Shale pore type identification method based on improved YOLOv8 model

Abstract

The invention discloses a shale pore type identification method based on an improved YOLOv8 model. The method comprises the following steps: acquiring a shale digital core image and annotation information thereof to construct a data set; carrying out improved YOLOv8 model training by utilizing the data set to obtain a trained improved YOLOv8 model; the improved YOLOv8 model comprises a backbone network, a neck network and a head network; and inputting a to-be-processed shale digital core image into the trained improved YOLOv8 model to obtain a corresponding shale pore type identification result. According to the method, through an improved YOLOv8 model, a C2fDCNv3 module and a content awareness recombination feature enhancement module are integrated, and the feature extraction and fusion capability of a shale pore complex structure is remarkably improved; and in combination with a dynamic detection head DyHead and an Inner-CIoU loss function, efficient and accurate shale pore type identification and bounding box regression prediction are realized.

Description

Technical Field

[0001] The present invention relates to the field of shale pore type identification, and more specifically to a shale pore type identification method based on an improved YOLOv8 model. Background Art

[0002] With the continuous development of unconventional oil and gas resources, the importance of shale oil and gas resources has become increasingly prominent. The pore structure of shale is complex, containing organic and inorganic pores at the micron to nanometer level, which puts higher requirements on the characterization of reservoir parameters. Traditional rock physics experiments face challenges in the characterization of shale reservoir parameters, and it is difficult to accurately describe these complex pore structures and their distribution characteristics.

[0003] In the existing technology, traditional image processing methods rely on manually designed feature extraction algorithms, such as texture analysis, morphological operations, etc. These methods often lack the ability to adapt to the diversity and complexity of shale pores, resulting in limited recognition accuracy. In addition, classification methods based on classical machine learning, such as support vector machines (SVM) or random forests (RF), although they can improve the classification effect to a certain extent, they usually require a lot of manual feature engineering and perform poorly when faced with high-dimensional data.

[0004] Deep learning models, especially convolutional neural networks (CNNs), have achieved remarkable results in many computer vision tasks. However, the standard CNN architecture has limitations when dealing with small targets with high variability such as shale pores. For example, the fixed receptive field limits its ability to learn features of different scales, and the traditional loss function is not accurate enough for the bounding box regression of irregularly shaped targets.

[0005] Therefore, how to design a shale pore type identification method based on the improved YOLOv8 model, which can make full use of the multi-scale information in shale digital core images, significantly improve the feature extraction and fusion capabilities of the complex structure of shale pores, and achieve efficient and accurate shale pore type identification and bounding box regression prediction is a problem that technicians in this field urgently need to solve. Summary of the invention

[0006] In view of this, the present invention provides a shale pore type identification method based on the improved YOLOv8 model, aiming to solve the shortcomings of traditional methods in shale pore type identification. By integrating deformable convolution modules, content-aware reconstruction feature enhancement, dynamic detection heads and optimized loss functions, efficient and accurate identification and analysis of shale pores are achieved, and the adaptability and robustness of the model are significantly improved.

[0007] In order to achieve the above object, the present invention adopts the following technical solution:

[0008] A shale pore type identification method based on an improved YOLOv8 model includes the following steps:

[0009] S1. Acquire shale digital core images and their annotation information to construct a data set;

[0010] S2. Performing training on an improved YOLOv8 model using the data set to obtain a trained improved YOLOv8 model; the training on an improved YOLOv8 model includes:

[0011] S21, extracting features from the shale digital core images in the data set based on the backbone network to generate feature maps of different sizes;

[0012] S22, using the neck network to perform feature fusion on the feature maps of different sizes to obtain a fused feature map;

[0013] S23, classifying and performing bounding box regression prediction on the fused feature map through a head network to identify the shale pore type and location;

[0014] S3. Input the shale digital core image to be processed into the trained improved YOLOv8 model to obtain the corresponding shale pore type recognition result.

[0015] Further, in S21, the backbone network includes a first convolution block, a second convolution block, a first CSP bottleneck module, a third convolution block, a first C2f_DCNv3 module, a fourth convolution block, a second C2f_DCNv3 module, a fifth convolution block, a third C2f_DCNv3 module and a spatial pyramid pooling module connected in sequence;

[0016] Among them, the first C2f_DCNv3 module, the second C2f_DCNv3 module and the spatial pyramid pooling module are connected to the neck network.

[0017] Furthermore, in the backbone network, each C2f_DCNv3 module includes a convolution, batch normalization and activation function combination unit, a splitting unit, a first bottleneck unit, a second bottleneck unit, a splicing unit, and a convolution, batch normalization and activation function combination unit connected in sequence.

[0018] Furthermore, each bottleneck unit includes two interconnected deformable convolutional networks DCNv3; the deformable convolutional network DCNv3 is expressed as:

[0019]

[0020] Among them, y(P 0 ) represents the output feature value of the deformable convolutional network DCNv3, G represents the number of spatial clustering groups, K represents the number of sampling points in each spatial clustering group, and wg represents the weight coefficient of the g-th group, m gk represents the weight coefficient of the k-th sampling point in the g-th group, x g represents the feature map of the g-th group, P 0 represents the base position in the feature map, p k the position of the k-th sampling point, Δp gk represents the offset of the k-th sampling point in the g-th group.

[0021] Furthermore, in the backbone network, the spatial pyramid pooling module includes an eighth convolutional block, a first max pooling layer, a second max pooling layer, a third max pooling layer connected in sequence, and a fifth splicing layer and a ninth convolutional block connected to each other;

[0022] Among them, the first max pooling layer, the second max pooling layer, and the third max pooling layer are jointly connected to the fifth splicing layer.

[0023] Furthermore, in the S22, the neck network includes a first splicing layer, a third CSP bottleneck module, a sixth convolutional block, a third splicing layer, a fourth CSP bottleneck module, a seventh convolutional block, a fourth splicing layer, and a fifth CSP bottleneck module connected in sequence; and a first content-aware recombination feature enhancement module, a second CSP bottleneck module, a second splicing layer, and a second content-aware recombination feature enhancement module;

[0024] Among them, the first splicing layer is also connected to the first content-aware recombination feature enhancement module and the second CSP bottleneck module in sequence; the second content-aware recombination feature enhancement module is connected to the second splicing layer and the second CSP bottleneck module in sequence; the second CSP bottleneck module is connected to the third splicing layer;

[0025] The first splicing layer, the second splicing layer, the second content-aware recombination feature enhancement module, and the fourth splicing layer are connected to the backbone network;

[0026] The third CSP bottleneck module, the fourth CSP bottleneck module, and the fifth CSP bottleneck module are connected to the head network.

[0027] Furthermore, in the neck network, each content-aware recombination feature enhancement module includes a content-aware recombination unit and an upsampling prediction unit;

[0028] Among them, the content-aware recombination unit includes a channel compressor, a content encoder, a pixel reorganizer, and a kernel normalizer connected in sequence.

[0029] Furthermore, in the S23, the head network includes a first dynamic detection head DyHead and a first detection layer connected to each other, a second dynamic detection head DyHead and a second detection layer connected to each other, and a third dynamic detection head DyHead and a third detection layer connected to each other;

[0030] The first dynamic detection head DyHead, the second dynamic detection head DyHead and the third dynamic detection head DyHead are connected to the neck network.

[0031] Furthermore, in the head network, each dynamic detection head DyHead is represented as:

[0032] W(F)=π C (π S (π L (F)·F)·F)·F

[0033] Among them, F represents the feature tensor, π L represents the scale-aware attention module, π S represents the spatial perception attention module, π C Represents the task-aware attention module.

[0034] Furthermore, in the improved YOLOv8 model, a binary cross entropy loss function is used for classification; an Inner-CIoU loss function is used for bounding box regression prediction;

[0035] Among them, the Inner-CIoU loss function is expressed as:

[0036] L Inner-CIoU =L CIoU +IoU-IoU Inner

[0037] Among them, L CIoU represents the CIoU loss function, IoU represents the overlap between the predicted box and the real box, IoU Inner Indicates the IoU value calculated by the auxiliary bounding box.

[0038] It can be seen from the above technical solution that, compared with the prior art, the technical solution of the present invention has the following beneficial effects:

[0039] 1. This method can more effectively capture multi-scale information in shale digital core images by introducing the C2f_DCNv3 module and the spatial pyramid pooling module into the backbone network. The C2f_DCNv3 module uses the deformable convolutional network DCNv3 to enhance the model's adaptability to complex shapes and structures, while the spatial pyramid pooling module ensures robustness to targets of different sizes, making the model perform better when dealing with fine and complex geological structures such as shale pores.

[0040] 2. The neck network uses a content-aware reorganization feature enhancement module combined with multiple CSP bottleneck modules to achieve effective fusion and reorganization from low-level features to high-level features. It not only helps to improve the quality of feature maps, but also better preserves important detail information, which is crucial for accurately identifying and locating shale pores.

[0041] 3. The head network uses a dynamic detection head DyHead, and combines scale, space and task-aware attention mechanisms to achieve efficient classification and bounding box regression prediction of shale pore types. The multi-level and multi-dimensional attention mechanism allows the model to focus on the most discriminative features, thereby improving recognition accuracy.

[0042] 4. The introduction of the Inner-CIoU loss function provides more precise guidance for bounding box regression prediction. By considering the IoU value calculated by the auxiliary border, this loss function can effectively reduce the difference between the predicted box and the true box, and improve the positioning accuracy of the model in dense pore areas. BRIEF DESCRIPTION OF THE DRAWINGS

[0043] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying creative work.

[0044] Figure 1 An improved YOLOv8 model structure framework diagram provided by an embodiment of the present invention;

[0045] Figure 2 A structural framework diagram of the C2f_DCNv3 module provided in an embodiment of the present invention;

[0046] Figure 3 A structural framework diagram of a bottleneck unit provided in an embodiment of the present invention;

[0047] Figure 4 A structural framework diagram of a spatial pyramid pooling module provided by an embodiment of the present invention;

[0048] Figure 5 A structural framework diagram of a content-aware reorganization feature enhancement module provided in an embodiment of the present invention;

[0049] Figure 6 A structural framework diagram of a dynamic detection head DyHead provided in an embodiment of the present invention;

[0050] Figure 7 A structural framework diagram of the detection layer provided by an embodiment of the present invention;

[0051] Figure 8 A convolutional block structure framework diagram provided for an embodiment of the present invention. DETAILED DESCRIPTION

[0052] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0053] This embodiment provides a shale pore type identification method based on an improved YOLOv8 model, comprising the following steps:

[0054] S1. Acquire shale digital core images and their annotation information to construct a data set;

[0055] S2. Using the data set to train an improved YOLOv8 model to obtain a trained improved YOLOv8 model;

[0056] like Figure 1 As shown, the improved YOLOv8 model training includes:

[0057] S21, extracting features from the shale digital core images in the data set based on the backbone network to generate feature maps of different sizes;

[0058] S22, using the neck network to perform feature fusion on the feature maps of different sizes to obtain a fused feature map;

[0059] S23, classifying and performing bounding box regression prediction on the fused feature map through a head network to identify the shale pore type and location;

[0060] S3. Input the shale digital core image to be processed into the trained improved YOLOv8 model to obtain the corresponding shale pore type recognition result.

[0061] This method improves the multi-scale information capture capability by introducing the C2f_DCNv3 module and the spatial pyramid pooling module. The neck network uses the content-aware reorganization feature enhancement module to achieve effective feature fusion. The head network uses the dynamic detection head DyHead combined with the multi-dimensional attention mechanism to improve the classification and bounding box regression accuracy. At the same time, the Inner-CIoU loss function is introduced to optimize the bounding box prediction, which significantly improves the accuracy and robustness of shale pore type recognition.

[0062] The following is a further detailed description of each step in the above method:

[0063] In this embodiment S1, a shale digital core image and its annotation information are obtained to construct a data set;

[0064] Specifically, digital core images of shale samples are first obtained through high-resolution scanning equipment to ensure that the images can clearly show the fine structure of shale pores. Subsequently, the obtained shale digital core images are carefully manually annotated to clearly distinguish different types of shale pores (such as organic pores and inorganic pores), and the specific locations and boundaries of the pores are marked. These annotation information will serve as supervisory signals for model training to ensure that the model can learn accurate pore characteristics. Finally, the annotated shale digital core images and their corresponding annotation information are integrated into a complete data set for subsequent model training and verification.

[0065] In this embodiment S2, the improved YOLOv8 model is trained using the data set to obtain a trained improved YOLOv8 model; the structure of the improved YOLOv8 model and the related training process are further described below:

[0066] Specifically, the backbone network includes a first convolution block, a second convolution block, a first CSP bottleneck module, a third convolution block, a first C2f_DCNv3 module, a fourth convolution block, a second C2f_DCNv3 module, a fifth convolution block, a third C2f_DCNv3 module, and a spatial pyramid pooling module, which are connected in sequence;

[0067] Among them, the first C2f_DCNv3 module, the second C2f_DCNv3 module and the spatial pyramid pooling module are connected to the neck network.

[0068] like Figure 2 As shown, each C2f_DCNv3 module in the backbone network includes a convolution, batch normalization and activation function combination unit, a splitting unit, a first bottleneck unit, a second bottleneck unit, a splicing unit, and a convolution, batch normalization and activation function combination unit connected in sequence.

[0069] like Figure 3 As shown, each bottleneck unit includes two interconnected deformable convolutional networks DCNv3; the deformable convolutional network DCNv3 is expressed as:

[0070]

[0071] Among them, y(P 0 ) represents the output feature value of the deformable convolutional network DCNv3, G represents the number of spatial clustering groups, K represents the number of sampling points in each spatial clustering group, and w g represents the weight coefficient of the gth group, m gk represents the weight coefficient of the kth sampling point in the gth group, x g represents the feature map of the g-th group, P0 represents the base position in the feature map, p k The position of the kth sampling point, Δp gk Indicates the offset of the kth sampling point in the gth group.

[0072] like Figure 4 As shown, the spatial pyramid pooling module in the backbone network includes the eighth convolution block, the first maximum pooling layer, the second maximum pooling layer, the third maximum pooling layer connected in sequence, and the fifth splicing layer and the ninth convolution block connected to each other;

[0073] Among them, the first maximum pooling layer, the second maximum pooling layer, and the third maximum pooling layer are jointly connected to the fifth splicing layer.

[0074] The backbone network structure design achieves efficient extraction of multi-scale features of shale pores by introducing the C2f_DCNv3 module and combining it with the deformable convolutional network DCNv3. Each C2f_DCNv3 module dynamically adjusts the sampling position through deformable convolution to better adapt to the complex morphology and scale changes of the pores, significantly improving the accuracy and robustness of feature extraction. At the same time, the spatial pyramid pooling module processes feature maps through multi-level maximum pooling layers, enhancing the model's perception of pores of different scales, thereby improving the overall detection accuracy and generalization ability.

[0075] Specifically, the neck network includes a first splicing layer, a third CSP bottleneck module, a sixth convolution block, a third splicing layer, a fourth CSP bottleneck module, a seventh convolution block, a fourth splicing layer, and a fifth CSP bottleneck module connected in sequence; and a first content-aware reorganization feature enhancement module, a second CSP bottleneck module, a second splicing layer, and a second content-aware reorganization feature enhancement module;

[0076] The first splicing layer is also connected to the first content-aware reorganization feature enhancement module and the second CSP bottleneck module in sequence; the second content-aware reorganization feature enhancement module is connected to the second splicing layer and the second CSP bottleneck module in sequence; the second CSP bottleneck module is connected to the third splicing layer;

[0077] The first splicing layer, the second splicing layer, the second content-aware reorganization feature enhancement module, and the fourth splicing layer are connected to the backbone network;

[0078] The third CSP bottleneck module, the fourth CSP bottleneck module, and the fifth CSP bottleneck module are connected to the head network.

[0079] like Figure 5 As shown, in the neck network, each content-aware reorganization feature enhancement module includes a content-aware reorganization unit and an upsampling prediction unit;

[0080] The content-aware reorganization unit includes a channel compressor, a content encoder, a pixel reorganizer, and a kernel normalizer which are connected in sequence.

[0081] The neck network improves the effect of feature fusion by introducing a content-aware reorganization feature enhancement module. The channel compressor in the module effectively reduces the computational cost, and the content encoder dynamically generates an adaptive kernel. Through the combination of a pixel reorganizer and a kernel normalizer, fine reorganization and enhancement of feature maps are achieved. It retains more semantic information, making feature maps of different scales more efficient in the fusion process, thereby enhancing the model's ability to identify and detect complex shale pore structures.

[0082] Specifically, the head network includes a first dynamic detection head DyHead and a first detection layer connected to each other, a second dynamic detection head DyHead and a second detection layer connected to each other, and a third dynamic detection head DyHead and a third detection layer connected to each other;

[0083] The first dynamic detection head DyHead, the second dynamic detection head DyHead and the third dynamic detection head DyHead are connected to the neck network.

[0084] like Figure 6 As shown, each dynamic detection head DyHead in the head network includes a scale-aware attention module, a space-aware attention module, and a task-aware attention module connected in sequence; each dynamic detection head DyHead is represented as:

[0085] W(F)=π C (π S (π L 9F)·F)·F)·F

[0086] Among them, F represents the feature tensor, π L represents the scale-aware attention module, π S represents the spatial perception attention module, π C Represents the task-aware attention module.

[0087] like Figure 7 As shown in the figure, each detection layer in the head network consists of multiple convolution blocks for the final shale pore type identification and location prediction. Specifically, the detection layer includes the tenth convolution block, the eleventh convolution block and the two-dimensional convolution layer for identifying the shale pore type; and the twelfth convolution block, the thirteenth convolution block and the two-dimensional convolution layer for predicting the location of shale pores. These convolution blocks gradually extract features through multi-layer convolution operations, and finally perform classification and regression prediction through the two-dimensional convolution layer to ensure that the model can efficiently and accurately identify and locate shale pores.

[0088] The neck network introduces a content-aware reorganization feature enhancement module. The channel compressor in the module effectively reduces the computational cost. The content encoder dynamically generates an adaptive kernel. In conjunction with the pixel reorganizer and kernel normalizer, the fine reorganization and enhancement of the feature map are achieved, retaining more semantic information. This makes the fusion process of feature maps of different scales more efficient, and enhances the model's ability to identify and detect complex shale pore structures. At the same time, the head network adopts multiple dynamic detection head DyHead structures, combined with scale-aware, space-aware, and task-aware attention modules, to further improve the model's recognition accuracy and positioning accuracy for shale pore types.

[0089] like Figure 8 As shown in Figure 1, each convolution block of the improved YOLOv8 model is composed of a two-dimensional convolutional layer, a two-dimensional batch normalization layer, and a weighted linear layer connected in sequence for feature extraction, data normalization, and nonlinear transformation.

[0090] Furthermore, in the traditional YOLOv8 network structure, CIoU is used to predict the bounding box regression task. The CIoU border regression considers the distance between the center point of the predicted box and the true box and the similarity of the border aspect ratio in the loss term, but its generalization is weak in practical applications, which is not conducive to improving the quality of detection results.

[0091] This embodiment uses Inner-IoU to improve the original CIoU. Inner-IoU calculates the IoU loss through the auxiliary border. The Inner-CIoU loss function is as follows:

[0092]

[0093] in, represents the center point inside the real box, (x c ,y c ) represents the center point inside the prediction box, w gt and h gt Represents the width and height of the real box, w and h represent the width and height of the predicted box, and the variable ratio represents the scale factor, which is used to control the scale of the auxiliary box;

[0094] They represent the left, right, top, and bottom boundaries of the real box generated by Inner-IoU, respectively. l , b r , b t , b b They represent the left, right, top, and bottom boundaries of the prediction box generated by Inner-IoU, respectively, and are calculated by the center point inside the box and the scale factor ratio.

[0095]

[0096] Final result: L Inner-CIoU =L CIoU +IoU-IoU Inner

[0097] L CIoU represents the CIoU loss function, IoU represents the overlap between the predicted box and the real box, IoU Inner Indicates the IoU value calculated by the auxiliary bounding box.

[0098] In the Inner-CIoU loss function, ratio is an adjustable hyperparameter. In the shale pore identification task, ratio = 0.75 is adjusted, the auxiliary bounding box size is smaller than the actual bounding box, and using a smaller-scale auxiliary bounding box to calculate the IoU loss will help the regression of high IoU samples. The Inner-CIoU loss function takes into account the advantages of both loss functions, while improving the quality of the detection results, it speeds up the convergence of the prediction box.

[0099] In this embodiment S3, the shale digital core image to be processed is input into the trained improved YOLOv8 model to obtain the corresponding shale pore type recognition result;

[0100] In step S3, when the shale digital core image to be processed is input into the trained improved YOLOv8 model, the model first uses its backbone network to perform deep feature extraction on the input image. Thanks to the efficient collaboration of the C2f_DCNv3 module and the spatial pyramid pooling module, it can capture the complex multi-scale characteristics of shale pores. These feature maps not only contain the subtle structural information of shale pores, but also adjust the sampling position through a carefully designed deformable convolution layer to adapt to pores of different shapes and sizes, thereby providing high-quality feature support for subsequent classification and positioning tasks.

[0101] Subsequently, the feature maps extracted by the backbone network enter the neck network for feature fusion and enhancement. The quality of the feature maps is further optimized through the content-aware reorganization feature enhancement module, which retains more semantic information and enables the effective integration of features of different scales. Finally, these optimized feature maps are processed by multiple dynamic detection heads DyHead in the head network, combining scale-aware, spatial-aware, and task-aware attention mechanisms to achieve high-precision recognition and location prediction of shale pore types.

[0102] Through this series of precisely designed processes, the input shale digital core images can be efficiently converted into accurate shale pore type identification results, greatly improving the efficiency and accuracy of shale pore analysis.

[0103] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0104] The above description of the disclosed embodiments enables one skilled in the art to implement or use the present invention. Various modifications to these embodiments will be apparent to one skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the present invention. Therefore, the present invention will not be limited to the embodiments shown herein, but rather to the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A shale pore type identification method based on an improved YOLOv8 model, characterized in that: The following steps are involved: S1. Acquire shale digital core images and their annotation information to construct a data set; S2. Using the data set to train an improved YOLOv8 model to obtain a trained improved YOLOv8 model; The improved YOLOv8 model training includes: S21, extracting features from the shale digital core images in the data set based on the backbone network to generate feature maps of different sizes; S22, using the neck network to perform feature fusion on the feature maps of different sizes to obtain a fused feature map; S23, classifying and performing bounding box regression prediction on the fused feature map through a head network to identify the shale pore type and location; S3. Input the shale digital core image to be processed into the trained improved YOLOv8 model to obtain the corresponding shale pore type recognition result.

2. The shale pore type identification method based on the improved YOLOv8 model according to claim 1 is characterized in that: In S21, the backbone network includes a first convolution block, a second convolution block, a first CSP bottleneck module, a third convolution block, a first C2f_DCNv3 module, a fourth convolution block, a second C2f_DCNv3 module, a fifth convolution block, a third C2f_DCNv3 module and a spatial pyramid pooling module connected in sequence; Among them, the first C2f_DCNv3 module, the second C2f_DCNv3 module and the spatial pyramid pooling module are connected to the neck network.

3. The shale pore type identification method based on the improved YOLOv8 model according to claim 2 is characterized in that: In the backbone network, each C2f_DCNv3 module includes a convolution, batch normalization and activation function combination unit, a splitting unit, a first bottleneck unit, a second bottleneck unit, a splicing unit, and a convolution, batch normalization and activation function combination unit connected in sequence.

4. The shale pore type identification method based on the improved YOLOv8 model according to claim 3 is characterized in that: Each bottleneck unit includes two interconnected deformable convolutional networks DCNv3; the deformable convolutional network DCNv3 is expressed as: Where y9P0) represents the output feature value of the deformable convolutional network DCNv3, G represents the number of spatial clustering groups, K represents the number of sampling points in each spatial clustering group, and w g represents the weight coefficient of the gth group, m gk represents the weight coefficient of the kth sampling point in the gth group, x g represents the feature map of the gth group, P0 represents the base position in the feature map, and p k The position of the kth sampling point, Δp gk Indicates the offset of the kth sampling point in the gth group.

5. The shale pore type identification method based on the improved YOLOv8 model according to claim 2 is characterized in that: In the backbone network, the spatial pyramid pooling module includes an eighth convolution block, a first maximum pooling layer, a second maximum pooling layer, a third maximum pooling layer connected in sequence, and a fifth splicing layer and a ninth convolution block connected to each other; Among them, the first maximum pooling layer, the second maximum pooling layer, and the third maximum pooling layer are jointly connected to the fifth splicing layer.

6. The shale pore type identification method based on the improved YOLOv8 model according to claim 1 is characterized in that: In S22, the neck network includes a first splicing layer, a third CSP bottleneck module, a sixth convolution block, a third splicing layer, a fourth CSP bottleneck module, a seventh convolution block, a fourth splicing layer, and a fifth CSP bottleneck module connected in sequence; and a first content-aware reorganization feature enhancement module, a second CSP bottleneck module, a second splicing layer, and a second content-aware reorganization feature enhancement module; The first splicing layer is also connected to the first content-aware reorganization feature enhancement module and the second CSP bottleneck module in sequence; the second content-aware reorganization feature enhancement module is connected to the second splicing layer and the second CSP bottleneck module in sequence; the second CSP bottleneck module is connected to the third splicing layer; The first splicing layer, the second splicing layer, the second content-aware recombinant feature enhancement module and the fourth splicing layer are connected to a backbone network; The third CSP bottleneck module, the fourth CSP bottleneck module, and the fifth CSP bottleneck module are connected to the head network.

7. The shale pore type identification method based on the improved YOLOv8 model according to claim 6 is characterized in that: In the neck network, each content-aware reorganization feature enhancement module includes a content-aware reorganization unit and an upsampling prediction unit; The content-aware reorganization unit includes a channel compressor, a content encoder, a pixel reorganizer, and a kernel normalizer which are connected in sequence.

8. The shale pore type identification method based on the improved YOLOv8 model according to claim 1 is characterized in that: In S23, the head network includes a first dynamic detection head DyHead and a first detection layer connected to each other, a second dynamic detection head DyHead and a second detection layer connected to each other, and a third dynamic detection head DyHead and a third detection layer connected to each other; The first dynamic detection head DyHead, the second dynamic detection head DyHead and the third dynamic detection head DyHead are connected to the neck network.

9. The shale pore type identification method based on the improved YOLOv8 model according to claim 8 is characterized in that: In the head network, each dynamic detection head DyHead is represented as: W(F)=π C (p S (p L (F)·F)·F)·F Among them, F represents the feature tensor, π L represents the scale-aware attention module, π S represents the spatial perception attention module, π C Represents the task-aware attention module.

10. The shale pore type identification method based on the improved YOLOv8 model according to claim 1 is characterized in that: In the improved YOLOv8 model, a binary cross entropy loss function is used for classification; an Inner-CIoU loss function is used for bounding box regression prediction; Among them, the Inner-CIoU loss function is expressed as: L Inner-CIoU =L CIoU +IoU-IoU Inner Among them, L CIoU represents the CIoU loss function, IoU represents the overlap between the predicted box and the real box, IoU Inner Indicates the IoU value calculated by the auxiliary bounding box.

Images