Machine vision-based borehole formation ai logging method and system

Abstract

The application discloses a drilling stratum AI logging method and system based on machine vision, collects multispectral drilling core image and multi-view ring scanning core surface image and depth identification information, generates a color consistent image group through light-color normalization registration; generates a lithology discrimination feature set through image partition splicing, lithology end member decomposition and grain size classification feature extraction; associates the lithology discrimination feature set with the depth identification information to establish a depth-lithology contrast index, identifies a stratum interface position through compaction compensation and cycle direction analysis, implements feature out-of-limit detection to generate a mutation trigger identification; extracts an interface transition zone feature slice according to the mutation trigger identification, constructs a depth-image analysis profile through multi-sampling point edge trend continuity evaluation and geometric distortion correction; finally, performs lithology partition boundary positioning and lithology classification comparison, and outputs a standardized image logging report, effectively improving the accuracy of lithology interface identification and the standardization degree of logging results.

Description

Technical Field

[0001] This invention relates to the field of geological exploration information processing technology, and in particular to a borehole strata AI logging method and system based on machine vision. Background Technology

[0002] Core logging is a core step in geological exploration to obtain stratigraphic and lithological information, and the logging results directly affect the reliability of subsequent stratigraphic division and resource evaluation. Due to factors such as core exposure oxidation, uneven lighting during acquisition, and transport deformation, core images acquired in the field generally suffer from color distortion and geometric aberration. Color consistency between different batches of images is difficult to guarantee, and directly interpreting lithology based on original images can easily introduce systematic errors.

[0003] In stratigraphic interface identification, traditional methods rely on manual, segment-by-segment visual comparison, which is inefficient and highly subjective, making it particularly difficult to identify gradually changing lithological transition zones. Existing image processing methods mostly perform threshold judgments based on a single feature dimension, lacking a depth compensation mechanism for stratigraphic compaction patterns. In deep strata, feature drift caused by compaction is easily misidentified as lithological interfaces, and the methods are also insufficiently resistant to interference from factors such as cutting marks and surface contamination. Facing the need for continuous logging of deep, long core samples, a new method is urgently needed to solve at least one of the above problems. Summary of the Invention

[0004] This invention discloses a machine vision-based AI-based logging method and system for borehole formations. It aims to eliminate color differences between images by performing illumination-color normalization registration on multispectral and multi-view core images; establish quantifiable lithological discrimination criteria by combining lithological endmember decomposition and grain size classification feature extraction; achieve adaptive positioning of formation interfaces by introducing diagenetic compaction compensation and cyclic direction analysis; and finally output a standardized image logging report after geometric distortion correction and lithological classification comparison, providing a complete technical path for automated logging of borehole formations.

[0005] The first aspect of this invention proposes a machine vision-based borehole formation AI logging method, comprising the following steps: Multispectral borehole core images and multi-view ring scan core surface images and depth marking information are acquired. Illumination-color normalization registration is performed on the multispectral borehole core images and multi-view ring scan core surface images to generate color-consistent image groups. The image is partitioned and stitched together using the color-homogenized image group to generate a stitched core image. The stitched core image is then subjected to lithological end-member decomposition to determine the lithological component ratio. Based on the lithological component ratio, grain size classification features are extracted to generate a lithological discrimination feature set. A depth-lithology comparison index is established by associating the lithology discrimination feature set with the depth identification information. Based on the depth-lithology comparison index, the continuous drop position of adjacent core segments is identified to generate an adaptive discrimination benchmark. Based on the adaptive discrimination benchmark, feature over-limit detection is performed on the lithology discrimination feature set to generate a sudden change triggering identifier. Based on the mutation triggering identifier, the interface transition zone feature slice is extracted. The geometric distortion correction coefficient is obtained by evaluating the edge direction continuity of the interface transition zone feature slice through multi-sampling points. The geometric distortion correction coefficient is then used to perform inverse perspective transformation and scale calibration to construct a depth-image analysis profile. Based on the depth-image analysis profile, lithological zoning boundaries are located to form a lithological confidence map. The lithological confidence map is then used for lithological classification and comparison to output a standardized image logging report.

[0006] A second aspect of this invention proposes a machine vision-based borehole formation AI logging system, comprising: The image acquisition module is used to acquire multispectral borehole core images and multi-view ring scan core surface images and depth marking information, and to perform illumination-color normalization registration on the multispectral borehole core images and multi-view ring scan core surface images to generate color-consistent image groups. The feature parsing module is used to generate a stitched core image by partitioning and stitching the image through the color-homogenized image group, to perform lithological end-member decomposition on the stitched core image to determine the proportion of lithological components, and to extract grain size classification features based on the proportion of lithological components to generate a lithological discrimination feature set. The image detection module is used to associate the lithology discrimination feature set with the depth identification information to establish a depth-lithology comparison index, identify the continuous drop position of adjacent core segments based on the depth-lithology comparison index to generate an adaptive discrimination benchmark, and perform feature over-limit detection on the lithology discrimination feature set according to the adaptive discrimination benchmark to generate a sudden change triggering identifier; The interface evaluation module is used to extract interface transition zone feature slices based on the mutation triggering identifier, perform multi-sampling point edge direction continuity evaluation on the interface transition zone feature slices to obtain geometric distortion correction coefficients, and use the geometric distortion correction coefficients to perform inverse perspective transformation and scale calibration to construct a depth-image analysis profile. The cataloging and output module is used to locate the lithological zoning boundaries based on the depth-image analysis profile to form a lithological confidence map, and to perform lithological classification and comparison on the lithological confidence map to output a standardized image cataloging report.

[0007] The beneficial effects of this invention are reflected in the following points: 1. The width of the exposed oxidation zone in the core exhibits an uneven annular color gradient on the end face due to differences in mineral activity. Based on this gradient feature, the width of the oxidation zone is extracted and a graded color difference compensation model is established. The color difference correction of each band of the multispectral image is accurate to each compensation level. After registration, the image information is transformed into quantifiable multi-dimensional lithological discrimination features through lithological end-member decomposition and grain size classification feature extraction, solving the problems of inconsistent colors in multi-source images and the difficulty in quantifying human visual features. 2. The lithological discrimination feature baseline of each depth segment is extracted from the depth-lithological comparison index. Depth compensation is performed according to the diagenetic compaction law to generate a compaction correction baseline. After removing the systematic drift of features caused by burial depth, the cycle direction analysis and difference drop point location are performed. The compaction correction baseline is adaptively divided into lithological segments that match the actual stratigraphic segments, establishing a discrimination benchmark that dynamically adjusts with depth. Based on this benchmark, feature limit detection is performed on the lithological discrimination feature set to distinguish between gradual and abrupt limit limits and generate abrupt trigger markers, realizing differentiated responses to different types of lithological interfaces. 3. The end face tilt and lateral torsion caused by core cutting and transportation are difficult to correct using external calibration targets. The continuity of the edge orientation of the lithological interface itself is used as the internal geometric reference. The perspective offset parameters are extracted through collinear grouping and the inverse mapping coefficient is solved to complete the self-correction of image geometric distortion without external calibration. Based on the classification comparison between the abundance angle cosine and the lithological standard abundance library, the automatic mapping from image features to standard lithological names is realized. The thickness distribution constraint and multi-index confidence assessment can effectively distinguish the interference response such as cutting marks and surface contamination from the real lithological interface, and output a standardized image logging report with reliability quantitative evidence. Attached Figure Description

[0008] The accompanying drawings illustrate specific examples of the technical solutions described in this invention and, together with the detailed embodiments, form part of the specification, serving to explain the technical solutions, principles, and effects of this invention.

[0009] Figure 1 This is a flowchart illustrating the machine vision-based AI logging method for borehole formations according to the present invention.

[0010] Figure 2 This is a structural block diagram of the machine vision-based drilling formation AI logging system of the present invention. Detailed Implementation

[0011] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0012] It should be understood that, when used in this application specification and the appended claims, the term "comprising" indicates the presence of the described features, integrals, steps, operations, elements and / or components, but does not exclude the presence or addition of one or more other features, integrals, steps, operations, elements, components and / or a collection thereof.

[0013] References to "one embodiment" or "some embodiments" as described in this specification mean that one or more embodiments of this application include a specific feature, structure, or characteristic described in connection with that embodiment. Therefore, the phrases "in one embodiment," "in some embodiments," "in other embodiments," "in still other embodiments," etc., appearing in different parts of this specification do not necessarily refer to the same embodiment, but rather mean "one or more, but not all, embodiments," unless otherwise specifically emphasized. The terms "comprising," "including," "having," and variations thereof mean "including but not limited to," unless otherwise specifically emphasized.

[0014] The technical solutions of the embodiments of this application will be described below.

[0015] like Figure 1 As shown, this embodiment of the invention provides a machine vision-based borehole formation AI logging method, including the following steps S110-S150: Step S110: Collect multispectral borehole core images and multi-view ring scan core surface images and depth marking information, and perform illumination-color normalization registration on the multispectral borehole core images and multi-view ring scan core surface images to generate a color-consistent image group.

[0016] Specifically, multispectral borehole core images and multi-view annular scan core surface images, along with depth marking information, were acquired. The multispectral borehole core images were acquired by a core scanner at 15nm intervals across 32 bands within the 430-895nm range, achieving a spatial resolution of 0.05mm / pixel. Each band of the multispectral borehole core image was stored independently and accompanied by a center wavelength marker. The multi-view annular scan core surface images were acquired by a line-scan camera at 10° steps during a 360° rotation of the core around its axis, capturing 36 frames of side views. The spatial resolution of the multi-view annular scan core surface images was aligned with that of the multispectral borehole core images to 0.05mm / pixel to ensure scale consistency during registration. Depth marking information was synchronously recorded at 0.1mm resolution by a bracket photoelectric encoder. The encoder synchronously wrote the depth marking information at the trigger moment of each frame acquisition of the multispectral borehole core image and the multi-view annular scan core surface image, ensuring the temporal binding of image frames and depth values. The effective acquisition area of ​​a single segment of multispectral borehole core image should not be less than 95% of the designed cross-sectional area. Multispectral borehole core images with a proportion lower than this are marked as partially missing and trigger re-acquisition. Core segments that still have partial missing markings after re-acquisition are excluded from the calculation of the oxide halo width parameter in subsequent sub-steps. The overlap rate of adjacent frames of multi-view ring scan core surface images should not be less than 30%. Multi-view ring scan core surface image segments with insufficient overlap rate trigger angle step densification re-acquisition. Depth identification information is marked with both core segment number and support displacement. The depth attribution of the depth identification information can be verified frame by frame through the mapping relationship between segment number and displacement.

[0017] In some embodiments, performing illumination-color normalization registration on the multispectral borehole core image and the multi-view ring scan core surface image to generate a color-consistent image set includes: extracting end-face color ring distribution features from the multispectral borehole core image to generate an image oxide halo width parameter; performing graded color difference compensation on the multispectral borehole core image based on the image oxide halo width parameter to generate a color difference compensation matrix; performing registration on the multispectral borehole core image and the multi-view ring scan core surface image based on the color difference compensation matrix to obtain a registered image pair; and performing color normalization processing on the registered image pair to generate a color-consistent image set.

[0018] The end-face color ring distribution features are extracted from multispectral borehole core images to generate the image oxidation halo width parameter. A color gradient band diffuses from the center to the outer edge in the end-face region of multispectral borehole core images. This gradient band is formed by the oxidation reaction that occurs when the core is exposed to air, and its width directly reflects the core exposure time and mineral activity. The end-face color ring distribution feature is defined as a comprehensive description of the color intensity and distribution width of this gradient band in each ring of pixels, and is the direct basis for extracting the image oxidation halo width parameter. The reflectance ratio of multispectral borehole core images in the near-infrared band (750–900 nm) to the visible red band (620–680 nm) shows a characteristic increase in the oxide zone region. The ratio is calculated outwards from each annular pixel ring, and the radius of the ring where the ratio first drops below 1.15 times the average of the unoxidized reference area at the outer edge is defined as the outer edge of the oxide zone. The unoxidized reference area at the outer edge is defined as the average reflectance ratio of the rings 5 ​​to 10 pixels from the edge of the end face. This sequence of ratios provides a quantitative representation of the annular color distribution characteristics of the end face. The central area of ​​the end face of the multispectral borehole core image is divided into 36 sectors with a 10° step size. The color gradient intensity of each sector is statistically analyzed segment by segment along the radial direction. Sectors with a color gradient standard deviation exceeding 0.08 are marked as non-uniform areas, and their local oxide zone radius is recorded separately. The image oxidation halo width parameter is defined as the difference between the weighted average of the oxidation zone radii of each sector and the radius of the central area. The weights are taken from the integral value of the color gradient of each sector. The image oxidation halo width parameter is stored in pixels. For non-uniform areas, a local variance field is added to the image oxidation halo width parameter to indicate the degree of unevenness in the distribution of end face color. The effective calculation area of ​​the end face of the multispectral borehole core image excludes the edge 5 pixels to avoid color anomalies caused by step cutting. At the same time, the core segments marked as partially missing in the first part are excluded to ensure the integrity of the end face.

[0019] A grading chromatic aberration compensation matrix was generated for multispectral borehole core images based on the image oxidation halo width parameter. The image oxidation halo width parameter divided the chromatic aberration regions of the multispectral borehole core images into three compensation levels: a width value less than 12 pixels was classified as light oxidation, 12 to 28 pixels as moderate oxidation, and more than 28 pixels as heavy oxidation. Areas with a local variance exceeding 0.01 in non-uniform regions were preferentially classified into a higher compensation level. The chromatic aberration in the light oxidation level was mainly caused by uneven illumination, while the heavy oxidation level also included a mineral chemical discoloration component. Compensation models were established for both types of components, and the parameters of the compensation models for each level were independently fitted and written into the corresponding columns of the chromatic aberration compensation matrix. Graded color difference compensation uses multispectral borehole core images within each compensation level as the object. The channel deviation between the measured reflectance and the standard reference spectrum is calculated band by band. The channel deviation is fitted as a linear correction function with the image oxidation halo width parameter w as the independent variable: ΔR_ij = k_ij·w + b_ij, where ΔR_ij is the reflectance correction amount (dimensionless) for the i-th band and j-th level, w is the image oxidation halo width parameter (unit: pixels), k_ij is the slope (unit: dimensionless / pixel), and b_ij is the intercept (dimensionless). k_ij increases with level j to reflect the characteristic of enhanced color difference as oxidation deepens. After fitting k_ij and b_ij for each band and level, the results are filled into the corresponding row and column positions of the color difference compensation matrix. The rows of the color difference compensation matrix correspond to the number of bands in the multispectral borehole core image, and the columns correspond to the three compensation levels. The dimension is the number of bands × 3. The correction coefficients at the boundaries of each level satisfy continuity constraints to prevent abrupt color difference changes across levels. The color difference compensation matrix is ​​calibrated based on the band configuration of the standard reference spectral library. When the reference spectral library is updated, the color difference compensation matrix must be refitted to maintain its match with the current acquisition device.

[0020] Based on the chromatic aberration compensation matrix, multispectral borehole core images and multi-view ring scan core surface images are registered to obtain registered image pairs. The registered image pair is the output storage structure of this sub-step. Each registered image pair record contains two items: a pixel coordinate transformation matrix describing the geometric correspondence between the two types of images, and image pixel data aligned to a unified coordinate system after registration. Both items are indispensable; the former is used for subsequent coordinate transformation, and the latter for statistical calculations in subsequent sub-steps. The chromatic aberration compensation matrix selects correction coefficients for the corresponding columns according to the compensation level of the core segment and applies them band by band to each channel of the multispectral borehole core image. After correction, the average chromatic difference between the two types of images in the 500–700 nm range is reduced from 18% to less than 4%. Before registration, the multi-view ring scan core surface image is fused into a cylindrical unfolded image according to angle weights. The spatial resolution of the cylindrical unfolded image is aligned with the multispectral borehole core image to 0.05 mm / pixel, and the scale difference between the two types of images is less than 0.5%. The registration region for the two types of images is the geometric boundary between the edge ring of the end face and the top and bottom edge bands of the cylindrical unfolded image. Registration is based on feature point matching within this boundary region. Feature points are detected within the boundary region by the SIFT operators of each type of image. After RANSAC removes outliers from the initial matched point pairs, at least 30 pairs of inliers are retained as the basis for affine transformation estimation. During the iterative optimization of pixel-level registration, the chromatic difference compensation matrix dynamically adjusts the compensation level judgment based on the feedback of normalized mutual information. When the normalized mutual information fails to improve for three consecutive iterations at the current level, the chromatic difference compensation matrix switches to the adjacent level and reapplies correction coefficients to break through local extrema. After registration convergence, the coordinate transformation matrix and the pixel data of the registered image are jointly written into the registered image pair. The registered image pair is deemed to have a valid registration based on a root mean square error of less than 0.8 pixels in the coordinate transformation and an inlier rate of at least 92% in the boundary region.

[0021] The registered image pairs are color-normalized to generate a color-consistent image group. In the overlapping areas of the multispectral borehole core images and the multi-view ring-scan core surface images within the registered image pairs, there are still systematic color deviations caused by differences in the acquisition equipment. Color normalization processing establishes a unified color reference space to address these equipment differences. Using the overlapping areas of each registered image pair as a reference, color normalization processing calculates the mean vector μ_m of each band of the multispectral borehole core image and the mean vector μ_s of the corresponding channel of the multi-view ring-scan core surface image. The mean difference Δμ = μ_m - μ_s, averaged in the R, G, and B channels, reflects the brightness difference between the two images. When the deviation exceeds 5 gray levels, translation correction is applied to the multi-view ring-scan core surface image; when the deviation is less than 5 gray levels, no translation is performed. The variance ratio σ_m² / σ_s² of the overlapping area of ​​the registered image pair reflects the contrast difference between the two images. When the variance ratio deviates from 1.0 by more than 0.15, scale correction is applied. The scale correction coefficient γ = σ_m / σ_s is calculated independently for each channel, where σ_m and σ_s are both in gray levels, and γ is a dimensionless ratio. Each color-homogenized image group contains color-aberration compensated and normalized multispectral borehole core images and multi-view ring scan core surface images. Color-homogenized image groups are managed by band-order indexing. The color difference between the multispectral borehole core images and the multi-view ring scan core surface images within a color-homogenized image group does not exceed 3 gray levels.

[0022] Step S120: Image partitioning and stitching are performed on the color-homogenized image group to generate a stitched core image. The lithological end-member decomposition is performed on the stitched core image to determine the lithological component ratio. Based on the lithological component ratio, grain size classification features are extracted to generate a lithological discrimination feature set.

[0023] Specifically, a mosaic core image is generated by partitioning and stitching images from color-homogenized image groups. The overlap width of the end faces of adjacent image groups in the color-homogenized image group is approximately 0.8 mm. The optimal stitching seam position is determined based on maximizing normalized mutual information in the overlap area, with a maximum stitching seam offset of no more than 0.3 mm. The color-homogenized image groups have already been depth-sorted based on the depth coordinates embedded in each image group during the S110 acquisition phase. During stitching, the depth range of each image group is directly used as the basis for dividing the stitching partitions. Each stitching partition contains 2 to 4 images from a continuous color-homogenized image group. The partition boundary is set at the point of highest overlap between adjacent image groups to ensure a natural transition. The end face coordinate systems of each color-homogenized image group within the stitching partition are aligned using rigid body transformation. The transformation parameters are estimated from the feature point matching results in the overlap area. After alignment, the root mean square error of the pixel coordinates in the overlap area is less than 0.5 pixels. When the alignment residual exceeds a threshold, feature point encryption matching is retried. After alignment, overlapping areas in the color-homogenized image set are eliminated using distance-weighted fusion to remove seams. The fusion transition width is 1.2 mm. The weighting coefficients of each pixel within the transition area are calculated linearly based on the pixel distance from the seam, ensuring continuous color without abrupt changes in the stitched core image within the transition area. In actual drilling operations, differences in core fragmentation rates lead to incomplete end faces in some segments of the color-homogenized image set. For color-homogenized image sets with an end face integrity rate below 85%, interpolation between adjacent complete segments is used to fill in the gaps within the stitched area. The interpolation results are marked with missing segments in the stitched core image for exclusion during subsequent spectral feature extraction. The spatial resolution of the stitched core image inherits 0.05 mm / pixel from the color-homogenized image set, and the depth coverage is recorded using the start and end values ​​of the embedded depth coordinates within the color-homogenized image set.

[0024] In some embodiments, the step of performing lithological endmember decomposition on the stitched core image to determine the proportion of lithological components includes: extracting spectral features from the stitched core image to generate spectral vectors for each pixel; performing clean pixel screening on the spectral vectors of each pixel to generate an image endmember spectral library; performing abundance unmixing processing on the spectral vectors of each pixel according to the image endmember spectral library to obtain a pixel endmember abundance matrix; and determining the proportion of lithological components in each core segment by statistically analyzing the proportion of each lithological component based on the pixel endmember abundance matrix.

[0025] Spectral features were extracted from the mosaic core image to generate spectral vectors for each pixel. The reflectance difference at 570 nm in the mosaic core image between areas with high concentrations of quartz minerals and areas rich in dark minerals reached 35%, and this difference constituted the main distinguishing signal for spectral feature extraction. The 32 band channels of the mosaic core image were strictly aligned according to pixel position. Dark current correction and flat-field correction were applied to each channel to eliminate inherent sensor noise. After correction, the noise-equivalent reflectance of each band in the mosaic core image was less than 0.5%. The 32-dimensional reflectance values ​​at each pixel position in the mosaic core image constituted the initial sequence of the spectral vectors for that pixel. After normalization of the maximum value of each band across the entire image, the value range of each dimension was unified to the interval between 0 and 1. The normalization method preserved the relative reflectance relationship between bands to support subsequent endmember selection. Pixels within a 3-pixel width of the edge of the mosaic core image suffer from reflectivity distortion due to the cutting step effect; therefore, the spectral vectors of these pixels are marked as invalid and excluded from the statistics. Pixels in the transition zone of the mosaic core image's seam are subject to slight spectral aliasing due to the fusion of multiple images; therefore, the spectral vectors of these pixels are marked with aliasing. During the endmember selection stage, the interference of aliased pixels on the endmember extraction results is reduced by lowering their PPI candidate weights. Pixel segments with missing markers in the mosaic core image are also excluded from the effective calculation range of spectral feature extraction. The dimension of each pixel's spectral vector is 32.

[0026] A clean pixel screening process is performed on the spectral vectors of each pixel to generate an image endmember spectral library. Clean pixels exhibit distinct spectral absorption valleys, while mixed pixels, due to the superposition of multiple minerals, have absorption valley depths 40% to 70% lower than clean pixels. This difference is used as a criterion to separate endmember candidate sets from the spectral vectors of each pixel. The spectral purity of each pixel spectral vector is quantified by the iterative Pixel Purity Index (PPI). Spectral vectors with PPI values ​​higher than the 85th percentile are considered clean pixel candidates. Pixels with aliasing markers have the lowest candidate priority when PPI values ​​are the same and do not participate in endmember vertex search. The candidate set typically accounts for 8% to 15% of the total number of spectral vectors of each pixel. The clean pixel candidate set is then used in a 32-dimensional feature space to determine endmember vertices using a maximum volume simplex search. These endmember vertices correspond to the initial members of the image endmember spectral library. The number of initial members is determined by the preset number of lithology categories in the core segment, typically set to 4 to 6. After the image endmember spectral library is established, the average reconstruction error between the spectral vector of each pixel and the endmember vector is used as the evaluation index. When the average reconstruction error exceeds 0.04, the number of endmembers is increased until the upper limit of 8 is reached to accommodate the diversity of endmembers in sections with complex lithological combinations. For the image endmember spectral library of different depth sections in the same borehole, after the number of endmembers is stabilized, the average endmember spectrum of adjacent sections is used for smoothing constraints to prevent the endmembers of adjacent sections from jumping due to random noise. The final number of members in the image endmember spectral library is determined by jointly minimizing the reconstruction error and the endmember redundancy.

[0027] The pixel endmember abundance matrix is ​​obtained by performing abundance unmixing on the spectral vectors of each pixel based on the image endmember spectral library. The endmember matrix E of the image endmember spectral library has a dimension of 32×N_e, where N_e is the number of effective endmembers in the image endmember spectral library. The abundance unmixing of each pixel spectral vector y is solved using a constrained least squares problem: min‖y-Ef‖ 2 The constraint condition is that all components of the abundance vector f are non-negative and their sum is 1. This constraint ensures the physical interpretability of the unmixing result. The spectral angle cosine values ​​of each endmember in the image endmember spectral library and the current pixel spectral vector y are pre-screened before unmixing. Endmembers with cosine values ​​below 0.6 are considered non-associated endmembers, and their abundance is directly set to zero in the unmixing calculation for this pixel to reduce computational load. The endmember matrix E used for abundance unmixing is taken from the image endmember spectral library after smoothing constraints on adjacent segments to ensure the depth continuity of the unmixing results across segments. The unmixing residual of each pixel spectral vector is defined as ‖y-Ef‖2. When the residual exceeds 0.06, the pixel is marked as a high residual pixel in the pixel endmember abundance matrix. High residual regions usually correspond to fracture fillers or contaminated minerals. High residual pixels undergo re-abundance unmixing after the image endmember spectral library is expanded. If the residual still exceeds 0.06, the pixel is retained in the pixel endmember abundance matrix and no longer iterated. The rows of the pixel endmember abundance matrix correspond to the pixel index of each pixel's spectral vector, the columns correspond to the endmember numbers of each pixel in the image endmember spectral library, and the matrix elements are the endmember abundance values ​​of the corresponding pixels.

[0028] The proportion of lithological components in each core segment is determined by statistically analyzing the percentage of each lithological component using the pixel endmember abundance matrix. The length of a single core segment is typically several centimeters to tens of centimeters. The mineral assemblage within the segment is relatively homogeneous on a macroscopic scale. Using the average abundance value of pixels within the segment to represent the lithological composition of that segment has a reasonable geological basis. The proportion of lithological components in each core segment is calculated using the mean vector of each endmember column in the pixel endmember abundance matrix within the core segment. The mean of the endmember column is the average abundance of the lithology corresponding to that endmember within the segment. The sum of the average abundance values ​​within each endmember segment is 1 to satisfy the abundance conservation constraint. The mean value of the pixel endmember abundance matrix in the segment corresponding to the quartz mineral endmember is 0.43, indicating that the core segment is dominated by quartz minerals. This mean value is directly used as the representative value of the quartz endmember in the proportion of lithological components in each core segment. In the pixel endmember abundance matrix, the abundance values ​​of high residual pixels are weighted less when calculating the proportion of each lithological component. The weight is the normalized inverse of the residual of the high residual pixels to reduce the interference of fracture contamination areas on the proportion of lithological components in each core segment. The spatial heterogeneity of the proportion of lithological components in each core segment is measured by the standard deviation of abundance within the segment. Core segments with an abundance standard deviation exceeding 0.15 are identified as lithological gradient segments. Gradual gradient segments are marked with gradient markers in the proportion of lithological components in each core segment to distinguish them from homogeneous lithological segments. The proportion of lithological components in each core segment is stored in the form of a multi-channel raster image. The number of channels is equal to the number of effective endmembers in the image endmember spectral library. The depth coordinates of the raster image are aligned with the depth label information to support subsequent cross-step depth correlation.

[0029] Grain size classification features were extracted based on lithological composition ratios to generate a lithological discrimination feature set. The abundance standard deviation within each endmember segment of the lithological composition ratio and the feldspar-quartz abundance ratio were used together as the criteria for grain size classification. The abundance standard deviation was divided into three grades according to preset thresholds: coarse-grained (standard deviation greater than 0.12), medium-grained (0.06 to 0.12), and fine-grained (below 0.06). The ratio of feldspar endmember abundance to quartz endmember abundance had a mean of 0.31 in fine-grained rock samples and a mean of 0.67 in coarse-grained rock samples. When the abundance standard deviation was within ±0.01 of the boundary between adjacent grades, this ratio was used to assist in determining the final grain size classification. The boundary thresholds between grades were calibrated based on standard rock thin section data to ensure that the consistency rate between the grain size classification results and the microscopic identification results was not less than 85%. Grain size classification results of gradient marker segments were supplemented with a gradient description field, indicating that the grain size of this segment showed a transitional change in the depth direction. The endmember mean vectors of grain size classification results and lithological component proportions together constitute the feature dimension of the lithology discrimination feature set. The endmember mean vector of each record is derived from the mean of each endmember segment of the lithological component proportion. The grain size classification label is derived from the joint classification result of abundance standard deviation and feldspar-quartz ratio. These two source fields are stored independently in the lithology discrimination feature set to support their respective quality verification. The feature dimension of the lithology discrimination feature set is the number of endmembers plus 2, where the two additions correspond to the integer code of the grain size classification label and the feldspar-quartz abundance ratio, respectively.

[0030] Step S130: Associate the lithological discrimination feature set with the depth identification information to establish a depth-lithological comparison index. Based on the depth-lithological comparison index, identify the continuous drop position of adjacent core segments to generate an adaptive discrimination benchmark. Based on the adaptive discrimination benchmark, perform feature limit detection on the lithological discrimination feature set to generate abrupt change triggering marker.

[0031] Specifically, a depth-lithology comparison index is established by associating the lithology discrimination feature set with depth identification information. The lithology discrimination feature set uses each core segment as a recording unit. The endmember mean vector, grain size classification label, and feldspar-quartz abundance ratio of each segment are stored according to segment number. The depth identification information records the start and end values ​​of each core segment at a resolution of 0.1 mm. The association between the two is performed segment-by-segment matching using the segment number as the primary key, ensuring accurate depth assignment for each lithology feature record. Some core segments in the lithology discrimination feature set have missing feature values ​​due to excessive fragmentation, which is particularly common when boreholes traverse fracture zones or dissolution cavities. These segments are filled by interpolation of the feature mean of adjacent valid segments when associating with the depth identification information. The interpolation result is marked with an interpolation marker in the depth-lithology comparison index to distinguish between the measured segment and the filled segment. The interpolation marker also records the continuous length of the filled segment; intervals with large continuous filled lengths are marked as areas of weak data quality in the index. The depth-lithology comparison index uses depth values ​​as the primary sorting key. The endmember mean vectors, grain size classification labels, and feldspar-quartz abundance ratios from the lithology discrimination feature set are stored as attribute fields for each depth record. The index's depth resolution inherits the 0.1mm precision of the depth identifier information. After the depth-lithology comparison index is established, the endmember mean vectors of adjacent segments in the lithology discrimination feature set undergo global depth continuity verification. Records with reversed or overlapping depth values ​​trigger a correction process to ensure that the index depth monotonically increases.

[0032] In some embodiments, the step of identifying the location of continuous drop-offs in adjacent core segments based on the depth-lithology comparison index and generating an adaptive discrimination benchmark includes: extracting lithology discrimination feature baselines for each depth segment from the depth-lithology comparison index; generating image feature compaction correction baselines by performing depth compensation according to the diagenetic compaction law based on the lithology discrimination feature baselines for each depth segment; extracting the difference between adjacent segments of the image feature compaction correction baselines to identify the location of continuous drop-offs in image features; and dividing the image feature compaction correction baselines into lithology segments based on the location of continuous drop-offs in image features to generate an adaptive discrimination benchmark.

[0033] The lithological discrimination feature baselines for each depth segment of the image were extracted from the depth-lithology comparison index. The endmember mean vectors of the records at each depth in the depth-lithology comparison index exhibit a step-like distribution along the depth direction, varying with lithology. The Euclidean distance between the endmember mean vectors of adjacent depth segments is less than 0.05 in the lithologically stable range and increases to above 0.15 in the lithological transition range. This distribution characteristic was used to verify the rationality of subsequent sliding window parameter settings. The mean value within the window of the endmember mean vector was calculated for each depth segment from the depth-lithology comparison index using a 1m sliding window. This window mean vector served as the initial value for the lithological discrimination feature baselines of each depth segment. The window step size was set to 0.5m to ensure 50% overlap between adjacent baseline segments, guaranteeing the continuity of baseline depth coverage. The weight of the interpolated fill segments in the depth-lithology comparison index was reduced to 0.3 during the calculation of the lithological discrimination feature baselines for each depth segment, to minimize the interference of the fractured segment fill values ​​on the baseline trend. The dimension of the lithological discrimination feature baseline of each depth segment image is consistent with the number of endmembers in the lithological discrimination feature set. The grain size classification label uses the mode statistics as the grain size representative value of the lithological discrimination feature baseline of each depth segment image. The feldspar-quartz abundance ratio uses the mean within the window as the ratio representative value of the depth segment and is stored synchronously with the baseline. The grain size representative value is used to determine whether the current depth segment is suitable for the clastic rock compaction model during subsequent depth compensation. The depth segment corresponding to the clastic rock grain size representative value is subjected to compaction correction by the porosity-related endmembers. Non-clastic rock segments skip the compaction correction and directly inherit the baseline value.

[0034] Based on the lithological discrimination feature baselines of images at each depth segment, depth compensation is performed according to the diagenetic compaction law to generate image feature compaction correction baselines. There is a systematic feature shift in the lithological discrimination feature baselines of images at each depth segment caused by the increasing pressure of the overlying strata between shallow and deep layers. This shift is particularly significant when the borehole traverses thick clastic rock strata. The reflectivity of the pore-related end-members systematically decreases with increasing burial depth. If this is not corrected, the overall lithological discrimination feature baseline of the deep strata will be too low, leading to misjudgment as exceeding the limit. The diagenetic compaction law describes the offset trend by the exponential decay of porosity with depth. The compaction compensation model is φ(z)=φ0·e^(-c·z), where φ0 is the surface reference porosity (dimensionless), c is the compaction coefficient (unit: 1 / m), z is the depth (unit: m), and e^(-c·z) is the dimensionless exponential decay term. The compaction correction factor is defined as φ0 / φ(z)=e^(c·z). The porosity-related endmember components in the endmember mean vector are multiplied by this factor to eliminate the compaction offset, while non-porosity-related endmember components are not corrected. The compensation amount is calculated by multiplying the porosity-related endmember components in the endmember mean vector of the lithological discrimination feature baseline of each depth segment by the compaction correction factor e^(c·z) at the corresponding depth. For depth segments corresponding to the representative values ​​of non-clastic rock grain size, this compensation is skipped, and the original baseline value is directly retained. The image feature compaction correction baseline applies a compaction correction factor segment by segment to the lithology discrimination feature baseline of each depth segment. After correction, the depth trend term of the image feature compaction correction baseline is removed, retaining only the characteristic fluctuations caused by lateral lithology variations. The depth resolution of the image feature compaction correction baseline is consistent with that of the lithology discrimination feature baseline of each depth segment. After correction, a residual trend test is performed on the image feature compaction correction baseline. When the absolute value of the residual trend slope exceeds a threshold, the compaction coefficient is recalibrated. After recalibration, depth compensation is re-performed to update the image feature compaction correction baseline.

[0035] For example, the step of extracting the difference between adjacent segments to identify the location of a sudden drop in image feature continuity based on the image feature compaction correction baseline includes: generating an image feature sampling sequence by segmenting the image feature compaction correction baseline according to a depth step size; extracting the difference between adjacent segments to generate an image feature difference sequence; identifying positive and negative rotation directions in the image feature difference sequence to generate image feature rotation direction identifiers; and locating the point of sudden drop in difference based on the image feature rotation direction identifiers to determine the location of a sudden drop in image feature continuity.

[0036] Image feature sampling sequences are generated by segmenting the image feature compaction correction baseline according to depth steps. The image feature compaction correction baseline describes the feature vectors at each depth as a continuous curve, but subsequent difference calculations and cycle identification both require discretized, equally spaced sequences. Segmented sampling is the key step in converting the continuous baseline into a discrete sequence that can be compared point-by-point. The fluctuation range of the endmember mean vector of the image feature compaction correction baseline is significantly higher in the carbonate and clastic interbedded sections than in the homogeneous sandstone sections. Too large a step size will cause the feature changes at thin-layer interfaces to be averaged and missed, while too small a step size will introduce too many noise points that interfere with cycle identification. A standard depth step size of 0.1m strikes a balance between the capture requirements of the two types of strata. The image feature compaction correction baseline is sampled segment by segment with a standard depth step size of 0.1m. The endmember mean vector of the image feature compaction correction baseline at that depth is taken as the sample value for each sampling point, and the coordinates of the sampling point correspond one-to-one with the depth value. When the Euclidean distance between adjacent sampling points in a segment with drastic feature changes exceeds a threshold, step size densification is triggered. The densification step size is half the standard step size, i.e., 0.05m. Sampling points in the densification interval are marked with densification markers in the image feature sampling sequence to distinguish them from standard step size sampling segments. These densification markers are used to eliminate amplitude deviations introduced by inconsistent step sizes during difference normalization. The image feature sampling sequence is arranged by depth index, and the sequence elements are the endmember mean vectors of each sampling point.

[0037] Image feature difference sequences are generated by performing adjacent difference extraction on the image feature sampling sequence. At the sandstone-mudstone interface in the borehole strata, the Euclidean distance of the endmember mean vectors of adjacent sampling points increases sharply, while the Euclidean distance of adjacent points remains low and stable within homogeneous sandstone or mudstone sections. The difference in the magnitude of the difference between these two types of intervals provides a clear magnitude reference for identifying the sudden drop in the image feature difference sequence. The Euclidean distance of the endmember mean vectors is calculated for each pair of adjacent sampling points in the image feature sampling sequence. The distance value is used as the difference element at that location to form the image feature difference sequence. The length of the image feature difference sequence is one element shorter than that of the image feature sampling sequence. Before being included in the image feature difference sequence, the difference between adjacent points in the densified step size segment of the image feature sampling sequence is multiplied by a step size normalization coefficient to eliminate the magnitude deviation introduced by inconsistent step sizes. The step size normalization coefficient is equal to the ratio of the standard step size to the densified step size, i.e., 2.0. This amplifies the difference in the densified segment to be comparable to the magnitude of the standard step size segment, ensuring the comparability of the image feature difference sequence across the entire depth range. The image feature difference sequence uses the depth midpoint value as the depth coordinate of each difference element. The magnitude of each element in the sequence directly reflects the degree of drastic lithological change at the corresponding depth.

[0038] Image feature difference sequences are used to identify positive and negative cycles, generating image feature cycle direction markers. Sedimentary strata exhibit periodicity in their formation, with alternating positive cycles (from coarse to fine grain size) and negative cycles (from fine to coarse grain size). This cycle structure manifests in the image feature difference sequence as a systematic increasing or decreasing trend in the difference amplitude. Without first identifying the cycle direction, abrupt drop determination cannot distinguish between the normal amplitude decline at the top of a cycle and the abrupt drop at the true interface, resulting in numerous misjudgments. The role of image feature cycle direction markers is to define the cycle background for abrupt drop point identification, ensuring that abrupt drop determination is only effective at cycle boundaries. For the image feature difference sequence, a linear regression slope of the difference within a sliding window is calculated. A positive cycle segment is marked when the slope is consistently positive for three or more consecutive windows, a negative cycle segment for three or more consecutive windows, and a stable segment when the absolute slope value is below a threshold. In actual borehole formations, these three types of segments alternate along the depth direction. The alternation frequency of positive and negative cycles reflects the period length of the sedimentary cycle. Stable segments typically correspond to thick layers of homogeneous sandstone or mudstone. The three types of segments are arranged in depth order to form the image feature cycle direction markers. The start and end depths of each segment and the mean difference within the segment are recorded synchronously. The mean difference within a segment reflects the overall intensity of change in that cycle segment. Segments with significantly higher amplitudes typically correspond to stratigraphic segments with drastic changes in lithological composition.

[0039] The location of abrupt drops in image feature continuity is determined by identifying points where the difference value suddenly drops, based on the image feature cycle direction markers. In borehole profiles, stratigraphic lithological interfaces typically exhibit a sharp drop in feature difference within a very short depth range. In the image feature cycle direction markers, the point where the difference value suddenly drops from a high level at the end of a negative cycle segment corresponds to a lithological interface. Similarly, a sudden drop immediately following a peak in the difference value at the end of a positive cycle segment indicates a lithological abrupt change. A pattern of a brief increase followed by a sudden drop in the difference value in a stable segment corresponds to the top and bottom interfaces of thin interlayers. All three types of sudden drop patterns are included in the identification range of locations where image feature continuity abrupt drops occur. In the image feature cycle direction identifier, the difference between the last element of each cycle segment and the first element of the next segment is defined as the inter-segment jump variable. A point is identified as a sudden drop in difference when the absolute value of the inter-segment jump variable satisfies |Δd_k|>μ_d+λ·σ_d, where k is the boundary number of adjacent cycle segments, μ_d and σ_d are the global mean and standard deviation of the image feature difference sequence, respectively, and λ is a sensitivity coefficient, typically ranging from 2.0 to 3.0, adjusted according to the complexity of the borehole formation. A smaller value is used to improve sensitivity when formation lithology changes frequently, and a larger value is used to reduce the false positive rate when the lithology is stable. When determining a sudden drop point, the absolute value of the inter-segment jump variable must also exceed 1.5 times the mean difference within that segment to exclude local fluctuations within the cycle from being misjudged as sudden drop points. The depth coordinates of the sudden drop point are recorded using the midpoint depth value of the corresponding difference element. The location of continuous sudden drops in image features is recorded using the depth coordinates of each sudden drop point and the corresponding jump variable amplitude.

[0040] Based on the locations of continuous abrupt drops in image features, the image feature compaction correction baseline is divided into lithological segments to generate an adaptive discrimination benchmark. The lithology varies at different depths in the borehole. If a uniform feature threshold is used to determine exceedances across the entire segment, it will inevitably lead to a misalignment in the judgment criteria between shallow soft rock segments and deep hard rock segments. The adaptive discrimination benchmark approach divides the baseline into several lithological segments based on the abrupt drop points. Each segment independently calculates its own feature center and fluctuation range. Exceedance judgments are only performed under the statistical parameters of its own segment, and the parameters are automatically updated according to the lithological segment. The depth coordinates of each abrupt drop point at the location of continuous abrupt drops in image features divide the image feature compaction correction baseline into several continuous intervals. The image feature compaction correction baseline segment between adjacent abrupt drop points corresponds to an independent lithological segment, and the feature variation amplitude within each lithological segment is significantly lower than the variation amplitude across segments. Points with small jump variable amplitudes are identified as weak interfaces, and the corresponding segment boundaries are marked as soft boundaries. Points with large jump variable amplitudes are marked as hard boundaries. The tolerance range for soft boundary segments is appropriately widened compared to hard boundary segments to avoid misjudging gradual transition zones. The mean vector of the image feature compaction correction baseline within each lithological segment serves as the feature center of that segment, and a multiple of the standard deviation vector serves as the upper limit of the allowable feature fluctuation for that segment. Exceeding this upper limit is considered an over-limit. The mean vector and standard deviation vector together constitute the adaptive discrimination benchmark parameters for that segment. The adaptive discrimination benchmark is managed using the lithological segment number as the primary key, and each segment records the depth boundary and corresponding statistical parameters.

[0041] In some embodiments, the step of performing feature limit detection on the lithological discrimination feature set according to the adaptive discrimination criterion to generate a mutation trigger identifier includes: arranging the lithological discrimination feature set in depth order to generate a depth-image feature sequence; scanning the depth-image feature sequence segment by segment according to the adaptive discrimination criterion to identify image limit-breaking feature segments; dividing the image limit-breaking feature segments into gradual limit-breaking and abrupt limit-breaking according to the limit-breaking gradient difference to generate image feature limit-breaking type parameters; and generating a mutation trigger identifier according to the image feature limit-breaking type parameters.

[0042] The lithology discrimination feature set is arranged in depth order to generate a depth-image feature sequence. Each segment of the lithology discrimination feature set is rearranged in ascending order of depth value based on the association between the segment number and depth identifier information. After rearrangement, the depth spacing between adjacent records inherits the 0.1mm resolution of the depth identifier information. Interpolated fill segments in the lithology discrimination feature set retain interpolation markers in the depth-image feature sequence. When the borehole traverses a fractured zone, the length of continuous fill segments may extend to tens of centimeters. Fractured segments with extremely low core recovery rates cannot provide effective lithology discrimination features. When the length of a continuous fill segment exceeds a threshold, it is marked as a long missing interval in the depth-image feature sequence. Long missing intervals are skipped during subsequent segment-by-segment scanning and do not participate in the limit-crossing judgment to avoid the fill value interfering with the interface recognition results. Each element of the depth-image feature sequence contains three attributes: the endmember mean vector of the lithology discrimination feature set, the grain size classification label, and the feldspar-quartz abundance ratio. The endmember mean vector and the grain size classification label are stored independently in the sequence to support multi-dimensional limit detection. The feldspar-quartz abundance ratio is stored synchronously with the record for quality verification during subsequent lithology classification comparison. The grain size classification label can trigger the identification of grain size abrupt change interfaces independently. The judgment condition is that the grain size classification label of adjacent segments jumps by more than one level and does not depend on the endmember mean vector to synchronously exceed the limit. The two limit detections are performed independently and then merged to generate a unified limit element label. The continuity of the lithology discrimination feature set after depth arrangement is measured by the mean depth difference between adjacent elements. During the field core extraction process, occasional drift of the bracket displacement encoder may introduce depth value anomalies. When the mean deviates from the standard resolution by more than a threshold, a depth value re-verification is triggered.

[0043] Based on an adaptive discrimination criterion, the depth-image feature sequence is scanned segment by segment to identify image over-limit feature segments. The depth boundaries of each lithological segment in the adaptive discrimination criterion divide the depth-image feature sequence into corresponding scanning intervals. The elements of the depth-image feature sequence within each scanning interval are compared element by element with the adaptive discrimination criterion parameters of that segment. When the Euclidean distance between the endmember mean vector of each element in the depth-image feature sequence and the mean vector of the corresponding segment in the adaptive discrimination criterion exceeds a threshold multiple of the standard deviation vector magnitude, it is determined to be an endmember dimension over-limit element. When the adjacent segments of the grain size classification label jump by more than one level, it is determined to be a grain size dimension over-limit element. After merging the two types of over-limit elements, they are merged into over-limit segments in the depth-image feature sequence according to the continuity criterion. When the number of consecutive over-limit elements reaches the minimum segment length threshold (usually set to 3 to 5 consecutive elements, determined by conversion based on the resolution of the depth label information and the minimum thickness of the target interface), they are merged into an image over-limit feature segment. The tolerance range of the soft boundary segment in the adaptive discrimination benchmark is appropriately expanded compared to the hard boundary segment to avoid misjudging the gradual transition zone near the weak interface as exceeding the limit. When the borehole traverses the lithological gradient zone, the exceeding elements of the depth-image feature sequence are dispersed, and isolated exceeding elements that do not meet the continuity criterion are not merged into the image exceeding feature segment. When the depth-image feature sequence reaches a long missing interval, it skips that interval and directly enters the next segment scan. The depth range of the image exceeding feature segment is recorded by the depth values ​​of the first and last elements of the exceeding segment.

[0044] For example, the step of dividing the image limit-crossing feature segments into gradual limit-crossing and abrupt limit-crossing based on the difference in limit-crossing gradient to generate image feature limit-crossing type parameters includes: extracting color abrupt change, texture abrupt change, and structural abrupt change from the image limit-crossing feature segments to generate an image abrupt change vector set; performing inter-dimensional synergy evaluation on the image abrupt change vector set to generate an image feature synergistic mutation degree; identifying single-dimensional dominant mutations and multi-dimensional synergistic mutations based on the image feature synergistic mutation degree to generate an image abrupt change pattern identifier; and distinguishing between gradual limit-crossing and abrupt limit-crossing based on the image abrupt change pattern identifier to obtain image feature limit-crossing type parameters.

[0045] Image mutation vector sets are generated by extracting color, texture, and structural abrupt changes from the image's over-limit feature segments. Interfaces of different lithologies in core profiles exhibit synchronous or differentiated changes in color, texture, and structure in the image. The three-dimensional abrupt changes at the hard interface between sandstone and mudstone are usually synchronous and significant, while alteration zones or mineralization halos often only change in hue, while texture and structure remain continuous. Extracting mutation values ​​dimensionally and incorporating them into the image mutation vector set aims to capture these multidimensional differences. The extraction of color, texture, and structural abrupt changes is based on the depth range of the image's over-limit feature segments, with image strips corresponding to the depth range extracted from the stitched core image generated by S120 as the calculation source. The color abrupt change in the image's out-of-limit feature segment is calculated using the color difference ΔE of the mean vector of the first and last endmembers of the out-of-limit segment in the visible light band. The texture abrupt change is characterized by the mean of the contrast difference of the gray-level co-occurrence matrix of several adjacent sampling points at the beginning and end of the out-of-limit segment in the corresponding positions of the stitched core image. The structural abrupt change is calculated using the mean edge density gradient of the corresponding area in the stitched core image within the out-of-limit segment. After normalization, each of the three indicators is stitched together in a fixed order of color, texture, and structure to form a single record of the image abrupt change vector set. The normalization benchmark for each indicator is taken as the high percentile value of the corresponding indicator in the full-depth out-of-limit feature segment. Normalization ensures that the three-dimensional indicators in the image abrupt change vector set have consistent dimensions and comparable amplitudes. The image abrupt change vector set is managed using the out-of-limit segment number as the primary key. Each record corresponds to a three-dimensional abrupt change description of an out-of-limit segment. The dimension of the image abrupt change vector set is fixed at 3, resulting in a simple and unified structure.

[0046] Inter-dimensional synergy assessment of image mutation vector sets generates image feature synergy mutation degree. At clear lithological hard interfaces in borehole cores, mutations in color, texture, and structure typically occur synchronously, exhibiting a highly synergistic change pattern in the three dimensions. In contrast, at gradual transition zones or single alteration boundaries, mutations often occur in only one or two dimensions, with lower synergy in the three dimensions. Inter-dimensional synergy assessment distinguishes hard and soft interfaces by quantifying this synergy. The synergy mutation degree of each record in the image mutation vector set is calculated as the ratio of the minimum to the maximum value for that transition segment. The synergy mutation degree ranges from 0 to 1; a higher value indicates a more balanced and synchronized 3D mutation amplitude. Records in the image mutation vector set with a significantly higher single-dimensional metric than the other two dimensions correspond to a single-dimensional dominant pattern. Records with similar 3D means and an image feature synergy mutation degree higher than 0.6 correspond to a multi-dimensional synergistic pattern. Records with a synergy mutation degree between 0.3 and 0.6 are classified as transitional patterns. Image feature co-mutation degree is combined with the range field of three-dimensional mutation amount. The larger the range, the higher the degree of three-dimensional mutation imbalance. When the range exceeds 0.3 times the normalized value range of each index, it tends to be judged as a single-dimensional dominant mode.

[0047] Image mutation pattern identifiers are generated by identifying unidimensional dominant mutations and multidimensional co-mutations based on the co-mutation degree of image features. The image mutation pattern identifier is a classification conclusion of the mutation nature of each boundary segment, answering the question: Is the color, texture, and structure of this boundary segment a synchronous mutation, a single-dimensional mutation, or something in between? These three properties correspond to three integer codes: multidimensional co-mutation, unidimensional dominant mutation, and transitional mode, respectively. The image mutation pattern identifier is the data field that records this coding conclusion. Boundary segments with a co-mutation degree higher than 0.6 are classified as multidimensional co-mutations and coded into this category, corresponding to typical features of overall switching in the three dimensions of mineral composition, grain size distribution, and bedding structure at lithological hard interfaces. Boundary segments with a co-mutation degree lower than 0.3 are classified as unidimensional dominant mutations and coded into this category, corresponding to soft interface features such as alteration zones or mineralization halos where only hue or texture changes. Boundary segments with a co-mutation degree between 0.3 and 0.6 are classified into transitional modes and labeled with independent codes, without being forcibly classified into unidimensional or multidimensional categories; these are commonly found in mixed sedimentary rocks or transitional sedimentary environments. Image mutation pattern identifiers are stored in integer code form along with the sequence number of the over-limit segment. Each over-limit segment corresponds to a unique pattern code record. The three types of codes exhaustively enumerate all mutation properties of the over-limit segment without omission or overlap.

[0048] Image feature limit-crossing type parameters are obtained by distinguishing between gradual and abrupt limit-crossing based on image mutation pattern identifiers. In the image mutation pattern identifiers, the mean values ​​of color, texture, and structural three-dimensional mutation amounts in the limit-crossing segments of multi-dimensional collaborative mutation type are significantly higher than those of single-dimensional dominant mutation type. This difference in magnitude constitutes the quantitative basis for distinguishing between gradual and abrupt limit-crossing, and the distinction between the two types of limit-crossing has a direct impact on the depth localization accuracy of subsequent interface transition zone feature slice extraction. In image mutation pattern identification, multi-dimensional cooperative mutation-type boundary crossings are classified as mutation-type boundary crossings, corresponding to a clear contact interface between two types of lithology; single-dimensional dominant mutation-type boundary crossings are classified as gradual boundary crossings, corresponding to a blurred interface where the proportion of lithological components gradually transitions, commonly seen in the grain size gradient sections of delta front and littoral facies deposits; transitional pattern boundary crossings are further distinguished based on the ratio of the boundary crossing depth range to the thickness of the adaptive discrimination benchmark segment. A ratio greater than 0.5 is classified as a gradual boundary crossing, and a ratio less than 0.5 is classified as a mutation-type boundary crossing. The specific value of the image feature cooperative mutation degree is also used for confirmation; a higher cooperative mutation degree (close to 0.6) tends to classify it as a mutation-type boundary crossing, while a lower degree (close to 0.3) tends to classify it as a gradual boundary crossing. The image feature boundary crossing type parameter uses the boundary crossing type code as the core field: 0 for gradual boundary crossing and 1 for mutation-type boundary crossing. The image feature boundary crossing type parameter also records the depth range of each boundary crossing segment and the pattern code of the image mutation pattern identifier.

[0049] Abrupt triggering identifiers are generated based on image feature exceedance type parameters. These parameters differentiate between gradual and abrupt changes in each exceedance segment. The abrupt triggering identifier transforms this property determination into an actionable trigger signal, determining whether each exceedance segment enters the interface extraction process. For abrupt exceedance segments, the exceedance gradient exceeds the adaptive discrimination benchmark tolerance range within a very short depth range, indicating a clear lithological interface and strong signal. When the type is coded as abrupt exceedance, a triggering identifier is directly generated. For gradual exceedance segments, the exceedance gradient typically accumulates slowly over a longer depth range. Not all gradual exceedances possess sufficient interface recognition value. When the type is coded as gradual exceedance, an additional triggering condition must be added: the depth range must exceed the corresponding segment thickness of the adaptive discrimination benchmark by 30%. Only when this condition is met is a triggering identifier generated. Very short gradual exceedance segments are not triggered to avoid noise interference. Differentiated triggering logic is used for the two types of exceedances to ensure that the abrupt triggering identifier provides a reasonable triggering determination in stratigraphic segments with varying lithological signal strengths. The mutation trigger identifier uses the out-of-limit segment number of the image feature out-of-limit type parameter as the association key. The trigger status is stored in binary code, with 1 for triggering and 0 for not triggering. The mutation trigger identifier corresponding to the out-of-limit segment of the transition mode class is marked as pending confirmation. It is distinguished from triggering and not triggering by independent encoding. The pending confirmation status retains the processing opportunity of the transition mode interface without forcibly triggering or excluding it.

[0050] Step S140: Extract interface transition zone feature slices based on mutation triggering identifiers, evaluate the edge direction continuity of the interface transition zone feature slices using multi-sampling points to obtain geometric distortion correction coefficients, and use the geometric distortion correction coefficients to perform inverse perspective transformation and scale calibration to construct a depth-image analysis profile.

[0051] Specifically, interface transition zone feature slices are extracted based on mutation trigger markers. Triggering states of 1 in the mutation trigger markers correspond to interface locations in the borehole profile where lithology undergoes a significant shift. Triggering states awaiting confirmation are also included in the extraction range to preserve processing opportunities for transitional interfaces. Both types of triggering states jointly determine the extraction depth range for interface transition zone feature slices. Based on the depth range of each triggering segment in the mutation trigger markers, corresponding image strips are extracted from the stitched core image, extending upwards and downwards by a fixed margin from the center depth of the triggering segment. The margin width covers a certain proportion of the thickness of the adaptive discrimination benchmark segment to ensure that the lithological background areas on both sides of the interface are included in the slice range. The spatial resolution of the interface transition zone feature slices inherits 0.05 mm / pixel from the stitched core image. The depth coordinate system of the slices is established with the depth of the first end of the triggering segment of the mutation trigger marker as the zero point. This local coordinate system facilitates the position indexing during subsequent evaluation of the continuity of edge orientation at multiple sampling points. The depth range of the interface transition zone feature slices corresponding to the gradual over-limit trigger segment in the mutation triggering identifier is appropriately expanded compared to the mutation over-limit segment to fully cover the entire depth range of the gradual transition. After extraction, the interface transition zone feature slices undergo signal-to-noise ratio (SNR) verification. Slices with an SNR below the threshold are triggered to be re-extracted from the original data of the color-homogenized image group. Backtracking the original data can avoid the accumulation and superposition of noise introduced by distance-weighted fusion and multiple resampling during the stitching process, thereby improving the slice SNR.

[0052] In some embodiments, the step of evaluating the continuity of edge direction at multiple sampling points of the interface transition zone feature slice to obtain geometric distortion correction coefficients includes: analyzing the edge gradient direction of sampling points based on the interface transition zone feature slice to generate an image edge principal direction sequence and image coordinate system reference parameters; performing collinearity grouping identification on the image edge principal direction sequence to generate image principal direction clustering parameters; determining image global perspective offset parameters based on the image principal direction clustering parameters and the image coordinate system reference parameters; and performing inverse mapping to solve the image global perspective offset parameters to generate geometric distortion correction coefficients.

[0053] Based on the feature slices of the interface transition zone, the gradient direction of the sampling points is analyzed to generate the main direction sequence of image edges and reference parameters of the image coordinate system. In the feature slices of the interface transition zone, the lithological interface appears as a continuous high-gradient edge band in the image. During the cutting and transportation of borehole cores, uneven stress causes end face tilting or lateral twisting. This geometric deformation makes the interface edge appear non-horizontal in the image. The analysis of the gradient direction of the sampling points is precisely to capture this deviation and provide directional basis for subsequent perspective correction. Sampling points of the interface transition zone feature slice are arranged in a uniform grid within the slice area, with the sampling point spacing matching the slice spatial resolution. At each sampling point, the horizontal gradient component Gx and the vertical gradient component Gy are calculated using the Sobel operator, and the gradient magnitude is... Sampling points exceeding the threshold are considered valid edge points. The principal direction angle of the edge is calculated using the four-quadrant arctangent function θ = arctan2(Gy,Gx) ​​mod 180°, with the value range limited to [0°, 180°) to represent the direction angle of an undirected edge, eliminating the directional ambiguity caused by arctan(Gy / Gx) when Gx < 0. The principal direction angle of a valid edge point is used as the direction record value of that point. The principal direction angles of each valid edge point are arranged according to the depth coordinates of the sampling points to form the image edge principal direction sequence. Each element of the image edge principal direction sequence stores the pixel coordinates (x,y) and direction angle θ of the corresponding valid edge point. The pixel coordinates (x,y) are used to support the collinear grouping straight line fitting in sub-step 2. The number of elements in the image edge principal direction sequence is equal to the total number of valid edge points verified by the gradient threshold in the interface transition zone feature slice. The image coordinate system reference parameters are based on the position of the origin of the pixel coordinate system and the direction of the coordinate axes of the feature slice of the interface transition zone. The origin of the coordinate system is set to the upper left pixel of the slice. The image coordinate system reference parameters and the image edge main direction sequence together serve as the spatial reference for subsequent collinear grouping.

[0054] Collinearity grouping identification is performed on the image edge principal direction sequence to generate image principal direction clustering parameters. Effective edge points of the same lithological interface in the image edge principal direction sequence tend to be consistent in direction angle, while the direction angles of different lithological interfaces or noise points are scattered. Collinearity grouping identification utilizes the degree of clustering of direction angles to group edge points belonging to the same interface into the same group, thereby extracting the principal direction parameters representing the overall interface orientation. The image edge principal direction sequence is statistically analyzed using a direction angle histogram to count the edge point distribution density in each direction. The histogram counts the number of effective edge points in each direction angle interval with an angular resolution of 1°. The peak interval corresponds to the principal direction of the interface orientation, and the peak width reflects the stability of the interface orientation; the narrower the peak, the more regular the interface orientation. Effective edge points in the image edge principal direction sequence whose direction angles fall into the peak interval are grouped into the same collinearity group. Member points of each collinearity group are used to perform least-squares linear fitting to calculate the representative direction within the group using their stored pixel coordinates (x, y). Member points whose fitting residuals exceed a threshold are removed as outliers. After removal, the fitting is repeated until the residuals meet the requirements. The image principal orientation clustering parameter records the representative orientation angle and the number of member points for each collinear group. The representative orientation angle of the image principal orientation clustering parameter comes from the least squares fitting result after convergence, and the number of member points reflects the edge point support strength of this direction in the slice. Collinear groups with a mean fitting residual exceeding a threshold are recorded in the image principal orientation clustering parameter but marked as low-quality groups, and their weights are reduced when calculating the global perspective offset parameter to reduce interference. The collinear group with the most member points is marked as the main interface orientation group in the image principal orientation clustering parameter, corresponding to the dominant lithological interface direction in the slice.

[0055] The global perspective offset parameter of the image is determined based on the image principal orientation clustering parameters and the image coordinate system reference parameters. The angle between the main interface orientation in the image principal orientation clustering parameters and the horizontal reference direction defined by the image coordinate system reference parameters reflects the overall tilt of the interface in the image. Deviations in the cutting angle of the borehole core end face and tilting caused by handling and flipping will both cause this angle to deviate from zero. The global perspective offset parameter is a quantitative description of this overall tilt. The difference between the representative orientation angle of each collinear group in the image principal orientation clustering parameters and the horizontal reference direction of the image coordinate system reference parameters is defined as the local perspective offset of each group. The local perspective offset of each group is weighted by the number of member points and by the reciprocal of the mean of the fitted residuals as a quality weight. These two weights are combined to calculate the directional offset component of the global perspective offset parameter, in degrees. The intercepts of the fitted lines for each collinear group along the image's horizontal axis are linearly regressed with the group's spatial depth position as the independent variable. The regression slope reflects the systematic lateral shift of the interface with depth, serving as the scale offset component of the global perspective offset parameter, measured in pixels per depth unit. A larger slope in the intercept linear trend indicates more significant lateral scale distortion within the image plane. The global perspective offset parameter consists of two components: orientation offset and scale offset. These two components together describe the perspective deformation of the interface transition zone feature slices within the image plane.

[0056] The geometric distortion correction coefficients are generated by inverse mapping of the global perspective offset parameters of the image. The direction offset and scale offset described by the global perspective offset parameters together constitute the perspective transformation in the image plane. In the scenario of core end face tilt correction, the perspective transformation is simplified to an affine transformation, and the perspective transformation matrix H is a 3×3 matrix: H=[cosψ,-η·sinψ,0;sinψ,η·cosψ,0;0,0,1], where ψ is the direction offset component (unit: rad, converted from degrees), and η is the anisotropic scaling coefficient (dimensionless) derived from the scale offset component. S is the scale offset component (unit: pixels / depth units), D is the depth span of the current interface transition zone feature slice (depth units), and W is the pixel width (pixels) along the horizontal axis of the slice image. When S approaches zero, η approaches 1; the third row maintains the projection unchanged. The inverse mapping solution uses the inverse of the perspective transformation matrix as the target, mapping the distorted pixel coordinates in the interface transition zone feature slice back to the theoretical standard coordinates. The standard coordinate system is defined with the core end face placed horizontally and the interface orientation parallel to the horizontal axis of the image as the reference. The directional offset component of the global perspective offset parameter of the image corresponds to the rotation component of the perspective transformation matrix, and the scale offset component corresponds to the orientation scaling component of the perspective transformation matrix. The two types of components jointly construct the complete perspective transformation matrix H, and the inverse mapping matrix H-1 is directly obtained by matrix inversion. The geometric distortion correction coefficients are stored in the elements of the inverse mapping matrix H-1. After applying the geometric distortion correction coefficients, the residual angle between the main interface orientation of the interface transition zone feature slice and the horizontal axis of the image should be lower than the correction accuracy threshold. Before applying the geometric distortion correction coefficients, a reversibility verification is performed. The verification method is to apply the correction coefficients to the feature slices of the original interface transition zone and then apply the forward transformation matrix H. The correction coefficients are considered valid when the root mean square error of the pixel coordinates between the restored image and the original image is lower than the threshold.

[0057] A depth-image analysis profile was constructed using inverse perspective transformation and scale calibration with geometric distortion correction coefficients. Each slice entered the inverse perspective transformation process in sequence according to the trigger segment depth recorded by the abrupt change trigger marker. Geometric distortion correction coefficients were applied to the interface transition zone feature slices in the form of an inverse mapping matrix. The inverse perspective transformation mapped the pixel coordinates of each slice from the distortion space to the standard coordinate system. After the transformation, the main interface orientation of each slice was aligned with the horizontal axis of the image, and the spatial pose of the core end face was restored to the theoretical level. Scale calibration used the actual length of each core segment recorded by the depth marker information as a benchmark, uniformly calibrating the pixel scale of the slices after the inverse perspective transformation to 0.05 mm / pixel. During the calibration process, the ratio of the pixel spacing along the depth direction to the actual depth spacing of each slice was used as the scale correction factor. When the scale correction factor deviated from the standard value by more than a threshold, bilinear interpolation resampling was performed on the slice to eliminate the scale deviation after the application of the geometric distortion correction coefficients, ensuring that the pixel scale of each slice after resampling was consistent with the calibration target. The depth-image analysis profile is constructed by stitching together all slices that have undergone inverse perspective transformation and scale calibration in depth order. The stitching boundaries of adjacent slices in the depth direction are aligned with the depth of the trigger segment boundary marked by abrupt change. Gradual blending is used at the stitching points to eliminate brightness abrupt changes. The depth axis resolution of the depth-image analysis profile inherits the 0.1mm accuracy of the depth marker information, while the image axis resolution inherits the 0.05mm / pixel accuracy after scale calibration. The resolutions of the two axes are recorded independently to support coordinate conversion during subsequent lithological zoning boundary location. After the depth-image analysis profile is constructed, a global depth continuity verification is performed. If the depth coordinate deviation at the stitching point of adjacent slices exceeds a threshold, slice realignment is triggered.

[0058] Step S150: Based on the depth-image analysis profile, lithological zoning boundaries are located to form a lithological confidence map. The lithological confidence map is then compared and classified to output a standardized image logging report.

[0059] In some embodiments, the step of locating lithological zoning boundaries and forming a lithological confidence map based on the depth-image analysis profile includes: extracting the thickness distribution and lithological transition features of each lithological zoning unit from the depth-image analysis profile; locating the boundaries of the lithological zoning units based on the thickness distribution and the lithological transition features to generate an image boundary candidate set; filtering false boundaries from the image boundary candidate set using the thickness distribution as a constraint criterion to obtain an effective image boundary set; and performing confidence assessment based on the effective image boundary set to form a lithological confidence map.

[0060] Thickness distribution and lithological transition characteristics of each lithological zoning unit are extracted from the depth-image analysis profile. The depth range of each lithological zoning unit in the depth-image analysis profile is initially defined by the lithological segment boundaries of the adaptive discrimination criterion. The color distribution and texture characteristics of the image within the zoning unit remain relatively stable in the depth direction, while obvious image feature jumps are observed at the unit boundaries. Thickness distribution is calculated based on the length of the depth range occupied by each lithological zoning unit in the depth-image analysis profile. When the borehole traverses thin interbedded strata, the thickness difference between adjacent zoning units is significant. Zoning units with extremely small thickness may correspond to only a few millimeters of thin interlayers. The statistical results of thickness distribution are stored in the records of each zoning unit as two fields: minimum thickness and mean thickness. The minimum thickness is used as a reasonable thickness threshold constraint for subsequent false boundary filtering, and the mean thickness is used as a reference for the reasonableness of thickness in subsequent confidence assessment. The lithological transition characteristics are described by the gradient intensity distribution of the depth-image analysis profile at the boundary of each partition unit. Boundaries with high gradient intensity correspond to clear lithological contact surfaces, while boundaries with low gradient intensity and wide distribution correspond to gradual transition zones. The lithological transition characteristics of the two types of boundaries are quantified by two indicators: the mean gradient of the boundary and the width of the gradient distribution. The extraction results are aligned with the depth coordinate system of the depth-image analysis profile using the partition unit number as the primary key.

[0061] Based on thickness distribution and lithological transition characteristics, a candidate set of image boundaries is generated for lithological zoning unit boundary localization. The depth intervals of each zoning unit in the thickness distribution determine the search range of the candidate boundary points, while the gradient intensity distribution in the lithological transition characteristics indicates the optimal position of the candidate boundary points within the search range. The combined constraint of these two factors ensures that boundary localization has both depth prior and relies on measured image features; relying solely on either could lead to candidate boundary points deviating from the actual interface position. Lithological zoning unit boundary localization treats the top and bottom ends of each zoning unit as independent objects. The peak depth of the gradient intensity at the bottom end of the previous unit is used as the candidate point for the top boundary, while the peak depth of the gradient intensity within the same unit is used as the candidate point for the bottom boundary. The peak depth is determined by calculating the row mean gradient of the depth-image analysis profile pixel by pixel within the boundary search range and taking the position of the maximum value. If the candidate depths at the top and bottom ends of the same boundary are inconsistent, the average of the two is taken as the final candidate depth. In lithological transition features, gradient transition zones with large gradient distribution widths are recorded as depth intervals rather than single depth points in the image boundary candidate set. The top and bottom depths of the gradient transition zone are taken as the depth coordinates where the gradient intensity drops to 50% of the peak value, and a transition zone marker is added to fully cover the depth range of the transition zone. The image boundary candidate set records the candidate depths of the top and bottom boundaries of each partition unit and the corresponding gradient intensities. Each record in the image boundary candidate set also stores a boundary type field—a clear interface records a single depth point, while a gradient transition zone records a depth interval and adds a transition zone marker. The candidate depth accuracy of the image boundary candidate set inherits the 0.1mm depth axis resolution of the depth-image analysis profile. The image boundary candidate set retains all candidate points before false boundary filtering to avoid missing real interfaces.

[0062] Using thickness distribution as a constraint, the image boundary candidate set is filtered to obtain the effective boundary set. Some candidate points in the image boundary candidate set originate from high-gradient responses formed in the image by local contamination, cutting marks, or transport scratches on the core surface. These false boundaries appear as isolated short-range high-gradient line segments in the depth-image analysis profile, which are fundamentally different in spatial extension from the continuous high-gradient bands of the real lithological interface. If not filtered, false boundaries will occupy recording positions in the confidence assessment and interfere with the confidence ranking of the real interface. Thickness distribution provides a reasonable thickness range constraint for each lithological partition unit. When the depth distance between two adjacent candidate boundaries in the image boundary candidate set is less than the minimum reasonable thickness threshold of the thickness distribution, the candidate boundary with a lower gradient intensity is judged as a false boundary and is removed. Candidate boundaries with abnormally small gradient distribution width in lithological transition features also trigger false boundary verification. The verification recalculates the average gradient of multiple rows at the boundary using the original data of the depth-image analysis profile. Candidate boundaries whose gradient intensity fails to pass the threshold after recalculation are marked as false boundaries. The effective boundary set of the image is constructed based on the candidate boundary set of the image after removing false boundaries. The depth coordinates, gradient intensity and transition zone marker fields of each effective boundary in the effective boundary set are fully inherited from the candidate boundary set of the image, and no new field loss is introduced due to the filtering operation. The depth coordinate accuracy of the effective boundary set of the image is consistent with that of the candidate boundary set of the image. The number of records in the effective boundary set of the image is equal to the total number of candidate boundaries filtered by false boundaries in the candidate boundary set of the image. The rationality of the number of records after filtering is verified with reference to the expected layer density of the borehole strata.

[0063] A lithological confidence map is generated based on the effective boundary set of the image. The confidence of each effective boundary in the effective boundary set is calculated by weighting three indicators: boundary gradient strength, boundary depth positioning accuracy, and the reasonableness of the thickness of the lithological zoning units on both sides of the boundary. The confidence formula is C=ω1·G_norm+ω2·P_norm+ω3·T_norm, where G_norm is the normalized value of the boundary gradient strength, P_norm is the normalized value of the depth positioning accuracy, and T_norm is the normalized value of the thickness reasonableness (calculated based on the deviation of the current unit thickness from the mean thickness). The weighting coefficients ω1, ω2, and ω3 are defaulted to 0.4, 0.3, and 0.3, respectively. For the confidence assessment of the gradient boundaries marked in the transition zone of the effective boundary set of the image, an additional indicator of the width of the transition zone depth interval is introduced. Gradient zones with moderate widths have higher confidence than those with abnormally large widths. Abnormally large widths usually indicate core disturbance or lithological contamination. The lithological confidence map uses the lithological zoning unit number as the primary key. The effective boundary depth, boundary confidence, thickness distribution, and lithological transition feature type of each unit are used as attribute fields. The endmember mean vector and grain size classification label of each zoning unit are mapped and associated with the segment number of the lithological discrimination feature set through the zoning unit depth interval and stored together in the lithological confidence map as input fields for subsequent lithological classification and comparison. Boundaries with confidence below the threshold are marked with low confidence, and the corresponding lithological zoning units are given a lower weight in subsequent classification and comparison. Zoning units with high confidence directly enter the classification process.

[0064] A standardized image logging report is generated by performing lithology classification comparison on the lithology confidence map. The lithology classification comparison uses the endmember mean vector of each lithological zoning unit in the lithology confidence map as input, and calculates the abundance angle cosine value cosα=(v·r) / (‖v‖·‖r‖) for each lithology category in the preset lithology standard abundance library, where v is the endmember mean vector of the zoning unit, r is the reference abundance vector, and α is the angle between the endmember mean vector and the reference abundance vector. The reference lithology category with the largest cosine value is determined as the classification result of that zoning unit. The lithology standard abundance library covers common lithology categories from borehole exploration, including six major categories: sandstone, mudstone, siltstone, limestone, dolomite, and igneous rocks. Each major category is further subdivided into 2 to 4 subcategories to cover differences in mineral assemblage. Each reference abundance vector is jointly calibrated using standard thin section identification and endmember unmixing. In the lithology confidence map, the first two matching lithology categories and cosine similarity of zoning units with low confidence markers are output during classification and comparison for manual review by geologists. High-confidence zoning units only output the best matching result. The lithology classification and comparison results, along with the boundary depth, confidence level, and thickness distribution fields of the lithology confidence map, constitute the core content of the standardized image logging report. The report is indexed by borehole number and depth interval, with each lithology zoning unit occupying one record in the report, including top and bottom depths, classified lithology name, boundary confidence level, and thickness. The standardized image logging report also includes a thumbnail image index of the depth-image analysis profile. Each zoning unit record is associated with the corresponding depth segment image via depth coordinates, supporting viewing of images at any depth segment. The report is output in a structured data format and is compatible with the borehole data exchange standards commonly used in the geological exploration industry.

[0065] To implement the machine vision-based borehole formation AI logging method corresponding to the above method embodiments, and to achieve the corresponding functions and technical effects, see [link to documentation]. Figure 2 , Figure 2 This paper illustrates a structural block diagram of a machine vision-based borehole formation AI logging system 200 provided in an embodiment of this application, including: Image acquisition module 201 is used to acquire multispectral borehole core images and multi-view ring scan core surface images and depth marking information, and to perform illumination-color normalization registration on the multispectral borehole core images and multi-view ring scan core surface images to generate color-consistent image groups. Feature parsing module 202 is used to generate a stitched core image by partitioning and stitching the image through the color-homogenized image group, to perform lithological end-member decomposition on the stitched core image to determine the lithological component ratio, and to extract grain size classification features based on the lithological component ratio to generate a lithological discrimination feature set. Image detection module 203 is used to associate the lithology discrimination feature set with the depth identification information to establish a depth-lithology comparison index, identify the continuous drop position of adjacent core segments based on the depth-lithology comparison index to generate an adaptive discrimination benchmark, and perform feature over-limit detection on the lithology discrimination feature set according to the adaptive discrimination benchmark to generate a sudden change triggering identifier; Interface evaluation module 204 is used to extract interface transition zone feature slices based on the mutation triggering identifier, perform multi-sampling point edge direction continuity evaluation on the interface transition zone feature slices to obtain geometric distortion correction coefficients, and use the geometric distortion correction coefficients to perform inverse perspective transformation and scale calibration to construct a depth-image analysis profile. The cataloging output module 205 is used to locate the lithological zoning boundaries based on the depth-image analysis profile to form a lithological confidence map, and to perform lithological classification comparison on the lithological confidence map to output a standardized image cataloging report.

[0066] The aforementioned machine vision-based borehole formation AI logging system 200 can implement the machine vision-based borehole formation AI logging method of the above method embodiments. The options in the above method embodiments are also applicable to this embodiment, and will not be detailed here. The remaining content of this application's embodiments can be referred to the content of the above method embodiments, and will not be repeated in this embodiment.

[0067] The above description is only a part or preferred embodiment of this application. Neither the text nor the drawings should limit the scope of protection of this application. All equivalent structural transformations made using the content of this application's specification and drawings under the overall concept of this application, or direct / indirect applications in other related technical fields, are included within the scope of protection of this application.

Claims

1. A machine vision-based AI-based borehole formation logging method, characterized in that, include: Multispectral borehole core images and multi-view ring scan core surface images and depth marking information are acquired. Illumination-color normalization registration is performed on the multispectral borehole core images and multi-view ring scan core surface images to generate color-consistent image groups. The image is partitioned and stitched together using the color-homogenized image group to generate a stitched core image. The stitched core image is then subjected to lithological end-member decomposition to determine the lithological component ratio. Based on the lithological component ratio, grain size classification features are extracted to generate a lithological discrimination feature set. A depth-lithology comparison index is established by associating the lithology discrimination feature set with the depth identification information. Based on the depth-lithology comparison index, the continuous drop position of adjacent core segments is identified to generate an adaptive discrimination benchmark. Based on the adaptive discrimination benchmark, feature over-limit detection is performed on the lithology discrimination feature set to generate a sudden change triggering identifier. Based on the mutation triggering identifier, the interface transition zone feature slice is extracted. The geometric distortion correction coefficient is obtained by evaluating the edge direction continuity of the interface transition zone feature slice through multi-sampling points. The geometric distortion correction coefficient is then used to perform inverse perspective transformation and scale calibration to construct a depth-image analysis profile. Based on the depth-image analysis profile, lithological zoning boundaries are located to form a lithological confidence map. The lithological confidence map is then used for lithological classification and comparison to output a standardized image logging report.

2. The method according to claim 1, characterized in that, The step of performing illumination-color normalization registration on the multispectral borehole core image and the multi-view ring scan core surface image to generate a color-consistent image group includes: The end-face color ring distribution features are extracted from the multispectral borehole core images to generate the image oxide halo width parameter; Based on the image oxide halo width parameter, a color difference compensation matrix is ​​generated by performing graded color difference compensation on the multispectral borehole core image; Based on the chromatic aberration compensation matrix, the multispectral borehole core image and the multi-view ring scan core surface image are registered to obtain a registered image pair. The registered image pairs are then subjected to color normalization processing to generate a color-consistent image group.

3. The method according to claim 1, characterized in that, The step of performing lithological end-member decomposition on the stitched core image to determine the lithological composition ratio includes: The spectral features of the stitched core images are extracted to generate spectral vectors for each pixel. The spectral vectors of each pixel are used to perform pure pixel screening to generate an image endmember spectral library; Based on the image endmember spectral library, the spectral vectors of each pixel are subjected to abundance unmixing processing to obtain the pixel endmember abundance matrix; The proportion of lithological components in each core segment is determined based on the percentage of each lithological component in the pixel end-member abundance matrix.

4. The method according to claim 1, characterized in that, The process of generating an adaptive discrimination criterion based on the depth-lithology comparison index to identify the location of continuous drops in adjacent core segments includes: Extract the lithological discrimination feature baselines of images for each depth segment from the depth-lithology comparison index; Based on the lithological discrimination feature baselines of the images at each depth segment, depth compensation is performed according to the diagenetic compaction law to generate image feature compaction correction baselines; The image feature compaction correction baseline is used to extract the difference between adjacent segments to identify the location of sudden drops in the continuity of image features; Based on the location of the sudden drop in the continuity of the image features, the image feature compaction correction baseline is divided into lithological segments to generate an adaptive discrimination benchmark.

5. The method according to claim 1, characterized in that, The step of performing feature over-limit detection on the lithological discrimination feature set based on the adaptive discrimination benchmark to generate a mutation trigger flag includes: The lithology discrimination feature set is arranged in depth order to generate a depth-image feature sequence; Based on the adaptive discrimination criterion, the depth-image feature sequence is scanned segment by segment to identify image feature segments that exceed the limit; The image out-of-limit feature segments are divided into gradual out-of-limit and abrupt out-of-limit types according to the difference in out-of-limit gradient, generating image feature out-of-limit type parameters; A mutation trigger identifier is generated based on the image feature over-limit type parameter.

6. The method according to claim 1, characterized in that, The step of evaluating the edge direction continuity of the interface transition zone feature slices using multi-sampling points to obtain geometric distortion correction coefficients includes: Based on the interface transition zone feature slices, the edge gradient direction of the sampling points is analyzed to generate the image edge main direction sequence and image coordinate system reference parameters; The image edge main direction sequence is subjected to collinearity grouping identification to generate image main direction clustering parameters; The global perspective offset parameters of the image are determined based on the image principal orientation clustering parameters and the image coordinate system reference parameters; The geometric distortion correction coefficients are generated by inverse mapping the global perspective offset parameters of the image.

7. The method according to claim 1, characterized in that, The process of locating lithological zoning boundaries and forming lithological confidence maps based on the depth-image analysis profile includes: The thickness distribution and lithological transition characteristics of each lithological zoning unit are extracted from the depth-image analysis profile. Based on the thickness distribution and the lithological transition characteristics, the boundary of the lithological partition unit is located to generate an image boundary candidate set; Using the thickness distribution as a constraint criterion, the image boundary candidate set is filtered to obtain the effective image boundary set; A lithology confidence map is generated based on the effective boundary set of the image and a confidence assessment is performed.

8. The method according to claim 4, characterized in that, The step of extracting adjacent segment differences from the image feature compaction correction baseline to identify locations of sudden drops in image feature continuity includes: Based on the image feature compaction correction baseline, segmented sampling is performed according to the depth step size to generate an image feature sampling sequence; The image feature sampling sequence is subjected to adjacent difference extraction to generate an image feature difference sequence; The positive and negative cycles are identified in the image feature difference sequence to generate image feature cycle direction identifiers; The location of the sudden drop in the continuity of image features is determined based on the point of sudden drop in the positioning difference of the image feature rotation direction identifier.

9. The method according to claim 5, characterized in that, The step of dividing the image out-of-limit feature segments into gradual out-of-limit and abrupt out-of-limit types according to the difference in out-of-limit gradient to generate image feature out-of-limit type parameters includes: The color mutation amount, texture mutation amount and structural mutation amount are extracted from the image over-limit feature segments to generate an image mutation vector set; The image mutation vector set is evaluated for inter-dimensional synergy to generate image feature synergy mutation degree; Based on the image feature co-mutation degree, single-dimensional dominant mutations and multi-dimensional co-mutations are identified to generate image mutation pattern identifiers; Based on the image mutation mode identifier, the gradual type of limit violation and the mutation type of limit violation are distinguished to obtain the image feature limit violation type parameter.

10. A machine vision-based borehole formation AI logging system, characterized in that, include: The image acquisition module is used to acquire multispectral borehole core images and multi-view ring scan core surface images and depth marking information, and to perform illumination-color normalization registration on the multispectral borehole core images and multi-view ring scan core surface images to generate color-consistent image groups. The feature parsing module is used to generate a stitched core image by partitioning and stitching the image through the color-homogenized image group, to perform lithological end-member decomposition on the stitched core image to determine the proportion of lithological components, and to extract grain size classification features based on the proportion of lithological components to generate a lithological discrimination feature set. The image detection module is used to associate the lithology discrimination feature set with the depth identification information to establish a depth-lithology comparison index, identify the continuous drop position of adjacent core segments based on the depth-lithology comparison index to generate an adaptive discrimination benchmark, and perform feature over-limit detection on the lithology discrimination feature set according to the adaptive discrimination benchmark to generate a sudden change triggering identifier; The interface evaluation module is used to extract interface transition zone feature slices based on the mutation triggering identifier, perform multi-sampling point edge direction continuity evaluation on the interface transition zone feature slices to obtain geometric distortion correction coefficients, and use the geometric distortion correction coefficients to perform inverse perspective transformation and scale calibration to construct a depth-image analysis profile. The cataloging and output module is used to locate the lithological zoning boundaries based on the depth-image analysis profile to form a lithological confidence map, and to perform lithological classification and comparison on the lithological confidence map to output a standardized image cataloging report.

Images