A Method for Identifying Minerals in Shale MAPS Based on QEMSCAN

Abstract

The present invention discloses a method for identifying shale MAPS minerals based on QEMSCAN, comprising: S1, performing multi-stage, cross-resolution pixel-level alignment processing on QEMSCAN images and MAPS images; S2, constructing a mapping relationship between fine-grained mineral categories in QEMSCAN images and coarse-grained mineral identification labels in MAPS images, and constructing a supervised training dataset based on the aligned QEMSCAN images and MAPS images; S3, training a mineral identification model for MAPS mineral identification based on the constructed supervised training dataset to learn the mapping relationship between MAPS images and mineral categories; S4, inputting the MAPS image to be identified into the trained mineral identification model to obtain the mineral category identification result. The method of the present invention does not require additional experimental steps. Except for obtaining the QEMSCAN image and BSE image of the sample with special equipment during model training, only a single MAPS image is needed subsequently to directly predict the mineral composition at its corresponding position, saving a large amount of cost and time, and the identification effect can meet the identification requirements of the production scenario.

Description

Technical Field

[0001] The present invention belongs to the technical field of image processing, and particularly relates to a method for identifying MAPS (a part of the large-field-of-view mosaic image of shale microscopy corresponding to the BSE image) minerals based on QEMSCAN (Comprehensive Automatic Mineral and Rock Detection Method). Background Art

[0002] Shale is a rock formed by the dehydration and cementation of clay. It is mainly composed of clay minerals (such as kaolinite and hydromica) and has obvious thin bedding structures. In the existing methods for identifying the mineral composition of shale, scanning electron microscopy and X-ray electron energy spectroscopy are mostly used. QEMSCAN effectively combines the two and can obtain a color map representing the mineral composition of shale more clearly, accurately, and intuitively. QEMSCAN (Quantitative Evaluation of Minerals by SCANning electronmicroscopy) quantitative analysis technology is a comprehensive automatic mineral and rock detection method, that is, quantitative evaluation of minerals by scanning electron microscopy. A high-energy electron beam accelerated by a pre-set raster scanning mode scans the surface of the mineral sample, thereby obtaining a two-dimensional color mineral distribution image of the sample end face, the mineral content of each measurement point (combining the gray level of the back-scattered electron (Back-ScatterredElectron, BSE) image and the intensity of the X-ray), and the category information. However, when using QEMSCAN to identify the mineral composition of shale, the cost is relatively high, the experimental period is long, and the efficiency is low. In recent years, significant breakthroughs have been made in the field of artificial intelligence. The image recognition technology based on the deep learning convolutional neural network is one of the popular directions in artificial intelligence. The color map of mineral composition and high-resolution images such as MAPS images obtained by QEMSCAN are input into the model as the input end of the convolutional neural network, and the model is continuously trained to achieve the expected experimental results. Summary of the Invention

[0003] Aiming at the above deficiencies in the prior art, the method for identifying MAPS minerals of shale based on QEMSCAN provided by the present invention solves the problems of high cost and low efficiency of the existing mineral identification methods.

[0004] In order to achieve the above invention purpose, the technical solution adopted by the present invention is: a method for identifying MAPS minerals based on shale QEMSCAN, comprising the following steps:

[0005] S1. Perform multi-stage and cross-resolution pixel-level alignment processing on the QEMSCAN image and the MAPS image;

[0006] S2. Construct the mapping relationship between the fine-grained mineral categories in the QEMSCAN images and the coarse-grained mineral identification labels in the MAPS images, and based on the aligned QEMSCAN images and MAPS images, construct a supervised training dataset;

[0007] S3. Train a mineral identification model for MAPS mineral identification based on the constructed supervised training dataset to learn the mapping relationship between MAPS images and mineral categories;

[0008] S4. Input the MAPS image to be identified into the trained mineral identification model to obtain the mineral category identification result;

[0009] Among them, the QEMSCAN image is a color image of the mineral aggregate dissemination characteristics obtained by the quantitative evaluation method of minerals by scanning electron microscopy, the BSE image is the backscattered electron imaging of the mineral sample, and the MAPS image is a part taken from the large-field-of-view mosaic image of the shale microscope corresponding to the BSE image.

[0010] Further, the step S1 is specifically:

[0011] S11. Obtain a number of BSE images, QEMSCAN images, and MAPS images;

[0012] Among them, the MAPS images include the 10th-layer MAPS image and the 16th-layer MAPS image;

[0013] S12. Extract the feature points in the BES image and the 10th-layer MAPS image, perform feature point matching, and perform multi-stage and cross-resolution pixel-level patch alignment on the QEMSCAN image and the 10th-layer MAPS image based on the homography matrix generated by the matching.

[0014] Further, the step S12 is specifically:

[0015] S12-1. Generate the difference-of-Gaussians pyramids of the BSE image and the 10th-layer MAPS image to construct a scale space;

[0016] S12-2. In the constructed scale space, perform extreme value detection on the difference-of-Gaussians pyramid space of the BSE image and the 10th-layer MAPS image, and then extract stable feature points in the BSE image and the 10th-layer MAPS image;

[0017] S12-3. Assign direction information to the extracted stable feature points;

[0018] S12-4. For the stable feature points extracted from the BSE image and the 10th-layer MAPS image, establish descriptors containing position, scale, and direction information;

[0019] S12-5. Use the 10th layer MAPS image as a template and the BSE image as a real-time image, and establish a descriptor set based on descriptors to represent the spatial mapping relationship between the 10th layer MAPS image and the BSE image;

[0020] S12-6. Based on the descriptor set, complete data search through the data structure of the kd tree to match and align the 10th layer MAPS image and the BSE, and use the RANSAC algorithm to eliminate incorrect matching points, and then generate a homography matrix;

[0021] S12-7. Based on the generated homography matrix, map the QEMSCAN image onto the 10th layer MAPS image to achieve pixel-level alignment of the QEMSCAN image and the MAPS image.

[0022] Further, the specific steps of step S2 are as follows:

[0023] S21. Obtain the aligned image when the QEMSCAN image and the 10th layer MAPS image are pixel-level aligned;

[0024] S22. Select the aligned images with the same resolution and quantity as the 16th layer MAPS image to construct a label atlas;

[0025] S23. Divide the 16th layer MAPS image into a training set and a test set;

[0026] S24. Use the label atlas, the training set, and the test set together as a supervised training data set.

[0027] Further, the mineral recognition model in step S3 includes an encoder and a decoder connected to each other.

[0028] Further, the encoder in the mineral recognition model includes a backbone network DCNN;

[0029] The output of the backbone network DCNN includes deep features and shallow features;

[0030] The deep features are input into a parallel first 1*1 convolutional layer, three 3*3 convolutional layers, and an average pooling layer, and then pass through a first fully connected layer and are input into a second 1*1 convolutional layer;

[0031] The shallow features are input into the decoder. The decoder includes a third 1*1 convolutional layer, a second fully connected layer, a first 4-fold upsampling layer, a second 3*3 convolutional layer, and a second 4-fold upsampling layer. The third 1*1 convolutional layer inputs the shallow features. The features output by the second 1*1 convolutional layer pass through the first 4-fold upsampling layer and are combined with the features output by the third 1*1 convolutional layer and then sequentially enter the second fully connected layer, the second 3*3 convolutional layer, and the second 4-fold upsampling layer to obtain the output image.

[0032] The beneficial effects of the present invention are as follows:

[0033] The present invention only needs to collect a large number of QEMSACN, BSE and corresponding MAPS pictures during model training. The QEMSCAN pictures are used as labels for the MAPS mineral recognition model, and the BSE images are used for the alignment of MAPS and QEMSCAN. When in use, no additional experimental steps are required. Except for the need for special equipment to obtain the QEMSCAN images and BSE images of mineral images initially, subsequent steps only require making a large number of sample data, placing them into the model for training. After achieving a high accuracy rate, the mineral composition of the BSE pictures can be directly predicted, thus no longer requiring the acquisition of their QEMSCAN pictures, saving a large amount of costs and time, with good recognition effects, greatly improved efficiency, and effectively reduced costs. Description of the Drawings

[0034] Figure 1 It is a flowchart of the shale MAPS mineral recognition method based on QEMSCAN provided by the present invention.

[0035] Figure 2 It is a network structure diagram of the mineral recognition model provided by the present invention.

[0036] Figure 3 It is a QEMSCAN picture provided by the present invention.

[0037] Figure 4 It is a BSE picture provided by the present invention.

[0038] Figure 5 It is the 10th layer MAPS picture provided by the present invention.

[0039] Figure 6 It is the 16th layer MAPS picture provided by the present invention. Detailed Embodiments

[0040] The following describes the detailed embodiments of the present invention to facilitate those skilled in the art to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the detailed embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions made using the concept of the present invention are within the scope of protection.

[0041] Embodiment 1:

[0042] The embodiment of the present invention provides a shale MAPS mineral recognition method based on QEMSCAN, as Figure 1 shown, including the following steps:

[0043] S1. Perform multi-stage and cross-resolution pixel-level alignment processing on QEMSCAN images and MAPS images;

[0044] S2. Construct the mapping relationship between the fine-grained mineral categories in QEMSCAN images and the coarse-grained mineral identification labels in MAPS images, and construct a supervised training dataset based on the aligned QEMSCAN images and MAPS images;

[0045] S3. Train a mineral recognition model for MAPS mineral recognition based on the constructed supervised training dataset to learn the mapping relationship between MAPS images and mineral categories;

[0046] S4. Input the MAPS image to be recognized into the trained mineral recognition model to obtain the mineral category recognition result;

[0047] Among them, the QEMSCAN image is a color image of the mineral aggregate dissemination characteristics obtained by the quantitative evaluation method of minerals by scanning electron microscopy. Each color represents a mineral. The BSE image is the backscattered electron imaging of the mineral sample. The MAPS image is a part of the large-field-of-view mosaic image of the shale microscope corresponding to the BSE image.

[0048] In the embodiment of the present invention, the QEMSCAN image contains six minerals, among which quartz (quartz) - pink, feldspar (potassium feldspar, sodium feldspar and anorthite) - blue, carbonate rocks (calcite, dolomite and ankerite) - green, clay (illite, chlorite and clinochlore) - red, pyrite (pyrite) - yellow, others (apatite, monazite, sphalerite, rutile, zircon, pore, organic and others) - black; both the QEMSCAN and BSE images are from the same field of view of the same mineral sample, and there is a one-to-one correspondence in the images; the field of view of the MAPS image is only a part of the BSE image.

[0049] In the embodiments of the present invention, the QEMSCAN image is a color image composed of pixel points that only has color information without texture and shape information, while the MAPS image has texture and shape information but no color information. Therefore, the QEMSCAN image and the MAPS image cannot be directly feature-matched. Since the BSE image has texture and shape information similar to that of the MAPS image, it can be feature-matched with the MAPS image. The QEMSCAN image and the BSE image are two corresponding images. The spatial mapping relationship between the BSE image and the MAPS image can be analogized to the relationship between the QEMSCAN image and the MAPS image, thus indirectly solving the problem that the QEMSCAN image and the MAPS image cannot be feature-matched. After the solution, the alignment algorithm is used to produce the training set and the test set, and through the mineral recognition model, the shale MAPS mineral recognition based on QEMSCAN is completed. Since the QEMSCAN image contains color information and only provides the mineral category information corresponding to the pixel points and cannot be directly recognized as a mineral feature; the BSE image and the MAPS image contain mineral image feature information such as shape and texture and can be directly used for recognition; the QEMSCAN image and the BSE image correspond one by one. Therefore, the pixel-level alignment of the MAPS and BSE images can be performed first to obtain the spatial mapping relationship, that is, the homography matrix, and then the homography matrix is mapped onto the QEMSCAN image to perform pixel-level alignment of the MAPS and QEMSCAN images.

[0050] Based on this, step S1 of the embodiments of the present invention is specifically as follows:

[0051] S11. Obtain a plurality of BSE images, QEMSCAN images, and MAPS images.

[0052] Figure 3 Among them, (a)-(f) are different QEMSCAN images used in the present invention; Figure 4 Among them, (a)-(f) are different BSE images used in the present invention;

[0053] Among them, the MAPS image includes the 10th-layer MAPS image and the 16th-layer MAPS image, Figure 5 Among them, (a)-(f) are different 10th-layer MAPS images; Figure 6 Among them, (a)-(f) are different 16th-layer MAPS images;

[0054] S12. Extract the feature points in the BES image and the 10th-layer MAPS image, perform feature point matching, and perform multi-stage and cross-resolution pixel-level alignment of the QEMSCAN image and the 10th-layer MAPS image based on the homography matrix generated by the matching.

[0055] In step S11 of the above embodiment, the 10th layer and the 16th layer represent the image pyramid, which is a kind of image multi-scale representation. It is an effective but conceptually simple structure for interpreting images at multiple resolutions. The image pyramid of an image is a set of image resolutions that gradually decrease in a pyramid shape (from bottom to top) and are derived from the same original image. It is obtained by successively downsampling until a certain termination condition is reached and the sampling stops. Comparing the images layer by layer to a pyramid, the higher the layer, the smaller the image and the lower the resolution. Since the experimental effect of the 10th layer of the MAPS image pyramid is the best, the 10th layer of the MAPS image is used for experiments during the experimental process of this embodiment.

[0056] In the above embodiment, step S12 is specifically as follows:

[0057] S12-1. Generate the difference-of-Gaussians pyramid of the BSE image and the 10th layer of the MAPS image, and construct the scale space;

[0058] S12-2. In the constructed scale space, perform extreme value detection on the difference-of-Gaussians pyramid space of the BSE image and the 10th layer of the MAPS image, and then extract stable feature points from the BSE image and the 10th layer of the MAPS image;

[0059] S12-3. Assign direction information to the extracted stable feature points;

[0060] S12-4. For the stable feature points extracted from the BSE image and the 10th layer of the MAPS image, establish descriptors containing position, scale, and direction information;

[0061] S12-5. Use the 10th layer of the MAPS image as the template and the BSE image as the real-time image, and establish a set of descriptors representing the spatial mapping relationship between the 10th layer of the MAPS image and the BSE image based on the descriptors;

[0062] S12-6. Based on the set of descriptors, complete data search through the data structure of the kd tree to match and align the 10th layer of the MAPS image and the BSE, and use the RANSAC algorithm to eliminate mis-matched points, and then generate a homography matrix;

[0063] S12-7. Based on the generated homography matrix, map the QEMSCAN image to the 10th layer of the MAPS image to achieve pixel-level alignment of the QEMSCAN image and the MAPS image.

[0064] Step S12-1 of this embodiment is specifically as follows:

[0065] Perform Gaussian blur on BES images and MAPS images at different scales, then perform downsampling while performing Gaussian filtering. Convolve the two images separately using Gaussian kernels of different scales, and then subtract them, that is, perform differencing.

[0066] Among them, the spatial function L(x, y, σ) at a certain scale in the image is obtained by convolving the variable parameter Gaussian function G(x, y, σ) with the original input image I(x, y). Among them, the Gaussian function G(x, y, σ) and the spatial function L(x, y, σ) are respectively:

[0067]

[0068] L(x, y, σ) = G(x, y, σ) * I(x, y)

[0069] The difference of Gaussian function D(x, y, σ) = [G(x, y, kσ) - G(x, y, σ)] * I(x, y)

[0070] = L(x, y, kσ) - L(x, y, σ)

[0071] In the formula, is the amplitude, x, y are position coordinates, σ is the scale space factor, regarded as a blurring quantity, the larger the value, the more blurred, k is a constant, representing the ratio of the scale factors of Gaussian blur between each layer, and * is the convolution operator, which applies the "Gaussian blur" G to the image I.

[0072] The specific steps of S12-2 in this embodiment are as follows:

[0073] 1) Accurately locate local extrema in the discrete space and find continuous space extrema through the iterative method; 2) Set a threshold and screen out extrema smaller than the set threshold; 3) Remove the extrema points that may be caused by edge effects during the image convolution process, thereby obtaining stable feature points.

[0074] In step S12-4 of this embodiment, establish a descriptor including position, scale and direction information, so that the image has the characteristics of being high and not changing with various changes.

[0075] In step S12-6 of this embodiment, the greater the difference between the two images, the more difficult it is to align. For example, it is very difficult to align an image with a resolution of 50000 * 50000 with an image with a resolution of 100 * 100. Therefore, it is necessary to eliminate incorrect matching points.

[0076] In the embodiment of the present invention, when training the mineral recognition model, 6 QEMSCAN images with a resolution of 3072*3124, 6 BSE images with a resolution of 3072*3124, 6 MAPS images of the 10th layer with a resolution of 781*781, and 6 groups of 13824 MAPS images of the 16th layer with a resolution of 1024*1024 are used.

[0077] Step S2 in the embodiment of the present invention is specifically as follows:

[0078] S21. Obtain the aligned images when the QEMSCAN images and the MAPS images of the 10th layer are pixel-level aligned;

[0079] S22. Screen out the aligned images with the same resolution and quantity as the MAPS images of the 16th layer to construct a label atlas;

[0080] S23. Divide the MAPS images of the 16th layer into a training set and a test set;

[0081] S24. Use the label atlas, the training set, and the test set together as a supervised training data set.

[0082] In step S22 of this embodiment, the mineral recognition model requires 13824 MAPS images of the 16th layer with a resolution of 1024*1024 as the test set and the training set, among which there are 12442 in the training set and 1382 in the test set. 13824 label images with a resolution of 1024*1024 are used during training, and the acquisition method is as follows: Align a generated MAPS image of the 10th layer with a resolution of 781*781 with the QEMSCAN, and crop it into 13824 images with a resolution of 1024*1024. First, combine all the MAPS images of the 16th layer, and the resolution of the combined image is 50000*50000. Then, enlarge the 781*781 MAPS image to 50000*50000, and crop it into a total of 13824 label images with a resolution of 1024*1024 to form a label atlas.

[0083] Since generating the label map and performing the alignment operation involve magnifying and rotating the image, the clarity of the image has been significantly reduced, and the RGB values corresponding to the minerals have thus changed from a specific value to a fluctuating range value. For example, the RGB value corresponding to quartz in the color matching table is (255, 192, 192), but after magnifying and rotating the image, the RGB value in the area of this mineral type becomes (200 - 255, 150 - 200, 120 - 203). Among them, the RGB value at the junction of the minerals is (200 - 255, 150 - 200, 120 - 203), with a large change range, and the other areas are (255, 192, 203). Since the pixels at the junction occupy a small number, they are discarded and set to the background white. During the experiment, the background area is discarded to increase the experimental accuracy.

[0084] The mineral recognition model in step S3 of the embodiment of the present invention includes an encoder and a decoder connected to each other, as Figure 2 shown. The encoder in the mineral recognition model includes a backbone network DCNN;

[0085] The output of the backbone network DCNN includes deep features and shallow features. The deep features are input to a parallel combination of a first 1*1 convolutional layer, 3 3*3 convolutional layers, and an average pooling layer, and then pass through a first fully connected layer and are input to a second 1*1 convolutional layer. The shallow features are input to the decoder. The decoder includes a third 1*1 convolutional layer, a second fully connected layer, a first 4-fold upsampling layer, a second 3*3 convolutional layer, and a second 4-fold upsampling layer. The third 1*1 convolutional layer inputs the shallow features. The features output by the second 1*1 convolutional layer and the features output by the third 1*1 convolutional layer enter the second fully connected layer, the second 3*3 convolutional layer, and the second 4-fold upsampling layer in sequence after passing through the first 4-fold upsampling layer, and then the output image is obtained.

[0086] In the embodiment of the present invention, the dilation rates of the 3 3*3 convolutional layers are 6, 12, and 18 in sequence.

[0087] The accuracy rate of the mineral recognition model in the embodiment of the present invention has reached 69%. The mineral composition of this area can be obtained from the prediction map. As the QEMSCAN samples increase and the quality improves, the accuracy rate of the model will be higher and higher, and the experimental effect will also be better and better.

[0088] In the description of the present invention, it should be understood that the orientation or positional relationship indicated by terms such as "center", "thickness", "upper", "lower", "horizontal", "top", "bottom", "inner", "outer", "radial", etc. is based on the orientation or positional relationship shown in the drawings. It is only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the device or element referred to must have a specific orientation, be constructed and operated in a specific orientation. Therefore, it should not be construed as a limitation to the present invention. In addition, the terms "first", "second", and "third" are only used for descriptive purposes and cannot be understood as indicating or implying relative importance or implicitly specifying the quantity of technical features. Therefore, the features defined by "first", "second", and "third" may explicitly or implicitly include one or more of such features.

Claims

1. A method for identifying shale MAPS minerals based on QEMSCAN, characterized in that, it includes the following steps: S1. Perform multi-stage and cross-resolution pixel-level alignment processing on QEMSCAN images and MAPS images; S2. Construct a mapping relationship between the fine-grained mineral categories in QEMSCAN images and the coarse-grained mineral identification labels in MAPS images, and construct a supervised training dataset based on the aligned QEMSCAN images and MAPS images; S3. Train a mineral identification model for MAPS mineral identification based on the constructed supervised training dataset to learn the mapping relationship between MAPS images and mineral categories; S4. Input the MAPS image to be identified into the trained mineral identification model to obtain the mineral category identification result; Among them, the QEMSCAN image is a color image of the mineral aggregate distribution characteristics obtained by the quantitative evaluation method of minerals by scanning electron microscopy, the BSE image is the backscattered electron imaging of the mineral sample, and the MAPS image is a part taken from the large-field mosaic image of the shale microscope corresponding to the BSE image; The specific step S1 is: S11. Obtain several BSE images, QEMSCAN images, and MAPS images; Among them, the MAPS images include the 10th layer MAPS image and the 16th layer MAPS image; S12. Extract the feature points in the BES image and the 10th layer MAPS image, perform feature point matching, and perform multi-stage and cross-resolution pixel-level alignment of the QEMSCAN image and the 10th layer MAPS image based on the homography matrix generated by the matching; The specific step S2 is: S21. Obtain the aligned images when the QEMSCAN image and the 10th layer MAPS image are pixel-level aligned; S22. Screen out the aligned images with the same resolution and quantity as the 16th layer MAPS image to construct a label atlas; S23. Divide the 16th layer MAPS image into a training set and a test set; S24. Use the label atlas, training set, and test set together as a supervised training dataset.

2. The method for identifying shale MAPS minerals based on QEMSCAN according to claim 1, characterized in that, the specific step S12 is: S12-1. Generate the difference of Gaussian pyramids of the BSE image and the 10th layer MAPS image to construct a scale space; S12-2. In the constructed scale space, perform extreme value detection on the difference of Gaussian pyramids of the BSE image and the 10th layer MAPS image, and then extract stable feature points in the BSE image and the 10th layer MAPS image; S12-3. Assign direction information to the extracted stable feature points; S12-4. Establish descriptors containing position, scale, and direction information for the stable feature points extracted from the BSE image and the 10th layer MAPS image; S12-5. Use the 10th layer MAPS image as a template and the BSE image as a real-time image to establish a descriptor set representing the spatial mapping relationship between the 10th layer MAPS image and the BSE image based on the descriptors; S12-6. Based on the descriptor subset, complete data search through the data structure of the kd-tree to match and align the 10th layer MAPS image and BSE, and use the RANSAC algorithm to eliminate incorrect matching points, and then generate a homography matrix; S12-7. Based on the generated homography matrix, map the QEMSCAN image onto the 10th layer MAPS image to achieve pixel-level alignment of the QEMSCAN image and the MAPS image.

3. The QEMSCAN-based shale MAPS mineral identification method according to claim 1, characterized in that, the mineral identification model in the step S3 includes a connected encoder and decoder.

4. The QEMSCAN-based shale MAPS mineral identification method according to claim 3, characterized in that, the encoder in the mineral identification model includes a backbone network DCNN; the output of the backbone network DCNN includes deep features and shallow features; the deep features are input into a parallel first 1*1 convolutional layer, three 3*3 convolutional layers, and an average pooling layer, and then input into a second 1*1 convolutional layer through a first fully connected layer; the shallow features are input into the decoder, and the decoder includes a third 1*1 convolutional layer, a second fully connected layer, a first 4-fold upsampling layer, a second 3*3 convolutional layer, and a second 4-fold upsampling layer. The third 1*1 convolutional layer inputs the shallow features, and the features output by the second 1*1 convolutional layer and the features output by the third 1*1 convolutional layer enter the second fully connected layer, the second 3*3 convolutional layer, and the second 4-fold upsampling layer in sequence after passing through the first 4-fold upsampling layer, and then an output image is obtained.

Images