Core rqd segmentation method based on frequency space feature enhancement and long distance dependence

Abstract

The application discloses a core RQD segmentation method based on frequency-space feature enhancement and long-distance dependence, comprising: spatial feature coding is performed on a core tank image to obtain output features, and a core feature map is determined; a morphological center axis of a core instance mask is extracted, and an accurate pixel length is calculated; a plurality of cores satisfying a core length threshold are screened, and a rock quality index in a region corresponding to the core tank is calculated. The application reduces technical bottlenecks such as loss of micro-fissure features, limited long-distance spatial dependence perception, discontinuous long-strip core segmentation and complex background interference in core segmentation; the average precision mean on the NEU-Rock dataset reaches 95.5%, which significantly reduces the parameter quantity and calculation amount compared with the prior art; and the core image can be converted into an accurate rock RQD index in real time and with high fidelity, thereby providing objective and robust digital support for mineral exploration and rock mass stability evaluation.

Description

Technical Field

[0001] This invention relates to the field of core RQD segmentation technology, and in particular to a core RQD segmentation method based on frequency-space feature enhancement and long-distance dependence. Background Technology

[0002] As the load-bearing structure of geological engineering projects, the structural characteristics, weathering state, and degree of fragmentation of rock masses directly determine the stability and safety of the engineering works. Among these parameters, the rock quality index (RQD), a key parameter for evaluating the engineering properties of rock masses, not only reflects the physical and mechanical properties of the rock mass, such as connectivity, permeability, and compressibility, but also indicates the integrity of the engineering geological body. Therefore, accurate calculation and analysis of RQD has significant theoretical and practical value for ensuring the construction safety and economic benefits of rock mass engineering projects.

[0003] Currently, RQD calculations still rely on traditional methods, primarily using manual recording by surveyors. This is not only time-consuming and labor-intensive, but also highly susceptible to human error in processing large volumes of data. With technological advancements, these traditional measurement methods can no longer meet the demands of current societal productivity. Therefore, transitioning core RQD calculations from manual to intelligent methods has become an urgent need in the field of geological engineering.

[0004] Although several deep learning-based methods have been developed to automatically calculate core RQD, existing automatic detection methods still face significant technical bottlenecks in practical applications, especially in real-world field environments:

[0005] Subpixel-level spatial feature dilution: Traditional deep learning models often use strided convolution or pooling layers for downsampling to reduce spatial resolution. During this process, minute structural features such as tiny cracks, bedding, and fracture zones, which occupy a very small proportion of pixels in the core image, are diluted, causing the model to lose key geometric features and texture information when processing high-level semantics.

[0006] Limited perception of long-range spatial dependencies: Conventional convolutional kernels (such as 3×3) have limited receptive fields, only able to capture textures in local areas. Due to the lack of global perception capabilities, long-range dependencies between features cannot be established, resulting in long strip cores often forming several discontinuous short strips after segmentation, which seriously interferes with the accurate extraction of core length and the quantification of RQD index.

[0007] Feature recognition bias under complex backgrounds and lighting interference: In field environments, the shooting angles of core boxes are inconsistent and the image resolution is low. When the core area is highly similar in color and stripe to the background of the core box, or when faced with uneven lighting or obstruction by debris, existing models struggle to effectively suppress background noise, leading to decreased focusing ability of the core area and poor accuracy and robustness of boundary regression. Therefore, it is of great significance to propose a core RQD segmentation method that can balance inference speed, complex background suppression, and long-distance texture association extraction to address the above problems. Summary of the Invention

[0008] Based on this, it is necessary to propose a core RQD segmentation method based on frequency-space feature enhancement and long-distance dependence to address the above problems.

[0009] A core RQD segmentation method based on frequency-space feature enhancement and long-range dependence, the method comprising:

[0010] S1: Obtain the core box image corresponding to the core box, and map it uniformly to the original core image in standard tensor format through preprocessing operations; perform initial spatial feature encoding on the original core image to obtain cross-stage convolutional features containing core edges, colors and shallow textures;

[0011] S2: The cross-stage convolutional features are demixed and shuffled into four independent and complementary first sub-feature maps using a spatial segmentation operator. The four first sub-feature maps are then concatenated along the channel dimension to obtain a first intermediate enhanced feature map. The first intermediate enhanced feature map is then convolved to obtain a downsampled and fully pixel-retained first spatial depth feature.

[0012] S3: Use a spatial segmentation operator to demix and shuffle the first spatial depth feature into four independent and complementary second sub-feature maps, and then concatenate the four second sub-feature maps in the channel dimension to obtain a second intermediate enhanced feature map. Finally, perform a convolution operation on the second intermediate enhanced feature map to obtain a second spatial depth feature that is downsampled and retains all pixels.

[0013] S4: Superimpose depthwise separable convolutions with different dilation rates on the second spatial depth features to construct multi-scale candidate features; perform global average pooling and max pooling operations on the multi-scale candidate features respectively, then concatenate and fuse them along the channel dimension before performing convolution operations, and perform element-wise multiplication based on the corresponding multi-scale candidate features to obtain large kernel selection features; sequentially perform spatial pyramid pooling fast operation and two-stage channel attention feature extraction operation on the large kernel selection features to obtain output features;

[0014] S5: Perform sampling space interpolation on the output feature and the large kernel selection feature to obtain the first merged feature; and perform purification and splicing operations and feature integration operations on the first merged feature in sequence to obtain the attention feature; determine the first attention cross-enhancement feature, the second attention cross-enhancement feature and the third attention cross-enhancement feature based on the second spatial depth feature and the attention feature;

[0015] S6: The first attention cross-enhancing feature, the second attention cross-enhancing feature, and the third attention cross-enhancing feature are obtained respectively under different grain sizes as shallow, medium and deep segmentation features. The shallow, medium and deep segmentation features are then merged to obtain a segmented core feature map.

[0016] S7: Perform instance-level separation on the core mask feature map to extract N independent core fragment instance masks existing in the core mask feature map, forming a core fragment instance mask set M={M1,M2,...,M...} N The morphological central axis of each core instance mask in the core fragment instance mask set is extracted using the skeleton extraction method, and the cumulative distance of the pixel skeleton points on the morphological central axis is calculated, which is the precise pixel length of the core instance mask.

[0017] S8: Extract the pixel width of the core box in the core box image based on the known standard physical width of the core box, and calculate the scale conversion factor; use the scale conversion factor to convert the precise pixel length into the physical length of the core in geological engineering; perform integrity screening on the cores in the core box based on the physical length of the core to obtain multiple cores that meet the core length threshold; calculate the rock quality index in the area corresponding to the core box based on the multiple cores that meet the threshold condition and the total drilling length, and output it in the form of a percentage.

[0018] In one embodiment, the cross-stage convolutional feature is implemented using the following expression:

[0019]

[0020] in, This represents cross-stage convolutional features; This is the original core image; Conv represents the convolution operation; This is a cross-stage local network feature extraction operation.

[0021] In one embodiment, the demixing and shuffling of the cross-stage convolutional features using a spatial segmentation operator results in four independent and complementary first sub-feature maps. These four first sub-feature maps are then concatenated along the channel dimension to obtain a first intermediate enhanced feature map. Finally, a convolution operation is performed on the first intermediate enhanced feature map to obtain a downsampled and fully pixel-preserving first spatial depth feature, achieved through the following expression:

[0022]

[0023]

[0024]

[0025]

[0026]

[0027]

[0028] in, , , and There are four first sub-feature maps; S is the resolution; This is the first intermediate enhanced feature map; This represents the first spatial depth feature.

[0029] In one embodiment, the second spatial depth features are superimposed with depthwise separable convolutions with different dilation rates to construct multi-scale candidate features; after performing global average pooling and max pooling operations on the multi-scale candidate features respectively, they are concatenated and fused along the channel dimension before performing convolution operations, and element-wise multiplication is performed based on the corresponding multi-scale candidate features to obtain large kernel selection features; the large kernel selection features are then sequentially subjected to spatial pyramid pooling fast operation and two-stage channel attention feature extraction operation to obtain the output features, which are implemented by the following expression:

[0030]

[0031]

[0032]

[0033]

[0034]

[0035]

[0036] in, , , and These are four independent and complementary second sub-feature maps; S is the resolution. This is the second intermediate enhanced feature map; This represents the second spatial depth feature.

[0037] In one embodiment, the output feature is implemented by the following expression:

[0038]

[0039]

[0040]

[0041] in, Multi-scale candidate features; This indicates that the kernel size is k. i The expansion rate is d i Depth-separable convolution operations; This is a second spatial depth feature; This is a large-nucleus selection feature with an extremely large receptive field; This represents the operation of concatenating and fusing the spatial saliency features of global average pooling and max pooling operations along the channel dimension. For activation functions; This indicates that spatial pyramid pooling is a fast operation; This indicates a two-stage channel attention feature extraction operation; These are the output characteristics of the backbone network.

[0042] In one embodiment, the attention feature is implemented by the following expression:

[0043]

[0044]

[0045] in, This is the first merging feature; This is an upsampling spatial interpolation operation; The first upsampled feature is obtained by upsampling the backbone network features; Attention features; This is a parameterless attention mechanism operation; This is a cross-stage local network feature extraction operation.

[0046] In one embodiment, the first attention cross-enhancement feature, the second attention cross-enhancement feature, and the third attention cross-enhancement feature are implemented by the following expression:

[0047]

[0048]

[0049]

[0050]

[0051]

[0052]

[0053] in, This is the second merging feature; This is the second upsampling feature; This is a first attention cross-enhancement feature; This is the fourth convolution feature; This is the third merging characteristic; This is a second attention cross-enhancement feature; This is the fifth convolution feature; This is the fourth merging feature; This is a third attention cross-enhancement feature.

[0054] In one embodiment, the segmented core feature map is implemented by the following expression:

[0055]

[0056]

[0057]

[0058]

[0059] in, This is a shallow segmentation feature; This is a mid-level segmentation feature; This represents deep segmentation features; This is a segmentation diagram of the core features.

[0060] In one embodiment, the precise pixel length is achieved by the following expression:

[0061]

[0062] in, For precise pixel length; K is the core instance mask M.i The total number of pixel skeleton points on the morphological central axis, ( , M is the mask for the j-th core instance. i The pixel coordinates of the pixel skeleton points on the morphological central axis.

[0063] In one embodiment, the rock quality index is achieved by the following expression:

[0064]

[0065]

[0066]

[0067]

[0068] in, This is the scale conversion factor; Standard physical width; The width is in pixels; The physical length of the core sample; For precise pixel length; This is a screening operation for core samples that meet the threshold conditions; This is the core length threshold; This refers to the drilling footage length. For rock quality indicators; This represents the total number of core samples taken in one advance.

[0069] The method of this invention reduces technical bottlenecks in core segmentation, such as loss of micro-fracture features, limited perception of long-distance spatial dependence, discontinuity in long strip core segmentation, and interference from complex backgrounds. The average accuracy mAp50 on the NEU-Rock dataset reaches 95.5%, which significantly reduces the number of parameters and computation compared to existing mainstream algorithms. It can convert core images into accurate rock RQD indicators in real time and with high fidelity, providing objective and robust digital support for mineral exploration and rock mass stability assessment. Attached Figure Description

[0070] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0071] in:

[0072] Figure 1This is an application environment diagram of a core RQD segmentation method based on frequency-space feature enhancement and long-distance dependence in one embodiment;

[0073] Figure 2 This is a flowchart of a core RQD segmentation method based on frequency-space feature enhancement and long-distance dependence in one embodiment;

[0074] Figure 3 This is a structural block diagram of a computer device in one embodiment. Detailed Implementation

[0075] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0076] To address the technical problems in the background art, this application provides a core RQD segmentation method based on frequency-space feature enhancement and long-distance dependence.

[0077] Figure 1 This is an application environment diagram of a core RQD segmentation method based on frequency-space feature enhancement and long-range dependence in one embodiment. (Refer to...) Figure 1This core RQD segmentation method based on frequency-space feature enhancement and long-range dependence is applied to a core RQD segmentation system based on frequency-space feature enhancement and long-range dependence. The system includes a terminal 110 and a server 120. The terminal 110 and server 120 are connected via a network. The terminal 110 can be a desktop terminal or a mobile terminal; the mobile terminal can be at least one of a mobile phone, tablet, or laptop. The server 120 can be a standalone server or a server cluster consisting of multiple servers.Terminal 110 is used to acquire the core box image corresponding to the core box, and to map it uniformly into the original core image in standard tensor format through preprocessing operations; the original core image is subjected to initial spatial feature encoding to obtain cross-stage convolutional features containing core edges, colors, and shallow textures. Server 120 is used to demix and shuffle the cross-stage convolutional features into four independent and complementary first sub-feature maps using a spatial segmentation operator, and to concatenate the four first sub-feature maps in the channel dimension to obtain a first intermediate enhanced feature map, and to perform a convolution operation on the first intermediate enhanced feature map to obtain downsampled and The first spatial depth feature retains all pixels; the first spatial depth feature is demixed and shuffled into four independent and complementary second sub-feature maps using a spatial segmentation operator, and the four second sub-feature maps are concatenated along the channel dimension to obtain a second intermediate enhanced feature map, and a convolution operation is performed on the second intermediate enhanced feature map to obtain a downsampled and fully pixel-retained second spatial depth feature; the second spatial depth feature is superimposed with depthwise separable convolutions with different dilation rates to construct multi-scale candidate features; the multi-scale candidate features are subjected to global average pooling and max pooling operations respectively, concatenated and fused along the channel dimension, and then convolved, and element-wise multiplication is performed based on the corresponding multi-scale candidate features to obtain a large kernel selection feature; the large kernel selection feature is sequentially subjected to spatial pyramid pooling fast operation and two-stage channel attention feature extraction operation to obtain output features; the output features and the large kernel selection feature are subjected to sampling spatial interpolation operation to obtain a first merged feature; the first merged feature is sequentially subjected to purification and concatenation operation and feature integration operation to obtain attention features; based on the second spatial depth feature and the attention feature, a first attention cross-enhancement feature and a second attention cross-enhancement feature are determined. Cross-enhanced features and third attention cross-enhanced features are used. The first attention cross-enhanced features, second attention cross-enhanced features, and third attention cross-enhanced features are obtained at different granularities for shallow, medium, and deep segmentation features. These features are then merged to obtain a segmented core feature map. Instance-level separation is performed on the core mask feature map to extract N independent core fragment instance masks, forming a core fragment instance mask set M = {M1, M2, ..., M}. NThe morphological central axis of each core instance mask in the core fragment instance mask set is extracted using a skeleton extraction method, and the cumulative distance of the pixel skeleton points on the morphological central axis is calculated, which is the precise pixel length of the core instance mask. The pixel width of the core box in the core box image is extracted based on the known standard physical width of the core box, thereby calculating the scale conversion factor. Using the scale conversion factor, the precise pixel length is converted into the physical length of the core in geological engineering. Based on the physical length of the core, the cores in the core box are screened for integrity to obtain multiple cores that meet the core length threshold. Based on the multiple cores that meet the threshold and the total drilling length, the rock quality index of the area corresponding to the core box is calculated and output as a percentage.

[0078] like Figure 2 As shown, in one embodiment, a core RQD segmentation method based on frequency-space feature enhancement and long-distance dependence is provided. This method can be applied to both terminals and servers; this embodiment illustrates its application to a terminal. The core RQD segmentation method based on frequency-space feature enhancement and long-distance dependence specifically includes the following steps:

[0079] S1: Obtain the core box image corresponding to the core box, and map it to the original core image in standard tensor format (uniformly scaled to 640×640 standard tensor input format) through preprocessing operations; perform initial spatial feature encoding on the original core image to obtain cross-stage convolution features containing core edges, colors and shallow textures;

[0080] S2: The cross-stage convolutional features are demixed and shuffled into four independent and complementary first sub-feature maps using a spatial segmentation operator. The four first sub-feature maps are then concatenated along the channel dimension to obtain a first intermediate enhanced feature map. The first intermediate enhanced feature map is then convolved to obtain a downsampled and fully pixel-retained first spatial depth feature.

[0081] S3: Use a spatial segmentation operator to demix and shuffle the first spatial depth feature into four independent and complementary second sub-feature maps, and then concatenate the four second sub-feature maps in the channel dimension to obtain a second intermediate enhanced feature map. Finally, perform a convolution operation on the second intermediate enhanced feature map to obtain a second spatial depth feature that is downsampled and retains all pixels.

[0082] S4: Superimpose depthwise separable convolutions with different dilation rates on the second spatial depth features to construct multi-scale candidate features; perform global average pooling and max pooling operations on the multi-scale candidate features respectively, then concatenate and fuse them along the channel dimension before performing convolution operations, and perform element-wise multiplication based on the corresponding multi-scale candidate features to obtain large kernel selection features; sequentially perform spatial pyramid pooling fast operation and two-stage channel attention feature extraction operation on the large kernel selection features to obtain output features;

[0083] S5: Perform sampling space interpolation on the output feature and the large kernel selection feature to obtain the first merged feature; and perform purification and splicing operations and feature integration operations on the first merged feature in sequence to obtain the attention feature; determine the first attention cross-enhancement feature, the second attention cross-enhancement feature and the third attention cross-enhancement feature based on the second spatial depth feature and the attention feature;

[0084] S6: The first attention cross-enhancing feature, the second attention cross-enhancing feature, and the third attention cross-enhancing feature are obtained respectively under different grain sizes as shallow, medium and deep segmentation features. The shallow, medium and deep segmentation features are then merged to obtain a segmented core feature map.

[0085] S7: Perform instance-level separation on the core mask feature map to extract N independent core fragment instance masks existing in the core mask feature map, forming a core fragment instance mask set M={M1,M2,...,M...} N The morphological central axis of each core instance mask in the core fragment instance mask set is extracted using the skeleton extraction method, and the cumulative distance of the pixel skeleton points on the morphological central axis is calculated, which is the precise pixel length of the core instance mask.

[0086] S8: Extract the pixel width of the core box in the core box image based on the known standard physical width of the core box, and calculate the scale conversion factor; use the scale conversion factor to convert the precise pixel length into the physical length of the core in geological engineering; perform integrity screening on the cores in the core box based on the physical length of the core to obtain multiple cores that meet the core length threshold; calculate the rock quality index in the area corresponding to the core box based on the multiple cores that meet the threshold condition and the total drilling length, and output it in the form of a percentage.

[0087] In one embodiment, the cross-stage convolutional feature is implemented using the following expression:

[0088]

[0089] in, ∈RS×S×C represents the cross-stage convolutional feature output after shallow feature refinement, which includes basic core texture information and prior features of small cracks. Its spatial resolution is 1 / 4 of the original image. R is the set of real numbers, S is the resolution, and C is the number of channels. This is the original core image; Conv represents the convolution operation, which includes standard convolution operations such as 2D convolution, batch normalization, and the SiLU nonlinear activation function. This is a cross-stage local network feature extraction operation used to achieve gradient enrichment flow in shallow features.

[0090] In one embodiment, the demixing and shuffling of the cross-stage convolutional features using a spatial segmentation operator results in four independent and complementary first sub-feature maps. These four first sub-feature maps are then concatenated along the channel dimension to obtain a first intermediate enhanced feature map. Finally, a convolution operation is performed on the first intermediate enhanced feature map to obtain a downsampled and fully pixel-preserving first spatial depth feature, achieved through the following expression:

[0091]

[0092]

[0093]

[0094]

[0095]

[0096]

[0097] in, , , and There are four first sub-feature maps; S is the resolution; This is the first intermediate enhanced feature map; This represents the first spatial depth feature.

[0098] In one embodiment, the second spatial depth features are superimposed with depthwise separable convolutions with different dilation rates to construct multi-scale candidate features; after performing global average pooling and max pooling operations on the multi-scale candidate features respectively, they are concatenated and fused along the channel dimension before performing convolution operations, and element-wise multiplication is performed based on the corresponding multi-scale candidate features to obtain large kernel selection features; the large kernel selection features are then sequentially subjected to spatial pyramid pooling fast operation and two-stage channel attention feature extraction operation to obtain the output features, which are implemented by the following expression:

[0099]

[0100]

[0101]

[0102]

[0103]

[0104]

[0105] in, , , and These are four independent and complementary second sub-feature maps; S is the resolution. This is the second intermediate enhanced feature map; This represents the second spatial depth feature.

[0106] In one embodiment, considering that the core sample, being a long cylinder, possesses extremely strong longitudinal continuity in its texture, this invention deploys an LSKBlock module at the deepest layer of the backbone network. This module constructs an ultra-large equivalent receptive field, preferably 23×23 or larger, through large kernel decomposition convolution and spatial selection mechanisms to adaptively capture the long-range spatial dependencies of the elongated core sample. Within the LSKBlock operation, firstly, by… Multi-scale candidate features are constructed by stacking depthwise separable convolutions with different dilation rates. After obtaining multi-scale candidate features Subsequently, the system performs global average pooling (AvgPool) and max pooling (MaxPool) operations in parallel along the spatial dimension to extract spatially salient features. These two types of salient features are then concatenated and fused along the channel dimension. Based on the fused spatially salient features, the network integrates these features using a standard 1×1 two-dimensional convolutional layer and introduces a non-linear activation function. (Preferably using the Sigmoid function) The feature values ​​are normalized to generate the final spatial attention map. Finally, the generated spatial attention map is compared with the corresponding multi-scale candidate features. Perform element-wise multiplication (noted as...) The modulated features at each scale are then accumulated and integrated along the channel or spatial axis to ultimately output the large kernel selection feature. :

[0107]

[0108]

[0109]

[0110] in, Multi-scale candidate features; This indicates that the kernel size is k. i The expansion rate is d i Depth-separable convolution operations; This is a second spatial depth feature; This is a large-nucleus selection feature with an extremely large receptive field; This represents the operation of concatenating and fusing the spatial saliency features of global average pooling and max pooling operations along the channel dimension. For activation functions; This indicates that spatial pyramid pooling is a fast operation; This indicates a two-stage channel attention feature extraction operation; These are the output characteristics of the backbone network.

[0111] In one embodiment, the attention feature is implemented by the following expression:

[0112]

[0113]

[0114] in, This is the first merging feature; This is an upsampling spatial interpolation operation; The first upsampled feature is obtained by upsampling the backbone network features; Attention features; This is a parameterless attention mechanism operation; This is a cross-stage local network feature extraction operation.

[0115] In one embodiment, the first attention cross-enhancement feature, the second attention cross-enhancement feature, and the third attention cross-enhancement feature are implemented by the following expression:

[0116]

[0117]

[0118]

[0119]

[0120]

[0121]

[0122] in, This is the second merging feature; This is the second upsampling feature; This is a first attention cross-enhancement feature; This is the fourth convolution feature; This is the third merging characteristic; This is a second attention cross-enhancement feature; This is the fifth convolution feature; This is the fourth merging feature; This is a third attention cross-enhancement feature.

[0123] In one embodiment, the segmented core feature map is implemented by the following expression:

[0124]

[0125]

[0126]

[0127]

[0128] in, This is a shallow segmentation feature; This is a mid-level segmentation feature; This represents deep segmentation features; This is a segmentation diagram of the core features.

[0129] In one embodiment, the precise pixel length is achieved by the following expression:

[0130]

[0131] in, For precise pixel length; K is the core instance mask M. i The total number of pixel skeleton points on the morphological central axis, ( , M is the mask for the j-th core instance. i The pixel coordinates of the pixel skeleton points on the morphological central axis.

[0132] In one embodiment, the connected-component labeling algorithm is used to analyze the fused core mask feature map X. MaskInstance-level separation was performed to extract the core mask feature map X. Mask The mask set M = {M1, M2, ..., M} of N independent core fragment instances exists in the sample. N For each core instance mask M in the set... i Since the core may be tilted or bent within the core box, directly calculating the circumscribed rectangle would lead to length measurement errors. Therefore, this invention preferably uses skeletonization to extract the core instance mask M. i The morphological centerline is determined, and the cumulative distance of the pixel skeleton points along this morphological centerline is calculated, i.e., the core instance mask M. i Precise pixel length :

[0133]

[0134]

[0135]

[0136]

[0137] in, This is the scale conversion factor; Standard physical width; The width is in pixels; The physical length of the core sample; For precise pixel length; This is a screening operation for core samples that meet the threshold conditions; This is the core length threshold; This refers to the drilling footage length. For rock quality indicators; This represents the total number of core samples taken in one advance.

[0138] Specifically, to address the issue of low regression sensitivity for slender targets like rock cores with aspect ratio distortion, an Inner-IoU loss function operator is introduced. This operator calculates the loss by constructing a smaller-scale auxiliary inscribed box inside the original bounding box. Leveraging the high sensitivity of the small-scale bounding box to positional deviations, it provides more efficient gradient guidance for the network during training. The specific process of bounding box optimization and loss calculation is as follows:

[0139] First, obtain the preset ground truth bounding boxes B in the current training batch. gt With the predicted box B output by the model pred Let the coordinates of the center points of the two be ( ). , )and( , The original width and height are ( ), , (w,h). This method introduces an adjustable scaling factor ratio∈[0.7,0.9]. By proportionally shrinking the original bounding box size along the center point, an auxiliary inscribed true bounding box is generated inside. :

[0140]

[0141]

[0142] The auxiliary inscribed frame physically corresponds to the core high-confidence region of the core entity, effectively removing the boundary interference of blurred core edges and background debris.

[0143] After constructing the auxiliary inline boxes, the system further calculates the intersection-union ratio (IU / R) between these two sets of low-scale inline boxes to more accurately capture minute alignment deviations in the central region. Specifically, it first needs to derive the absolute geometric boundaries (i.e., the left edge) of the true and predicted inline boxes in two-dimensional space based on the center point coordinates and scaled width and height of the auxiliary inline boxes. right edge upper edge bottom edge ).

[0144] Boundary coordinates of the auxiliary inscribed bounding box ( )as follows:

[0145]

[0146]

[0147]

[0148]

[0149] Boundary coordinates of the auxiliary inscribed true bounding box ( )as follows:

[0150]

[0151]

[0152]

[0153]

[0154] Subsequently, using the two sets of boundary coordinates mentioned above, the spatial intersection area of ​​the auxiliary inscribed ground truth box and the auxiliary inscribed predicted box is extracted. The intersection area is calculated by multiplying the maximum overlap distance between the two elements on the horizontal and vertical axes (or 0 if there is no overlap).

[0155]

[0156] The area of ​​the union is equal to the sum of the independent areas of the two auxiliary inscribed boxes minus the aforementioned area of ​​the intersection.

[0157]

[0158] After obtaining the precise areas of intersection and union, the ratio of the inscribed intersection and union is then calculated. The mathematical formula for its calculation is as follows:

[0159]

[0160] Finally, the internal intersection-union ratio calculated above is used as an example. The global loss function introduced into the bounding box In this approach, gradient dynamics generated by the auxiliary path drive the parameter updates of the original bounding box. The expression is as follows:

[0161]

[0162] in, The underlying bounding box regression loss is used (e.g., CIoU loss). This Inner-IoU constraint mechanism makes the algorithm more sensitive to pixel-level offsets of the core boundary, ensuring the accuracy of the core mask feature map X output by the model. Mask The geometric long axis closely matches the physical core, laying a solid geometric foundation for subsequent high-precision RQD calculations.

[0163] Through the complete closed-loop processing described above, this invention not only solves the technical pain points of traditional manual surveying methods, which are time-consuming, labor-intensive, and prone to subjective errors, but also overcomes the shortcomings of existing deep learning algorithms in segmenting micro-fractures and long strip-shaped rock cores in complex field backgrounds. This system can convert rock core images acquired by visual sensors into accurate RQD engineering quantification indicators in real time and with high fidelity, providing objective and reliable digital geological evidence for subsequent rock core mechanical assessment, engineering geological modeling, and disaster early warning.

[0164] The size and quality of the dataset are crucial factors determining the performance of deep learning algorithms. Given the limited availability and low resolution of publicly available datasets for core box images, we constructed the NEURock dataset through extensive collection and organization. The NEURock dataset includes two categories: rock and box, and after data augmentation, contains 2000 high-quality rock images. These images were all taken manually in a mining area in Tangshan City, Hebei Province.

[0165] To expand the dataset and improve model robustness, we performed data augmentation on the original rock images, mainly in the following three aspects:

[0166] Geometric transformations. Cropping, mirroring, and scaling are used to simulate the geometric changes in images caused by various factors in real-world image acquisition devices. While rotation is a common data augmentation technique, we did not apply it to the NEURock dataset. This is because our constructed images are typically full-size images of the core boxes, containing features from both the rock and box categories. Excessive rotation would significantly impact the directional features of horizontal fractures in the core, thus interfering with model training.

[0167] Random noise. We simulated real-world conditions such as sensor malfunction and poor shooting environment by adding salt-and-pepper noise and Gaussian noise to the images. In terms of noise scale, to simulate the extremely harsh dust obstruction in a field mining area or the severe aging of an industrial camera, we used random salt-and-pepper noise in the range of 0.4-0.5 and random Gaussian noise in the range of 0.1-0.2.

[0168] For core box image segmentation, traditional Labelme annotation tools, which only support regular rectangular bounding boxes, are insufficient for accurately describing complex morphological targets such as joints and bedding in core samples. Therefore, this solution employs Labelme software for pixel-level annotation. First, the polygon tool is used to refine the target edges to obtain accurate shape information. Second, a categorized hierarchical annotation method is used, labeling core samples and core boxes as independent categories, effectively reducing visual interference during the annotation process and ensuring annotation quality against complex backgrounds. After annotation, a script converts the initial JSON format files into a standard format suitable for training the segmentation model.

[0169] To further validate the effectiveness of the proposed method, we compared it with other state-of-the-art methods using our own dataset. 1600 images were used for model training. The test and validation sets consisted of the remaining 100 and 300 images, respectively.

[0170] We begin by providing a detailed overview of the experimental setup and model performance evaluation metrics. Then, we compare our methods with other state-of-the-art approaches on the NEURock dataset and visualize the results. Finally, we validate the effectiveness of each module using ablation experiments and evaluate the synergistic effects among the modules.

[0171] The experiments in this paper were conducted using the PyTorch 2.4.0 framework on Windows 10, with Python 3.8 as the programming language. The computer's CPU was an Intel Core i7 7700K, the GPU was a GeForce GTX 1660 Ti, and the CUDA version was 12.6. YOLO-Rock RQD was trained for 300 epochs using the AdamW optimizer, with an initial learning rate of 1×10⁻⁶. 3 And gradually decrease during training, cosine annealing to 1×10 2 We also used methods such as Mosaic, Mixup, and HSV to enhance training, and the specific hyperparameter settings are shown in Table 1.

[0172] Table 1. Hyperparameter settings of the proposed method on the NEU-Rock dataset.

[0173]

[0174] To quantitatively evaluate the effectiveness of the proposed method and compare it with other methods, we selected three widely used evaluation metrics: Precision, Recall, and the mean Average Precision (mAP). Precision measures the accuracy of the prediction results by calculating the proportion of truly positive samples among those predicted as positive by the model. Higher accuracy means fewer false positives and more reliable results. Recall measures the model's ability to capture all relevant instances; it calculates the proportion of truly positive samples that the model successfully predicts. The calculations for Precision and Recall are shown below.

[0175]

[0176]

[0177] mAP, calculated by the area under the Precision-Recall curve, reflects the overall performance of a model in terms of both precision and recall. When calculating mAP, the first step is to determine the decision threshold for the Intersection over Union (IoU). This threshold can be set as a single value, such as 0.50 (mAP50); or it can be set as a range, such as from 0.50 to 0.95, increasing in increments of 0.05 (mAP[50,95]). In the latter case, the mAP value corresponding to each IoU threshold within that range needs to be calculated separately, and then these results are averaged. Generally, a match is considered successful only when the IoU value between a predicted bounding box and a ground truth bounding box is greater than or equal to 0.5; any match below this threshold is considered unsuccessful.

[0178] To evaluate the effectiveness of the proposed method, we compared it with five state-of-the-art algorithms on the NEURock dataset, as shown in Table 2. The algorithms used in the experiments included YOLOv8, YOLOv9, YOLOv10, YOLOv11, and YOLOv12.

[0179] Experiments on datasets demonstrate that the YOLO-Rock RQD method of this invention outperforms five other state-of-the-art methods. Compared to YOLOv12, the proposed method reduces the number of parameters by 0.5M and the computational cost by 0.9G, achieving higher accuracy while maintaining a lightweight model. Furthermore, Mask Precision and mAP[50,95] are improved by 0.6% and 3.1%, respectively.

[0180] Table 2 Performance comparison of different methods on the NEU-Rock dataset

[0181]

[0182] The segmentation results from different methods show that YOLO-Rock RQD is more effective in detecting and segmenting smaller cracks in rock cores, with a lower probability of missed detections. This superior performance is attributed to the synergistic effect of SPD-Conv, SimAM, and LSKBlock, which we designed based on the characteristics of rock images. This ensures the algorithm's real-time performance while enhancing its response to fine cracks and blurred boundaries.

[0183] In this experiment, we arranged and combined the key modules used in YOLO-Rock RQD and verified the effectiveness of the proposed method and the collaborative efficiency among multiple modules on the NEU-Rock dataset, as shown in Table 3.

[0184] Table 3 Ablation experimental results on the NEU-Rock dataset

[0185]

[0186] Experimental results show that when the three proposed modules (SPD-Conv, SimAM, and LSKBlock) are deployed simultaneously, the model exhibits the best segmentation performance, with mAP50 and mAP[50,95] reaching 95.5% and 69.3%, respectively. When the three modules are deployed separately, LSKBlock performs best, thanks to its application of wide-range context aggregation, long-range dependency capture, and adaptive spatial receptive field selection, achieving a precision of 93.8%. It is worth noting that the application of the SPD-Conv module reduces the number of model parameters and computational cost. This reduction is due to SPD-Conv replacing the computationally redundant standard convolutional modules in the original YOLOv11, rather than performing lightweighting within SPD-Conv itself. This also reduces the model complexity while retaining the ability to preserve the original pixels.

[0187] Further observation of the combined deployment of the two modules revealed that the deployment strategy with LSKBlock achieved the highest accuracy. When deployed in conjunction with SimAM and SPD-Conv, the model accuracy was 0.1% and 0.3% higher, respectively, compared to deploying LSKBlock alone. This further validates the complementarity between the modules.

[0188] The reason why LSKBlock expands the receptive field while reducing the computational cost lies primarily in the combined application of a large kernel decomposition mechanism and depthwise separable convolution. While the parameter count and computational cost of conventional convolution kernels typically increase quadratically with kernel size, LSKBlock does not directly use a dense large kernel. Instead, it simulates an ultra-large receptive field by cascading multiple depthwise separable convolutions (DW-Conv) with different expansion rates. This decomposition strategy breaks down the originally complex large-scale matrix operations into multiple sets of lightweight linear operations, achieving an equivalent receptive field of over 23×23 while maintaining low floating-point computational cost. This design not only ensures the ability to model long-range dependencies in the longitudinal texture of elongated rock cores but also avoids the surge in computational resources associated with conventional large kernel convolution.

[0189] Furthermore, the significant reduction in computational cost is also attributed to the synergistic optimization effect between modules. Experimental data shows that when LSKBlock and the SPD-Conv module are deployed together, SPD-Conv replaces the computationally redundant standard convolutional modules in the original YOLO structure through a spatial-depth mapping mechanism, thereby greatly simplifying the number of parameters in the feature map downsampling process. According to ablation experiments, after integrating all improved modules such as LSKBlock, the overall number of model parameters decreased from 3.13M in YOLOv12 to 2.63M, and the computational cost decreased from 11.5G to 10.6G, achieving the optimal state of model lightweighting. This demonstrates that LSKBlock effectively filters out invalid computations through an adaptive spatial selection mechanism when dealing with complex long-distance feature associations, achieving a substantial improvement in algorithm efficiency while improving core segmentation accuracy.

[0190] This application integrates SPD-Conv spatial-depth mapping convolution, LSKBlock large-kernel adaptive receptive field modeling, SimAM parameterless saliency refinement, and Inner-IoU assisted boundary regression operator to construct a closed-loop core segmentation system from sub-pixel feature extraction to macroscopic index quantification, achieving high-precision real-time quantitative assessment of core fragmentation under complex field geological conditions. In the downsampling stage, the scheme utilizes the full-pixel preservation mechanism of SPD-Conv to solve the problem of micro-fracture feature dilution caused by traditional strided convolution. The LSKBlock-constructed ultra-large equivalent receptive field (preferably 23×23 or larger) enables long-range dependency modeling of the longitudinal texture of elongated cores, effectively eliminating the phenomenon of interrupted segmentation results for originally continuous objects in the segmentation mask. Simultaneously, the SimAM module uses the analytical solution of the three-dimensional energy function to accurately suppress background interference from the core box and purify the core signal without increasing the model's computational load. Combined with the geometric constraints of Inner-IoU on slender targets, the system significantly improves the fit between the segmentation boundary and the physical ontology. Experimental results show that the present invention achieves an mAP50 of 95.5% on the NEU-Rock dataset. Compared with the existing YOLOv12 algorithm, it reduces the number of parameters and computation while achieving a balanced improvement in segmentation accuracy and RQD calculation accuracy. It completely overcomes the technical pain points of large subjective errors and time-consuming and labor-intensive manual surveying mode, and provides objective and robust technical support for digital mine construction and rock mass stability assessment.

[0191] Figure 3 An internal structural diagram of a computer device in one embodiment is shown. This computer device can specifically be a terminal or a server. Figure 3As shown, the computer device includes a processor, memory, and network interface connected via a system bus. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system and may also store a computer program. When executed by the processor, this computer program enables the processor to implement a core RQD segmentation method based on frequency-space feature enhancement and long-range dependence. The internal memory may also store a computer program, which, when executed by the processor, enables the processor to implement the core RQD segmentation method based on frequency-space feature enhancement and long-range dependence. Those skilled in the art will understand that... Figure 3 The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0192] Those skilled in the art will understand that all or part of the processes in the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments described above. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0193] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0194] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A core RQD segmentation method based on frequency-space feature enhancement and long-distance dependence, characterized in that, The method includes: S1: Obtain the core box image corresponding to the core box, and map it uniformly to the original core image in standard tensor format through preprocessing operations; perform initial spatial feature encoding on the original core image to obtain cross-stage convolutional features containing core edges, colors and shallow textures; S2: The cross-stage convolutional features are demixed and shuffled into four independent and complementary first sub-feature maps using a spatial segmentation operator. The four first sub-feature maps are then concatenated along the channel dimension to obtain a first intermediate enhanced feature map. The first intermediate enhanced feature map is then convolved to obtain a downsampled and fully pixel-retained first spatial depth feature. S3: Use a spatial segmentation operator to demix and shuffle the first spatial depth feature into four independent and complementary second sub-feature maps, and then concatenate the four second sub-feature maps in the channel dimension to obtain a second intermediate enhanced feature map. Finally, perform a convolution operation on the second intermediate enhanced feature map to obtain a second spatial depth feature that is downsampled and retains all pixels. S4: Superimpose depthwise separable convolutions with different dilation rates on the second spatial depth features to construct multi-scale candidate features; perform global average pooling and max pooling operations on the multi-scale candidate features respectively, then concatenate and fuse them along the channel dimension before performing convolution operations, and perform element-wise multiplication based on the corresponding multi-scale candidate features to obtain large kernel selection features; sequentially perform spatial pyramid pooling fast operation and two-stage channel attention feature extraction operation on the large kernel selection features to obtain output features; S5: Perform sampling space interpolation on the output feature and the large kernel selection feature to obtain the first merged feature; and perform purification and splicing operations and feature integration operations on the first merged feature in sequence to obtain the attention feature; determine the first attention cross-enhancement feature, the second attention cross-enhancement feature and the third attention cross-enhancement feature based on the second spatial depth feature and the attention feature; S6: The first attention cross-enhancing feature, the second attention cross-enhancing feature, and the third attention cross-enhancing feature are obtained respectively under different grain sizes as shallow, medium and deep segmentation features. The shallow, medium and deep segmentation features are then merged to obtain a segmented core feature map. S7: Perform instance-level separation on the core mask feature map to extract N independent core fragment instance masks existing in the core mask feature map, forming a core fragment instance mask set M={M1,M2,...,M...} N The morphological central axis of each core instance mask in the core fragment instance mask set is extracted using the skeleton extraction method, and the cumulative distance of the pixel skeleton points on the morphological central axis is calculated, which is the precise pixel length of the core instance mask. S8: Extract the pixel width of the core box in the core box image based on the known standard physical width of the core box, and calculate the scale conversion factor; use the scale conversion factor to convert the precise pixel length into the physical length of the core in geological engineering; perform integrity screening on the cores in the core box based on the physical length of the core to obtain multiple cores that meet the core length threshold; calculate the rock quality index in the area corresponding to the core box based on the multiple cores that meet the threshold condition and the total drilling length, and output it in the form of a percentage.

2. The core RQD segmentation method based on frequency-space feature enhancement and long-distance dependence according to claim 1, characterized in that, The cross-stage convolutional features are implemented using the following expression: in, This represents cross-stage convolutional features; This is the original core image; Conv represents the convolution operation; This is a cross-stage local network feature extraction operation.

3. The core RQD segmentation method based on frequency-space feature enhancement and long-distance dependence according to claim 1, characterized in that, The method of demixing and shuffling the cross-stage convolutional features using a spatial segmentation operator to obtain four independent and complementary first sub-feature maps, concatenating the four first sub-feature maps along the channel dimension to obtain a first intermediate enhanced feature map, and performing a convolution operation on the first intermediate enhanced feature map to obtain a downsampled and fully pixel-preserving first spatial depth feature is achieved through the following expression: in, , , and There are four first sub-feature maps; S is the resolution; This is the first intermediate enhanced feature map; This represents the first spatial depth feature.

4. The core RQD segmentation method based on frequency-space feature enhancement and long-distance dependence according to claim 1, characterized in that, The second spatial depth feature is superimposed with depthwise separable convolutions with different dilation rates to construct multi-scale candidate features. These multi-scale candidate features are then subjected to global average pooling and max pooling operations, concatenated and fused along the channel dimension, and then convolved again. Element-wise multiplication based on the corresponding multi-scale candidate features yields large kernel selection features. The large kernel selection features are then sequentially subjected to spatial pyramid pooling and two-stage channel attention feature extraction operations to obtain the output features, implemented through the following expression: in, , , and These are four independent and complementary second sub-feature maps; S is the resolution. This is the second intermediate enhanced feature map; This represents the second spatial depth feature.

5. The core RQD segmentation method based on frequency-space feature enhancement and long-distance dependence according to claim 1, characterized in that, The output feature is achieved through the following expression: in, Multi-scale candidate features; This indicates that the kernel size is k. i The expansion rate is d i Depth-separable convolution operations; This is a second spatial depth feature; This is a large-nucleus selection feature with an extremely large receptive field; This represents the operation of concatenating and fusing the spatial saliency features of global average pooling and max pooling operations along the channel dimension. For activation functions; This indicates that spatial pyramid pooling is a fast operation; This indicates a two-stage channel attention feature extraction operation; These are the output characteristics of the backbone network.

6. The core RQD segmentation method based on frequency-space feature enhancement and long-distance dependence according to claim 1, characterized in that, The attention feature is implemented through the following expression: in, This is the first merging feature; This is an upsampling spatial interpolation operation; The first upsampled feature is obtained by upsampling the backbone network features; Attention features; This is a parameterless attention mechanism operation; This is a cross-stage local network feature extraction operation.

7. The core RQD segmentation method based on frequency-space feature enhancement and long-distance dependence according to claim 6, characterized in that, The first attention cross-enhancement feature, the second attention cross-enhancement feature, and the third attention cross-enhancement feature are implemented by the following expressions: in, This is the second merging feature; This is the second upsampling feature; This is a first attention cross-enhancement feature; This is the fourth convolution feature; This is the third merging characteristic; This is a second attention cross-enhancement feature; This is the fifth convolution feature; This is the fourth merging feature; This is a third attention cross-enhancement feature.

8. The core RQD segmentation method based on frequency-space feature enhancement and long-distance dependence according to claim 1, characterized in that, The segmented core feature map is achieved through the following expression: in, This is a shallow segmentation feature; This is a mid-level segmentation feature; This represents deep segmentation features; This is a segmentation diagram of the core features.

9. The core RQD segmentation method based on frequency-space feature enhancement and long-distance dependence according to claim 1, characterized in that, The precise pixel length is achieved through the following expression: in, For precise pixel length; K is the core instance mask M. i The total number of pixel skeleton points on the morphological central axis, ( , M is the mask for the j-th core instance. i The pixel coordinates of the pixel skeleton points on the morphological central axis.

10. The core RQD segmentation method based on frequency-space feature enhancement and long-distance dependence according to claim 1, characterized in that, The rock quality index is achieved through the following expression: in, This is the scale conversion factor; Standard physical width; The width is in pixels; The physical length of the core sample; For precise pixel length; This is a screening operation for core samples that meet the threshold conditions; This is the core length threshold; This refers to the drilling footage length. For rock quality indicators; This represents the total number of core samples taken in one advance.

Images