A cross-modal alignment method for rock thin section images

Abstract

The present invention relates to the field of image processing technology, and specifically to a cross-modal alignment method for rock thin section images. First, preprocessing of the aligned cross-modal rock thin section image dataset is performed. Based on the collected dataset, the MeanShift‑Canny rock core contour detection algorithm is used to extract the effective core contour common to the two modalities. Feature descriptors are formed based on the extracted core contours, and feature points are matched to complete the rock thin section image feature alignment and extract the effective core contour common to the two modalities. Based on the extracted core contours, the present invention takes corner points with obvious angles as feature points. After extracting feature points of different modal images respectively, the main direction and auxiliary direction of the feature points are extracted, thereby combining the corner point coordinates and response value intensity as descriptors to form the judgment basis of the subsequent matching algorithm. Finally, the feature points are matched to achieve rock thin section feature alignment.

Description

Technical Field

[0001] The present invention relates to the field of image processing technology, and in particular to a cross-modal alignment method for rock thin section images. Background Art

[0002] Research on semantic segmentation of rock thin section images based on cross-modal datasets often requires separate alignment of information between modalities during training. This is due to misalignments in spatial and temporal dimensions between modalities caused by error calibration methods, data acquisition methods, or sensor defects. In cross-modal datasets consisting of multiple rock thin section polarization images, spatial consistency issues exist among polarization images of different modalities, as the image acquisition process under different optical conditions is affected by factors such as lens jitter, sample displacement, and sensor defects. This impacts network training and model accuracy, and can also interfere with the interpretation of semantic segmentation results by rock thin section identification practitioners. Traditional image matching algorithms typically rely on edge detection and feature point matching to address cross-modal alignment of rock thin sections. However, the complexity and diversity of rock thin section images make these algorithms poorly performant in practical applications. To address these issues, a cross-modal rock thin section modal alignment algorithm that utilizes data label information is proposed.

[0003] In summary, it is necessary to propose a cross-modal alignment method for rock thin section images to solve the above problems. Summary of the Invention

[0004] The object of the present invention is to provide a cross-modal alignment method for rock thin section images to solve the problems raised in the above background technology.

[0005] To achieve the above object, the present invention provides the following technical solutions:

[0006] A cross-modal alignment method for rock thin section images, comprising the following steps:

[0007] S10. Preprocessing of cross-modal rock thin section image dataset for alignment:

[0008] Manual semantic segmentation was performed on rock thin-section polarization images acquired under single polarization and cross polarization conditions. The quartz regions in the images were gradually marked using manual annotation tools to provide segmentation label information for subsequent image analysis and model training.

[0009] S20. Based on the collected data set, use the MeanShift-Canny rock core contour detection algorithm to extract the effective core contour common to the two modes:

[0010] Applying mean shift filtering to smooth the image, the mean shift filtering smoothes the image by considering the neighborhood of each pixel in multidimensional space, thereby preserving edge features;

[0011] Check whether the filtering process has converged according to the preset convergence conditions. If the changes in the image meet the preset convergence conditions, the process is considered to be completed and the subsequent steps are entered. If it has not converged, the mean shift filtering is continued until the convergence conditions are met.

[0012] Apply Gaussian filtering to further smooth the image. Gaussian filtering filters out the high-frequency components of the image and retains the low-frequency components of the image. While reducing noise interference, it also blurs the edges to a certain extent.

[0013] The gradient magnitude of the Gaussian filtered image is calculated to detect edges in the image. The edge area is emphasized by calculating the gradient magnitude of each pixel. The non-maximum suppression technique is used to process the gradient image to remove the gradient intensity of non-edge areas, ensuring that the edge is retained with a single pixel width. Finally, only the local maximum value is retained as the edge.

[0014] Double threshold selection is applied to detect edges. Two thresholds, high and low, are set to distinguish strong edges, weak edges, and non-edges. Strong edges are directly retained, while weak edges are retained only when they are connected to strong edges. The image is then binarized to generate a contour map and morphological dilation and multi-scale contour fusion are performed.

[0015] S30. Form a feature descriptor based on the extracted core contour:

[0016] The Shi-Tomasi corner detection algorithm is used to calculate the corner points in the image. The important corner point areas in the image are determined based on the response values ​​of the corner points. The response value classification threshold is set according to the specific conditions of the image and the requirements of corner point detection. Points with high response values ​​are selected as key corner points.

[0017] Take the key corner points with obvious angles in the contour as feature points;

[0018] Define the descriptor of feature points in the image;

[0019] S40. Match feature points to complete rock slice image feature alignment:

[0020] First, the FLANN algorithm is used to find the K nearest neighbor feature points of each feature point in the reference image corresponding to the image to be matched;

[0021] Then, the RANSAC algorithm is used to calculate the homography matrix while eliminating the wrong matching pairs;

[0022] Preferably, after the smoothing process in step S20, adaptive mean shift filtering is performed, the steps of which are as follows:

[0023] S2011. Analyze the image content complexity and automatically adjust the bandwidth parameters of the mean shift filter according to the texture density and contrast in the image;

[0024] S2012. Implementing adaptive mean shift filtering to dynamically adjust the filter strength according to local image features;

[0025] S2013. Evaluate the filtering effect by comparing the edge clarity of the image before and after filtering, and adjust the filtering times or parameters until the optimal effect is achieved.

[0026] Preferably, the specific operations of binarizing the image to generate a contour map and performing morphological dilation in step S20 are: converting the edge information of the image into a black and white format, performing a morphological dilation operation on the binarized contour map, and enhancing the edge features by expanding the edge pixels to fill the small gaps in the detected edges.

[0027] Preferably, the specific steps of forming the feature descriptor are as follows:

[0028] The multi-scale contour fusion in step S20 is as follows:

[0029] S2021. Perform multi-scale morphological dilation on the binary contour image to generate contour images of different scales;

[0030] S2022. Use image fusion technology to fuse multi-scale contour images to enhance the integrity and coherence of contours.

[0031] S2023. Optimize the fused contour image by edge refinement method to ensure the accuracy and clarity of the contour.

[0032] Preferably, the step S30 takes key corner points with obvious angles in the contour as feature points, and the steps are as follows:

[0033] S3011. Set the significant angle threshold to 45 degrees. When the angle between two adjacent edges on the contour is greater than 45 degrees, it is considered to have a significant angle. Select key feature points based on the contrast and texture complexity image features of the feature points.

[0034] S3012. Perform weighted processing on the selected feature points, assigning different weights according to their importance and stability in the image;

[0035] S3013. Update the feature point descriptor and incorporate the weighted information into the descriptor.

[0036] Preferably, the steps of defining the descriptors of the feature points in the image in step S30 are as follows:

[0037] S3021. The starting edge and the ending edge at the corner point are understood as the two directions with the most white pixels in the neighborhood of the feature point, and are defined as the main direction and the auxiliary direction respectively;

[0038] S3022. Combine the response value intensity information of the coordinates of the feature point to form a descriptor of the feature point.

[0039] Preferably, in step S20, the upper threshold is 75% of the maximum value of the gradient amplitude, which is used to determine a strong edge, and the lower threshold is 25% of the maximum value of the gradient amplitude, which is used to determine a weak edge.

[0040] Preferably, in step S40, when erroneous matching is caused by noise, illumination change, or image distortion, a RANSAC algorithm robustness estimation method is used to eliminate the erroneous matching pairs.

[0041] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention designs a MeanShift-Canny rock core contour detection algorithm for rock slice scenes by improving the Canny algorithm, and extracts the effective core contour common to the two modes; then, based on the extracted core contour, the corner points with obvious angles are taken as feature points. After extracting the feature points of different modal images respectively, the main direction and auxiliary direction of the feature points are extracted, and the corner point coordinates and response value intensity are combined as descriptors to form the judgment basis of the subsequent matching algorithm; finally, the feature points are matched to realize the feature alignment of the rock slice. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 The single polarization image of the rock slice of the present invention is shown;

[0043] Figure 2 The figure shows the contour extraction result of the MeanShift-Canny algorithm of the present invention;

[0044] Figure 3 shows a corner point detection result diagram of the present invention;

[0045] Figure 4 The fan-shaped area of ​​the present invention is rotated around the corner point to determine the corner point direction diagram;

[0046] Figure 5 The figure shows the result of feature point matching of the present invention. DETAILED DESCRIPTION

[0047] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0048] See also Figure 1 、 Figure 2 、 Figure 3 、 Figure 4 as well as Figure 5 The present invention proposes a cross-modal alignment method for rock thin section images, the method comprising the following steps:

[0049] S10. Preprocessing of cross-modal rock thin section image dataset for alignment:

[0050] Manual semantic segmentation was performed on rock thin-section polarization images acquired under single polarization and cross polarization conditions. The quartz regions in the images were gradually marked using manual annotation tools to provide segmentation label information for subsequent image analysis and model training.

[0051] S20. Based on the collected data set, use the MeanShift-Canny rock core contour detection algorithm to extract the effective core contour common to the two modes:

[0052] Applying mean shift filtering to smooth the image, the mean shift filtering smoothes the image by considering the neighborhood of each pixel in multidimensional space, thereby preserving edge features;

[0053] Check whether the filtering process has converged according to the preset convergence conditions. If the changes in the image meet the preset convergence conditions, the process is considered to be completed and the subsequent steps are entered. If it has not converged, the mean shift filtering is continued until the convergence conditions are met.

[0054] Apply Gaussian filtering to further smooth the image. Gaussian filtering filters out the high-frequency components of the image and retains the low-frequency components of the image. While reducing noise interference, it also blurs the edges to a certain extent.

[0055] The gradient magnitude of the Gaussian filtered image is calculated to detect edges in the image. The edge area is emphasized by calculating the gradient magnitude of each pixel. The non-maximum suppression technique is used to process the gradient image to remove the gradient intensity of non-edge areas, ensuring that the edge is retained with a single pixel width. Finally, only the local maximum value is retained as the edge.

[0056] Double threshold selection is applied to detect edges. Two thresholds, high and low, are set to distinguish strong edges, weak edges, and non-edges. Strong edges are directly retained, while weak edges are retained only when they are connected to strong edges. The image is then binarized to generate a contour map and morphological dilation and multi-scale contour fusion are performed.

[0057] S30. Form a feature descriptor based on the extracted core contour:

[0058] The Shi-Tomasi corner detection algorithm is used to calculate the corner points in the image. The important corner point areas in the image are determined based on the response values ​​of the corner points. The response value classification threshold is set according to the specific conditions of the image and the requirements of corner point detection. Points with high response values ​​are selected as key corner points.

[0059] Take the key corner points with obvious angles in the contour as feature points;

[0060] Define the descriptor of feature points in the image;

[0061] S40. Match feature points to complete rock slice image feature alignment:

[0062] First, the FLANN algorithm is used to find the K nearest neighbor feature points of each feature point in the reference image corresponding to the image to be matched;

[0063] Then, the RANSAC algorithm is used to calculate the homography matrix while eliminating the wrong matching pairs;

[0064] In this embodiment, it should be noted that after the smoothing process is performed in step S20, adaptive mean shift filtering is performed, and the steps are as follows:

[0065] S2011. Analyze the image content complexity and automatically adjust the bandwidth parameters of the mean shift filter according to the texture density and contrast in the image;

[0066] S2012. Implementing adaptive mean shift filtering to dynamically adjust the filter strength according to local image features;

[0067] S2013. Evaluate the filtering effect by comparing the edge clarity of the image before and after filtering, and adjust the filtering times or parameters until the optimal effect is achieved.

[0068] In this embodiment, it should also be noted that the specific operations of binarizing the image to generate a contour map and performing morphological dilation in step S20 are: converting the edge information of the image into a black and white format, performing a morphological dilation operation on the binarized contour map, and enhancing the edge features by expanding the edge pixels to fill the small gaps in the detected edges.

[0069] In this embodiment, it should be noted that the specific steps of forming the feature descriptor are as follows:

[0070] The multi-scale contour fusion in step S20 is as follows:

[0071] S2021. Perform multi-scale morphological dilation on the binary contour image to generate contour images of different scales;

[0072] S2022. Use image fusion technology to fuse multi-scale contour images to enhance the integrity and coherence of contours.

[0073] S2023. Optimize the fused contour image by edge refinement method to ensure the accuracy and clarity of the contour.

[0074] In this embodiment, it should be noted that the step S30 takes key corner points with obvious angles in the contour as feature points, and the steps are as follows:

[0075] S3011. Set the significant angle threshold to 45 degrees. When the angle between two adjacent edges on the contour is greater than 45 degrees, it is considered to have a significant angle. Select key feature points based on the contrast and texture complexity image features of the feature points.

[0076] S3012. Perform weighted processing on the selected feature points, assigning different weights according to their importance and stability in the image;

[0077] S3013. Update the feature point descriptor and incorporate the weighted information into the descriptor.

[0078] In this embodiment, it should be noted that the steps of defining the descriptors of the feature points in the image in step S30 are as follows:

[0079] S3021. The starting edge and the ending edge at the corner point are understood as the two directions with the most white pixels in the neighborhood of the feature point, and are defined as the main direction and the auxiliary direction respectively;

[0080] S3022. Combine the response value intensity information of the coordinates of the feature point to form a descriptor of the feature point.

[0081] In this embodiment, it should be noted that in step S20, the upper threshold value is 75% of the maximum gradient amplitude value, which is used to determine a strong edge, and the lower threshold value is 25% of the maximum gradient amplitude value, which is used to determine a weak edge.

[0082] In this embodiment, it should also be noted that in step S40, when erroneous matching is caused by noise, illumination change, or image distortion, the RANSAC algorithm robustness estimation method is used to eliminate the generated erroneous matching pairs.

[0083] Example

[0084] In practical applications, the cross-modal alignment method for rock thin section images specifically includes the following steps:

[0085] (1) Preprocessing of cross-modal rock thin section image datasets for alignment;

[0086] The dataset used in this paper contains two modal information: single polarization and cross polarization of rock slices. It is mainly taken from sandstone rock samples from Sichuan Basin and Ordos Basin. The rock storage layers in the two areas are mainly composed of quartz and feldspar. Figure 1 As shown;

[0087] This dataset was prepared under the guidance of professional geologists. Polarized images of rock thin sections from the two locations were obtained using standard operating procedures of electron polarizing microscopes under single polarization and crossed polarization optical conditions.

[0088] A total of 369 single-polarization images and 369 cross-polarization images of rock thin sections were obtained, and 738 semantic segmentation labels were manually created. The main segmentation object was quartz in the image. The specific process is as follows:

[0089] Slice the rock thinly and observe it under a polarizing microscope;

[0090] The acquisition of single polarization image and orthogonal polarization image of rock thin section is achieved by superposition of polarization filters. Two polarization filters are placed between the light source and the object to be detected and between the objective lens and the eyepiece. When only a single polarization filter is used, a single polarization image can be obtained. When two polarization filters are used at the same time and made perpendicular to each other, an orthogonal offset is formed to obtain an orthogonal polarization image.

[0091] There are single and double refracting bodies in rock slice images. We can judge them by changing the direction of light and observing whether the optical properties such as refractive index and absorptivity of the material change.

[0092] If changes occur, it is a birefringent body (such as the identification target quartz of the present invention), otherwise it is a single refractive body (such as pores, non-crystalline particles, etc.);

[0093] (2) Based on the collected data set, the MeanShift-Canny rock core contour detection algorithm is used to extract the effective core contour common to the two modes;

[0094] First, the rock thin section image is smoothed and segmented by using mean shift filtering to blur the texture details of the rock area and better preserve the core contour features, thereby reducing the noise and useless rock texture information in the rock thin section image.

[0095] The implementation process of mean shift is based on statistical methods to find the pixel with the highest frequency in a specific context in the image. The basic process in this invention is as follows:

[0096] (2.1) Traverse and select each pixel point in the RGB input image, set the currently selected point to m0, and loop through the drift calculation from step (2.2) to step (2.4);

[0097] (2.2) Establishing the iteration space: Taking pixel point m0 ​​as the starting center, establish a physical space of size r p And the color space size is r c Drift window;

[0098] (2.3) Obtaining the mass point: Calculate the drift vectors of all pixels in the drift window relative to m0, and take the mean of the color vectors relative to m0 as the center of mass m1 of this window;

[0099] (2.4) Drift iteration: Using the center of mass m1 as the new starting center, repeat steps (2.2) and (2.3) to obtain the next center of mass m2;

[0100] This process is iterated to the nth time, satisfying the starting center point m n-1 Equal to the center of mass m of the drift window at this time n When the condition is met, the iteration is terminated and the pixel value of the original center point m0 ​​in the input image is set to the iteration end point m n The color value at

[0101] Mean shift can merge similar color regions and remove small color regions. However, due to the large number of fine rock textures and dissolved regions in rock thin section images, it is difficult for the mean shift algorithm to process them all at once.

[0102] The present invention sets the number of mean shift rounds i and uses multiple mean shift algorithms to improve the interference of useless feature information and further extract key contour information;

[0103] And by adjusting the physical space size r p and color space size r c To compare the smoothing effect, the results are shown in the attached Figure 2 As shown;

[0104] (3) Form a feature descriptor based on the extracted core contour;

[0105] First, calculate the corner points in the image using the Shi-Tomasi corner detection algorithm;

[0106] Determine the important corner areas in the image based on the response values ​​of the corner points, and select those points with higher response values ​​as key corner points;

[0107] Then take the key corner points with obvious angles in the contour as feature points;

[0108] Finally, the starting and ending edges at the corner point are understood as the two directions with the most white pixels in the neighborhood of the feature point, which are defined as the main direction and auxiliary direction respectively;

[0109] Finally, the coordinates of the feature point, the intensity of the response values ​​in these two directions and other information are combined to form the descriptor of the feature point;

[0110] (4) Match feature points and complete the alignment of rock slice image features;

[0111] First, the FLANN algorithm is used to find the K nearest neighbor feature points of each feature point in the reference image corresponding to the image to be matched;

[0112] Specifically, the Hamming distance between feature point descriptors is calculated, and a difference threshold is set to determine whether the matched feature point pairs originate from the core contour at the same location;

[0113] Then use the RANSAC algorithm to calculate the optimal homography matrix. The specific steps are as follows:

[0114] (4.1) Reduce noise and select 4 pairs of non-collinear point pairs from all matched corner point pairs, calculate the homography matrix H, and record it as model N;

[0115] (4.2) Calculate the error of all remaining matching points to model N. If the error is within the threshold, add it to the set D and use it as the best set Db;

[0116] (4.3) Repeat steps (4.1) and (4.2). If the number of matching corner point pairs in the set D obtained in the new round of calculation is greater than the current optimal set Db, update Db = D;

[0117] (4.4) The operation ends when the number of repetitions exceeds the set maximum value, otherwise the loop continues;

[0118] The matrix H corresponding to the optimal set Db is the optimal homography matrix output by this algorithm.

[0119] Use this matrix to perform geometric transformation on the image to be matched, and finally obtain the matched image, as shown in the attached figure. Figure 5 shown.

[0120] While embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions, and variations may be made to these embodiments without departing from the principles and spirit of the invention, and that the scope of the invention is defined by the appended claims and their equivalents.

Claims

1. A cross-modal alignment method for rock thin section images, characterized in that: The method comprises the following steps: S10. Preprocessing of cross-modal rock thin section image dataset for alignment: Manual semantic segmentation was performed on rock thin-section polarization images acquired under single polarization and cross polarization conditions. The quartz regions in the images were gradually marked using manual annotation tools to provide segmentation label information for subsequent image analysis and model training. S20. Based on the collected data set, use the MeanShift-Canny rock core contour detection algorithm to extract the effective core contour common to the two modes: Applying mean shift filtering to smooth the image, the mean shift filtering smoothes the image by considering the neighborhood of each pixel in multidimensional space, thereby preserving edge features; Check whether the filtering process has converged according to the preset convergence conditions. If the changes in the image meet the preset convergence conditions, the process is considered to be completed and the subsequent steps are entered. If it has not converged, the mean shift filtering is continued until the convergence conditions are met. Apply Gaussian filtering to further smooth the image. Gaussian filtering filters out the high-frequency components of the image and retains the low-frequency components of the image. While reducing noise interference, it also blurs the edges to a certain extent. The gradient magnitude of the Gaussian filtered image is calculated to detect edges in the image. The edge area is emphasized by calculating the gradient magnitude of each pixel. The non-maximum suppression technique is used to process the gradient image to remove the gradient intensity of non-edge areas, ensuring that the edge is retained with a single pixel width. Finally, only the local maximum value is retained as the edge. Double threshold selection is applied to detect edges. Two thresholds, high and low, are set to distinguish strong edges, weak edges, and non-edges. Strong edges are directly retained, while weak edges are retained only when they are connected to strong edges. The image is then binarized to generate a contour map and morphological dilation and multi-scale contour fusion are performed. S30. Form a feature descriptor based on the extracted core contour: The Shi-Tomasi corner detection algorithm is used to calculate the corner points in the image. The important corner point areas in the image are determined based on the response values ​​of the corner points. The response value classification threshold is set according to the specific conditions of the image and the requirements of corner point detection. Points with high response values ​​are selected as key corner points. Take the key corner points with obvious angles in the contour as feature points; Define the descriptor of feature points in the image; S40. Match feature points to complete rock slice image feature alignment: First, the FLANN algorithm is used to find the K nearest neighbor feature points of each feature point in the reference image corresponding to the image to be matched; Then, the RANSAC algorithm is used to calculate the homography matrix while eliminating the wrong matching pairs.

2. A cross-modal alignment method for rock thin section images according to claim 1, characterized in that: After the smoothing process in step S20, adaptive mean shift filtering is performed, and the steps are as follows: S2011. Analyze the image content complexity and automatically adjust the bandwidth parameters of the mean shift filter according to the texture density and contrast in the image; S2012. Implementing adaptive mean shift filtering to dynamically adjust the filter strength according to local image features; S2013. Evaluate the filtering effect by comparing the edge clarity of the image before and after filtering, and adjust the filtering times or parameters until the optimal effect is achieved.

3. The cross-modal alignment method for rock thin section images according to claim 1, characterized in that: In step S20, the image is binarized to generate a contour map and morphological dilation is performed. The specific operations are: converting the edge information of the image into black and white format, performing a morphological dilation operation on the binarized contour map, and enhancing the edge features by expanding the edge pixels to fill the small gaps in the detected edges.

4. The cross-modal alignment method for rock thin section images according to claim 1, characterized in that: The specific steps of forming the feature descriptor are as follows: The multi-scale contour fusion in step S20 is as follows: S2021. Perform multi-scale morphological dilation on the binary contour image to generate contour images of different scales; S2022. Use image fusion technology to fuse multi-scale contour images to enhance the integrity and coherence of contours. S2023. Optimize the fused contour image by edge refinement method to ensure the accuracy and clarity of the contour.

5. The cross-modal alignment method for rock thin section images according to claim 1, characterized in that: The step S30 takes key corner points with obvious angles in the contour as feature points, and the steps are as follows: S3011. Set the significant angle threshold to 45 degrees. When the angle between two adjacent edges on the contour is greater than 45 degrees, it is considered to have a significant angle. Select key feature points based on the contrast and texture complexity image features of the feature points. S3012. Perform weighted processing on the selected feature points, assigning different weights according to their importance and stability in the image; S3013. Update the feature point descriptor and incorporate the weighted information into the descriptor.

6. The cross-modal alignment method for rock thin section images according to claim 1, characterized in that: The steps of defining the descriptors of the feature points in the image in step S30 are as follows: S3021. The starting edge and the ending edge at the corner point are understood as the two directions with the most white pixels in the neighborhood of the feature point, and are defined as the main direction and the auxiliary direction respectively; S3022. Combine the response value intensity information of the coordinates of the feature point to form a descriptor of the feature point.

7. The cross-modal alignment method for rock thin section images according to claim 1, characterized in that: In step S20 , the upper threshold is 75% of the maximum value of the gradient amplitude, which is used to determine a strong edge, and the lower threshold is 25% of the maximum value of the gradient amplitude, which is used to determine a weak edge.

8. The cross-modal alignment method for rock thin section images according to claim 1, characterized in that: In step S40, when erroneous matching is caused by noise, illumination change, or image distortion, the RANSAC algorithm robustness estimation method is used to eliminate the erroneous matching pairs.

Images