Deep learning-based coal seam separation boundary extraction method

Abstract

The invention relates to the technical field of coal seam image processing, and particularly discloses a coal seam separation boundary extraction method based on deep learning, which is used for solving the problem that pseudo-edges, fractures and horizon drifts easily occur in interface segmentation and edge extraction due to interference of water film fogging, reflective distortion and texture gradual change on a composite roof roadway drill hole wall image. Comprising the steps of hole wall image preprocessing and cylindrical surface expansion correction, near-far contrast difference calculation and equivalent driving quantity thermodynamic diagram generation, imaging quality and contour continuity weight construction, multi-source feature fusion depth segmentation and boundary contour extraction, contour and mask write-back weight self-updating and closed-loop iteration convergence control. According to the method, through near-far side comparison driving thermodynamic diagrams, imaging quality and contour continuity weight constraints and in combination with depth segmentation closed-loop iteration, the boundary contour of the potential structural plane of the separation layer is stably extracted, and false edges, fractures and horizon drifts caused by fogging water films and crack drill marks are reduced.

Description

Technical Field

[0001] This invention relates to the field of coal seam image processing technology, and in particular to a method for extracting coal seam delamination boundaries based on deep learning. Background Technology

[0002] In coal mine roadways with complex roofs, determining the critical value of delamination requires drilling at typical monitoring sections at a certain distance from the working face gate. Combining hole depth with the columnar relationship of lithology, key interfaces such as sandstone, mudstone, carbonaceous mudstone, coal seams, and weak interlayers are identified. Multi-point displacement gauges are then anchored at the upper and lower plates of these interfaces to invert the delamination amount from the displacement difference between adjacent fixed points. Simultaneously, it is necessary to distinguish the delamination response differences inside and outside the anchorage zone to determine the critical value for dynamic updating of the gate position and anchorage zones. However, existing borehole recorder images, after washing, are still susceptible to residual water film and cooling fogging, localized reflections and uneven lighting, cylindrical geometric distortion caused by slight camera contact with the wall, and gradual changes in the grayscale texture of sandy mudstone and carbonaceous mudstone. Due to the influence of the current image processing, existing methods typically employ binary segmentation by setting a fixed threshold for the grayscale of the entire image or a local hole wall, or by determining the edge on a single side of the candidate interface based solely on the grayscale gradient amplitude exceeding the threshold. This is combined with connected component filtering and morphological opening and closing operations to obtain the interface contour. However, there is a lack of joint statistics and robust calculations for the lithological contrast difference between the upper and lower sides of the candidate interface. Furthermore, there is a lack of constraints on imaging quality and contour continuity, resulting in broken virtual edges, false edges, layer drift, or missed detections in the interface contour, which are confused with strong gradient structures such as fracture drill marks. This leads to mismatch of anchored layers at fixed points or incorrect displacement attribution, causing deviations in the calculation of delamination amount and the determination of delamination critical values. It is difficult to support the automated, stable, and reproducible determination of the delamination critical value of composite roof. Summary of the Invention

[0003] The technical problem to be solved by the present invention is to provide a method for extracting coal seam delamination boundaries based on deep learning. The method generates a driving heat map by equivalent conversion of the logarithmic mean difference between the near and far lithological contrast between the upper and lower sides of the candidate interface, and guides the deep segmentation network to stably extract the boundary contour of the potential delamination structural surface by combining the back-write iteration mechanism of imaging quality and contour continuity constraints.

[0004] To achieve the above objectives, the present invention provides the following technical solution:

[0005] A deep learning-based method for extracting coal seam delamination boundaries includes:

[0006] Step 1: Obtain the borehole wall image and perform cylindrical surface unfolding, defogging, and water film correction.

[0007] Step 2: Calculate the lithological contrast difference on both sides of the candidate interface. The near side is the strip sampling area close to the boundary on one side of the candidate interface, and the far side is the strip sampling area on the same side along the boundary normal direction at a preset distance from the near side and far from the boundary. The equivalent contrast driving amount is converted into the logarithmic mean difference. Then, the equivalent driving amount obtained by each candidate boundary pixel is laid back into the image to form a driving heat map to guide the segmentation of the potential delamination structure surface boundary.

[0008] Step 3: Calculate the image quality weight and the contour continuity weight;

[0009] Step 4: After fusing the multi-scale features of the driving heat map and the hole wall image, input the data into the deep learning segmentation network, output the interface mask, and perform contour extraction on the interface mask to obtain the boundary contour of the potential delamination structure surface.

[0010] Step 5: Update the contour continuity weight by writing back the boundary contour, and update the imaging quality weight by writing back the interface mask.

[0011] Step 6: Iterate through steps 1 to 5 until the convergence threshold is met or the upper limit of iterations is reached.

[0012] It should be noted that converting the cylindrical surface imaging of the borehole wall into an approximate planar coordinate expression reduces geometric distortion caused by wall eccentricity and viewing angle changes, making the position and shape of the interface more consistent in different depth frames, which facilitates subsequent stable segmentation and contour tracking. Defogging and water film correction weaken the brightness drift and false high gradients caused by residual water film from rinsing, cooling fogging, and local reflections, improving the discernibility of interlayer textures and grayscale differences, and reducing the probability of false edges, broken edges, and missed detections. The interface discrimination is improved from a single-point gradient to a local statistical comparison of the upper and lower walls, using the difference between the lithology near the interface and the background to establish a more robust criterion, reducing the interference of local strong gradient structures such as fractures and drill marks on interface localization. Logarithmic domain equalization is performed on the contrast difference between the near and far sides. This approach suppresses contrast distortion caused by extreme bright spots or low-contrast areas, making the driving force more stable and its dynamic range more controllable under strong interference and gradually changing lithology conditions, thereby reducing boundary jitter and drift. It transforms one-dimensional or local interface contrast information into pixel-level guiding information, providing explicit interface priors for the segmentation network, improving the ability to focus on narrow structures such as thin coal lines and weak interlayers, and reducing the probability of non-interface textures being missegmented as interfaces. It integrates driving heatmaps and multi-scale features into the deep learning segmentation network, while utilizing explicit driving priors and deep semantic texture information to take into account both large-scale lithological zoning and small-scale bedding details, improving the integrity of the interface mask and the accuracy of boundary positioning, and reducing missegmentation and omissions caused by relying solely on textures or gradients.

[0013] As a further aspect of the present invention, in step 2, the lithological contrast difference includes gray intensity statistical difference, gray fluctuation statistical difference, gradient energy statistical difference, and texture statistical feature difference. The gray intensity statistical difference is the difference in gray mean or gray median between the near-side and far-side zonal sampling areas. The gray fluctuation statistical difference is the difference in gray variance between the near-side and far-side zonal sampling areas. The gradient energy statistical difference is the difference in gradient amplitude between the near-side and far-side zonal sampling areas. The texture statistical feature difference is the difference in texture contrast, texture energy, texture entropy, or texture homogeneity between the near-side and far-side zonal sampling areas.

[0014] It should be noted that the interface discrimination is expanded from the common single grayscale threshold binarization or single gradient peak edge extraction in existing technologies to a multi-dimensional comparative description that simultaneously characterizes the differences in brightness level, brightness fluctuation, edge energy, and texture statistics in the near and far strip sampling areas on the same side of the candidate interface. This makes the interface discrimination no longer dependent on the accidental response of a certain type of feature when it is affected by water film reflection, fogging and whitening, or uneven lighting. When grayscale gradient causes threshold failure or fracture drill marks cause gradient false triggering, grayscale fluctuation and texture statistics can still provide complementary constraints, thereby reducing the probability of false edges, broken edges, and missed detections, reducing layer drift, and improving the stable identification ability of weak interfaces such as thin coal lines and soft interlayers.

[0015] As a further aspect of the present invention, in step 2, for each side of the candidate interface, the statistical difference of gray intensity, the statistical difference of gray fluctuation, the statistical difference of gradient energy, and the statistical difference of texture features are normalized with the same caliber to obtain four types of normalized contrast differences. Based on the four types of normalized contrast differences, pairwise interaction terms are constructed, including interaction terms between gray intensity statistical difference and gradient energy statistical difference, interaction terms between gray intensity statistical difference and texture statistical difference, interaction terms between gray fluctuation statistical difference and gradient energy statistical difference, and interaction terms between gray fluctuation statistical difference and texture statistical difference. The consistency measure of the four types of contrast differences is calculated based on the pairwise interaction terms, and the coupling weights corresponding to the four types of contrast differences are determined according to the consistency measure. The four types of normalized contrast differences are weighted and fused according to the coupling weights to obtain the comprehensive lithological contrast difference of the current side.

[0016] It should be noted that the multi-source lithological correlation features are first normalized to the same caliber to avoid single features dominating the discrimination due to differences in dimensions and dynamic range. Then, the co-occurrence relationships of intensity and gradient, intensity and texture, fluctuation and gradient, and fluctuation and texture are explicitly characterized through pairwise interaction. A consistency metric is used to determine whether each feature supports or conflicts with each other in the current pixel neighborhood. This adaptively generates coupling weights to increase the weight of reliable features and decrease the weight of conflicting features, making the comprehensive lithological correlation difference more consistent with the joint change law of the hanging wall and footwall of the real interface. This reduces the false edges, broken edges and stratigraphic drift caused by interference from fogging, water film reflection, gradual lithology and fracture drill marks, and makes the heat map formed by subsequent driving volume backlay more continuous and stable.

[0017] As a further aspect of the present invention, in step 2, the equivalent comparison driving quantity includes: a first-side equivalent driving component calculated based on the logarithmic mean difference between the near-side and far-side lithological comparison differences on one side of the candidate interface; a second-side equivalent driving component calculated based on the logarithmic mean difference between the near-side and far-side lithological comparison differences on the other side of the candidate interface; and a comprehensive equivalent comparison driving quantity obtained by weighting the first-side equivalent driving component and the second-side equivalent driving component; wherein when the near-side or far-side lithological comparison difference is lower than a preset lower limit, it is replaced by the preset lower limit; when the near-side or far-side lithological comparison difference is greater than a preset upper limit, it is replaced by the preset upper limit.

[0018] It should be noted that the candidate interface is logarithmically equivalently converted to near-far band contrast difference on both sides of the upper and lower plates and then weighted and converged. This avoids the bias and drift caused by the existing method of only taking one-sided threshold or single-point gradient. At the same time, the distortion of the conversion between low-contrast weak interface and high-contrast reflective peak is suppressed by replacing the lower limit of contrast difference and limiting the upper limit of contrast difference. This ensures that the driving force maintains a stable dynamic range and continuous distribution under the conditions of atomization, water film, uneven illumination and lithological gradation, and reduces false edge breakage and missed detection.

[0019] As a further aspect of the present invention, in step 2, the method for obtaining the driving heat map includes:

[0020] Step 21: Using the candidate interface boundary pixels as indexes, construct near-side and far-side zonal sampling areas on both sides of the candidate interface, and calculate the lithological contrast difference in each near-side and far-side zonal sampling area.

[0021] Step 22: The near-side lithological contrast difference and the far-side lithological contrast difference on each side of the candidate interface are converted into the logarithmic mean difference to obtain the current equivalent driving component. The equivalent driving components on both sides of the candidate interface are weighted and converged to obtain the comprehensive equivalent contrast driving quantity.

[0022] Step 23: Assign the comprehensive equivalent contrast driving value to the corresponding candidate interface boundary pixel to form a boundary pixel driving value set, and extend the boundary pixel driving value to a preset bandwidth along the candidate interface boundary normal direction to form a boundary band driving map.

[0023] Step 24: Perform spatial interpolation, smoothing and normalization on the boundary zone driving map to map the boundary zone driving map into a pixel-level driving heat map of the same size as the hole wall image.

[0024] It should be noted that the contrast difference is calculated by sampling within the near and far bands of the upper and lower plates using candidate boundary pixels as indices. After logarithmic domain conversion and bilateral weighted convergence, pixel-level driving values ​​are obtained. Then, the values ​​are extended along the normal, interpolated and smoothed, and normalized to form a full-image driven heatmap. This expands the interface evidence from discrete boundary points to continuous banded priors and aligns it with the image at the same scale. This reduces the edge breakage, false edges, and layer drift caused by the fragmented edge response that relies solely on thresholds or gradients, and improves the traceability of narrow interfaces such as weak coal lines and soft interlayers.

[0025] As a further aspect of the present invention, in step 1, the cylindrical surface unfolding, defogging, and water film correction adopt an adaptive processing method based on comprehensive lithological contrast difference, specifically including: generating a pixel-level contrast difference distribution map from the comprehensive lithological contrast difference obtained in the previous iteration; determining the update amount of the unfolding axis and unfolding radius of the cylindrical surface unfolding according to the contrast difference distribution map, and completing the iterative correction of the unfolding parameters; constructing a fogging intensity map and a water film disturbance map according to the contrast difference distribution map, and performing pixel-level defogging correction on the borehole wall image according to the fogging intensity map and pixel-level water film attenuation correction on the borehole wall image according to the water film disturbance map, and the correction coefficients of defogging correction and water film attenuation correction are spatially adaptively updated according to the contrast difference distribution map.

[0026] It should be noted that the previous round of comprehensive lithological contrast difference is transformed into a pixel-level contrast difference distribution map. This map is used to iteratively correct the cylinder development axis and radius, and simultaneously generate a fogging intensity map and a water film disturbance map. Spatially adaptive defogging and water film attenuation correction are applied to the borehole wall image. This avoids overcorrection or undercorrection caused by existing fixed development parameters and global defogging or uniform filtering under conditions of wall-attached distortion, local fogging, and reflective water film. The correction intensity is finely adjusted according to the distribution of local contrast anomalies and maintains the consistency of subsequent interface features.

[0027] As a further aspect of the present invention, in step 1, constructing a fog intensity map and a water film disturbance map based on the pixel-level contrast difference distribution map and performing pixel-level defogging correction and water film attenuation correction includes: performing scale normalization and smoothing processing on the pixel-level contrast difference distribution map to obtain a contrast difference baseline map; mapping the low-contrast regions of the contrast difference baseline map to fog candidate regions and calculating a fog intensity map, the fog intensity map being jointly determined by the low contrast degree, local brightness uniformity, and high-frequency texture energy attenuation; mapping the abnormally high-contrast regions of the contrast difference baseline map and their neighborhoods to water film candidate regions and calculating a water film disturbance map, the water film disturbance map being jointly determined by the high contrast peak degree, local specular highlight ratio, and gradient direction consistency; generating a pixel-level transmittance map based on the fog intensity map and performing pixel-level defogging correction on the hole wall image in conjunction with global background brightness estimation; generating a pixel-level attenuation coefficient map based on the water film disturbance map and performing pixel-level intensity attenuation and highlight suppression correction on the hole wall image.

[0028] It should be noted that, in response to the problems of existing hole wall images that only use global dehazing or simple filtering, which cannot distinguish between low contrast of fogging and abnormal high contrast of water film reflection, and easily misinterpret strong gradients such as cracks and drill marks as interference, leading to local overcorrection or undercorrection and thus destroying the texture consistency of the interface, a contrast difference baseline map is generated by contrast difference distribution map. The fogging intensity map is determined by low contrast degree, brightness uniformity and high frequency energy attenuation, and the water film disturbance map is determined by high contrast peak, specular high brightness ratio and gradient direction consistency. Based on this, a transmittance map and attenuation coefficient map are generated to perform pixel-level dehazing and water film attenuation correction on the hole wall image. This achieves partition recognition, partition correction and intensity adaptive update of the two types of degradation, fogging and water film, improving the interface separability and boundary stability after correction.

[0029] As a further aspect of the present invention, in step 1, the pixel values ​​of the contrast difference reference image are normalized and limited to the range of 1 to 1. The low contrast region is the region where the pixel value of the contrast difference reference image is ≤0.35, and the abnormally high contrast region is the region where the pixel value of the contrast difference reference image is ≥0.75 and the increment relative to the mean of its neighborhood is ≥0.15. The neighborhood is a 5×5 to 15×15 pixel window centered on the current pixel. The local brightness uniformity is characterized by the ratio of the brightness standard deviation to the brightness mean within the window and is limited to ≤0.12. The high-frequency texture energy attenuation is characterized by the ratio of high-frequency energy to mid-low-frequency energy within the window and is limited to ≤0.6. The proportion of local specular highlights is limited to the proportion of pixels with brightness ≥98th percentile of the overall image brightness within the window, which is between 0.05 and 0.35. The gradient direction consistency is characterized by the circular square error of the gradient direction within the window and is limited to ≤0.4.

[0030] It should be noted that by normalizing the contrast difference benchmark image and using explicit threshold windows and statistical constraints to limit the low-contrast fogging area, the abnormally high-contrast water film area and their criterion range, the identification of fogging and water film disturbance is transformed from empirical judgment to reproducible numerical judgment, reducing threshold drift and misclassification, and avoiding over-correction or under-correction of defogging and highlight suppression near the boundary, thereby stabilizing the extraction of pore wall texture and interface contrast features.

[0031] As a further aspect of the present invention, in step 4, the input to the deep learning segmentation network includes: the borehole wall image processed in step 1, the pixel-level driving heat map obtained in step 2, the imaging quality weight map and contour continuity weight map obtained in step 3 and mapped to the same size as the borehole wall image, as well as the fogging intensity map and water film disturbance map; and, a recommendation model based on prior information of the local coal mine is introduced into the deep learning segmentation network. The prior information of the local coal mine includes coal seam number, burial depth, dip angle, composite roof lithology sequence, key interface borehole depth labeling table, anchor The cable anchorage length and anchorage section location, and the distance between the roadway and the gate are considered. The recommended model outputs a priori map of the interface layer aligned with the borehole depth coordinates and a priori vector of the interface category based on the prior information of the local coal mine. The priori map of the interface layer is used as an additional input channel by splicing the borehole wall image and the driving heat map at the input end of the segmentation network encoder. The prior vector of the interface category is used in the feature bottleneck layer between the encoder and the decoder to perform scaling and bias modulation on the feature channels, and at the jump connection between the encoder and the decoder, it is used to perform weighted fusion on the decoded feature channels.

[0032] It should be noted that, based on the conventional segmentation network which only uses the borehole wall image as a single source input, a multi-channel prior is added, including a driving heat map, an imaging quality weight map, a contour continuity weight map, a fogging intensity map, and a water film disturbance map, which are used together with the borehole wall image as encoder input. At the same time, a recommendation model based on prior information such as local coal seam number, burial depth and dip angle, composite roof lithology sequence, key interface borehole depth labeling, anchor cable anchorage section location and distance from the gate is introduced. The output borehole depth-aligned stratigraphic prior map is spliced ​​at the encoder end, and the output category prior vector is channel-modulated at the bottleneck layer and weighted fusion is performed at the jump connection. This solves the problem that relying solely on image texture is prone to false edges, broken edges, stratigraphic drift, and boundary instability caused by confusion between the inner and outer interfaces of the anchorage zone under the interference of fogging water film reflection, lithological grayscale gradation, and fracture drill marks. This makes the interface segmentation constrained at both the spatial stratigraphic and category levels and more closely fits the borehole depth and geological structure, thereby improving the stability, continuity, and reproducibility of the extraction of potential structural surface boundaries.

[0033] Compared with the prior art, the technical effects of the present invention are as follows:

[0034] This invention constructs a joint statistical lithological contrast difference between near-side and far-side sampling areas on both sides of the candidate interface after cylindrical expansion and defogging water film correction of composite top plate hole wall images. Then, it generates a pixel-level equivalent contrast driving quantity by logarithmic domain equivalent conversion to form a driving heat map. At the same time, it introduces imaging quality weight and contour continuity weight to constrain the driving and segmentation results and iterates and updates the results by contour and mask back-writing. This enables the depth segmentation network to stably output the interface mask and its boundary contour of the potential delamination structure surface under multi-source interference, improves the robustness and reproducibility of interface identification, and supports the automatic discrimination of delamination critical value dynamically updated according to the gate position and anchoring zone difference. Attached Figure Description

[0035] Figure 1 This is a flowchart of the method of the present invention;

[0036] Figure 2 This is an image of the original hole wall in this invention;

[0037] Figure 3 This is a diagram showing the unfolded cylindrical surface of the present invention.

[0038] Figure 4 This is an image after defogging and water film correction according to the present invention;

[0039] Figure 5 This invention drives the thermal map;

[0040] Figure 6 This is the interface mask of the present invention;

[0041] Figure 7 This is a diagram showing the result of contour stabilization in this invention. Detailed Implementation

[0042] The technical solutions in the embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0043] like Figure 1 As shown, the present invention proposes a deep learning-based method for extracting coal seam delamination boundaries, comprising:

[0044] Step 1: Obtain the borehole wall image and perform cylindrical surface unfolding, defogging, and water film correction.

[0045] Step 2: Calculate the lithological contrast difference on both sides of the candidate interface. The near side is the strip sampling area close to the boundary on one side of the candidate interface, and the far side is the strip sampling area on the same side along the boundary normal direction at a preset distance from the near side and far from the boundary. The equivalent contrast driving amount is converted into the logarithmic mean difference. Then, the equivalent driving amount obtained by each candidate boundary pixel is laid back into the image to form a driving heat map to guide the segmentation of the potential delamination structure surface boundary.

[0046] Step 3: Calculate the image quality weight and the contour continuity weight;

[0047] Step 4: After fusing the multi-scale features of the driving heat map and the hole wall image, input the data into the deep learning segmentation network, output the interface mask, and perform contour extraction on the interface mask to obtain the boundary contour of the potential delamination structure surface.

[0048] Step 5: Update the contour continuity weight by writing back the boundary contour, and update the imaging quality weight by writing back the interface mask.

[0049] Step 6: Iterate through steps 1 to 5 until the convergence threshold is met or the upper limit of iterations is reached.

[0050] In practice, in typical monitoring sections underground in coal mines, anchor cable drilling rigs are used to drill holes in the middle of the roadway roof. The hole depth is determined according to the design value, and the hole depth and lithological changes are recorded synchronously by a rock stratum detection recorder during drilling. After drilling, the hole wall is cleaned and drained to cool it using the drilling rig's water supply or a special flushing device. Then, the drilling camera and the threaded mounting rod are combined to form a probe assembly, equipped with a depth encoder or depth measuring wheel to align the image frame with the hole depth. The hole wall video is acquired by raising or lowering the machine at a constant speed. The acquired video is connected to an intrinsically safe industrial control computer or an explosion-proof edge computing box underground, and the cylindrical surface is unfolded, defogging and water film correction are performed to generate a contrast difference distribution map and a driving heat map. A deep learning segmentation network is deployed on the edge computing box or ground server, and the interface mask and boundary contour are output and iteratively updated by combining imaging quality and contour continuity weights. Finally, the identified key interface hole depth positions are sent to the on-site construction personnel for anchoring the multi-point displacement gauge fixing points on the upper and lower plates of the interface and for subsequent delamination monitoring.

[0051] It should be noted that in step 2, the lithological contrast difference includes the statistical difference in gray intensity, the statistical difference in gray fluctuation, the statistical difference in gradient energy, and the statistical difference in texture features. The statistical difference in gray intensity is the difference in the mean gray value or the difference in the median gray value between the near-side and far-side zonal sampling areas. The statistical difference in gray fluctuation is the difference in the variance of gray values ​​between the near-side and far-side zonal sampling areas. The statistical difference in gradient energy is the difference in the mean gradient amplitude between the near-side and far-side zonal sampling areas. The statistical difference in texture features is the difference in texture contrast, texture energy, texture entropy, or texture homogeneity between the near-side and far-side zonal sampling areas.

[0052] In practice, an intrinsically safe borehole camera, under the constraint of the mounting rod and centering wheel, uniformly scans and acquires video along the borehole wall. The depth encoder synchronously outputs a borehole depth trigger signal. The video and borehole depth data are sent to the downhole edge computing box via an explosion-proof gateway. After unfolding each frame and completing defogging and water film correction, the edge computing box calls the candidate interface detection module to generate a boundary pixel index. According to the index, it automatically generates near-side and far-side strip-shaped sampling areas along the boundary normal direction in the image. Within the sampling area, the gray-level statistics unit calculates the gray-level mean or median and calculates the difference, the fluctuation statistics unit calculates the gray-level variance and calculates the difference, the gradient calculation unit calculates the difference of the gradient amplitude mean, and the texture statistics unit extracts the texture contrast, energy, entropy, or homogeneity and calculates the difference, obtaining four types of lithological contrast differences and writing them into the cache for subsequent driving quantity conversion and heat map generation modules to call.

[0053] It should be noted that in step 2, for each side of the candidate interface, the statistical difference of gray intensity, the statistical difference of gray fluctuation, the statistical difference of gradient energy, and the statistical difference of texture features are normalized with the same caliber to obtain four types of normalized contrast differences. Based on the four types of normalized contrast differences, pairwise interaction terms are constructed, including interaction terms between gray intensity statistical difference and gradient energy statistical difference, gray intensity statistical difference and texture statistical difference, gray fluctuation statistical difference and gradient energy statistical difference, and gray fluctuation statistical difference and texture statistical difference. The consistency measure of the four types of contrast differences is calculated based on the pairwise interaction terms, and the coupling weights corresponding to the four types of contrast differences are determined according to the consistency measure. The four types of normalized contrast differences are weighted and fused according to the coupling weights to obtain the comprehensive lithological contrast difference of the current side.

[0054] In specific implementation, after the downhole edge computing box completes the calculation of the four types of lithological contrast differences in the near and far side strip sampling areas, it calls the normalization unit to normalize the gray intensity statistical difference, gray fluctuation statistical difference, gradient energy statistical difference, and texture statistical feature difference with the same caliber according to the preset range or sliding window quantile, thus obtaining the four types of normalized contrast differences; the interaction construction unit generates pairwise interaction terms for the normalized contrast differences pixel by pixel, corresponding to the interaction feature maps of intensity and gradient, intensity and texture, fluctuation and gradient, and fluctuation and texture, respectively; the consistency measurement unit calculates the consistency index for each interaction feature map in the neighborhood of the candidate interface and forms a consistency measurement value; the weight determination unit generates coupling weights for the four types of normalized contrast differences based on the consistency measurement and applies normalization constraints; the fusion unit calculates the weighted sum of the four types of normalized contrast differences according to the coupling weights to obtain the current side comprehensive lithological contrast difference, and writes it into the driving quantity conversion and heat map generation module for subsequent boundary segmentation guidance.

[0055] It should be noted that in step 2, the equivalent comparison driving quantity includes: the first equivalent driving component calculated based on the logarithmic mean difference between the near-side and far-side lithological comparison differences on one side of the candidate interface; the second equivalent driving component calculated based on the logarithmic mean difference between the near-side and far-side lithological comparison differences on the other side of the candidate interface; and the comprehensive equivalent comparison driving quantity obtained by weighting the first equivalent driving component and the second equivalent driving component. When the near-side or far-side lithological comparison difference is lower than a preset lower limit, it is replaced by the preset lower limit; when the near-side or far-side lithological comparison difference is greater than a preset upper limit, it is replaced by the preset upper limit.

[0056] In specific implementation, after obtaining the near-side lithological contrast difference and the far-side lithological contrast difference on both sides of the candidate interface, the downhole edge computing box first applies threshold constraints to each contrast difference by the limiting unit: those below the preset lower limit are set to the lower limit value, and those above the preset upper limit are set to the upper limit value. Then, the conversion unit performs logarithmic domain equivalent conversion on the near-side contrast difference and the far-side contrast difference on each side of the candidate interface to obtain the first-side equivalent driving component and the second-side equivalent driving component. The weighted aggregation unit performs weighted fusion on the equivalent driving components on both sides according to the preset weight coefficient or the weight output by the consistency metric, outputs the comprehensive equivalent contrast driving quantity, and binds the comprehensive equivalent contrast driving quantity and the boundary pixel index to the cache for subsequent backlay and heatmap generation calls.

[0057] It should be noted that, in step 2, the method for obtaining the driving heatmap includes:

[0058] Step 21: Using the candidate interface boundary pixels as indexes, construct near-side and far-side zonal sampling areas on both sides of the candidate interface, and calculate the lithological contrast difference in each near-side and far-side zonal sampling area.

[0059] Step 22: The near-side lithological contrast difference and the far-side lithological contrast difference on each side of the candidate interface are converted into the logarithmic mean difference to obtain the current equivalent driving component. The equivalent driving components on both sides of the candidate interface are weighted and converged to obtain the comprehensive equivalent contrast driving quantity.

[0060] Step 23: Assign the comprehensive equivalent contrast driving value to the corresponding candidate interface boundary pixel to form a boundary pixel driving value set, and extend the boundary pixel driving value to a preset bandwidth along the candidate interface boundary normal direction to form a boundary band driving map.

[0061] Step 24: Perform spatial interpolation, smoothing and normalization on the boundary zone driving map to map the boundary zone driving map into a pixel-level driving heat map of the same size as the hole wall image.

[0062] In specific implementation, the downhole edge computing box outputs a set of boundary pixel indices from the candidate interface detection module on the expanded and corrected borehole wall image. The sampling construction unit generates near-side and far-side strip sampling areas on the hanging wall and footwall respectively, centered on each boundary pixel and along the boundary normal direction. The statistical calculation unit calculates the lithological contrast difference in each sampling area. The conversion unit performs logarithmic domain equivalent conversion on the near-far contrast difference on each side to obtain the equivalent driving component on the current side. The weighting unit aggregates the equivalent driving components on both sides into a comprehensive equivalent contrast driving quantity according to a preset or consistent weight. The assignment and expansion unit writes the comprehensive driving quantity into the corresponding boundary pixel to form a driving value set and expands it along the normal direction to a preset bandwidth to generate a boundary zone driving map. Subsequently, the interpolation and smoothing unit performs spatial interpolation completion, low-pass smoothing, and normalization on the boundary zone driving map, mapping it into a pixel-level driving heat map of the same size as the borehole wall image and outputting it to the depth segmentation network as a guide channel.

[0063] It should be noted that in step 1, the cylindrical surface unfolding, defogging, and water film correction adopt an adaptive processing method based on the comprehensive lithological contrast difference. Specifically, this includes: generating a pixel-level contrast difference distribution map from the comprehensive lithological contrast difference obtained in the previous iteration; determining the update amount of the unfolding axis and unfolding radius of the cylindrical surface unfolding based on the contrast difference distribution map, and completing the iterative correction of the unfolding parameters; constructing a fogging intensity map and a water film disturbance map based on the contrast difference distribution map, and performing pixel-level defogging correction on the borehole wall image according to the fogging intensity map and pixel-level water film attenuation correction on the borehole wall image according to the water film disturbance map. The correction coefficients of the defogging correction and the water film attenuation correction are spatially adaptively updated according to the contrast difference distribution map.

[0064] In practice, after completing one round of interface identification, the comprehensive lithological contrast difference output by the fusion module is mapped to a contrast difference distribution map of the same size as the borehole wall unfolded map and written into the shared cache. The parameter update unit uses the high consistency interface band in the contrast difference distribution map as a constraint to iteratively correct the axial offset and equivalent radius update of the cylindrical surface unfolded and writes back the unfolded parameters. Subsequently, the degradation modeling unit generates a fogging intensity map and a water film disturbance map based on the contrast difference distribution map, and calculates the pixel-level transmittance and pixel-level attenuation coefficient respectively. The correction execution unit defogs the borehole wall image pixel by pixel according to the transmittance and suppresses the water film brightness and intensity attenuation pixel by pixel according to the attenuation coefficient. Each correction coefficient is adaptively adjusted according to the spatial change of the contrast difference distribution map and then output to the next round of step 2 processing.

[0065] It should be noted that in step 1, constructing a fog intensity map and a water film perturbation map based on the pixel-level contrast difference distribution map and performing pixel-level defogging correction and water film attenuation correction includes: performing scale normalization and smoothing on the pixel-level contrast difference distribution map to obtain a contrast difference baseline map; mapping the low-contrast regions of the contrast difference baseline map to fog candidate regions and calculating the fog intensity map, which is determined by the low contrast degree, local brightness uniformity, and high-frequency texture energy attenuation; mapping the abnormally high-contrast regions of the contrast difference baseline map and their neighborhoods to water film candidate regions and calculating the water film perturbation map, which is determined by the high contrast peak degree, local specular highlight ratio, and gradient direction consistency; generating a pixel-level transmittance map based on the fog intensity map and performing pixel-level defogging correction on the hole wall image in conjunction with global background brightness estimation; generating a pixel-level attenuation coefficient map based on the water film perturbation map and performing pixel-level intensity attenuation and highlight suppression correction on the hole wall image.

[0066] In specific implementation, the downhole edge computing box reads the pixel-level contrast difference distribution map from the shared cache. The normalization and smoothing unit normalizes it according to a preset range and uses a sliding window to smooth it to obtain a contrast difference baseline map. The fogging determination unit filters out low-contrast pixels in the baseline map and calculates the fogging intensity map by combining the ratio of the standard deviation and mean of the brightness in the same window, as well as the ratio of high-frequency texture energy to mid-low-frequency energy. The water film determination unit filters out abnormally high-contrast peak pixels in the baseline map and calculates the water film disturbance map by statistically analyzing the proportion of bright pixels and the circular square error of the gradient direction in their neighborhood. The transmittance generation unit maps the fogging intensity map to pixel-level transmittance and the background estimation unit obtains the global background brightness. The defogging execution unit restores the contrast of the borehole wall image pixel by pixel based on this. The attenuation coefficient generation unit maps the water film disturbance map to pixel-level attenuation coefficient. The high brightness suppression unit suppresses local specular high brightness according to the attenuation coefficient and compensates for the intensity attenuation. The corrected image is then output to enter the subsequent interface contrast difference calculation and heat map generation process.

[0067] It should be noted that in step 1, the pixel values ​​of the contrast difference reference image are normalized and limited to the range of 1 to 1. The low contrast region is the region where the pixel value of the contrast difference reference image is ≤0.35, and the abnormally high contrast region is the region where the pixel value of the contrast difference reference image is ≥0.75 and the increment relative to its neighborhood mean is ≥0.15. The neighborhood is a 5×5 to 15×15 pixel window centered on the current pixel. The local brightness uniformity is represented by the ratio of the brightness standard deviation to the brightness mean within the window and is limited to ≤0.12. The high-frequency texture energy attenuation is represented by the ratio of high-frequency energy to mid-low frequency energy within the window and is limited to ≤0.6. The proportion of local specular highlights is limited to the proportion of pixels with brightness ≥98th percentile of the overall image brightness within the window, which is between 0.05 and 0.35. The gradient direction consistency is represented by the circular square error of the gradient direction within the window and is limited to ≤0.4.

[0068] In specific implementation, the downhole edge computing box normalizes the reference image pixel by pixel and limits it to the range of zero to one. In the sliding window statistics module, the neighborhood mean is calculated for a neighborhood of 5×5 to 15×15 pixels. When the reference image pixel value is ≤0.35, it is judged as a candidate pixel for fogging. The ratio of the standard deviation of brightness to the mean within the window is further calculated and is ≤0.12, and the ratio of high frequency energy to medium and low frequency energy is ≤0.6 to confirm the degree of fogging. When the reference image pixel value is ≥0.75 and the increase in relative neighborhood mean is ≥0.15, it is judged as a candidate pixel for water film. The proportion of pixels with brightness not lower than the 98th percentile of the whole image within the window is calculated and is between 0.05 and 0.35. At the same time, the gradient direction circular square error is calculated and is ≤0.4 to determine the water film disturbance intensity. The judgment results are used to generate fogging intensity map and water film disturbance map.

[0069] It should be noted that in step 4, the input to the deep learning segmentation network includes: the borehole wall image processed in step 1, the pixel-level driving heat map obtained in step 2, the imaging quality weight map and contour continuity weight map obtained in step 3 and mapped to the same size as the borehole wall image, as well as the fogging intensity map and water film disturbance map; and a recommendation model based on prior information of the local coal mine is introduced into the deep learning segmentation network. The prior information of the local coal mine includes coal seam number, burial depth, dip angle, composite roof lithology sequence, key interface borehole depth labeling table, and anchor cable anchor. The model considers the anchorage length, anchorage section location, and distance from the gate in the roadway. Based on prior information from the local coal mine, the recommended model outputs a priori map of the interface layer aligned with the borehole depth coordinates, as well as a priori vector of the interface category. The priori map of the interface layer is used as an additional input channel by splicing the borehole wall image and the driving heat map at the encoder input of the segmentation network. The prior vector of the interface category is used in the feature bottleneck layer between the encoder and the decoder to perform scaling and bias modulation on the feature channels, and at the jump connection between the encoder and the decoder to perform weighted fusion on the decoded feature channels.

[0070] In practice, the underground edge computing box stacks the borehole wall unfolding and correction map output in step 1, the pixel-level driven heat map generated in step 2, the imaging quality weight map and contour continuity weight map generated and resampled to the same size in step 3, as well as the fogging intensity map and water film disturbance map, into a network input tensor and sends it to the deep learning segmentation network deployed on the edge computing box or ground server. At the same time, prior information is imported from the mine geological database and construction parameter ledger, including coal seam number, burial depth, dip angle, composite roof lithology sequence, key sub-coal seams, etc. The interface hole depth annotation, anchor cable anchorage length and anchorage segment position, and distance from the gate are used by the recommendation model to generate a layer prior map and a category prior vector aligned with the hole depth coordinates. The layer prior map is concatenated with the above multi-channel input tensor at the encoder input end as an additional channel and enters the first convolutional layer. The category prior vector modulates the encoded features in the feature bottleneck layer through channel scaling and biasing, and is used as a fusion weight at each level of skip connection to weight the decoded features. Finally, the interface mask is output and the contour extraction unit generates the boundary contour of the potential structural surface of the delamination.

[0071] To clearly illustrate the method proposed in this invention, the original hole wall image is used. Figure 2 To process, Figure 2 The original borehole wall frame acquired by the borehole camera downhole is affected by residual water film, local reflection, uneven lighting, and cylindrical surface distortion caused by slight adhesion to the wall. The borehole wall texture shows a coexistence of gray-scale gradients and bright spots. The interfaces between fractures, drill marks, and real lithology are visually strong gradient structures. If the traditional method directly uses fixed threshold binarization or one-sided gradient operator to extract the edges, it is easy to mistake the edge of the reflective water film for the interface or cut off the real interface, which will cause stratigraphic drift and missed detection. This is the interference source that this invention focuses on suppressing in steps 1 and 2. Figure 3 The core of this method is to map the annular hole wall texture to planar coordinates after unfolding the cylindrical surface of the hole wall. This allows the circumferential information near the same depth to be continuously traced on the image and provides a unified geometric reference for near-far band sampling of the candidate interface boundary pixel index-normal direction. At the same time, after unfolding, the depth information of the hole depth encoder can be aligned with the horizontal and vertical coordinates of the image, reducing the impact of geometric stretching caused by wall eccentricity on the boundary position determination. This also prevents subsequent contrast difference statistics from being interfered with by circumferential projection aliasing, thereby improving the localizability and reproducibility of candidate boundaries from the source. Figure 4Corresponding to the defogging and water film correction output in step 1, pixel-level adaptive correction is used to correct the overall low contrast caused by fogging and the local abnormally high contrast caused by the specular reflection of the water film. After correction, the texture details of the borehole wall are more balanced, the differences in sandstone and mudstone textures that were originally smoothed out by fogging are restored, the bright peaks of the water film are suppressed so as not to generate false edges, and the edge sharpness of the fractures and drill marks is maintained but the brightness is abnormally reduced. This provides a more stable data distribution for the intensity, fluctuation, gradient and texture statistics of the near and far strip sampling areas in step 2, making the contrast difference closer to the real lithological differences. Figure 5 The pixel-level driven heatmap obtained in step 2 can be understood as the equivalent contrast driving quantity after logarithmic domain conversion of the near-far contrast difference on both sides of the candidate interface back to the guiding prior of image formation. Bright responses are concentrated in the abrupt change zone of hole wall texture and obvious circumferential structure, while the responses to the fog flat area and reflective peak area are weakened by amplitude limiting and weight constraints, so that the interface evidence changes from a single-point gradient to a continuous band-shaped energy distribution, reducing the problem of traditional operators starting edges everywhere on noise points, and providing an additional input channel for deep learning segmentation network, making it easier for the network to focus along the real structural surface rather than being pulled by the cracks and drill marks. Figure 6 The diagram shows the interface mask output in step 4. The mask separates the region that may belong to the boundary of the potential delamination structure from the background. During training and inference, it simultaneously integrates multi-scale features of the borehole wall image with driving heatmap, imaging quality weight, and contour continuity weight, so that the network reduces confidence in low-quality imaging areas and enhances coherence in contour continuous areas. Compared with threshold segmentation or unilateral gradient edges, the mask has stronger connectivity and regional consistency, and can stably extract the narrow interfaces of weak coal lines and soft interlayers from the gray-scale gradient background, providing structured input for subsequent contour extraction. Figure 7 This reflects the stable result of the mask after connectivity constraints and morphological correction, corresponding to the iterative convergence driven by updating the contour continuity weight by writing back the boundary contour and updating the imaging quality weight by writing back the interface mask in step 5: by suppressing scattered small pieces and broken fragments, the continuous band response of the main structural surface is strengthened, so that the boundary used for contour extraction has fewer virtual edges and holes; when the contour is written back to the hole depth coordinates and used for multi-point displacement meter fixed point anchorage layer position verification, the probability of fixed point mismatch and displacement assignment error can be reduced, thus making the discrimination of delamination amount and delamination critical value more stable and reproducible.

[0072] After aligning the boundary contour of the potential delamination structural surface extracted by this invention with the borehole depth coordinates and lithological columnar section, the key interfaces of sandstone, mudstone, carbonaceous mudstone, coal seam, and weak interlayers are quickly located at typical cross-sections of the displacement gate. Based on this, the anchorage layer position of the multi-point displacement gauge fixed point is checked and corrected, so that the displacement difference between adjacent fixed points can be stably attributed to the same structural surface. Furthermore, the displacement response difference inside and outside the anchorage zone is distinguished by combining the anchorage length and anchorage section position. Comparable delamination amount-borehole depth curves and critical criterion inputs are formed at different displacement gate positions, thereby reducing the miscalculation of delamination amount and misjudgment of critical value caused by interface omission, drift, or false edge. This makes the zoning, dynamic, and reproducible analysis of the critical value of composite roof delamination more reliable, supporting the adjustment of support parameters and the updating of early warning thresholds.

[0073] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0074] Finally: The above description is only a specific embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for extracting coal seam delamination boundaries based on deep learning, characterized in that, include: Step 1: Obtain the borehole wall image and perform cylindrical surface unfolding, defogging, and water film correction. Step 2: Calculate the lithological contrast difference on both sides of the candidate interface. The near side is the strip sampling area close to the boundary on one side of the candidate interface, and the far side is the strip sampling area on the same side along the boundary normal direction at a preset distance from the near side and far from the boundary. The equivalent contrast driving amount is converted into the logarithmic mean difference. Then, the equivalent driving amount obtained by each candidate boundary pixel is laid back into the image to form a driving heat map to guide the segmentation of the potential delamination structure surface boundary. Step 3: Calculate the image quality weight and the contour continuity weight; Step 4: After fusing the multi-scale features of the driving heat map and the hole wall image, input the data into the deep learning segmentation network, output the interface mask, and perform contour extraction on the interface mask to obtain the boundary contour of the potential delamination structure surface. Step 5: Update the contour continuity weight by writing back the boundary contour, and update the imaging quality weight by writing back the interface mask. Step 6: Iterate through steps 1 to 5 until the convergence threshold is met or the upper limit of iterations is reached.

2. The method for extracting coal seam delamination boundaries based on deep learning according to claim 1, characterized in that, In step 2, the lithological contrast difference includes the statistical difference in gray intensity, the statistical difference in gray fluctuation, the statistical difference in gradient energy, and the statistical difference in texture features. The statistical difference in gray intensity is the difference in the mean gray value or the difference in the median gray value between the near-side and far-side zonal sampling areas. The statistical difference in gray fluctuation is the difference in the variance of gray values ​​between the near-side and far-side zonal sampling areas. The statistical difference in gradient energy is the difference in the mean gradient amplitude between the near-side and far-side zonal sampling areas. The statistical difference in texture features is the difference in texture contrast, texture energy, texture entropy, or texture homogeneity between the near-side and far-side zonal sampling areas.

3. The method for extracting coal seam delamination boundaries based on deep learning according to claim 2, characterized in that, In step 2, for each side of the candidate interface, the statistical difference of gray intensity, the statistical difference of gray fluctuation, the statistical difference of gradient energy, and the statistical difference of texture features are normalized with the same caliber to obtain four types of normalized contrast differences. Based on the four types of normalized contrast differences, pairwise interaction terms are constructed. The pairwise interaction terms include the interaction terms between the statistical difference of gray intensity and the statistical difference of gradient energy, the interaction terms between the statistical difference of gray intensity and the statistical difference of texture features, the interaction terms between the statistical difference of gray fluctuation and the statistical difference of gradient energy, and the interaction terms between the statistical difference of gray fluctuation and the statistical difference of texture features. The consistency measure of the four types of comparison differences is calculated based on the pairwise interaction terms, and the coupling weights corresponding to the four types of comparison differences are determined according to the consistency measure. The four types of normalized comparison differences are weighted and fused according to the coupling weights to obtain the comprehensive lithological comparison difference of the current side.

4. The method for extracting coal seam delamination boundaries based on deep learning according to claim 2, characterized in that, In step 2, the equivalent comparison driving quantity includes: a first-side equivalent driving component calculated based on the logarithmic mean difference between the near-side and far-side lithological comparison differences on one side of the candidate interface; a second-side equivalent driving component calculated based on the logarithmic mean difference between the near-side and far-side lithological comparison differences on the other side of the candidate interface; and a comprehensive equivalent comparison driving quantity obtained by weighting the first-side and second-side equivalent driving components. When the near-side or far-side lithological comparison difference is lower than a preset lower limit, it is replaced by the preset lower limit; when the near-side or far-side lithological comparison difference is greater than a preset upper limit, it is replaced by the preset upper limit.

5. The method for extracting coal seam delamination boundaries based on deep learning according to claim 4, characterized in that, In step 2, the method for obtaining the driving heatmap includes: Step 21: Using the candidate interface boundary pixels as indexes, construct near-side and far-side zonal sampling areas on both sides of the candidate interface, and calculate the lithological contrast difference in each near-side and far-side zonal sampling area. Step 22: The near-side lithological contrast difference and the far-side lithological contrast difference on each side of the candidate interface are converted into the logarithmic mean difference to obtain the current equivalent driving component. The equivalent driving components on both sides of the candidate interface are weighted and converged to obtain the comprehensive equivalent contrast driving quantity. Step 23: Assign the comprehensive equivalent contrast driving value to the corresponding candidate interface boundary pixel to form a boundary pixel driving value set, and extend the boundary pixel driving value to a preset bandwidth along the candidate interface boundary normal direction to form a boundary band driving map. Step 24: Perform spatial interpolation, smoothing and normalization on the boundary zone driving map to map the boundary zone driving map into a pixel-level driving heat map of the same size as the hole wall image.

6. The method for extracting coal seam delamination boundaries based on deep learning according to claim 3, characterized in that, In step 1, the cylindrical surface unfolding, defogging, and water film correction adopt an adaptive processing method based on the comprehensive lithological contrast difference. Specifically, this includes: generating a pixel-level contrast difference distribution map from the comprehensive lithological contrast difference obtained in the previous iteration; determining the update amount of the unfolding axis and unfolding radius of the cylindrical surface unfolding based on the contrast difference distribution map, and completing the iterative correction of the unfolding parameters; constructing a fogging intensity map and a water film disturbance map based on the contrast difference distribution map, and performing pixel-level defogging correction on the borehole wall image according to the fogging intensity map and pixel-level water film attenuation correction on the borehole wall image according to the water film disturbance map, and the correction coefficients of the defogging correction and water film attenuation correction are spatially adaptively updated according to the contrast difference distribution map.

7. The method for extracting coal seam delamination boundaries based on deep learning according to claim 6, characterized in that, In step 1, constructing a fog intensity map and a water film perturbation map based on the pixel-level contrast difference distribution map and performing pixel-level defogging correction and water film attenuation correction includes: performing scale normalization and smoothing on the pixel-level contrast difference distribution map to obtain a contrast difference baseline map; mapping the low-contrast regions of the contrast difference baseline map to fog candidate regions and calculating the fog intensity map, which is determined by the low contrast degree, local brightness uniformity, and high-frequency texture energy attenuation; mapping the abnormally high-contrast regions of the contrast difference baseline map and their neighborhoods to water film candidate regions and calculating the water film perturbation map, which is determined by the high contrast peak degree, local specular highlight ratio, and gradient direction consistency; generating a pixel-level transmittance map based on the fog intensity map and performing pixel-level defogging correction on the hole wall image in conjunction with global background brightness estimation; generating a pixel-level attenuation coefficient map based on the water film perturbation map and performing pixel-level intensity attenuation and highlight suppression correction on the hole wall image.

8. The method for extracting coal seam delamination boundaries based on deep learning according to claim 7, characterized in that, In step 1, the pixel values ​​of the contrast difference reference image are normalized and limited to the range of 1 to 1. The low contrast region is the region where the pixel value of the contrast difference reference image is ≤0.35, and the abnormally high contrast region is the region where the pixel value of the contrast difference reference image is ≥0.75 and the increment relative to the mean of its neighborhood is ≥0.

15. The neighborhood is a 5×5 to 15×15 pixel window centered on the current pixel. The local brightness uniformity is represented by the ratio of the brightness standard deviation to the brightness mean within the window and is limited to ≤0.

12. The high-frequency texture energy attenuation is represented by the ratio of high-frequency energy to mid-low frequency energy within the window and is limited to ≤0.

6. The proportion of local specular highlights is limited to the proportion of pixels with brightness ≥98th percentile of the overall image brightness within the window, which is between 0.05 and 0.

35. The gradient direction consistency is represented by the circular square error of the gradient direction within the window and is limited to ≤0.

4.

9. The method for extracting coal seam delamination boundaries based on deep learning according to claim 1, characterized in that, In step 4, the inputs to the deep learning segmentation network include: the borehole wall image processed in step 1, the pixel-level driving heat map obtained in step 2, the imaging quality weight map and contour continuity weight map obtained in step 3 and mapped to the same size as the borehole wall image, as well as the fogging intensity map and water film disturbance map. Furthermore, a recommendation model based on prior information from the local coal mine is introduced into the deep learning segmentation network. The prior information from the local coal mine includes coal seam number, burial depth, dip angle, composite roof lithology sequence, key interface borehole depth labeling table, anchor cable anchorage length and anchorage section position, and roadway distance from the gate. The recommendation model outputs an interface layer prior map aligned with the borehole depth coordinates and an interface category prior vector based on the prior information from the local coal mine. The interface layer prior map is concatenated with the borehole wall image and driving heat map at the encoder input of the segmentation network as an additional input channel. The interface category prior vector is used in the feature bottleneck layer between the encoder and decoder to perform scaling and bias modulation on the feature channels, and at the jump connection between the encoder and decoder to perform weighted fusion on the decoded feature channels.

Images