Fracture segmentation deep learning method based on logging imaging

Abstract

The invention relates to a crack segmentation deep learning method based on logging imaging. The method comprises the steps of data acquisition and preprocessing; optimizing a target detection algorithm, specifically including adding an MSF attention mechanism module, replacing a network loss function, replacing with a more lightweight network, and optimizing a pyramid pooling layer; and carrying out an ablation experiment, and screening out an electrical imaging fracture segmentation network model with optimal performance by combining comprehensive index analysis of experimental data. According to the invention, an attention mechanism module-MSF is designed, so that crack characteristics can be accurately represented and background interference can be effectively suppressed. By optimizing an original network structure, including improving a pyramid pooling network layer to integrate higher-scale features, replacing a more appropriate loss function and replacing the pyramid pooling network layer with a more lightweight skeleton network, the overall generalization ability and robustness of the model are remarkably improved through the optimization, and a model foundation is laid for a subsequent imaging crack segmentation task.

Description

Technical Field

[0001] The present invention relates to the field of intelligent segmentation of imaging well logging fractures, and in particular to a deep learning method for fracture segmentation based on well logging imaging. Background Art

[0002] The fracture response characteristics of traditional logging are very complex, with low vertical resolution. It can only roughly describe the area where the fracture is located, and cannot directly display the fracture, let alone accurately calculate the fracture parameter information. However, imaging logging has high resolution, which can present fracture information in detail and intuitively, and can calculate fracture width, inclination, dip and density and other parameter information through image information and mathematical methods. For a long time, logging experts have strived to automatically identify and segment fractures from logging images efficiently and accurately, and then calculate parameter information, classify fracture grades through specific parameter information, determine the most developed fracture layers in the formation, and classify reservoir grades. Therefore, it is important to provide a reliable image segmentation method to realize intelligent and automatic detection of fractures and extract fracture information.

[0003] In view of the challenges faced in the electrical imaging fracture segmentation task, such as insufficient feature representation and severe background interference, as well as the considerations of limited computing resources and deployment platform requirements in practical applications, optimizing the accuracy and efficiency of imaging fracture segmentation and identification to meet the strict requirements of well logging interpretation has become a key issue that needs to be solved urgently. Summary of the invention

[0004] The purpose of the present invention is to provide a deep learning method for fracture segmentation based on well logging imaging to solve the technical problems raised in the background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions: A deep learning method for fracture segmentation based on well logging imaging, the method comprising: Step 1: Data collection and preprocessing; Step 2: Optimize the target detection algorithm, including adding the MSF attention mechanism module, changing the network loss function, changing to a lightweight network, and optimizing the pyramid pooling layer; Step 3: Conduct ablation experiments and analyze the comprehensive indicators of experimental data to select the electrical imaging crack segmentation network model with the best performance.

[0006] As a further technical solution of the present invention, the data collection and preprocessing steps in step 1 include: Collect the required electro-imaging fracture image data of different lithologies from the target oilfield area to be processed and establish a dataset. In terms of formation lithology, it covers any different lithologies in sedimentary rocks, igneous rocks, and metamorphic rocks, and selects electro-imaging fracture images from different regions as the dataset. Each image corresponds to a well section depth of 2m, and the total covered well section length reaches the ten-thousand-meter level.

[0007] In the label production link of data preprocessing, in order to ensure the fineness of label production, ensure that the number of calibration points for each fracture always remains above 20, and for the first time, the fracture labels are classified in detail, specifically covering eight categories: high-conductivity fractures, local fractures, dissolution fractures, semi-filled fractures, feather-induced fractures, vertical-induced fractures, high-resistance fractures, and faults.

[0008] Data preprocessing includes data augmentation, eliminating electro-imaging white stripes, generating full-well section images, image cutting, label production, and label format conversion. Among them, data augmentation includes image flipping, adding Gaussian noise, and adding salt-and-pepper noise; the method of eliminating electro-imaging white stripes is automatically processed by the Ciflog2.1 software, and a full-well section image is generated after eliminating the stripes. For key well sections, two fine cutting ratios of 1:10 and 1:20 are used to cut the full-well section image, and then label production is carried out on the cut image, and finally the conversion of the label data format is completed.

[0009] As a further technical solution of the present invention, in step two, the steps of adding the MSF attention mechanism module include; In the spatial dimension, a fine spatial attention module is designed. The spatial attention module accurately locates the key regions of fractures in the image. Here, the key regions refer to the black sine and cosine curve regions of fractures in well logging imaging, and higher weights are assigned to these regions, thereby improving the accuracy and robustness of well logging imaging fracture recognition; in order to capture the correlation between different positions of the feature map, a position attention module is introduced. The position attention module realizes the effective integration and enhancement of fracture features at different positions by calculating the similarity between any two positions of the feature map.

[0010] As a further technical solution of the present invention, in step two, the steps of replacing the network loss function include; Bounding box regression loss based on auxiliary bounding boxes Inner-IoU , by using a scale factor to control the generation of auxiliary bounding boxes, calculating the loss, and accelerating the convergence of training. Currently, this method can be integrated into the existing IoU -based loss function. The loss function performs well in small target detection tasks and detection and segmentation tasks with complex backgrounds. Its formula is as follows: ； ； ； ； ； ； ； ； where represents the offset or distance of the left boundary of the ground truth box relative to the reference point (such as the left boundary of the grid cell or the left boundary of the anchor box); represents the offset or distance of the right boundary of the ground truth box relative to the reference point (such as the right boundary of the grid cell or the right boundary of the anchor box), and represent the coordinate values of the center point x, y of the ground truth box, and represent the width and height values of the ground truth box respectively, and the scaling factor ratio takes values in the range of [0.5, 1.5]; represents the offset or distance of the upper boundary of the ground truth box relative to the reference point (such as the upper boundary of the grid cell or the upper boundary of the anchor box); represents the offset or distance of the lower boundary of the ground truth box relative to the reference point (such as the lower boundary of the grid cell or the lower boundary of the anchor box); represents the offset or distance of the left boundary of the anchor box relative to the reference point (such as the left boundary of the grid cell or the left boundary of the anchor box); 、 、 represent the offsets or distances of the right, upper, and lower boundaries of the anchor box relative to the reference point respectively; represents the intersection area of two bounding boxes; union represents the union area of two bounding boxes.

[0011] Focaler-IoU Reconstruct the original IoU loss through linear interval mapping to achieve the purpose of focusing on easy and hard samples. The specific formula is as follows: ; represents the IoU based on the improved focal loss function, and adjusts the attention degree of different samples by adjusting the values of d and u, and the value adjustment range is [0, 1]; LetInner-IoU The loss function and Focaler-IoU the advantages of the loss function are summarized and combined and applied to the electrical imaging crack segmentation task.

[0012] As a further technical solution of the present invention, in step two, the steps of replacing the lightweight network include; YOLOv8-seg Combined with C2f-DualConv the lightweight network, DualConv combines 3×3 group convolutions and 1×1 point convolutions to solve the problems of cross-channel communication and information preservation in the original input feature map, where the 3×3 group convolution groups the input channels for convolution operations; at the same time, the group convolution extracts features within each channel group; the 1×1 point convolution is mainly used for cross-channel communication, which linearly combines the feature maps output by the group convolution, adjusts the number of channels, and realizes information interaction between different channels.

[0013] As a further technical solution of the present invention, in step two, the steps of optimizing the pyramid pooling layer include; In the application scenario for the electrical imaging crack segmentation task, YOLOv8-seg the original spatial pyramid pooling layer SPPF in the network is replaced with a spatial pyramid pooling layer combined with cross-stage partial convolution SPPCSPC。

[0014] The term explanations obtained in the present invention are as follows: YOLOv8-seg : is YOLO an algorithm specifically for object segmentation in the eighth-generation version of the MSF : a name of an attention mechanism, mainly focusing on multi-scale features; IoU : intersection over union, a commonly used metric in computer vision tasks such as object detection and image segmentation, used to measure the overlap degree between two regions; Inner-IoU: A new way of calculating the loss function, which is improved on the basis of the traditional IoU ; Focaler-IoU: A new way of calculating the loss function, which combines Focal Loss and IoU ideas; DualConv : a lightweight convolution module, whose main purpose is to reduce the number of parameters and computational amount of the model while ensuring the model performance, thereby improving the running efficiency of the model; SPPCSPC: A new type of pyramid pooling network module, which combines spatial pyramid pooling ( SPP ), and cross-stage partial convolution (CSP The idea of

[0015] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. The dataset constructed for the task of electrical imaging crack detection and segmentation has significant advantages, including the richness of data volume, wide universality, fineness of label production, and comprehensiveness of crack label types. These datasets prepared in advance provide data support for the subsequent model training process, ensuring the efficiency, robustness, and accuracy of the model training process.

[0016] 2. An attention mechanism module - MSF is designed, which performs excellently in the task of electrical imaging crack segmentation, can accurately represent crack features, and effectively suppress background interference.

[0017] 3. The original network structure is optimized, including improving the pyramid pooling network layer to integrate higher-scale features, replacing with a more suitable loss function, and replacing with a more lightweight backbone network. These optimizations significantly improve the overall generalization ability and robustness of the model.

[0018] 4. The ablation experiment proves the robustness and efficiency of the method in the task of imaging crack segmentation. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Figure 1 is a flow chart of a deep learning method for crack segmentation based on logging imaging.

[0020] Figure 2 is a flow chart of data acquisition and preprocessing.

[0021] Figure 3 is a structural diagram of the MSF attention mechanism model in an embodiment of the present invention.

[0022] Figure 4 is a comparison chart of network performance with various attention mechanisms added in an embodiment of the present invention.

[0023] Figure 5 is a comparison chart of loss function performance in an embodiment of the present invention.

[0024] Figure 6 is a comprehensive heat map of crack recognition and segmentation of the present invention.

[0025] Figure 7 is the first part of the oilfield electrical imaging logging crack segmentation in the application example of an embodiment of the present invention.

[0026] Figure 8 is the second part of the oilfield electrical imaging logging crack segmentation in the application example of an embodiment of the present invention.

[0027] Figure 9This is the third part of the figure for fracture segmentation in the application example of the embodiment of the present invention in oilfield electrical imaging logging.

[0028] Figure 10 This is the fourth part of the figure for fracture segmentation in the application example of the embodiment of the present invention in oilfield electrical imaging logging. Detailed implementation manners

[0029] In order to make the technical problems, technical solutions and beneficial effects to be solved by the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0030] Please refer to Figure 1 , the present invention provides a deep learning method for fracture segmentation based on logging imaging, and the method includes: Step 1, data acquisition and preprocessing; Please refer to Figure 2 , data acquisition and preprocessing: Collect electrical imaging fracture image data of different lithologies required from the target oilfield area to be processed and establish a data set. Among them, the formation lithology can cover any different lithologies in sedimentary rocks, igneous rocks and metamorphic rocks; establish a data set with wide range and representativeness to ensure that the trained model has stronger universality and persuasiveness, and can perform effective fracture detection and segmentation under different types of oilfields and formation conditions; secondly, in order to effectively avoid the overfitting problem caused by insufficient data volume or improper data augmentation operations during model training, and further improve the credibility of the model, the present invention carefully selects 5000 electrical imaging fracture images from different regions as the data set, and each image corresponds to a well section depth of about 2m, with a total covered well section length reaching the ten-thousand-meter level; for key well sections, the key well sections specifically refer to well sections with excellent selected well conditions, no obvious hole enlargement phenomenon observed in the measured well section, high imaging quality, and fully meeting the requirements of electrical imaging interpretation. Two fine cutting ratios of 1:10 and 1:20 are adopted, aiming to improve the accuracy of the model through this strategy. Finally, 3579 labels are made from the high-quality images among these 5000 images for training. Here, the high-quality images refer to fracture images with less background noise in logging images and clear fractures. By training the model with this large-scale and high-quality electrical imaging data set, not only can the excellent performance of the model in fracture detection and segmentation tasks be ensured, but also the persuasiveness and credibility of the model can be significantly improved; In the label production process, the present invention does not adopt a blind or random approach. Instead, based on the accurate identification and interpretation of the results of electrical imaging fractures by Schlumberger experts, fracture labels of different categories are produced. During the production process, we strictly carried out the calibration work to ensure that the number of calibration points for each fracture always remained above 20, thus following and conforming to the fracture atlas standard of Schlumberger to the greatest extent. This production process ensures the accuracy and reliability of the labels, providing a high-quality guiding basis for subsequent model training.

[0031] In the process of label production, the present invention for the first time carried out a detailed classification of fracture labels, specifically covering eight categories: continuous fractures, discontinuous fractures, dissolution fractures, semi-filled fractures, plume-induced fractures, vertical-induced fractures, high-resistance fractures, and faults. This refined classification not only significantly enhances the applicability and accuracy of the deep learning model in practical applications, but also provides strong support for the in-depth understanding and analysis of underground fracture structures. Through this method, researchers can more accurately determine the fracture type and then conduct a more comprehensive and detailed evaluation of the reservoir space. The overall data processing flow is shown in the figure. Data preprocessing includes data augmentation, eliminating electrical imaging white bands, generating full-wellbore images, image cutting, label production, and label format conversion. Among them, data augmentation includes image flipping, adding Gaussian noise, and adding salt-and-pepper noise. The method of eliminating electrical imaging white bands is automatically processed by the Ciflog2.1 software. After eliminating the bands, full-wellbore images are generated, and then the full-wellbore images are cut according to 1:10 and 1:20. Subsequently, labels are produced for the cut images, and finally, the conversion of the label data format is completed.

[0032] Step 2: Optimize the object detection algorithm: The specific content includes adding the MSF attention mechanism module, replacing the network loss function, replacing it with a lightweight network, and optimizing the pyramid pooling layer. The structural diagram of the MSF attention mechanism module is as Figure 3 shown; The steps of adding the MSF attention mechanism module include; In the spatial dimension, we designed a fine spatial attention module. The spatial attention module accurately locates the key regions of fractures in the image and assigns higher weights to these regions. Here, the key regions refer to the black sine and cosine curve regions of fractures in well logging imaging, thus improving the accuracy and robustness of well logging imaging fracture recognition. To capture the correlation between different positions of the feature map, a position attention module is introduced. The position attention module realizes the effective integration and enhancement of fracture features at different positions by calculating the similarity between any two positions of the feature map. By organically integrating the channel attention module, spatial attention module, and position attention module, the attention module not only achieves comprehensive extraction and representation of image information but also significantly improves the performance of electrical imaging crack recognition or segmentation tasks. In addition, the module also has the ability of multi-scale feature fusion, which can make full use of feature information at different scales to further improve the accuracy and efficiency of crack recognition. The experimental result figures are as shown in Figure 4 , where BP represents the crack recognition accuracy rate, MP represents the crack segmentation accuracy rate, BP@0.5 represents the average precision when the recognition intersection over union (IoU) threshold is 0.5, MP@0.5 represents the average precision when the segmentation IoU threshold is 0.5, BP@0.5:0.95 represents the average precision when the recognition IoU threshold is between 0.5 and 0.95, MP@0.5:0.95 represents the average precision when the segmentation IoU threshold is between 0.5 and 0.95, and the other compared labels are the current mainstream attention mechanism modules.

[0033] The steps to replace the network loss function include: YOLOv8-seg The original loss function of the network is calculated based on IoU. The calculation of IoU uses the intersection of the predicted bounding box (A) and the ground truth bounding box (B) divided by their union. The higher the value of IoU, the higher the degree of overlap between the two, indicating that the model prediction is more accurate. Its formula is: ; Bounding box regression loss based on the auxiliary bounding box Inner-IoU , by using a scaling factor to control the generation of the auxiliary bounding box, calculating the loss, and accelerating the convergence of training. Currently, this method can be integrated into the existing loss function based on IoU . The loss function performs well in small object detection tasks and detection and segmentation tasks with complex backgrounds. Its formula is as follows: ; ; ; ; ; ; ; ; where represents the offset or distance of the left boundary of the ground truth bounding box relative to the reference point; represents the offset or distance of the right boundary of the ground truth bounding box relative to the reference point, and Represent the coordinate values of the center point x and y of the ground truth box, and represent the width and height values of the ground truth box respectively, and the scale factor ratio takes values in the range of [0.5, 1.5]; represents the offset or distance of the upper boundary of the ground truth box relative to the reference point; represents the offset or distance of the lower boundary of the ground truth box relative to the reference point; represents the offset or distance of the left boundary of the anchor box relative to the reference point; 、 、 represent the offsets or distances of the right, upper, and lower boundaries of the anchor box relative to the reference point respectively; represents the intersection area of two bounding boxes; union represents the union area of two bounding boxes; Focaler-IoU Reconstruct the original IoU loss through linear interval mapping to achieve the purpose of focusing on easy and difficult samples. The specific formula is as follows: ; represents the IoU based on the improved focal loss function. In the formula, the attention degree of different samples is adjusted by adjusting the values of d and u, and the value adjustment range is [0, 1]; Combine Inner-IoU the loss function and Focaler-IoU the advantages of the loss function are summarized and combined and applied to the electrical imaging crack segmentation task. Through experiments, it is proved that the new loss function is more sensitive to the electrical imaging dataset and has stronger generalization ability than the original loss calculation method. The specific comparison experiments are as Figure 5 .

[0034] The steps to replace the lightweight network include: Network model lightweighting refers to reducing the number of model parameters, computational complexity, and model size by optimizing the internal structure of the network module, so as to improve its operating efficiency in computing resources-limited environments such as embedded systems and mobile devices while ensuring the model performance. Lightweight backbone networks are currently widely applied to object detection and image segmentation tasks. YOLOv8-segIn the original network, an excessive number of parameters may lead to overfitting of the model and a reduction in generalization ability. Through the operation of this lightweight network model, not only can overfitting of the model be effectively prevented, but also the comprehensive performance of the model can be improved. Secondly, the lightweight design can also significantly reduce the model volume, thereby reducing the demand for storage space. During the process of electrical imaging crack detection, the lightweight YOLO segmentation network can process videos with low latency while maintaining a high detection accuracy, which is crucial for real-time segmentation and detection of electrical imaging cracks.

[0035] YOLOv8-seg The methods for lightweighting the network model include simplifying the model structure, using lightweight convolutional structures, optimizing the feature pyramid, and replacing with lighter activation functions; the present invention mainly focuses on replacing the convolutional structure of the original network, such as separable convolution, grouped convolution, etc. These convolutional operations can significantly reduce the computational amount and the number of parameters while maintaining the model performance.

[0036] YOLOv8-seg Combined with C2f-DualConv a lightweight network, DualConv Combining 3×3 grouped convolution and 1×1 point convolution solves the problems of cross-channel communication and information preservation in the original input feature map. Among them, the 3×3 grouped convolution groups the input channels for convolution operations. Compared with the traditional convolution method, it can significantly reduce the number of convolution kernels, thereby reducing the network parameters. At the same time, the grouped convolution extracts features within each channel group, which can retain the local information of the input features to a certain extent and contribute to improving the feature expression ability. The 1×1 point convolution is mainly used for cross-channel communication. It can perform a linear combination of the feature maps output by the grouped convolution, adjust the number of channels, and achieve information interaction between different channels, thereby enriching the feature expression. Through the 1×1 point convolution, the feature information of different channels can be effectively fused without introducing too many parameters, improving the performance of the network.

[0037] By combining YOLOv8-seg with the C2f-DualConv lightweight network, the method of the present invention has the following advantages in object detection and image segmentation tasks: Improving network accuracy: The C2f-DualConv module realizes effective cross-channel communication while retaining the original input feature information, enabling the network to better capture the features of the target, thereby improving the accuracy of detection and segmentation.

[0038] Reducing network parameters: The use of 3×3 grouped convolution reduces the number of convolution kernels, reduces the complexity of the network, thereby reducing the number of parameters to be learned and alleviating the storage burden of the model.

[0039] Reduce computational cost: Due to the reduction of network parameters, the computational volume during the forward and backward propagations of the model is correspondingly reduced, thereby reducing the computational cost and improving the training and inference efficiency of the model.

[0040] Shorten the inference time: The reduction of the computational cost significantly improves the inference speed of the model in practical applications, enabling it to give detection and segmentation results faster and meeting the application scenarios with high real-time requirements. The steps to optimize the pyramid pooling layer include: In the application scenario of the electrical imaging crack segmentation task, SPPF Although multi-scale features of the image can be extracted through max-pooling operations of different scales, in electrical imaging crack data, the morphology, orientation, and scale of cracks vary complexly. SPPF It may not be able to fully capture the long-range dependencies between crack features, and some key context information may be lost during the feature fusion process, resulting in the segmentation accuracy being affected. SPPCSPC The layer combines the idea of cross-stage partial convolution ( CSP ). Cross-stage partial convolution can segment and fuse the feature maps, enhancing the diversity and expressiveness of features while reducing the computational volume. In electrical imaging crack segmentation, crack features of different scales can be more effectively extracted. For example, for tiny cracks and large cracks, SPPCSPC their feature information can be captured from multiple scales and levels, improving the sensitivity to crack features. Through the design of cross-stage partial convolution, SPPCSPC it can better fuse the feature information of different stages, solving the SPPF problem of information loss that may occur during the feature fusion process. In electrical imaging data, crack features are often correlated with the surrounding geological background, and SPPCSPC can more effectively integrate this context information, making the segmentation results more accurate and complete. By YOLOv8-seg replacing the original spatial pyramid pooling layer in the network SPPF with a spatial pyramid pooling layer combined with cross-stage partial convolution SPPCSPC , the model can more accurately identify and segment the crack area, reducing the situations of misjudgment and missed judgment, and being more efficient in the use of computing resources, shortening the training and inference time of the model.

[0041] Step 3: Conduct ablation experiments, and through the comprehensive index analysis of the experimental data, screen out the electrical imaging crack segmentation network model with the best performance.

[0042] To verify the effectiveness of the model, we conducted ablation experiments. By systematically removing or replacing certain components in the model, we were able to meticulously analyze the impact of each part on the overall performance. This series of experiments not only helped us understand the key elements of the model but also gradually optimized the model structure by comparing the performance metrics under different configurations. Finally, through the comprehensive index analysis of the experimental data, we selected the electro-imaging crack segmentation network model with the best performance. Figure 6 It is the comprehensive heat map for crack identification and segmentation in the module. This model performs excellently in terms of accuracy, robustness, and efficiency, laying a solid foundation for our subsequent research and applications.

[0043] Practical application: When applying the selected electro-imaging crack segmentation network model with the best performance to the electro-imaging crack segmentation of an oilfield, it can be seen that the segmentation effect is good. The blue segmentation area and the green area represent different crack types, such as Figures 7 - 10 shown.

[0044] It should be noted that in this article, the term "including" or any of its other variants is intended to cover non-exclusive inclusion, so that a process, method, article, or device including a series of elements not only includes those elements but also includes other elements not explicitly listed, or also includes elements inherent to such a process, method, article, or device. Without more limitations, an element defined by the statement "including a..." does not exclude the existence of another identical element in the process, method, article, or device including that element.

[0045] The above are only the preferred embodiments of the present invention and do not limit the patent scope of the present invention. Any equivalent structural or equivalent process transformations made using the specification and drawings of the present invention, or directly or indirectly applied in other related technical fields, are similarly included in the patent protection scope of the present invention.

Claims

1. A deep learning method for fracture segmentation based on logging imaging, characterized in that, The method includes: Step 1, data collection and preprocessing; Step 2, optimizing the object detection algorithm: including adding an MSF attention mechanism module, replacing the network loss function, replacing with a lightweight network, and optimizing the pyramid pooling layer; Step 3, conducting ablation experiments, and screening out the electro-imaging fracture segmentation network model with the best performance by combining the comprehensive indicators analysis of the experimental data.

2. The deep learning method for fracture segmentation based on logging imaging according to claim 1, wherein The data collection and preprocessing steps in Step 1 include: Collecting electro-imaging fracture image data of different lithologies required from the target oilfield area to be processed and establishing a dataset; in terms of formation lithology, covering any different lithologies among sedimentary rocks, igneous rocks, and metamorphic rocks, and selecting electro-imaging fracture images from different regions as the dataset, each image corresponding to a well section depth of 2m, with a total covered well section length reaching the ten-thousand-meter level; In the label production link of data preprocessing, the fracture labels were classified in detail for the first time, specifically covering eight categories: high-conductivity fractures, local fractures, dissolution fractures, semi-filled fractures, feather-induced fractures, vertical-induced fractures, high-resistance fractures, and faults; Data preprocessing includes data augmentation, eliminating electro-imaging white stripes, generating full-well-section images, image cutting, label production, and label format conversion; among them, data augmentation includes image flipping, adding Gaussian noise, and adding salt-and-pepper noise; the method of eliminating electro-imaging white stripes is automatically processed by the Ciflog2.1 software, and full-well-section images are generated after eliminating the stripes; for key well sections, the full-well-section images are cut with two fine cutting ratios of 1:10 and 1:20, and then label production is carried out on the cut images, and finally the conversion of the label data format is completed.

3. The deep learning method for fracture segmentation based on logging imaging according to claim 1, wherein In Step 2, the steps of adding an MSF attention mechanism module include: In the spatial dimension, a fine spatial attention module is designed. The spatial attention module accurately locates the key areas of fractures in the image and assigns higher weights to these areas, thereby improving the accuracy and robustness of logging imaging fracture recognition; in order to capture the correlation between different positions of the feature map, a position attention module is introduced. The position attention module realizes the effective integration and enhancement of fracture features at different positions by calculating the similarity between any two positions of the feature map.

4. The deep learning method for fracture segmentation based on logging imaging according to claim 3, wherein In Step 2, the steps of replacing the network loss function include; Bounding Box Regression Loss Based on Auxiliary Bounding Boxes Inner-IoU , by using a scale factor to control the generation of auxiliary bounding boxes, calculating the loss, and accelerating the convergence of training, it can be integrated into existing IoU -based loss functions. The loss function performs well in small object detection tasks and detection and segmentation tasks with complex backgrounds. Its formula is as follows: ； ； ； ； ； ； ； ； Among them represents the offset or distance of the left boundary of the ground truth box relative to the reference point; represents the offset or distance of the right boundary of the ground truth box relative to the reference point, and represents the coordinate values of the center point x, y of the ground truth box, and represent the width and height values of the ground truth box respectively, and the scale factor ratio takes values between [0.5, 1.5]; represents the offset or distance of the upper boundary of the ground truth box relative to the reference point; represents the offset or distance of the lower boundary of the ground truth box relative to the reference point; represents the offset or distance of the left boundary of the anchor box relative to the reference point; 、 、 represent the offsets or distances of the right, upper, and lower boundaries of the anchor box relative to the reference point respectively; represents the intersection area of two bounding boxes; union represents the union area of two bounding boxes; Focaler-IoU Reconstruct the original through linear interval mapping IoU Loss, to achieve the purpose of focusing on difficult and easy samples, and the specific formula is as follows: ； Indicates the IoU based on the improved catenary loss function. In the formula, by adjusting d and u the values of, the attention degrees of different samples are adjusted, and the value adjustment range is [0, 1]; By combining Inner-IoU the loss function and Focaler-IoU the loss function to form a set of loss functions more suitable for the imaging crack segmentation task will greatly improve the overall segmentation accuracy and the generalization ability of the model.

5. The deep learning method for fracture segmentation based on logging imaging according to claim 3, wherein In Step 2, the steps of replacing the lightweight network include; YOLOv8-seg Combined with C2f-DualConv a lightweight network, DualConv combines 3×3 grouped convolutions and 1×1 pointwise convolutions to address the issues of cross-channel communication and information preservation in the original input feature map. Among them, the 3×3 grouped convolution groups the input channels for convolution operations; meanwhile, the grouped convolution extracts features within each channel group; the 1×1 pointwise convolution is mainly used for cross-channel communication, which linearly combines the feature maps output by the grouped convolution, adjusts the number of channels, and realizes information interaction between different channels.

6. The deep learning method for fracture segmentation based on logging imaging according to claim 3, characterized in that In Step 2, the steps of optimizing the pyramid pooling layer include; In the application scenario for the task of electrical imaging crack segmentation, YOLOv8-seg the original Spatial Pyramid Pooling Layer (SPPF) in the network is replaced with a Spatial Pyramid Pooling Layer that combines cross-stage partial convolution SPPCSPC .

Images