithm is invariant to features under changes in image scaling, rotation, and translation, this invention uses the SIFT algorithm for image feature point alignment and image information fusion.

[0054] S2: Data preprocessing, including feature point alignment and fusion of multi-scale images to create a time-series dataset. Specifically, S2 includes the following steps:

[0055] S21: Upsample the low-resolution two-dimensional core slice image to a resolution of 3µm according to the optimal registration ratio, while retaining the original low-resolution image information.

[0056] S22: Align the upsampled low-resolution two-dimensional core slice image with the high-resolution two-dimensional core slice image using SIFT feature points, and then fuse the aligned image information with a resolution of 0.3µm into the image with a resolution of 3µm.

[0057] S23: Use the original corresponding regions of the preserved images as model input, fuse the aligned high-resolution image information and reconstruct it in the low-resolution image as a label set, and create a time series dataset based on the preserved original low-resolution image samples.

[0058] The time series dataset is generated in the following ways:

[0059] First, the label set is scanned in sequence from front to back. Using a sliding window strategy, three images are taken sequentially to form a sequence, i.e., one sample. For example, the first three images are taken as the first sequence, the second four as the second sequence, the third five as the third sequence, and so on until the (N-2)th to the Nth image is taken as the last sequence, forming a total of N-2 samples, thus generating the time series dataset.

[0060] Because CT slice data possesses a certain sequential relationship, which characterizes the continuity and distribution of pores and sutures in space, this invention employs an EDVR model capable of super-resolution reconstruction of sequential images to learn this sequential relationship and ensure that the generated high-resolution image also possesses continuity and integrity. This model fully utilizes information between adjacent frames and generates super-resolution images. The model structure is as follows: Figure 3 As shown.

[0061] S3: Input the time-series dataset into the training of the sequence image super-resolution reconstruction (EDVR) model to obtain the mapping relationship between the fused high-resolution image and the original low-resolution image. Specifically, S3 includes the following steps:

[0062] S31: Create a time series dataset using a sliding window and input the time series dataset into the model.

[0063] The specific process of model input is as follows: First, the label set is scanned from front to back, and three images are taken sequentially using a sliding window strategy to form a sequence, i.e., one sample. For example, the first three images are taken as the first sequence, the second four images as the second sequence, the third five images as the third sequence, and so on until the (N-2)th image is taken as the last sequence, forming a total of N-2 samples, thus generating the time series dataset.

[0064] The input consists of three consecutive frames, each representing a sample. These frames are labeled as frame t-1, t, and t+1, where frame t is the reference frame (the frame the model needs to upsample (the super-resolution object)) and frames t-1 and t+1 are the support frames. The input frames are then processed by a deblurring module consisting of several residual blocks and convolutional layers with a stride of 2 to extract image features and improve the quality of subsequent inter-frame alignment.

[0065] S32: By using pyramid-cascaded deformable convolutional layers, i.e. PCD inter-frame alignment, the accuracy of inter-frame alignment is improved, thereby helping to facilitate subsequent fusion SR reconstruction.

[0066] PCD inter-frame alignment is a coarse-to-fine, top-down process. This module is based on a deformable convolutional network (DCN), using multiple deformable convolutional networks and convolutional networks cascaded to form a pyramid structure. In this embodiment, the pyramid has 3 layers. The PCD network structure diagram is shown below. Figure 4 As shown.

[0067] Specifically, S32 includes the following steps:

[0068] S321: Take the two-dimensional core slice sequence images from the same time window (in this example, one reference frame and two support frames), and output their respective feature blocks through convolution as the feature information of the first layer. Among them, F t+i , i∈[-1,1] represents 3 frames of images, where F t The reference frame is used as the reference frame, and the rest are support frames; a convolution operation with a stride of 2 is used to obtain the downsampled feature information of this layer. In one layer s∈[1,S], S=3, feature information of 3 levels is generated, and their image size is attenuated by a factor of 2.

[0069] S322: Starting from the top layer L3, the feature blocks of the 2D core slice reference frame image of this layer are stitched and fused to output the offset of this layer. Then, the offset and the feature blocks of the 2D core slice support frame image of this layer are deformably convolved to output the feature blocks of the aligned 2D core slice reference frame image of this layer. The superscript "3" indicates the third level of the pyramid.

[0070] S323: Feature map of a 2D core slice reference frame after aligning the offset generated in the third layer with that layer. Upsampling is performed using bilinear interpolation and the data is then fed to the next layer, i.e., the second layer. The offset of the second layer is obtained not only from the two input feature blocks of this layer, but also from the offset of the third layer. Furthermore, the aligned feature blocks output by this layer are not only from the output of the deformable convolution of this layer, but also depend on the aligned feature blocks of the third layer. Then, the data is fed to the first layer according to the following formulas (1) and (2) until the feature blocks of the aligned two-dimensional core slice support frame image are output.

[0071]

[0072]

[0073] in(.) ↑2 This indicates that bilinear interpolation is used for upsampling by a factor of 2; DConv(.) indicates deformable convolution; g(.) and h(.) indicate general convolution processes; [.] indicates concat (fusion);

[0074] S324: Outside the pyramid structure, the feature blocks of the first layer's two-dimensional core slice reference frame image and The feature information from the first layer is concatenated and fused to output an offset. Then, this offset is used in conjunction with... The first layer of feature information is used for deformable convolution to output two-dimensional core slices supporting frame feature blocks (F). t+i ) 1 The aligned version, This allows for further adjustments to the alignment feature blocks.

[0075] To fully utilize the temporal relationships between frames and the spatial relationships within frames, this fusion module introduces a temporal attention mechanism in both time and space. The TSA fusion module assigns pixel-level aggregation weights to each frame, improving the effectiveness and efficiency of the fusion process.

[0076] S33: Utilizing inter-frame temporal relationships and intra-frame spatial relationships, TSA fusion is performed. This fusion method assigns pixel-level aggregation weights to each frame, improving the effectiveness and efficiency of the fusion. Specifically, S33 includes the following steps:

[0077] S331: Perform convolutional learning on the input two-dimensional core slice frame image to obtain an embedding space, then calculate the feature information similarity between the input two-dimensional core slice support frame image and the two-dimensional core slice reference frame image in the embedding space, and obtain the feature block after fusion with temporal attention.

[0078] The specific calculation method is as follows: Temporal attention weights (pixel-level) for each frame are obtained by multiplying the weights with the reference frame and then applying a sigmoid function. These weights are then multiplied with three consecutive input frames. The multiplication results are then subjected to continuous convolution to obtain the feature block F fused with the temporal attention. fusion The formula is shown below:

[0079]

[0080]

[0081]

[0082] Where sigmoid(.) indents the weights between 0 and 1 to increase training stability; ⊙ represents the dot product, which is multiplied element by element; [.,.] represents concat; Conv() represents convolution. and They have the same size; they support frame feature information. Reference frame feature information:

[0083] S332: Incorporate the feature block F with temporal attention fusion The system is placed within a pyramid-shaped spatial attention architecture. First, alignment is achieved using two convolutional downsampling operations (Conv) to obtain F0 and F1 respectively. Then, starting from the top layer and working downwards, operations such as upsampling, addition, and dot product are used to obtain the sum F. fusion Spatial attention F2 of the same size is used, and finally, the feature block fused with spatial attention is output by element-wise multiplication, i.e., F. The specific expression of spatial attention is as follows:

[0084] F0 = Conv(F fusion ), F1 = Conv(F fusion (6)

[0085]

[0086]

[0087] in(.) T Indicates upsampling; and F fusion They have the same size.

[0088] S34: The feature block F after spatial and temporal attention fusion is fed into the SR network for reconstruction, and a high-resolution two-dimensional core slice support frame image is reconstructed. The reconstruction method is bicubic interpolation.

[0089] S35: The reconstructed 2D core slice image is passed through an upsampling layer. Subpixel convolution is used to increase the feature block size, generating a high-resolution (HR) image. The predicted image is then concatenated with a 2D core slice reference frame image obtained from the input through upsampling, outputting its super-resolution image. This concatenation acts as a regularization term, forcing the network to learn residual information, i.e., how to reconstruct a higher-quality reference frame image. The goal of this invention is to optimize the MSE loss function to make the super-resolution structure approximate the reference frame image. By training with N-2 samples, the optimal mapping relationship between the low-resolution 2D core slice image and the super-resolution image is learned.

[0090] S4: Based on the mapping relationship, the unfused low-resolution image is reconstructed into a high-resolution image using the EDVR model to obtain the corresponding structural similarity and peak signal-to-noise ratio indices, and then the model is tested.

[0091] Specifically, after training the model sequentially with N-2 samples from the label set, the peak signal-to-noise ratio (PSNR) and structural similarity (SSIM) values ​​obtained were both very high, at 38 and 0.97 respectively, indicating that the super-resolution image generation effect was good and the model successfully trained the mapping relationship from low-resolution images to high-resolution images. Therefore, 2N+2 two-dimensional slice images of coal and rock sample cores were made into 2N sequences, i.e., 2N samples, and sequentially fed into the EDVR model for training, resulting in 2N high-resolution images displaying a larger field of view.

[0092] S5: Use Avico software to perform three-dimensional reconstruction on the two-dimensional slices to obtain a three-dimensional plunger-shaped coal core filled with detailed information about pores and fractures.

[0093] In summary, this invention, based on deep learning, can quickly and accurately obtain high-resolution images with a large field of view. Compared with traditional methods, this invention has the advantages of cost reduction and efficiency improvement, and does not require a large amount of manual operation, thus reducing human error.

[0094] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for multi-scale fusion and super-resolution reconstruction of digital core sequence images of coal and rock, characterized by: Includes the following steps, S1: Obtain the dataset, including low-resolution and high-resolution images; S2: Data preprocessing, performing feature point alignment and fusion of multi-scale images, and creating time series datasets; S3: Input the time series dataset into the training of the sequence image super-resolution reconstruction EDVR model to obtain the mapping relationship between the fused high-resolution image and the original low-resolution image. S3 includes the following steps. S31: Create a time series dataset using a sliding window and input the time series dataset into the model; S32: Inter-frame alignment via pyramid-cascaded deformable convolutional layers, S32 includes the following steps. S321: The two-dimensional core slice sequence images under the same time window are input and their respective feature blocks are output through convolution as the feature information of the first layer; the feature information of the downsampled layer is obtained by using a convolution operation with a stride of 2, and the image size is attenuated by a factor of 2; S322: Starting from the top layer L3, the feature blocks of the two-dimensional core slice reference frame image of this layer are stitched and fused to output the offset of this layer. Then, the offset and the feature blocks of the two-dimensional core slice support frame image of this layer are deformably convolved to output the feature blocks of the two-dimensional core slice reference frame image after alignment of this layer. S323: The offset generated by the third layer and the feature map of the two-dimensional core slice reference frame after alignment with the third layer are upsampled and fed to the second layer using bilinear interpolation. The offset of the second layer is obtained not only from the two input feature blocks of the layer, but also from the offset of the third layer. In addition, the aligned feature blocks output by the second layer are not only from the output of the deformable convolution of the layer, but also depend on the aligned feature blocks of the third layer, and so on, until the feature blocks of the frame image of the aligned two-dimensional core slice are output. S324: Outside the pyramid structure, the feature blocks of the first layer's two-dimensional core slice reference frame image and the feature information of the first layer are spliced ​​and fused to output an offset. Then, this offset and the feature information of the first layer are used to perform deformable convolution to output an aligned version of the two-dimensional core slice support frame feature blocks. S33: Utilize the temporal relationship between frames and the spatial relationship within frames to perform TSA fusion and assign pixel-level aggregation weights to each frame. S34: The feature block F after spatial and temporal attention fusion is fed into the SR network for reconstruction, and a high-resolution two-dimensional core slice support frame image is reconstructed. S35: The reconstructed 2D core slice image is passed through an upsampling layer. The size of the feature block is increased by using subpixel convolution to generate an HR-level image. The predicted image is concatenated with the 2D core slice reference frame image obtained from the input end through upsampling to output its super-resolution image. S4: Based on the mapping relationship, the unfused low-resolution image is reconstructed into a high-resolution image using the EDVR model to obtain the corresponding structural similarity and peak signal-to-noise ratio indices, and the model is tested. S5: Use Avico software to perform three-dimensional reconstruction on the two-dimensional slices to obtain a three-dimensional plunger-shaped coal core filled with detailed information about pores and fractures.

2. The method for multi-scale fusion and super-resolution reconstruction of digital core sequence images of coal and rock as described in claim 1, characterized in that: S1 includes the following steps: S11: Acquire low-resolution images with a pixel diameter of 9µm and high-resolution images with a pixel diameter of 0.3µm through CT scanning; S12: The ratio of the number of high-resolution two-dimensional slice images to the number of low-resolution two-dimensional slice images is 1:

3.

3. The method for multi-scale fusion and super-resolution reconstruction of digital core sequence images of coal and rock as described in claim 1 or 2, characterized in that: S2 includes the following steps: S21: Upsample the low-resolution image at the optimal registration ratio while preserving the original low-resolution image information; S22: Align the upsampled low-resolution image with the high-resolution image using SIFT feature points; S23: The aligned high-resolution image information is fused and reconstructed into a label set in the low-resolution image. A time series dataset is created based on the preserved original low-resolution image samples.

4. The method for multi-scale fusion and super-resolution reconstruction of digital core sequence images of coal and rock as described in claim 1 or 2, characterized in that: S33 Includes the following steps, S331: Perform convolutional learning on the input two-dimensional core slice frame image to obtain an embedding space, then calculate the similarity of feature information between the input two-dimensional core slice support frame image and the two-dimensional core slice reference frame image in the embedding space, and obtain the feature block after fusion with temporal attention. S332: The feature blocks that incorporate temporal attention are placed into a pyramid-shaped spatial attention architecture. First, alignment is performed using two convolutional downsampling operations. Then, starting from the top layer, the feature blocks that combine spatial attention are output from top to bottom through upsampling, addition, and dot product operations.

5. The method for multi-scale fusion and super-resolution reconstruction of digital core sequence images of coal and rock as described in claim 1 or 2, characterized in that: The reconstruction method in S34 is bicubic interpolation.

6. An apparatus, characterized in that: Run the multi-scale fusion and super-resolution reconstruction method for coal and rock digital core sequence images as described in any one of claims 1 to 5.

7. An apparatus comprising a memory, a processor, and an algorithm stored in the memory and executable on the processor, characterized in that: When the processor executes the computer program, it implements the multi-scale fusion and super-resolution reconstruction method for coal and rock digital core sequence images as described in any one of claims 1 to 5.

8. A computer-readable storage medium storing a computer algorithm, characterized in that, When the computer algorithm is executed by the processor, it implements the multi-scale fusion and super-resolution reconstruction method for coal and rock digital core sequence images as described in any one of claims 1 to 5.

Images