A deep learning-based grading method for core fluorescence brightness and area calculation

Abstract

The present invention discloses a grading method for calculating the fluorescence brightness and area of ​​rock cores based on deep learning. The method comprises the following steps: S1, forming a labeled fluorescence region dataset and an unlabeled fluorescence region dataset; S2, preprocessing the rock core fluorescence images in the labeled fluorescence region dataset and the unlabeled fluorescence region dataset; S3, segmenting and labeling the fluorescence regions; S4, constructing a semi-supervised model; S5, quantitatively calculating the brightness and area of ​​the fluorescence regions using an improved deep convolutional neural network model based on the labeled data and pseudo-labeled data in the extended dataset; S6, outputting the fluorescence brightness grade and geological property evaluation results of the rock core; S7, storing the grading results and related calculation data in a database, and generating an analysis report. The present invention can improve the efficiency of large-scale data processing while improving the accuracy of fluorescence brightness and area calculation, and can dynamically adapt to the characteristics of different rock core samples.

Description

Technical Field

[0001] The present invention relates to the field of core technology, and in particular to a grading method for core fluorescence brightness and area calculation based on deep learning. Background Art

[0002] In existing technologies, the calculation of core fluorescence brightness and area mainly relies on traditional manual observation and empirical judgment. Traditional methods usually require geological engineers to evaluate the fluorescence brightness and coverage area of ​​the core surface with the naked eye or with the help of simple physical instruments. In this process, the observation results are easily affected by the observer's subjective factors, resulting in poor accuracy and consistency of the results. Manual observation is inefficient and not scalable in large-scale resource assessment.

[0003] In order to improve the accuracy and consistency of calculations, some technical solutions have introduced automated image processing technology. Automated image processing technology usually uses the brightness and contrast information of fluorescence images combined with some basic image segmentation algorithms to perform preliminary calculations of the brightness and area of ​​core fluorescence. However, existing image processing methods still have obvious technical defects when dealing with complex image backgrounds and refined identification of fluorescent areas. On the one hand, simple image segmentation algorithms are difficult to cope with the strong interference between fluorescent areas and backgrounds, especially in core samples with more complex fluorescence brightness changes. Over-segmentation or missed segmentation is prone to occur, resulting in large errors in the calculation of the brightness and area of ​​the fluorescent area. On the other hand, existing image processing algorithms lack the ability to process large-scale data sets. When processing a large number of unlabeled core images, the algorithm's generalization ability is insufficient, which can easily lead to inaccurate grading results.

[0004] In addition, existing technologies also have shortcomings in the classification of fluorescence brightness levels and the evaluation of geological characteristics. Most solutions rely on preset brightness and area thresholds for simple classification and evaluation, and are unable to dynamically adjust according to the diversity of fluorescent areas, resulting in low reliability of the results. At the same time, existing technologies are difficult to adapt to the geological differences of different core samples, resulting in a lack of sufficient flexibility and adaptability in practical applications.

[0005] In summary, the existing technologies have the following main shortcomings: First, manual observation and simple image processing techniques are highly subjective and lack precision; second, existing image segmentation algorithms have poor robustness in complex backgrounds, resulting in large errors in the calculation of fluorescence brightness and area; finally, existing fluorescence brightness grading and geological property evaluation methods are difficult to adapt to the different characteristics of core samples and lack the ability to dynamically adjust, which affects their application in large-scale resource assessment. Summary of the Invention

[0006] One purpose of the present invention is to propose a deep learning-based grading method for core fluorescence brightness and area calculation. The present invention can improve the accuracy of fluorescence brightness and area calculation while improving the efficiency of large-scale data processing and can dynamically adapt to the characteristics of different core samples.

[0007] A deep learning-based core fluorescence brightness and area calculation grading method according to an embodiment of the present invention includes the following steps:

[0008] S1. Use imaging equipment to acquire multi-angle and multi-resolution core fluorescence images to form a dataset of labeled fluorescence areas and a dataset of unlabeled fluorescence areas;

[0009] S2, preprocessing the core fluorescence images in the labeled fluorescence area dataset and the unlabeled fluorescence area dataset;

[0010] S3. Use the pre-trained improved deep convolutional neural network model to analyze the pre-processed labeled fluorescence region dataset, automatically identify the fluorescent regions in the core fluorescence image, and segment and label the fluorescent regions;

[0011] S4. Build a semi-supervised model to predict different enhanced versions of the same unlabeled core fluorescence image using the consistency regularization technique for the unlabeled fluorescence region dataset. Apply a consistency loss to ensure that the prediction results of the core fluorescence image under different enhancement conditions remain consistent. Generate pseudo labels for the unlabeled data, and merge the pseudo-label data with the labeled fluorescence region dataset to form an extended dataset.

[0012] S5. Based on the labeled data and pseudo-labeled data in the extended dataset, the brightness and area of ​​the fluorescent area are quantified using an improved deep convolutional neural network model.

[0013] S6. Input the calculated results of fluorescence brightness and area into the trained semi-supervised deep learning model, grade the extended data set through the semi-supervised deep learning model, and output the fluorescence brightness level and geological property evaluation results of the core;

[0014] S7. The grading results and related calculation data are stored in a database, and an analysis report is generated. The report includes information on fluorescence brightness, area, and grading results.

[0015] Optionally, the S1 step specifically includes:

[0016] S11. Use a multi-angle rotating imaging device to obtain a core fluorescence image. The multi-angle rotating imaging device captures the core fluorescence images at different angles and integrates them into a core fluorescence image set I. m,n , m represents the imaging angle of the core fluorescence image, and n represents the core fluorescence images with different resolutions;

[0017] S12. Label the acquired core fluorescence image set to form a labeled fluorescence region dataset:

[0018] D l ={(I m,n ,L m,n )|m∈M,n∈N};

[0019] Among them, I m,n is the core fluorescence image, L m,n is the label of the fluorescence area corresponding to the core fluorescence image, M is the imaging angle set of the core fluorescence image, and N is the resolution set;

[0020] S13. Based on the acquired core fluorescence image set, construct an unlabeled fluorescence area dataset:

[0021] D u ={I m,n |m∈M,n∈N}.

[0022] Optionally, the step S2 specifically includes:

[0023] S21. Use Gaussian filtering algorithm to annotate the fluorescence area dataset D l and the unlabeled fluorescence region dataset D u The core fluorescence images in the denoising process are processed to obtain the denoised core fluorescence image set I′ m,n ;

[0024] S22, the denoised core fluorescence image set I′ m,n Brightness and contrast are adjusted, and the contrast of the core fluorescence image is optimized using the histogram equalization method to make the fluorescent area in the core fluorescence image clear. The adjusted core fluorescence image set is:

[0025] I″ m,n =HE(I′ m,n );

[0026] Among them, HE represents the histogram equalization operation, I″ m,n represents a set of core fluorescence images after brightness and contrast adjustment;

[0027] S23, adjusting the core fluorescence image set I″ m,n Perform image enhancement processing to obtain the enhanced core fluorescence image set I″′ m,n ;

[0028] S24. Output the preprocessed labeled fluorescence region dataset:

[0029] D′ l ={(I″′ m,n,L m,n )|m∈M,n∈N};

[0030] Output the preprocessed unlabeled fluorescence area dataset:

[0031] D′ u ={I″′ m,n |m∈M,n∈N}.

[0032] Optionally, the S3 step specifically includes:

[0033] S31. Analyze the pre-processed labeled fluorescence region dataset using a pre-trained improved deep convolutional neural network model. The improved deep convolutional neural network model includes a fluorescence region guided attention mechanism module. The attention mechanism module generates a weight matrix A by identifying the region with concentrated fluorescence brightness in the core fluorescence image. m,n , the weight matrix emphasizes the areas with large changes in fluorescence brightness, so that the convolution operation focuses on the fluorescent area and ignores background interference:

[0034]

[0035] Among them, A m,n (x,y) represents the weight value at the coordinate (x,y), I″′ m,n (x, y) is the brightness value of the core fluorescence image after preprocessing, G(x, y) is the local gradient of the fluorescence area, Z m,n is the normalization factor, the sum of all weights in the fluorescent area is 1;

[0036] S32, the core fluorescence image I″′ m,n Input the improved deep convolutional neural network model and extract the preliminary feature representation F m,n :

[0037]

[0038] Among them, F m,n (x,y) represents the eigenvalue at the coordinate (x,y), W i k ,j is the weight of convolution kernel k, b k is the bias term;

[0039] S33, the attention mechanism module guided by the fluorescent area for the preliminary feature representation F m,n Perform weighted processing to generate weighted feature map F′ m,n :

[0040] F′ m,n (x,y)=A m,n (x,y)·F m,n (x,y);

[0041] Among them, A m,n (x,y) is the fluorescence area weight, F m,n (x,y) is the initial feature representation, F′ m,n (x,y) represents the eigenvalue after weighting by the weight matrix;

[0042] S34, based on weighted feature map F' m,n By improving the segmentation module in the deep convolutional neural network model, the fluorescent area is segmented using the pixel-level segmentation algorithm to generate the segmented fluorescent area image S m,n , the segmentation process is optimized by the energy minimization function:

[0043] E(S m,n )=∑ x,y ((S m,n (x,y)-F′ m,n (x,y)) 2 +λ∑ (x′,y′)∈N(x,y) |S m,n (x,y)-S m,n (x′,y′)|);

[0044] Among them, S m,n (x, y) represents the value of the segmented fluorescence region image at the coordinate (x, y), F′ m,n (x, y) is the weighted feature representation, N(x, y) is the neighborhood pixel set, and λ is the smoothness parameter;

[0045] S35, the segmented fluorescence area image S m,n Perform annotation and match the segmentation results with the annotated fluorescence area dataset to generate an annotated fluorescence area dataset:

[0046]

[0047] Among them, δ(S m,n (x, y), T) is the decision function, which divides the segmented fluorescence area S into m,n Mark, L m,n is the generated annotation label matrix, D″ l This is the final annotated fluorescence region dataset.

[0048] Optionally, the S4 step specifically includes:

[0049] S41. Build a semi-supervised deep learning model to annotate the fluorescence area dataset D″ l and the unlabeled fluorescence region dataset D′ uFor training, the semi-supervised model includes a main network and an auxiliary network. The main network is used to classify and segment the input core fluorescence image, and the auxiliary network is used for consistency regularization.

[0050] S42. For the unlabeled fluorescence region dataset D′ u Core fluorescence image I″′ m,n , generate multiple different enhanced versions Each enhanced version is generated by different core fluorescence image transformation operations, including rotation, scaling and noise addition;

[0051] S43. Input the enhanced core fluorescence image version into the main network of the semi-supervised model for prediction and obtain the corresponding prediction results. Represents the prediction result of the i-th enhanced version:

[0052]

[0053] Among them, P i m,n (x,y) represents the predicted output of the i-th enhanced version at the coordinate (x,y); f θ is the activation function of the main network, and the parameter θ represents the weight and bias of the network; is the weight of the convolution kernel applied to the pixel position (p,q) in the i-th enhanced version; I″′ i (xp,yq) m,n is the input image value of the pixel position (xp, yq) in the i-th enhanced version;

[0054] S44, in the consistency regularization process by applying the consistency loss function L cons To keep the prediction results of different enhanced versions consistent, the loss function is defined as:

[0055]

[0056] in, represents the average prediction result of all enhanced versions, k is the number of enhanced versions;

[0057] S45. Generate pseudo labels for the unlabeled fluorescence region dataset based on the prediction results after consistency regularization

[0058]

[0059] Among them, arg max is a function that selects the maximum value in the prediction result, which is used to determine the category of the pseudo label;

[0060] S46. Generate pseudo-label dataset and the labeled fluorescence region dataset D″ l Merge to form an extended data set:

[0061]

[0062] Optionally, the step S5 specifically includes:

[0063] S51, based on the labeled fluorescence area dataset D″ l and pseudo-labeled data The brightness B of the core fluorescence area is analyzed using the improved convolutional neural network model. m,n (x,y) and area A m,n Perform quantitative calculation, brightness B m,n (x,y) is the brightness value of the pixel (x,y), area A m,n is the total number of pixels in the fluorescent area;

[0064] S52, using the improved convolutional neural network model to input the core fluorescence image I″′ m,n Perform pixel-by-pixel calculations to obtain the brightness prediction value of each pixel

[0065]

[0066] in, is the weight of the brightness prediction convolution kernel, b b is the bias term, is the brightness prediction value at pixel (x, y);

[0067] S53, the brightness value B of each fluorescent area m,n (x,y) is accumulated to calculate the total brightness of the fluorescent area

[0068]

[0069] in, It represents the total brightness value of the fluorescent area in the core fluorescence image;

[0070] S54. Calculate the area A based on the total number of pixels in the fluorescent area. m,n :

[0071]

[0072] in, is the indicator function, when the pixel brightness When the value exceeds the set threshold T, it is recorded as 1, otherwise it is recorded as 0, which is used to count the total area of ​​the fluorescent area.

[0073] Optionally, the step S6 specifically includes:

[0074] S61, the fluorescence brightness B m,n (x,y) and fluorescence area A m,n The calculation results are input into the trained semi-supervised deep learning model, and the semi-supervised deep learning model is used to calculate the extended dataset D ext Perform grading and output the fluorescence brightness level and geological characteristics evaluation results of the core samples;

[0075] S62, by improving the fully connected layer of the deep convolutional neural network model to measure the fluorescence brightness B m,n (x,y) and area A m,n Perform feature fusion to generate feature vector F m,n :

[0076] F m,n =αB m,n (x,y)+βA m,n ;

[0077] Among them, α and β are the weight coefficients of brightness and area, respectively, which are used to adjust the relative importance between features. m,n is the fused feature vector;

[0078] S63, the fused feature vector F m,n Passed to the classification layer, the trained semi-supervised model outputs the fluorescence brightness level G m,n :

[0079]

[0080] Among them, G m,n is the predicted fluorescence brightness level, P represents the number of feature maps, and W (p) is the weight matrix, is the pth feature map value, γ is the adjustment coefficient, Indicates the binding area A m,n and brightness B m,n The nonlinear function, b (p) is the bias term, σ is the activation function;

[0081] S64, conduct a comprehensive evaluation of geological characteristics, according to the fluorescence brightness level G m,n and the fluorescence area A m,n The calculation results generate the geological characteristics evaluation results R m,n :

[0082]

[0083] Among them, R m,n is the geological characteristic evaluation result, Q is the characteristic dimension, λ is the weight attenuation factor, Represents the gradient change of area in the feature dimension, represents the influence function of brightness level, λ is the control factor, Ψ(G m,n ,A m,n ) represents the comprehensive influence of brightness and area.

[0084] S65, the fluorescence brightness level of the core is G m,n and geological characteristics evaluation results R m,n The output is the final result, which is used for resource assessment and geological analysis.

[0085] The beneficial effects of the present invention are:

[0086] (1) The present invention introduces an improved deep convolutional neural network model. In the process of calculating fluorescence brightness and area, a pixel-by-pixel refined calculation method is adopted to accurately identify the fluorescent area under complex background. The attention mechanism is used to guide the convolution operation. The improved deep convolutional neural network model can perform weighted processing on the fluorescent area with obvious brightness changes, ignore background interference, and ensure the accuracy and consistency of calculation. Compared with traditional image segmentation algorithms, this method can better process core images with unclear brightness contrast, significantly improve the quantization accuracy of fluorescence brightness and area, and ensure robustness in different complex scenes.

[0087] (2) The present invention adopts a semi-supervised deep learning model and generates pseudo labels through consistency regularization technology, so that the semi-supervised deep learning model can make full use of unlabeled data during training. Unlike traditional fully supervised learning that relies on large-scale labeled data, the present invention ensures the consistency of prediction results under different enhancement conditions by constraining the enhanced version of unlabeled data, thereby generating high-quality pseudo labels, greatly improving data utilization, and reducing dependence on manually labeled data. This pseudo-label generation strategy can improve the generalization ability of the model and still maintain a high grading accuracy when processing unlabeled core data.

[0088] (3) The present invention introduces a feature fusion mechanism to dynamically fuse the calculation results of fluorescence brightness and fluorescence area, and combines it with a semi-supervised learning model to grade the geological characteristics of core samples. Through the feature weighting mechanism, it can adaptively adjust the importance of brightness and area in geological characteristic evaluation, and use nonlinear functions to deeply fuse these features to generate more accurate core fluorescence brightness levels and geological characteristic evaluation results. Different from the traditional fixed threshold division method, the present invention can flexibly divide and evaluate the grades according to the diversity of core samples, so that it can provide a more reliable basis for geological analysis in practical applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0089] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0090] Figure 1 This is a flowchart of a deep learning-based core fluorescence brightness and area calculation grading method proposed by the present invention;

[0091] Figure 2 This is a schematic diagram of using a convolutional neural network to segment the fluorescence area in a deep learning-based core fluorescence brightness and area calculation grading method proposed in the present invention. DETAILED DESCRIPTION

[0092] The present invention will now be described in further detail with reference to the accompanying drawings, which are simplified schematic diagrams that illustrate the basic structure of the present invention in a schematic manner.

[0093] refer to Figure 1-2 A deep learning-based core fluorescence brightness and area calculation method is proposed, comprising the following steps:

[0094] S1. Use imaging equipment to acquire multi-angle and multi-resolution core fluorescence images to form a dataset of labeled fluorescence areas and a dataset of unlabeled fluorescence areas;

[0095] S2, preprocessing the core fluorescence images in the labeled fluorescence area dataset and the unlabeled fluorescence area dataset;

[0096] S3. Use the pre-trained improved deep convolutional neural network model to analyze the pre-processed labeled fluorescence region dataset, automatically identify the fluorescent regions in the core fluorescence image, and segment and label the fluorescent regions;

[0097] S4. Build a semi-supervised model to predict different enhanced versions of the same unlabeled core fluorescence image using the consistency regularization technique for the unlabeled fluorescence region dataset. Apply a consistency loss to ensure that the prediction results of the core fluorescence image under different enhancement conditions remain consistent. Generate pseudo labels for the unlabeled data, and merge the pseudo-label data with the labeled fluorescence region dataset to form an extended dataset.

[0098] S5. Based on the labeled data and pseudo-labeled data in the extended dataset, the brightness and area of ​​the fluorescent area are quantified using an improved deep convolutional neural network model.

[0099] S6. Input the calculated results of fluorescence brightness and area into the trained semi-supervised deep learning model, grade the extended data set through the semi-supervised deep learning model, and output the fluorescence brightness level and geological property evaluation results of the core;

[0100] S7. The grading results and related calculation data are stored in a database, and an analysis report is generated. The report includes information on fluorescence brightness, area, and grading results.

[0101] In this embodiment, step S1 specifically includes:

[0102] S11. Use a multi-angle rotating imaging device to obtain a core fluorescence image. The multi-angle rotating imaging device captures the core fluorescence images at different angles and integrates them into a core fluorescence image set I. m,n , m represents the imaging angle of the core fluorescence image, and n represents the core fluorescence images with different resolutions;

[0103] S12. Label the acquired core fluorescence image set to form a labeled fluorescence region dataset:

[0104] D l ={(I m,n ,L m,n )|m∈M,n∈N};

[0105] Among them, I m,n is the core fluorescence image, L m,n is the label of the fluorescence area corresponding to the core fluorescence image, M is the imaging angle set of the core fluorescence image, and N is the resolution set;

[0106] S13. Based on the acquired core fluorescence image set, construct an unlabeled fluorescence area dataset:

[0107] D u ={I m,n |m∈M,n∈N}.

[0108] In this embodiment, step S2 specifically includes:

[0109] S21. Use Gaussian filtering algorithm to annotate the fluorescence area dataset D l and the unlabeled fluorescence region dataset D u The core fluorescence images in the denoising process are processed to obtain the denoised core fluorescence image set I′ m,n ;

[0110] S22, the denoised core fluorescence image set I′ m,n Brightness and contrast are adjusted, and the contrast of the core fluorescence image is optimized using the histogram equalization method to make the fluorescent area in the core fluorescence image clear. The adjusted core fluorescence image set is:

[0111] I″ m,n =HE(I′ m,n );

[0112] Among them, HE represents the histogram equalization operation, I″ m,n represents a set of core fluorescence images after brightness and contrast adjustment;

[0113] S23, adjusting the core fluorescence image set I″ m,n Perform image enhancement processing to obtain the enhanced core fluorescence image set I″′ m,n ;

[0114] S24. Output the preprocessed labeled fluorescence region dataset:

[0115] D′ l ={(I″′ m,n ,L m,n )|m∈M,n∈N};

[0116] Output the preprocessed unlabeled fluorescence area dataset:

[0117] D′ u ={I″′ m,n |m∈M,n∈N}.

[0118] In this embodiment, step S3 specifically includes:

[0119] S31. Analyze the pre-processed labeled fluorescence region dataset using a pre-trained improved deep convolutional neural network model. The improved deep convolutional neural network model includes a fluorescence region guided attention mechanism module. The attention mechanism module generates a weight matrix A by identifying the region with concentrated fluorescence brightness in the core fluorescence image. m,n , the weight matrix emphasizes the areas with large changes in fluorescence brightness, so that the convolution operation focuses on the fluorescent area and ignores background interference:

[0120]

[0121] Among them, A m,n (x,y) represents the weight value at the coordinate (x,y), I″′ m,n (x, y) is the brightness value of the core fluorescence image after preprocessing, G(x, y) is the local gradient of the fluorescence area, Z m,n is the normalization factor, the sum of all weights in the fluorescent area is 1;

[0122] S32, the core fluorescence image I″′ m,n Input the improved deep convolutional neural network model and extract the preliminary feature representation F m,n :

[0123]

[0124] Among them, F m,n(x,y) represents the eigenvalue at the coordinate (x,y), is the weight of convolution kernel k, b k is the bias term;

[0125] S33, the attention mechanism module guided by the fluorescent area for the preliminary feature representation F m,n Perform weighted processing to generate weighted feature map F′ m,n :

[0126] F′ m,n (x,y)=A m,n (x,y)·F m,n (x,y);

[0127] Among them, A m,n (x,y) is the fluorescence area weight, F m,n (x,y) is the initial feature representation, F′ m,n (x,y) represents the eigenvalue after weighting by the weight matrix;

[0128] S34, based on weighted feature map F' m,n By improving the segmentation module in the deep convolutional neural network model, the fluorescent area is segmented using the pixel-level segmentation algorithm to generate the segmented fluorescent area image S m,n , the segmentation process is optimized by the energy minimization function:

[0129] E(S m,n )=∑ x,y ((S m,n (x,y)-F′ m,n (x,y)) 2 +λ∑ (x′,y′)∈N(x,y) |S m,n (x,y)-S m,n (x′,y′)|);

[0130] Among them, S m,n (x, y) represents the value of the segmented fluorescence region image at the coordinate (x, y), F′ m,n (x, y) is the weighted feature representation, N(x, y) is the neighborhood pixel set, and λ is the smoothness parameter;

[0131] S35, the segmented fluorescence area image S m,n Perform annotation and match the segmentation results with the annotated fluorescence area dataset to generate an annotated fluorescence area dataset:

[0132]

[0133] Among them, δ(S m,n(x, y), T) is the decision function, which divides the segmented fluorescence area S into m,n Mark, L m,n is the generated annotation label matrix, D″ l This is the final annotated fluorescence region dataset.

[0134] In this embodiment, step S4 specifically includes:

[0135] S41. Build a semi-supervised deep learning model to annotate the fluorescence area dataset D″ l and the unlabeled fluorescence region dataset D′ u For training, the semi-supervised model includes a main network and an auxiliary network. The main network is used to classify and segment the input core fluorescence image, and the auxiliary network is used for consistency regularization.

[0136] S42. For the unlabeled fluorescence region dataset D′ u Core fluorescence image I″′ m,n , generate multiple different enhanced versions Each enhanced version is generated by different core fluorescence image transformation operations, including rotation, scaling and noise addition;

[0137] S43. Input the enhanced core fluorescence image version into the main network of the semi-supervised model for prediction and obtain the corresponding prediction results. Represents the prediction result of the i-th enhanced version:

[0138]

[0139] in, Represents the predicted output of the i-th enhanced version at coordinate (x, y); f θ is the activation function of the main network, and the parameter θ represents the weight and bias of the network; is the weight of the convolution kernel applied to the pixel position (p,q) in the i-th enhanced version; I″′ i (xp,yq) m,n is the input image value of the pixel position (xp, yq) in the i-th enhanced version;

[0140] S44, in the consistency regularization process by applying the consistency loss function L cons To keep the prediction results of different enhanced versions consistent, the loss function is defined as:

[0141]

[0142] in, represents the average prediction result of all enhanced versions, k is the number of enhanced versions;

[0143] S45. Generate pseudo labels for the unlabeled fluorescence region dataset based on the prediction results after consistency regularization

[0144]

[0145] Among them, arg max is a function that selects the maximum value in the prediction result, which is used to determine the category of the pseudo label;

[0146] S46. Generate pseudo-label dataset and the labeled fluorescence region dataset D″ l Merge to form an extended data set:

[0147]

[0148] In this embodiment, step S5 specifically includes:

[0149] S51, based on the labeled fluorescence area dataset D″ l and pseudo-labeled data The brightness B of the core fluorescence area is analyzed using the improved convolutional neural network model. m,n (x,y) and area A m,n Perform quantitative calculation, brightness B m,n (x,y) is the brightness value of the pixel (x,y), area A m,n is the total number of pixels in the fluorescent area;

[0150] S52, using the improved convolutional neural network model to input the core fluorescence image I″′ m,n Perform pixel-by-pixel calculations to obtain the brightness prediction value of each pixel

[0151]

[0152] in, is the weight of the brightness prediction convolution kernel, b b is the bias term, is the brightness prediction value at pixel (x, y);

[0153] S53, the brightness value B of each fluorescent area m,n (x,y) is accumulated to calculate the total brightness of the fluorescent area

[0154]

[0155] in, It represents the total brightness value of the fluorescent area in the core fluorescence image;

[0156] S54. Calculate the area A based on the total number of pixels in the fluorescent area. m,n :

[0157]

[0158] in, is the indicator function, when the pixel brightness When the value exceeds the set threshold T, it is recorded as 1, otherwise it is recorded as 0, which is used to count the total area of ​​the fluorescent area.

[0159] In this embodiment, step S6 specifically includes:

[0160] S61, the fluorescence brightness B m,n (x,y) and fluorescence area A m,n The calculation results are input into the trained semi-supervised deep learning model, and the semi-supervised deep learning model is used to calculate the extended dataset D ext Perform grading and output the fluorescence brightness level and geological characteristics evaluation results of the core samples;

[0161] S62, by improving the fully connected layer of the deep convolutional neural network model to measure the fluorescence brightness B m,n (x,y) and area A m,n Perform feature fusion to generate feature vector F m,n :

[0162] F m,n =αB m,n (x,y)+βA m,n ;

[0163] Among them, α and β are the weight coefficients of brightness and area, respectively, which are used to adjust the relative importance between features. m,n is the fused feature vector;

[0164] S63, the fused feature vector F m,n Passed to the classification layer, the trained semi-supervised model outputs the fluorescence brightness level G m,n :

[0165]

[0166] Among them, G m,n is the predicted fluorescence brightness level, P represents the number of feature maps, and W (p) is the weight matrix, is the pth feature map value, γ is the adjustment coefficient, Indicates the binding area A m,n and brightness B m,n The nonlinear function, b (p) is the bias term, σ is the activation function;

[0167] S64, conduct a comprehensive evaluation of geological characteristics, according to the fluorescence brightness level G m,n and the fluorescence area A m,n The calculation results generate the geological characteristics evaluation results R m,n :

[0168]

[0169] Among them, R m,n is the geological characteristic evaluation result, Q is the characteristic dimension, λ is the weight attenuation factor, Represents the gradient change of area in the feature dimension, represents the influence function of brightness level, λ is the control factor, Ψ(G m,n ,A m,n ) represents the comprehensive influence of brightness and area.

[0170] S65, the fluorescence brightness level of the core is G m,n and geological characteristics evaluation results R m,n The output is the final result, which is used for resource assessment and geological analysis.

[0171] Example 1:

[0172] Example 1 is applied to a large oil field in western China. On September 10, 2023, during resource exploration at a large oil field in western China, a geological team analyzed the fluorescence brightness and area of ​​500 core samples collected. The core samples were obtained by drilling equipment from a reservoir at a depth of 3,500 meters. The fluorescence brightness of the core samples in the reservoir varies complexly and is subject to background noise interference. Traditional methods have difficulty accurately segmenting the fluorescent areas. The geological engineering team decided to use the deep learning-based core fluorescence brightness and area calculation and grading method of the present invention to automatically analyze these core samples.

[0173] On September 12, 2023, the geological team used high-precision imaging equipment to image the fluorescent area of ​​the core at different angles and resolutions, generating an initial core fluorescence image dataset. Through the multi-angle rotation function of the imaging equipment, the geological team captured fluorescence images of each core under different lighting conditions. The data volume reached 1TB, and a preliminary labeled dataset and an unlabeled dataset were formed. Geological engineers found that the brightness changes in the fluorescent area of ​​the core were more obvious, but due to the presence of background noise, traditional image processing tools found it difficult to accurately identify the fluorescent area.

[0174] To improve recognition accuracy, the team decided to input the data into the deep convolutional neural network model of the present invention for analysis. By preprocessing the images of 500 core samples, the system removed background noise and performed contrast enhancement, making the boundaries of the fluorescent area clearer. The preprocessed core fluorescence images were input into the improved deep convolutional neural network model, which used the attention mechanism guided by the fluorescent area to successfully identify and segment the fluorescent area in the core.

[0175] On September 15, 2023, the system generated pseudo-labeled data for 100 unlabeled core samples. These pseudo-labels were generated using consistency regularization techniques, ensuring consistent fluorescence brightness and area calculations under varying image enhancement conditions. This pseudo-labeled data was then fused with the annotated dataset, expanding the dataset to 1,200 core samples.

[0176] That same day, the geological team trained a semi-supervised deep learning model using the expanded dataset. Training took 36 hours on the GPU cluster at the oilfield data center. After completion, the system began automatically calculating and grading fluorescence brightness and area on 300 newly collected core samples.

[0177] At 9:00 AM on September 17, the team selected core sample numbered "XY00123" for detailed analysis. The fluorescence image of this sample showed significant brightness areas, but also significant noise interference in the background area. Through processing with the improved convolutional neural network model, the fluorescent area was accurately segmented, and the total fluorescence brightness was calculated to be 3500 cd / m 2 , the fluorescent area is 200mm 2 The system further inputs these calculation results into a semi-supervised learning model for grading, and generates a report with a fluorescence brightness level of "A2," indicating that the core sample has moderate oil and gas content.

[0178] On September 18, the geological team graded the core sample numbered "XY00456". The fluorescence area of ​​this sample was relatively complex and the brightness variation range was large. Through automated calculation, the system concluded that the fluorescence brightness of this sample was 4900cd / m 2 , area is 310mm 2 , and was finally graded as “A1”, indicating that it has a high oil and gas content.

[0179] To verify the effectiveness of the method of the present invention, the geological team compared the fluorescence brightness and area grading results of 50 core samples with those of traditional manual methods. The following Table 1 shows the comparative data:

[0180] Table 1 Partial comparison data of the present invention and the traditional manual method

[0181]

[0182] As can be seen from the data in Table 1 above, the traditional method takes an average of 27.7 minutes to grade each sample, while the method of the present invention takes only 1.23 minutes, significantly improving efficiency. Furthermore, the grading consistency of the traditional manual method is 77.7%, while that of the method of the present invention reaches 97%. The calculated errors of fluorescence brightness and area are reduced to ±1.5% and ±2%, respectively, far below the error range of traditional methods. For sample number "XY00456," for example, the fluorescence brightness error obtained by the traditional manual method is ±10%, while the error of the method of the present invention is only ±2%. In area calculation, the error of the traditional method is ±14%, while that of the method of the present invention is only ±2.5%. These data clearly demonstrate the significant improvement in the accuracy of fluorescence brightness and area calculation achieved by the present invention.

[0183] The application of this invention demonstrates a high level of automation and data processing capabilities in oilfield exploration. In particular, it can quickly and accurately segment fluorescent areas when processing complex fluorescence images and generate reliable fluorescence brightness and area calculations. By applying a semi-supervised deep learning model, the system can generate high-quality pseudo-labels even with large amounts of unlabeled data, significantly improving data utilization and grading accuracy. The successful application in this oilfield project provides reliable technical support for the geological team in subsequent exploration, significantly enhancing the efficiency and accuracy of resource assessment.

[0184] Through the simulation process of this embodiment, we can see the strong application potential of the present invention in large-scale core analysis. It solves the problems existing in traditional methods such as large subjective judgment errors, poor grading consistency, and low processing efficiency, and provides a strong technical guarantee for reservoir evaluation in oil and gas exploration.

[0185] The present invention introduces an improved deep convolutional neural network model. During the calculation of fluorescence brightness and area, a pixel-by-pixel refined calculation method is adopted to accurately identify fluorescent areas in complex backgrounds. The attention mechanism is used to guide the convolution operation. The improved deep convolutional neural network model can perform weighted processing on fluorescent areas with obvious brightness changes, ignore background interference, and ensure the accuracy and consistency of calculation. Compared with traditional image segmentation algorithms, this method can better process core images with unclear brightness contrast, significantly improve the quantization accuracy of fluorescence brightness and area, and ensure robustness in different complex scenes.

[0186] The present invention adopts a semi-supervised deep learning model and generates pseudo labels through consistency regularization technology, so that the semi-supervised deep learning model can fully utilize unlabeled data during training. Unlike traditional fully supervised learning that relies on large-scale labeled data, the present invention ensures the consistency of prediction results under different enhancement conditions by constraining the enhanced version of the unlabeled data, thereby generating high-quality pseudo labels, greatly improving data utilization and reducing dependence on manually labeled data. This pseudo-label generation strategy can improve the generalization ability of the model and still maintain a high level of grading accuracy when processing unlabeled core data.

[0187] The present invention introduces a feature fusion mechanism to dynamically fuse the calculation results of fluorescence brightness and fluorescence area, and combines it with a semi-supervised learning model to grade the geological characteristics of core samples. Through the feature weighting mechanism, it can adaptively adjust the importance of brightness and area in geological characteristic evaluation, and use nonlinear functions to deeply fuse these features to generate more accurate core fluorescence brightness levels and geological characteristic evaluation results. Unlike the traditional fixed threshold classification method, the present invention can flexibly classify and evaluate core samples according to their diversity, so that it can provide a more reliable basis for geological analysis in practical applications.

[0188] The above description is only a preferred specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any technician familiar with the technical field, within the technical scope disclosed by the present invention, who makes equivalent replacements or changes based on the technical solution and inventive concept of the present invention, should be covered by the scope of protection of the present invention.

Claims

1. A deep learning-based core fluorescence brightness and area calculation grading method, characterized in that: The steps include: S1. Use imaging equipment to acquire multi-angle and multi-resolution core fluorescence images to form a dataset of labeled fluorescence areas and a dataset of unlabeled fluorescence areas; S2, preprocessing the core fluorescence images in the labeled fluorescence area dataset and the unlabeled fluorescence area dataset; S3. Use the pre-trained improved deep convolutional neural network model to analyze the pre-processed labeled fluorescence region dataset, automatically identify the fluorescent regions in the core fluorescence image, and segment and label the fluorescent regions; S4. Build a semi-supervised model to predict different enhanced versions of the same unlabeled core fluorescence image using the consistency regularization technique for the unlabeled fluorescence region dataset. Apply a consistency loss to ensure that the prediction results of the core fluorescence image under different enhancement conditions remain consistent. Generate pseudo labels for the unlabeled data, and merge the pseudo-label data with the labeled fluorescence region dataset to form an extended dataset. S5. Based on the labeled data and pseudo-labeled data in the extended dataset, the brightness and area of ​​the fluorescent area are quantified using an improved deep convolutional neural network model. S6. Input the calculated results of fluorescence brightness and area into the trained semi-supervised deep learning model, grade the extended data set through the semi-supervised deep learning model, and output the fluorescence brightness level and geological property evaluation results of the core; S7. The grading results and related calculation data are stored in a database, and an analysis report is generated. The report includes information on fluorescence brightness, area, and grading results.

2. The method for grading core fluorescence brightness and area calculation based on deep learning according to claim 1 is characterized in that: The S1 step specifically includes: S11. Use a multi-angle rotating imaging device to obtain a core fluorescence image. The multi-angle rotating imaging device captures the core fluorescence images at different angles and integrates them into a core fluorescence image set I. m,n , m represents the imaging angle of the core fluorescence image, and n represents the core fluorescence images with different resolutions; S12. Label the acquired core fluorescence image set to form a labeled fluorescence region dataset: D l ={(I m,n ,L m,n )|m∈M,n∈N}; Among them, I m,n is the core fluorescence image, L m,n is the label of the fluorescence area corresponding to the core fluorescence image, M is the imaging angle set of the core fluorescence image, and N is the resolution set; S13. Based on the acquired core fluorescence image set, construct an unlabeled fluorescence area dataset: D u ={I m,n |m∈M,n∈N}。 3. The method for grading core fluorescence brightness and area calculation based on deep learning according to claim 1 is characterized in that: The S2 step specifically includes: S21, using Gaussian filtering algorithm to annotate the fluorescent area dataset D l and the unlabeled fluorescence region dataset D u The core fluorescence images in the denoising process are processed to obtain the denoised core fluorescence image set I′ m,n ; S22, the denoised core fluorescence image set I′ m,n Brightness and contrast are adjusted, and the contrast of the core fluorescence image is optimized using the histogram equalization method to make the fluorescent area in the core fluorescence image clear. The adjusted core fluorescence image set is: I″ m,n =HE(I′ m,n ); Among them, HE represents the histogram equalization operation, I″ m,n represents a set of core fluorescence images after brightness and contrast adjustment; S23, adjusting the core fluorescence image set I″ m,n Perform image enhancement processing to obtain the enhanced core fluorescence image set I″′ m,n ; S24. Output the preprocessed labeled fluorescence region dataset: D′ l ={(I″′ m,n ,L m,n )|m∈M,n∈N}; Output the preprocessed unlabeled fluorescence area dataset: D′ u ={I″′ m,n |m∈M,n∈N}。 4. The method for grading core fluorescence brightness and area calculation based on deep learning according to claim 1 is characterized in that: The S3 step specifically includes: S31. Analyze the pre-processed labeled fluorescence region dataset using a pre-trained improved deep convolutional neural network model. The improved deep convolutional neural network model includes a fluorescence region guided attention mechanism module. The attention mechanism module generates a weight matrix A by identifying the region with concentrated fluorescence brightness in the core fluorescence image. m,n , the weight matrix emphasizes the areas with large changes in fluorescence brightness, so that the convolution operation focuses on the fluorescent area and ignores background interference: Among them, A m,n (x,y) represents the weight value at the coordinate (x,y), I″′ m,n (x, y) is the brightness value of the core fluorescence image after preprocessing, G(x, y) is the local gradient of the fluorescence area, Z m,n is the normalization factor, the sum of all weights in the fluorescent area is 1; S32, the core fluorescence image I″′ m,n Input the improved deep convolutional neural network model and extract the preliminary feature representation F m,n : Among them, F m,n (x,y) represents the eigenvalue at the coordinate (x,y), is the weight of convolution kernel k, b k is the bias term; S33, the attention mechanism module guided by the fluorescent area for the preliminary feature representation F m,n Perform weighted processing to generate weighted feature map F′ m,n : F′ m,n (x,y)=A m,n (x,y)·F m,n (x,y); Among them, A m,n (x,y) is the fluorescence area weight, F m,n (x,y) is the initial feature representation, F′ m,n (x,y) represents the eigenvalue after weighting by the weight matrix; S34, based on weighted feature map F' m,n By improving the segmentation module in the deep convolutional neural network model, the fluorescent area is segmented using the pixel-level segmentation algorithm to generate the segmented fluorescent area image S m,n , the segmentation process is optimized by the energy minimization function: E(S m,n )=∑ x,y ((S m,n (x,y)-F′ m,n (x,y)) 2 +λ∑ (x′,y′)∈N(x,y) |S m,n (x,y)-S m,n (x′,y′)|); Among them, S m,n (x, y) represents the value of the segmented fluorescence region image at the coordinate (x, y), F′ m,n (x, y) is the weighted feature representation, N(x, y) is the neighborhood pixel set, and λ is the smoothness parameter; S35, the segmented fluorescence area image S m,n Perform annotation and match the segmentation results with the annotated fluorescence area dataset to generate an annotated fluorescence area dataset: Among them, δ(S m,n (x, y), T) is the decision function, which divides the segmented fluorescence area S into m,n Mark, L m,n is the generated annotation label matrix, D″ l This is the final annotated fluorescence region dataset.

5. The method for grading core fluorescence brightness and area calculation based on deep learning according to claim 1 is characterized in that: The S4 step specifically includes: S41. Build a semi-supervised deep learning model to annotate the fluorescence area dataset D″ l and the unlabeled fluorescence region dataset D′ u For training, the semi-supervised model includes a main network and an auxiliary network. The main network is used to classify and segment the input core fluorescence image, and the auxiliary network is used for consistency regularization. S42. For the unlabeled fluorescence region dataset D′ u Core fluorescence image I″′ m,n , generate multiple different enhanced versions Each enhanced version is generated by different core fluorescence image transformation operations, including rotation, scaling and noise addition; S43. Input the enhanced core fluorescence image version into the main network of the semi-supervised model for prediction and obtain the corresponding prediction results. Represents the prediction result of the i-th enhanced version: in, Represents the predicted output of the i-th enhanced version at coordinate (x, y); f θ is the activation function of the main network, and the parameter θ represents the weight and bias of the network; is the weight of the convolution kernel applied to the pixel position (p,q) in the i-th enhanced version; I″′ i (xp,yq) m,n is the input image value of the pixel position (xp, yq) in the i-th enhanced version; S44, in the consistency regularization process by applying the consistency loss function L cons To keep the prediction results of different enhanced versions consistent, the loss function is defined as: in, represents the average prediction result of all enhanced versions, k is the number of enhanced versions; S45. Generate pseudo labels for the unlabeled fluorescence region dataset based on the prediction results after consistency regularization Among them, arg max is a function that selects the maximum value in the prediction result, which is used to determine the category of the pseudo label; S46. Generate pseudo-label dataset and the labeled fluorescence region dataset D″ l Merge to form an extended data set:

6. The method for grading core fluorescence brightness and area calculation based on deep learning according to claim 1 is characterized in that: The S5 step specifically includes: S51, based on the labeled fluorescence area dataset D″ l and pseudo-labeled data The brightness B of the core fluorescence area is analyzed using the improved convolutional neural network model. m,n (x,y) and area A m,n Perform quantitative calculation, brightness B m,n (x,y) is the brightness value of the pixel (x,y), area A m,n is the total number of pixels in the fluorescent area; S52, using the improved convolutional neural network model to input the core fluorescence image I″′ m,n Perform pixel-by-pixel calculations to obtain the brightness prediction value of each pixel in, is the weight of the brightness prediction convolution kernel, b b is the bias term, is the brightness prediction value at pixel (x, y); S53, the brightness value B of each fluorescent area m,n (x,y) is accumulated to calculate the total brightness of the fluorescent area in, It represents the total brightness value of the fluorescent area in the core fluorescence image; S54. Calculate the area A based on the total number of pixels in the fluorescent area. m,n : in, is the indicator function, when the pixel brightness When the value exceeds the set threshold T, it is recorded as 1, otherwise it is recorded as 0, which is used to count the total area of ​​the fluorescent area.

7. The method for grading core fluorescence brightness and area based on deep learning according to claim 1, characterized in that: The S6 step specifically includes: S61, the fluorescence brightness B m,n (x,y) and fluorescence area A m,n The calculation results are input into the trained semi-supervised deep learning model, and the semi-supervised deep learning model is used to calculate the extended dataset D ext Perform grading and output the fluorescence brightness level and geological characteristics evaluation results of the core samples; S62, by improving the fully connected layer of the deep convolutional neural network model to measure the fluorescence brightness B m,n (x,y) and area A m,n Perform feature fusion to generate feature vector F m,n : F m,n =αB m,n (x,y)+βA m,n ; Among them, α and β are the weight coefficients of brightness and area, respectively, which are used to adjust the relative importance between features. m,n is the fused feature vector; S63, the fused feature vector F m,n Passed to the classification layer, the trained semi-supervised model outputs the fluorescence brightness level G m,n : Among them, G m,n is the predicted fluorescence brightness level, P represents the number of feature maps, and W (p) is the weight matrix, is the pth feature map value, γ is the adjustment coefficient, Indicates the binding area A m,n and brightness B m,n The nonlinear function, b (p) is the bias term, σ is the activation function; S64, conduct a comprehensive evaluation of geological characteristics, according to the fluorescence brightness level G m,n and the fluorescence area A m,n The calculation results generate the geological characteristics evaluation results R m,n : Among them, R m,n is the geological characteristic evaluation result, Q is the characteristic dimension, λ is the weight attenuation factor, Represents the gradient change of area in the feature dimension, represents the influence function of brightness level, λ is the control factor, Ψ(G m,n ,A m,n ) represents the comprehensive influence of brightness and area; S65, the fluorescence brightness level of the core is G m,n and geological characteristics evaluation results R m,n The output is the final result, which is used for resource assessment and geological analysis.

Images