CN-121982331-A - Deep learning-based coal bed separation layer boundary extraction method

CN121982331ACN 121982331 ACN121982331 ACN 121982331ACN-121982331-A

Abstract

The invention relates to the technical field of coal seam image processing, and particularly discloses a coal seam separation layer boundary extraction method based on deep learning, which is used for solving the problems that boundary segmentation and edge extraction are easy to generate false edges, fracture and horizon drift caused by water film fogging, reflection distortion and texture gradual change interference of a composite roof roadway drilling hole wall image, and comprises hole wall image pretreatment and cylindrical surface expansion correction, near-far contrast difference calculation and equivalent driving quantity thermodynamic diagram generation, imaging quality and contour continuity weight construction, depth segmentation and boundary contour extraction of multi-source feature fusion, weight self-updating and closed loop iteration convergence control of contour and mask write-back; according to the invention, the boundary contour of the potential structural surface of the separation layer is stably extracted by combining the near-far side contrast driving thermodynamic diagram, imaging quality and contour continuity weight constraint and by combining the depth segmentation closed loop iteration, so that the false edge, fracture and horizon drift caused by a foggy water film and crack drilling mark are reduced.

Inventors

TANG DIANGUI
SHI CHANGKUN
ZHU FENGLONG
LI HUIPING
GAO DONGSHAN
LIANG JUNWEI
XU WEICHAO
ZHU ZHIHENG

Assignees

崇信县百贯沟煤业有限公司
中国矿业大学

Dates

Publication Date: 20260505
Application Date: 20260122

Claims (9)

1. The method for extracting the coal bed separation layer boundary based on deep learning is characterized by comprising the following steps of: step 1, acquiring a borehole wall image, and executing cylindrical surface expansion, defogging and water film correction; Step 2, calculating near-side lithology contrast difference and far-side lithology contrast difference at two sides of a candidate interface, wherein the near side is a strip-shaped sampling area of which one side is clung to a boundary, the far side is a strip-shaped sampling area of which the same side is separated from the near side by a preset distance along the normal direction of the boundary and away from the boundary, converting the strip-shaped sampling area into equivalent contrast driving quantity in a logarithmic average difference mode, and paving the equivalent driving quantity obtained by pixels of each section of candidate boundary back into an image to form a driving thermodynamic diagram so as to guide boundary segmentation of a potential structural surface of a separation layer; step3, calculating an imaging quality weight and a contour continuity weight; step 4, inputting the fused driving thermodynamic diagram and the multi-scale features of the hole wall image into a deep learning segmentation network, outputting an interface mask, and performing contour extraction on the interface mask to obtain a boundary contour of a potential structural plane of the separation layer; step 5, updating the contour continuity weight by writing back the boundary contour, and updating the imaging quality weight by writing back the interface mask; and 6, iteratively repeating the steps 1 to 5 until the convergence threshold is met or the iteration upper limit is reached.
2. The method according to claim 1, wherein in the step 2, the lithology contrast difference comprises a gray level intensity statistical difference, a gray level fluctuation statistical difference, a gradient energy statistical difference and a texture statistical feature difference, wherein the gray level intensity statistical difference is a gray level mean difference or a gray level median difference between the near side strip-shaped sampling region and the far side strip-shaped sampling region, the gray level fluctuation statistical difference is a gray level variance difference between the near side strip-shaped sampling region and the far side strip-shaped sampling region, the gradient energy statistical difference is a gradient amplitude mean difference between the near side strip-shaped sampling region and the far side strip-shaped sampling region, and the texture statistical feature difference is a texture contrast difference, a texture energy difference, a texture entropy difference or a texture homogeneity difference between the near side strip-shaped sampling region and the far side strip-shaped sampling region.
3. The method for extracting the coal seam separation layer boundary based on deep learning according to claim 2 is characterized in that in step 2, gray level intensity statistical differences, gray level fluctuation statistical differences, gradient energy statistical differences and texture statistical characteristic differences are respectively subjected to same-caliber normalization aiming at each side of a candidate interface to obtain four types of normalized contrast differences, two-by-two interaction items are constructed based on the four types of normalized contrast differences, each two interaction items comprise interaction items of gray level intensity statistical differences and gradient energy statistical differences, interaction items of gray level intensity statistical differences and texture statistical characteristic differences, interaction items of gray level fluctuation statistical differences and gradient energy statistical differences and interaction items of gray level fluctuation statistical differences and texture statistical characteristic differences, consistency metrics of the four types of contrast differences are calculated according to the two-by-two interaction items, coupling weights corresponding to the four types of contrast differences are determined according to the consistency metrics, and the four types of normalized contrast differences are subjected to weighted fusion according to the coupling weights to obtain the comprehensive lithology contrast differences of the current side.
4. The method for extracting the coal seam separation boundary based on deep learning according to claim 2, wherein in the step 2, the equivalent contrast driving quantity comprises a first side equivalent driving component obtained by converting a proximal lithology contrast difference and a distal lithology contrast difference on one side of a candidate interface in a logarithmic average difference form, a second side equivalent driving component obtained by converting a proximal lithology contrast difference and a distal lithology contrast difference on the other side of the candidate interface in a logarithmic average difference form, and a comprehensive equivalent contrast driving quantity obtained by weighting the first side equivalent driving component and the second side equivalent driving component, wherein when the proximal lithology contrast difference or the distal lithology contrast difference is lower than a preset lower limit, the preset lower limit is replaced, and when the proximal lithology contrast difference or the distal lithology contrast difference is higher than a preset upper limit, the preset upper limit is replaced.
5. The method for extracting a coal seam delamination boundary based on deep learning as claimed in claim 4, wherein in step 2, the method for acquiring the driving thermodynamic diagram comprises: Step 21, respectively constructing a near-side band-shaped sampling area and a far-side band-shaped sampling area at two sides of a candidate interface by taking the boundary pixels of the candidate interface as indexes, and calculating lithology contrast differences in each near-side band-shaped sampling area and each far-side band-shaped sampling area; Step 22, converting the near side lithology contrast difference and the far side lithology contrast difference at each side of the candidate interface in a logarithmic average difference form to obtain a current side equivalent driving component, and carrying out weighted convergence on the equivalent driving components at two sides of the candidate interface to obtain comprehensive equivalent contrast driving quantity; Step 23, assigning the comprehensive equivalent contrast driving quantity to the corresponding candidate interface boundary pixels to form a boundary pixel driving value set, and expanding the boundary pixel driving value to a preset bandwidth along the normal direction of the candidate interface boundary to form a boundary band driving diagram; and step 24, performing spatial interpolation, smoothing and normalization on the boundary band driving map, and mapping the boundary band driving map into a pixel-level driving thermodynamic diagram with the same size as the hole wall image.
6. The method for extracting the coal seam separation layer boundary based on deep learning is characterized in that in the step 1, a cylindrical surface unfolding, defogging and water film correction adopts an adaptive processing mode based on comprehensive lithology contrast difference, and specifically comprises the steps of generating a pixel-level contrast differential layout from the comprehensive lithology contrast difference obtained by the previous iteration, determining the updating quantity of an unfolding axis and an unfolding radius of the cylindrical surface unfolding according to the contrast differential layout, finishing iterative correction of unfolding parameters, respectively constructing an atomization intensity map and a water film disturbance map according to the contrast differential layout, executing pixel-level defogging correction on a hole wall image according to the atomization intensity map, executing pixel-level water film attenuation correction on the hole wall image according to the water film disturbance map, and performing spatial adaptive updating of correction coefficients of the defogging correction and the water film attenuation correction along with the contrast differential layout.
7. The method for extracting the coal seam delamination boundary based on deep learning according to claim 6, wherein in step 1, constructing an atomization intensity map and a water film disturbance map according to a pixel-level contrast difference map and performing pixel-level defogging correction and water film attenuation correction comprises performing scale normalization and smoothing on the pixel-level contrast difference map to obtain a contrast difference reference map, mapping a low contrast region of the contrast difference reference map to an atomization candidate region and calculating an atomization intensity map, wherein the atomization intensity map is jointly determined by low contrast degree, local brightness uniformity and high-frequency texture energy attenuation, mapping an abnormally high contrast region of the contrast difference reference map and a neighborhood thereof to a water film candidate region and calculating a water film disturbance map, wherein the water film disturbance map is jointly determined by high contrast peak degree, local mirror surface highlight duty ratio and gradient direction consistency, generating a pixel-level transmittance map according to the atomization intensity map and performing pixel-level defogging correction on a hole wall image in combination with global background brightness estimation, and generating a pixel-level attenuation coefficient map according to the water film disturbance map and performing pixel-level intensity attenuation and suppression correction on the hole wall image.
8. The method for extracting the coal seam separation boundary based on deep learning according to claim 7 is characterized in that in the step 1, pixel values of a contrast difference reference image are normalized and limited to be within a range of 1 to 1, a low contrast area is an area with the pixel value of the contrast difference reference image being less than or equal to 0.35, an abnormally high contrast area is an area with the pixel value of the contrast difference reference image being more than or equal to 0.75 and an increment of a neighborhood mean value being more than or equal to 0.15, the neighborhood is a 5X 5 to 15X 15 pixel window with a current pixel as a center, local brightness uniformity is characterized by a ratio of a standard deviation of brightness in the window to the average of the brightness and is limited to be less than or equal to 0.12, high-frequency texture energy attenuation is characterized by a ratio of high-frequency energy in the window to medium-low-frequency energy and is limited to be less than or equal to 0.6, local mirror surface highlight proportion is limited to be a pixel proportion of a 98 th percentile of brightness in the window is more than or equal to 0.05 to 0.35, gradient direction uniformity is characterized by a circle variance in the gradient direction in the window is limited to be less than or equal to 0.4.
9. The method according to claim 1, wherein in the step 4, the input of the deep learning segmentation network comprises a borehole wall image processed in the step 1, a pixel-level driving thermodynamic diagram obtained in the step 2, an imaging quality weight diagram and a contour continuity weight diagram which are obtained in the step 3 and are mapped to be the same size as the hole wall image, and an atomization strength diagram and a water film disturbance diagram, a recommendation model based on local coal mine priori information is introduced into the deep learning segmentation network, the local coal mine priori information comprises a coal seam number, a buried depth, an inclination angle, a composite roof lithology sequence, a key boundary surface hole depth marking table, an anchor cable anchoring length and an anchor section position, a roadway distance shifting gate distance, and the recommendation model outputs an interface horizon priori diagram aligned with hole depth coordinates according to the local coal mine priori information, wherein the interface horizon prior diagram is spliced with the hole wall image according to a channel at the input end of a segmentation network encoder, the interface horizon diagram is used as an additional input channel, and the feature layer between the encoder and a decoder is used for performing bias modulation on the feature channel and performing jump-scaling on the feature channel.

Description

Deep learning-based coal bed separation layer boundary extraction method Technical Field The invention relates to the technical field of coal seam image processing, in particular to a coal seam separation layer boundary extraction method based on deep learning. Background In a composite roof roadway for coal mining, a separation critical value is determined by arranging drilling holes on a typical monitoring section which is a certain distance away from a working surface shifting gate, locking a critical interface such as sandstone, mudstone, carbonaceous mudstone, a coal line and a weak interlayer by combining hole depth and lithology columnar relation, anchoring a multi-point displacement meter fixed point on an upper and lower plate of the interface to invert the structure surface separation layer quantity by adjacent fixed point displacement difference, and meanwhile distinguishing separation layer response differences inside and outside an anchor cable anchoring area to realize critical value judgment on dynamic updating of the shifting gate position and the anchoring partition, however, the wall image of the conventional rock stratum drilling recorder is still easy to be influenced by residual water film and cooling mist, partial reflection and illumination uneven, the geometric distortion caused by slight adherence of a camera and the gradient gradation gradient of the sandstone, the conventional image processing generally adopts binary segmentation of the whole or partial hole wall setting a fixed threshold, or the boundary is obtained by only determining the edge according to the fact that the gradient amplitude exceeds the threshold value on one side of a candidate interface, and the boundary contour profile is matched with a communication domain screening and a morphological opening and closing operation, and the like, and the boundary profile is difficult to calculate the boundary profile and the stability of the interface is difficult to be combined with the critical value of the separation layer gradient and the fault tolerance of the separation layer, and the boundary profile is difficult to be compared with the boundary profile and the fault tolerance of the boundary profile and the fault of the boundary is difficult to be calculated, and the boundary profile is difficult to cause to be caused by the fault and the boundary profile and the fault-stability. Disclosure of Invention The invention aims to solve the technical problem of providing a coal bed separation layer boundary extraction method based on deep learning, which is characterized in that a driving thermodynamic diagram is generated through logarithmic average difference equivalent conversion of near-far-side lithology contrast differences of upper and lower discs of a candidate interface, and a write-back iteration mechanism of imaging quality and contour continuity constraint is combined to guide a deep segmentation network to stably extract boundary contours of potential structural planes of separation layers. In order to achieve the above purpose, the present invention provides the following technical solutions: a coal bed separation layer boundary extraction method based on deep learning comprises the following steps: step 1, acquiring a borehole wall image, and executing cylindrical surface expansion, defogging and water film correction; Step 2, calculating near-side lithology contrast difference and far-side lithology contrast difference at two sides of a candidate interface, wherein the near side is a strip-shaped sampling area of which one side is clung to a boundary, the far side is a strip-shaped sampling area of which the same side is separated from the near side by a preset distance along the normal direction of the boundary and away from the boundary, converting the strip-shaped sampling area into equivalent contrast driving quantity in a logarithmic average difference mode, and paving the equivalent driving quantity obtained by pixels of each section of candidate boundary back into an image to form a driving thermodynamic diagram so as to guide boundary segmentation of a potential structural surface of a separation layer; step3, calculating an imaging quality weight and a contour continuity weight; step 4, inputting the fused driving thermodynamic diagram and the multi-scale features of the hole wall image into a deep learning segmentation network, outputting an interface mask, and performing contour extraction on the interface mask to obtain a boundary contour of a potential structural plane of the separation layer; step 5, updating the contour continuity weight by writing back the boundary contour, and updating the imaging quality weight by writing back the interface mask; and 6, iteratively repeating the steps 1 to 5 until the convergence threshold is met or the iteration upper limit is reached. The method is characterized in that the imaging of the cylindrical surface of the hole wall is converted into the expres