CN115564706A - Method and device for automatic identification of rock composition - Google Patents

Abstract

The present invention proposes a method and device for automatic identification of rock components, which relate to the field of petroleum technology, including: preparing rock thin section samples; obtaining full-field microscopic images under orthogonal polarized light and single polarized light; obtaining QEMSCAN output images; performing image alignment; and performing synchronization Slicing and numbering; counting the position and proportion of pores and pores in all single polarized sub-images; performing superpixel region segmentation, saving the segmented images of rock mineral particles containing rock mineral particles separately, and splicing the remaining images according to the number Finally, it is used as the background image; the automatic identification model of rock composition is trained; the image of the rock thin section to be tested is recognized by the trained automatic identification model of rock composition, and the rock composition of the rock thin section to be tested is obtained; the corresponding cementation in the background image is obtained through the output image of QEMSCAN The composition and content of the substance and the heterogeneous part. The invention greatly improves the working efficiency and greatly improves the accuracy of rock composition analysis.

Description

Method and device for automatic identification of rock composition

technical field

The invention relates to the field of petroleum technology, in particular to a method and device for automatic identification of rock components.

Background technique

At present, rock thin section identification is an important petroleum geological research experiment. In this experiment, the rock samples were ground to 0.03mm to make rock thin slices. The experimenters manually identified the optical characteristics of the minerals under a polarizing microscope, obtained the composition and content data of rock minerals and interstitials, and named the rocks according to relevant standards. . However, there are many kinds of rock minerals and they are easy to be altered and dissolved. Therefore, the identification of rock thin sections is time-consuming and difficult, requiring the identification personnel to have high professional ability. In addition, the identification of rock thin sections is the estimation of multi-fields by experimenters under a polarizing microscope, and there are subjective differences in the identification results.

In order to improve the accuracy of identification of rock thin sections and the efficiency of experimental identification, the QEMSCAN experiment can identify the composition of rock minerals and interstitials and estimate the content of 0.03mm rock thin sections, but it cannot be used for homogeneous and polycrystalline mineral particles. Therefore, there is a lack of accuracy in the identification of mineral particle components.

Contents of the invention

The invention proposes a method and device for automatic identification of rock components, which can improve the accuracy and efficiency of rock component identification by combining QEMSCAN and microscopic images of rock thin sections.

The invention provides a method for automatic identification of rock composition, comprising:

Obtaining single polarized light microscopic images at different positions of the rock thin section sample and crossed polarized light microscopic images of the rock thin section at different positions and at different angles; using the QEMSCAN method to obtain the QEMSCAN output image of the rock thin section sample;

Preprocessing the single polarized light microscopic image, the crossed polarized light microscopic image and the QEMSCAN output image respectively to obtain a preprocessed single polarized light full field microscopic image and a preprocessed crossed polarized light full field microscopic image Image and the preprocessed QEMSCAN output image; the preprocessing includes aligning the single polarized full-field microscopic image, the crossed polarized full-field microscopic image and the QEMSCAN output image;

Synchronously slice and number the preprocessed single-polarized full-field microscopic image and the preprocessed crossed-polarized full-field microscopic image to obtain multiple numbered cross-polarized sub-images and multiple numbered single-polarized sub-images ;

Determine the aperture position information and aperture ratio in the numbered single-polarized sub-image;

Perform superpixel segmentation on the numbered single-polarized light sub-image and the numbered cross-polarized light sub-image in combination with the hole position information to obtain a rock grain image containing rock grains, except for the rock grain image The images of are spliced according to the numbers to obtain the background image;

Using the rock particle image as a training sample data set to train an automatic identification model of rock components to be trained to obtain a trained automatic identification model for rock components;

Obtain the image of the rock thin section to be tested;

The image of the rock thin section to be tested is input to the trained rock composition automatic identification model, and the rock composition of the rock thin section to be tested is output;

Combined with the alignment results, the cement components in the background image and the heterogeneous components in the background image were determined through the output image of QEMSCAN.

In an optional embodiment, the single polarized light microscopic image, the crossed polarized light microscopic image and the QEMSCAN output image are respectively preprocessed to obtain the preprocessed single polarized light full field microscopic image, and the preprocessed The final cross-polarized full-field microscopic image and the preprocessed QEMSCAN output image; the preprocessing includes aligning the single-polarized full-field microscopic image, the crossed crossed polarized full-field microscopic image and the QEMSCAN output image ,include:

The single-polarized light microscopic images and crossed-polarized light microscopic images of the same sample and at the same angle are spliced separately to obtain single-polarized light full-field microscopic images and cross-polarized full-field microscopic images respectively;

Combining the QEMSCAN output image, aligning the single-polarized global microscopic image and the orthogonal polarized global microscopic image to obtain the aligned single-polarized global microscopic image and the aligned orthogonal polarized global microscopic image.

In an optional implementation manner, determining the aperture position and aperture ratio in the numbered single-polarized sub-image includes:

Preprocessing the multiple numbered single-polarized sub-images respectively and then converting them to the HSV color space to obtain a single-polarized sub-image under the HSV color space corresponding to each numbered single-polarized sub-image;

Determining the proportion of blue pixels in each single-polarized light sub-image under the HSV color space and recording the position information of the blue pixels, wherein the proportion of the blue pixels is the proportion of the aperture , the position information of the blue pixel is the position information of the aperture.

In an optional embodiment, superpixel segmentation is performed on the numbered single-polarization sub-image and the numbered cross-polarization sub-image in combination with the hole position information to obtain a rock particle image containing rock particles, The images other than the rock particle image are spliced according to the number to obtain the background image, including:

Step S1, combining the perforation position information to remove the perforations in the numbered single-polarized photon sub-images to obtain the single-polarized photon sub-images after the perforations are removed; initialize and segment the single-polarized photon sub-images after the perforations are removed to obtain single-polarized photon sub-images multiple initially segmented regions of polarized sub-images;

Step S2, in the numbered orthogonally polarized sub-images, obtain A first segmented area image data corresponding to a plurality of initial segmented areas of the single-polarized sub-image, and convert the first segmented area image data in the RGB color space Each channel below is quantified into B levels respectively, and N feature dimensions are obtained; the distribution of statistical pixels in the N feature dimensions is performed and the feature space is normalized, and A is generated according to the image data of the first segmented region. A color histogram and a color mean value generated according to the image data of the first segmented region;

Step S3, acquiring second segmented region image data corresponding to multiple initial segmented regions of the single-polarized light sub-image after the holes are removed, and determining each grayscale histogram generated according to the second segmented region image data;

Step S4, generating a corresponding region adjacency graph according to the adjacency relationship between multiple initially segmented regions, wherein each region in the region adjacency graph includes corresponding region adjacency graph image data, and the region adjacency graph image The data includes a color histogram generated according to the image data of the first segmented region, a color mean value generated according to the image data of the first segmented region, and a corresponding grayscale histogram generated according to the image data of the second segmented region;

Step S5, determine the similarity h _c of the orthogonally polarized sub-images according to the following formula (1):

Among them, H _m represents the color histogram generated from the image data of the first region adjacency graph corresponding to region m in the region adjacency graph, and H _m (i) represents the image data of the first region adjacency graph corresponding to region m in the region adjacency graph. The normalization results corresponding to i feature dimensions; H _n represents the color histogram generated from the image data of the second region adjacency graph corresponding to region n in the region adjacency graph; H _n (i) represents the color histogram corresponding to region n in the region adjacency graph The normalization result corresponding to the image data of the second region adjacency graph in the i-th feature dimension; region n in the region adjacency graph is an adjacent node of region m;

Step S6, determine the boundary distance h _e between the image data of the first region adjacency graph corresponding to region m in the region adjacency graph and the image data of the second region adjacency graph corresponding to region n according to the following formula (2):

h _e =||u _m -u _n || ² , (2);

Among them, u _m represents the color mean value of area m; u _n represents the color mean value of area n;

Step S7, in the cross-polarized sub-image, according to the following formula (3), determine the distance metric value of the first region adjacency graph image data corresponding to region m in the region adjacency graph and the second region adjacency graph image data corresponding to region n D:

D=p*hc+(1-p)*he, (3);

Among them, p is a constant between 0-1;

Step S8, according to the following formula (4), determine the gray similarity h _a of the single polarized sub-image after removing the holes:

Among them, F _m represents the grayscale histogram of the image data of the first region adjacency graph corresponding to region m in the region adjacency graph; F _m (j) represents the image data of the first region adjacency graph corresponding to region m in the region adjacency graph at jth gray histogram under intensity values; F _n represents the gray histogram of the second region adjacency graph image data corresponding to region n in the region adjacency graph; F _n (j) represents the second region adjacency graph image data corresponding to region n The gray histogram of the region adjacency graph image data under the jth intensity value;

Step S9, according to the following formula (5), it is determined that the similarity between the region m and the region n in the region adjacency graph is merged into h:

Among them, K represents the number of shooting angles of the numbered cross-polarized light microscopic image at the same position, and D(k) represents the image data corresponding to area m in the cross-polarized sub-image area adjoining figure at the kth shooting angle and the distance metric of the image data corresponding to region n; b is a constant;

Step S10, judging whether the degree of similarity merged is greater than the preset threshold, if so, merge the region m and region n into a third region, and combine the image data of the first region adjacency graph corresponding to region m and the second region adjacency graph corresponding to region n The image data is merged into the third image data; the color histogram, color mean value and grayscale histogram corresponding to the third area are determined; the area adjacency graph is updated according to the third area;

Step S11, repeating step S5 to step S10 until there is no third image data to be merged to obtain the combined total segmented area;

Step S12, under the crossed polarized image, perform two-category labeling on the total image data corresponding to the merged total segmented area to obtain the foreground rock particle image and the background image of other impurities; combine the foreground rock particle image and the Other impurity images in the background are used as the input of the rock composition automatic recognition model to be trained for training, and the trained rock composition automatic recognition model is obtained;

Step S13, using the trained rock composition automatic recognition model to classify the rock thin section image to be tested to obtain the segmented area of rock particles, and extract the corresponding first segmented area image data on the orthogonal polarized sub-image according to the segmented area of rock particles The rock particle image containing the rock particle, the orthogonally polarized sub-images containing the rock particle image are spliced according to the numbers to obtain the background image.

In an optional embodiment, the rock composition automatic identification model is an EfficientDet model.

In an optional embodiment, adopt the QEMSCAN method to obtain the QEMSCAN output image of the rock thin section sample, including:

The rock thin section sample is scanned by QEMSCAN method, and the backscattering map and mineral composition content map of the rock thin section are obtained.

In an optional embodiment, combined with the alignment results, the cement components in the background image and the heterogeneous components in the background image are determined through the QEMSCAN output image, including:

Get the position of the remaining data in the background image;

According to the alignment results, the remaining data positions in the background image except for the positions of pores and fractures and the segmentation results in the mineral composition content map are regarded as cement and matrix, and the number of each pixel value in the remaining data positions is counted in the mineral composition content map to obtain The composition and content of cement and heterogeneous groups.

In a second aspect, the present invention provides a device for automatic identification of rock composition, comprising:

The first acquisition module is used to acquire single polarized light microscopic images at different positions of the rock thin section and crossed polarized light microscopic images of the rock thin section at different positions and angles; the QEMSCAN method is used to acquire the rock thin section samples QEMSCAN output image;

The preprocessing module is used to preprocess the single polarized light microscopic image, the crossed polarized light microscopic image and the QEMSCAN output image respectively, to obtain the preprocessed single polarized light global microscopic image, the preprocessed Orthogonal polarization global microscopic image and the preprocessed QEMSCAN output image; the preprocessing includes aligning the single polarized global microscopic image, the orthogonal polarized global microscopic image and the QEMSCAN output image;

The slicing module is used for synchronously slicing and numbering the preprocessed single polarized global microscopic image and the preprocessed orthogonal polarized global microscopic image to obtain multiple numbered orthogonal polarized sub-images and multiple numbered The single polarized photon image of ;

Aperture information determination module, configured to determine the perforation position information and perforation ratio in the numbered single-polarized sub-image;

The superpixel segmentation module is used to perform superpixel segmentation on the numbered single-polarized light sub-image and the numbered orthogonally polarized light sub-image in combination with the hole position information to obtain a rock particle image containing rock particles, and Images other than the rock particle images are spliced according to numbers to obtain a background image;

The training module is used to use the rock particle image as a training sample data set to train the rock composition automatic recognition model to be trained, and obtain the trained rock composition automatic recognition model;

The second acquisition module is used to acquire images of rock thin sections to be measured;

The first rock composition determination module is used to input the image of the rock thin section to be tested into the trained rock composition automatic recognition model, and output the rock composition of the rock thin section to be tested;

The second rock composition determination module is used to combine the alignment results to determine the cement composition in the background image and the matrix composition in the background image through the QEMSCAN output image.

The rock composition automatic identification method and device of the present invention have the following beneficial effects:

1. Combining the microscopic image intelligent recognition with the QEMSCAN method solves the problem that the traditional rock thin section image intelligent recognition cannot identify the interstitial objects, realizes the accurate identification of rock thin section components, and greatly improves the accuracy while reducing the time for experts;

2. The present invention improves the accuracy of automatic identification of mineral particles and interstitial components of sandstone slices, realizes all identification and quantitative output of sandstone slice components, and finally realizes rock classification and naming;

3. The merging method of the superpixel regions in the segmentation process of the present invention integrates the characteristic information of the polarized sequence images, which can better solve the over-segmentation phenomenon during the initial segmentation (SLIC), and realize the segmentation of different superpixel regions in the same particle. Merge, and use the classification model to separate the foreground and background, to achieve the ideal segmentation effect on mineral particles.

Additional features and advantages of the invention will be set forth in the description which follows, and in part will be apparent from the description, or may be learned by practice of the invention. The objectives and other advantages of the invention will be realized and attained by the structure particularly pointed out in the written description and appended drawings.

In order to make the above-mentioned objects, features and advantages of the present invention more comprehensible, preferred embodiments will be described in detail below together with the accompanying drawings.

Description of drawings

In order to more clearly illustrate the specific implementation of the present invention or the technical solutions in the prior art, the following will briefly introduce the accompanying drawings that need to be used in the specific implementation or description of the prior art. Obviously, the accompanying drawings in the following description The drawings show some implementations of the present invention, and those skilled in the art can obtain other drawings based on these drawings without any creative work.

Fig. 1 is the flow chart of rock component automatic identification method of the present invention;

Fig. 2 is the image taken under different angles of single polarized light and crossed polarized light at the same position of the present invention;

Fig. 3 is the QEMSCAN output image schematic diagram of the present invention;

Fig. 4 is a schematic diagram of the alignment of the global microscopic image under crossed polarized light and the global microscopic image under single polarized light according to the present invention;

Fig. 5 is the structural representation of EfficientDet model of the present invention;

Fig. 6 is a schematic diagram of the structure of the automatic identification device for rock composition of the present invention.

In the figure: 10-first acquisition module; 20-preprocessing module; 30-slicing module; 40-perforation information determination module; 50-superpixel segmentation module; 60-training module; 70-second acquisition module; 80 - a first rock composition determination module; 90 - a second rock composition determination module.

detailed description

In order to make the purpose, technical solutions and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below in conjunction with the accompanying drawings. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. the embodiment. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

In order to facilitate the understanding of this embodiment, a method for automatically identifying rock components disclosed in the present invention will be described in detail below through an embodiment.

Referring to Fig. 1, the present embodiment provides a method for automatic identification of rock composition, comprising the following steps:

Step S10, obtaining single polarized light microscopic images at different positions of the rock thin section sample and crossed polarized light microscopic images of the rock thin section at different positions and angles; using the QEMSCAN method to obtain the QEMSCAN output image of the rock thin section sample.

Specifically, rock thin section samples are prepared; single polarized light microscopic images at different positions of the rock thin section samples are obtained, and crossed polarized light microscopic images of the rock thin section samples are obtained at different positions and angles. As shown in Figure 2, the specific method is: place the rock thin section sample under an optical microscope, and take images under single polarized light and crossed polarized light through the microscope camera system, where the shooting angle of single polarized light at the same position is 0 degrees, as As shown in Figure 2(a), a total of 5 cross-polarized images were taken, and the angles were: 0°36°, 72°, 108°, and 144°, as shown in Figure 2(b1)-(b5), the rock There are six single polarized light microscopic images and crossed polarized light microscopic images at the same position of the sheet sample; among them, (b1) is the crossed polarized light (0 degree), (b2) is the crossed polarized light (36 degree), (b3 ) is crossed polarized light (72 degrees), (b4) is crossed polarized light (108 degrees), (b5) is crossed polarized light (144 degrees). Move the preset step length and repeat the above operations to obtain a total of six single polarized light microscopic images and crossed polarized light microscopic images at the next position until all the mineral regions of the rock thin section sample are taken.

The QEMSCAN method is used to scan the rock thin section samples, and the QEMSCAN output image is obtained. The QEMSCAN output image includes the rock thin section backscattering map and the mineral composition content map. The mineral composition content map is shown in Figure 3. Because the current QEMSCAN method has the following disadvantages: it is impossible to divide each particle separately, and the accuracy is not enough. For the production process, there are impurities covering the surface of the particles, which will cause QEMSCAN to identify errors. Therefore, in the subsequent steps of this example, the analysis results of rock particles by QEMSCAN are not adopted, so the effective information used in the subsequent steps is the analysis results of QEMSCAN on the composition and content of cement and matrix.

Step S20, preprocessing the single polarized light microscopic image, the crossed polarized light microscopic image and the QEMSCAN output image respectively, to obtain the preprocessed single polarized light full field microscopic image, the preprocessed crossed polarized light full field microscopic image and the preprocessed Processed QEMSCAN output image; preprocessing includes aligning single-polarization full-field micrographs, crossed-polarization full-field micrographs, and QEMSCAN output images.

Specifically, the preprocessing in this embodiment at least includes splicing and alignment, that is, splicing the single-polarized light microscopic images and cross-polarized light microscopic images belonging to the same sample and at the same angle to obtain the corresponding single-polarized light full-field microscopic images. Image and cross-polarized full-field microscopic image; then, align the single-polarized full-field microscopic image, crossed-polarized full-field microscopic image and QEMSCAN output image; as shown in Figure 4, the upper and lower pictures on the left side of Figure 4 are aligned Later partial display. The specific method can be:

The single-polarized full-field microscopic image, the orthogonal single-polarized full-field microscopic image, and the backscattering image of the rock thin section are smoothed, denoised, and illuminated to equalize. Cross-polarized full-field microscopic image and preprocessed backscattered image of rock thin section;

The FAST algorithm is used to extract the feature points of the three preprocessed images, and the feature points are described by the Brief algorithm to obtain the corresponding description feature points;

Perform similarity matching on the descriptive feature points corresponding to the single-polarization full-field microscopic image, cross-polarized full-field microscopic image, and rock thin section backscattering image, and obtain the rough matching results of the feature points;

Based on the rough matching results, a homography matrix for pixel mapping is obtained using the RANSAC method;

Based on the homography matrix, all pixels in the cross-polarized full-field microscopic image and the backscattered image of the rock thin section are mapped to the single-polarized full-field microscopic image to complete the alignment.

Step S30, synchronously slice and number the preprocessed single-polarized global microscopic image and the preprocessed orthogonal polarized global microscopic image to obtain multiple numbered single-polarized sub-images and multiple numbered orthogonal Polarized subimages.

Specifically, the aligned single-polarized global microscopic images and cross-polarized global microscopic images are sliced synchronously, in the order of "original image name-row number-column number" from left to right and from top to bottom Number the sub-images after slicing, and obtain several sub-images under single-polarization and cross-polarization respectively; for example: "img_num1&num2", where img is the name of the original image for slicing, num1 is the number of rows to which it belongs, and num2 is the sub-image to which it belongs number of columns. The size of each sliced sub-image is less than or equal to 1024*1024. Synchronous slicing means that different images have the same slicing method and size at the same location.

The sub-images under the cross-polarized light after several slices and the sub-images under the single-polarized light after several slices are numbered respectively, to obtain multiple numbered cross-polarized light sub-images and multiple numbered single-polarized light sub-images.

Step S40, determining the aperture position information and aperture ratio in the numbered single-polarized sub-images.

Here, under the single polarized light image, the casting liquid is blue, and the rock particles will not appear blue, so the specific method of this step is: smoothing and filtering the sub-image under single polarized light and then converting it into HSV Color space; count the number of blue pixels, and take the proportion of blue pixels as the proportion of apertures in the sub-image under single polarized light after the slice; save the position information of blue pixels and use it as the aperture position , and then convert the sub-image under single polarized light into RGB color space. Among them, the value of blue in HSV space is: H(100-124), S(43-255), V(46-255).

Step S50, combining the numbered single-polarization sub-image and the numbered orthogonal-polarization sub-image with superpixel segmentation to obtain the rock particle image containing rock particles, and splicing the images other than the rock particle image according to the number , to get the background image.

In this embodiment, by using superpixel segmentation, the distance from pixel to pixel is shortened, which guarantees the processing speed of the subsequent processing of the histogram and the like. The concrete steps of this embodiment are:

Step N1, remove the apertures in the numbered single-polarized photon images in combination with the aperture position information, and obtain the single-polarized photon images after the apertures are removed; initialize and segment the single-polarized photon images after removing the apertures, and obtain single-polarized photons multiple initially segmented regions of the image;

Here, the SLIC algorithm is used to initially segment the single-polarized photon sub-image after removing the holes.

Step N2, acquire A first segmented region image data corresponding to a plurality of initial segmented regions of the single polarized sub-image in the numbered orthogonally polarized sub-image, and convert the first segmented region image data in the RGB color space Each channel is quantified into B levels respectively, and N feature dimensions are obtained; the distribution of pixels in the N feature dimensions is counted and the feature space is normalized, and A color histogram generated based on the image data of the first segmented area and A color mean value generated according to the image data of the first segmented region;

Here, the value of B is 16 and the value of N is 4096.

Step N3, acquiring second segmented area image data corresponding to a plurality of initial segmented areas of the single-polarized light sub-image after the holes are removed, and determining each grayscale histogram generated according to the second segmented area image data;

Step N4, generate a corresponding region adjacency graph according to the adjacency relationship between multiple initially segmented regions, wherein each region in the region adjacency graph includes corresponding region adjacency graph image data, and the region adjacency graph image data includes A color histogram generated from the image data of the segmented area, a color mean value generated from the image data of the first segmented area, and a corresponding grayscale histogram generated from the image data of the second segmented area;

Step N5, determine the similarity h _c of the orthogonally polarized sub-images according to the following formula (1):

Among them, H _m represents the color histogram generated from the image data of the first region adjacency graph corresponding to region m in the region adjacency graph, and H _m (i) represents the image data of the first region adjacency graph corresponding to region m in the region adjacency graph. The normalization results corresponding to i feature dimensions; H _n represents the color histogram generated from the image data of the second region adjacency graph corresponding to region n in the region adjacency graph; H _n (i) represents the color histogram corresponding to region n in the region adjacency graph The normalization result corresponding to the image data of the second region adjacency graph in the i-th feature dimension; region n in the region adjacency graph is an adjacent node of region m;

Here, the region adjacency graph image data includes first region adjacency graph image data, second region adjacency graph image data, and the like for each of the divided regions m, n, and the like.

Step N6, determine the boundary distance h _e between the image data of the first region adjacency graph corresponding to region m in the region adjacency graph and the image data of the second region adjacency graph corresponding to region n according to the following formula (2):

h _e =||u _m -u _n || ² , (2);

Among them, u _m represents the color mean value of area m; u _n represents the color mean value of area n;

Step N7, in the cross-polarized sub-image, determine the distance metric value of the first region adjacency graph image data corresponding to region m in the region adjacency graph and the second region adjacency graph image data corresponding to region n according to the following formula (3) D:

D=p*hc+(1-p)*he, (3);

Among them, p is a constant between 0-1;

Step N8, according to the following formula (4), determine the gray similarity h _a of the single polarized sub-image after removing the holes:

Among them, F _m represents the grayscale histogram of the image data of the first region adjacency graph corresponding to region m in the region adjacency graph; F _m (j) represents the image data of the first region adjacency graph corresponding to region m in the region adjacency graph at jth gray histogram under intensity values; F _n represents the gray histogram of the second region adjacency graph image data corresponding to region n in the region adjacency graph; F _n (j) represents the second region adjacency graph image data corresponding to region n The gray histogram of the region adjacency graph image data under the jth intensity value;

Step N9, according to the following formula (5), determine the similarity between region m and region n in the region adjacency graph and merge them into h:

Among them, K represents the number of shooting angles of the numbered cross-polarized light microscopic image at the same position, and D(k) represents the image data corresponding to area m in the cross-polarized sub-image area adjoining figure at the kth shooting angle and the distance metric of the image data corresponding to region n; b is a constant;

Step S10, judging whether the degree of similarity merged is greater than the preset threshold, if so, merge the region m and region n into a third region, and combine the image data of the first region adjacency graph corresponding to region m and the second region adjacency graph corresponding to region n The image data is merged into the third image data; the color histogram, color mean value and grayscale histogram corresponding to the third area are determined; the area adjacency graph is updated according to the third area;

Step N11, repeating step N5 to step N10 until there is no third image data to be merged, and the combined total segmented area is obtained;

Step N12, under the crossed polarized image, perform two-category labeling on the total image data corresponding to the merged total segmented area to obtain the foreground rock particle image and the background other impurity image; the foreground rock particle image and the background other impurity image are used as the The input of the trained rock composition automatic recognition model is trained to obtain the trained rock composition automatic recognition model;

Step N13, using the trained rock composition automatic recognition model to classify the rock thin section image to be tested to obtain the segmented area of rock particles, and extract the corresponding first segmented area image data on the orthogonal polarized sub-image according to the segmented area of rock particles For the rock particle image containing the rock particle, the orthogonally polarized sub-images containing the rock particle image are spliced according to the numbers to obtain K background images under the orthogonally polarized light.

Here, as shown in Fig. 5, the automatic identification model of rock composition is the EfficientDet model. Part A in Figure 5 is the multi-channel feature extraction part of the image. Use the EfficientNet network as the backbone network, select the backbone type of the EfficientNet-B0 network, and set the hyperparameters, learning rate, Batch size, and training rounds (Epochs) of the training network. , optimizers (such as SGD, Adam), etc. EfficientNet-B0 includes 1 Conv(3×3), 1 MBConv1(3×3), 2 MBConv6(3×3), 2 MBConv6(5×5), 3 MBConv6(3×3), 3 One MBConv6(5×5), four MBConv6(5×5), one MBConv6(3×3), one Conv(1×1), one Pooling layer, one FC layer. Among them, MBConv contains the residual structure, first use the 1×1 convolution to perform the dimension-up operation, and then perform the 3×3 or 5×5 convolution, and then add the attention mechanism about the channel, and then use the 1×1 convolution The dimensionality reduction operation is performed, and then stacked with the residual structure. The activation function of MBConv uses the Swish function and is standardized using Batch Normalization.

Part B of the EfficientDet model is a multi-channel feature fusion part, which is composed of multiple BiFPN networks. The BiFPN network can learn the importance of different input features, and at the same time repeatedly apply top-down and bottom-up multi-size feature fusion. The multi-level feature fusion idea of FPN and PANet enables information to flow in both top-down and bottom-up directions while using rules and efficient connections. BiFPN adds additional weights to each input and lets the network learn how important each input feature is. BiFPN uses Fast normalized fusion (fast normalized fusion) to realize further extraction and fusion of features.

In step S60, the rock particle image is used as a training sample data set to train the automatic recognition model of rock composition to be trained, and the trained automatic recognition model of rock composition is obtained.

Step S70, acquiring the image of the rock thin section to be tested.

Step S80, inputting the image of the rock thin section to be tested into the trained rock composition automatic identification model, and outputting the rock composition of the rock thin section to be tested.

Step S90, combining the alignment results, determining the cement components in the background image and the heterogeneous components in the background image through the QEMSCAN output image.

Specifically, the position of the remaining data in the background image is obtained, and according to the alignment result, the remaining data positions in the background image except for the positions of pores and fractures and the positions of the segmentation results in the mineral composition content map are used as cement and matrix parts, and in the mineral composition content map Count the number of each pixel value in the remaining data position in , and then get the composition and content of the cement and the matrix.

The final output of this embodiment includes the type of target rock, the composition and content of cement and matrix, and the proportion of pores and fractures.

In the specific implementation process, the target rock can be processed in the same way as the rock thin section sample to obtain the corresponding types of images, and then the orthogonal polarized image data corresponding to the segmented area after the target rock is merged is used as the binary classification model after training. Input, the rock particle segmentation area can be obtained, and the orthogonal polarized image containing the rock particle corresponding to the target rock can be used as the input of the trained multi-classification model to obtain the composition of the rock particle.

In one embodiment of the present invention, the concrete steps of preparing rock thin section sample are:

①. Cut the prepared rock specimens into thin plates with a slicer according to the required orientation. If the rocks are hard and dense enough, they can be cut into rock slices with an average length and width of 3-5 cm;

②. After the cut rock slices are rinsed with water, put them on the iron grinding disc of the grinder for rough grinding, medium grinding, and fine grinding. After reaching 0.1mm, use very fine diamond powder (800 No.) net grinding, after reaching about 0.03mm, polish to make the polished surface very smooth and complete, then rinse with clean water, and adjust the oven temperature to 47°C for 12 hours;

③. Paste the smooth surface on the glass slide with Canadian gum, and continue to grind the other side of the test piece after solidification, and grind the glass piece to become translucent;

④. Put the ground flat, smooth and translucent rock on the glass plate, use the finest corundum (No. 800), press and rub it with the belly of your finger until it is about 0.03mm thick, and wash the rock slices with an appropriate thickness. Dry to obtain rock thin slice samples.

This embodiment has the following beneficial effects:

1. Preprocess the acquired single polarized light microscopic image, crossed polarized light microscopic image and QEMSCAN output image respectively, and then slice and number the preprocessed single polarized light microscopic image and crossed polarized light microscopic image synchronously, Calculate the position information and proportion of pores and fractures in the numbered single-polarization sub-images, and combine the position information of pores and fractures to perform superpixel segmentation on the numbered single-polarization sub-images and orthogonal polarization sub-images to obtain the Rock particle image; use the rock particle image to train the rock composition automatic recognition model to be trained; then use the trained rock composition automatic recognition model to identify the rock composition of the rock thin section to be tested; combined with QEMSCAN to determine the cement composition and background in the background image Impurity components in the image; this embodiment combines the intelligent recognition of microscopic images with the QEMSCAN method, which solves the problem that the traditional intelligent recognition of rock slice images cannot identify interstitial objects, realizes accurate identification of rock slice components, and greatly reduces the time spent by experts. Improved accuracy;

2. In this embodiment, the accuracy rate of automatic identification of mineral particles and interstitial components of sandstone slices is improved, and all identification and quantitative output of sandstone slice components are realized, and rock classification and naming are finally realized;

3. The merging method of the superpixel regions in the segmentation process of this embodiment integrates the characteristic information of the polarized sequence images, which can better solve the over-segmentation phenomenon during the initial segmentation (SLIC, that is, the linear iterative clustering superpixel algorithm), Realize the merging of different superpixel segmentation regions in the same particle, and use the classification model to separate the foreground and background, so as to achieve the ideal segmentation effect on mineral particles.

To sum up, the present invention realizes the automatic identification of rock composition, greatly improves the accuracy of the experiment and saves the experiment time, and can effectively assist experts in the analysis of rock composition in a short time, which greatly improves the work efficiency and greatly improves the The accuracy of rock composition analysis is improved.

Referring to Fig. 6, a kind of rock component automatic identification device provided in the present embodiment comprises the following modules:

The first acquisition module 10 is used to acquire the single polarized light microscopic images at different positions of the rock thin section and the crossed polarized light microscopic images of the rock thin section at different positions and angles; the QEMSCAN method is used to acquire the QEMSCAN of the rock thin section samples output image;

The preprocessing module 20 is used to preprocess the single polarized light microscopic image, the crossed polarized light microscopic image and the QEMSCAN output image respectively, so as to obtain the preprocessed single polarized light full field microscopic image and the crossed polarized light full field microscopic image after preprocessing. Microimages and preprocessed QEMSCAN output images; preprocessing includes aligning single-polarized full-field microscopic images, cross-polarized full-field microscopic images, and QEMSCAN output images;

Slicing module 30, for synchronously slicing and numbering the preprocessed single-polarized full-field microscopic image and the preprocessed orthogonally polarized full-field microscopic image to obtain multiple numbered orthogonally polarized sub-images and multiple numbers After the single polarized photon image;

Aperture information determination module 40, used to determine the aperture position information and aperture ratio in the numbered single-polarized sub-image;

The superpixel segmentation module 50 is used to perform superpixel segmentation on the numbered single-polarized sub-images and the numbered orthogonally polarized sub-images in combination with the position information of the pores to obtain a rock particle image containing rock particles. The images of are spliced according to the numbers to obtain the background image;

The training module 60 is used to use the rock particle image as a training sample data set to train the rock composition automatic recognition model to be trained, and obtain the trained rock composition automatic recognition model;

The second acquiring module 70 is used to acquire the image of the rock thin section to be tested;

The first rock composition determination module 80 is used to input the image of the rock thin section to be tested into the trained rock composition automatic identification model, and output the rock composition of the rock thin section to be tested;

The second rock composition determination module 90 is configured to combine the alignment results to determine the cement composition in the background image and the matrix composition in the background image through the QEMSCAN output image.

The implementation principles and technical effects of the device provided by the embodiment of the present invention are the same as those of the foregoing method embodiment. For brief description, for the parts not mentioned in the device embodiment, reference may be made to the corresponding content in the foregoing method embodiment. The rock composition automatic recognition device provided by the embodiment of the present invention has the same technical features as the rock composition automatic recognition method provided by the above embodiment, so it can also solve the same technical problem and achieve the same technical effect.

Finally, it should be noted that: the above-described embodiments are only specific implementations of the present invention, used to illustrate the technical solutions of the present invention, rather than limiting them, and the scope of protection of the present invention is not limited thereto, although referring to the foregoing The embodiment has described the present invention in detail, and those skilled in the art should understand that any person familiar with the technical field can still modify the technical solutions described in the foregoing embodiments within the technical scope disclosed in the present invention Changes can be easily thought of, or equivalent replacements are made to some of the technical features; and these modifications, changes or replacements do not make the essence of the corresponding technical solutions deviate from the spirit and scope of the technical solutions of the embodiments of the present invention, and should be included in the scope of the present invention within the scope of protection.

Claims (8)

1. A rock composition automatic identification method is characterized by comprising the following steps:

acquiring single-polarization microscopic images of different positions of a rock slice sample and orthogonal-polarization microscopic images of the rock slice at different positions and different angles; acquiring a QEMSCAN output image of the rock slice sample by adopting a QEMSCAN method;

respectively preprocessing the single-polarization microscopic image, the orthogonal polarization microscopic image and the QEMSCAN output image to obtain a preprocessed single-polarization global microscopic image, a preprocessed orthogonal polarization global microscopic image and a preprocessed QEMSCAN output image; the preprocessing comprises aligning the single-polarization global microscopic image, the orthogonal-polarization global microscopic image and the QEMSCAN output image;

synchronously slicing and numbering the preprocessed single polarization global microscopic image and the preprocessed orthogonal polarization global microscopic image to obtain a plurality of numbered single polarization photon images and a plurality of numbered orthogonal polarization sub-images;

determining the hole seam position information and the hole seam ratio in the numbered single-polarization photon images;

combining the hole seam position information to perform super-pixel segmentation on the numbered single polarization sub-images and the numbered orthogonal polarization sub-images to obtain rock particle images containing rock particles, and splicing the images except the rock particle images according to the numbers to obtain background images;

taking the rock particle image as a training sample data set to train an automatic rock component identification model to be trained so as to obtain a trained automatic rock component identification model;

acquiring an image of a rock slice to be detected;

inputting the image of the rock slice to be tested into the trained rock component automatic identification model, and outputting the rock component of the rock slice to be tested;

and determining the cement component in the background image and the miscellaneous base component in the background image through the QEMSCAN output image by combining the alignment result.

2. The automatic rock component identification method according to claim 1, wherein the mono-polarization microscopic image, the orthogonal-polarization microscopic image and the qems can output image are respectively preprocessed to obtain a preprocessed mono-polarization global microscopic image, a preprocessed orthogonal-polarization global microscopic image and a preprocessed qems can output image; the preprocessing includes aligning the single-polarization gamut microscopic image, the orthogonal-polarization gamut microscopic image, and the qems scan output image, including:

respectively splicing the single-polarization microscopic image and the orthogonal-polarization microscopic image of the same sample at the same angle to respectively obtain a single-polarization global microscopic image and an orthogonal-polarization global microscopic image;

and aligning the single-polarization global microscopic image and the orthogonal polarization global microscopic image by combining the QEMSCAN output image to obtain the aligned single-polarization global microscopic image and the aligned orthogonal polarization global microscopic image.

3. The method for automatically identifying rock compositions according to claim 1, wherein the determining of the hole seam position and the hole seam ratio in the numbered single polarization photon images comprises:

respectively preprocessing the numbered single polarization photon images and converting the preprocessed single polarization photon images into an HSV color space to obtain single polarization photon images in the HSV color space corresponding to the numbered single polarization photon images;

determining the proportion of blue pixel points in the single-polarization photon image in each HSV color space and recording the position information of the blue pixel points, wherein the proportion of the blue pixel points is the hole seam proportion, and the position information of the blue pixel points is the hole seam position information.

4. The method according to claim 1, wherein the step of performing superpixel segmentation on the numbered single polarization photon images and the numbered orthogonal polarization sub-images by combining with the hole seam position information to obtain rock particle images containing rock particles, and the step of splicing the images except the rock particle images according to numbers to obtain background images comprises the steps of:

step S1, removing the hole seams in the numbered single polarization photon images by combining the hole seam position information to obtain single polarization photon images with the hole seams removed; performing initialization segmentation on the single-polarization photon image after the hole seams are removed to obtain a plurality of initial segmentation areas of the single-polarization photon image;

s2, obtaining A first segmentation area image data corresponding to a plurality of initial segmentation areas of the single-polarization photon image from the numbered orthogonal polarization sub-images, and quantizing each channel of the first segmentation area image data in an RGB color space into B levels respectively to obtain N characteristic dimensions; counting the distribution of pixels in N characteristic dimensions and carrying out characteristic space normalization to obtain A color histograms generated according to the first segmentation region image data and a color mean value generated according to the first segmentation region image data;

s3, acquiring second segmentation area image data corresponding to a plurality of initial segmentation areas of the single-polarization photon image without the apertures, and determining each gray level histogram generated according to the second segmentation area image data;

step S4, generating corresponding region adjacency graphs according to the adjacent relation among a plurality of initial segmentation regions, wherein each region in the region adjacency graphs comprises corresponding region adjacency graph image data, and the region adjacency graph image data comprises a color histogram generated according to the first segmentation region image data, a color mean value generated according to the first segmentation region image data and a corresponding gray level histogram generated according to the second segmentation region image data;

step S5, determining the similarity h of the orthogonal polarization sub-images according to the following formula (1) _c ：

Wherein H _m Representing a color histogram, H, generated in the region adjacency graph from the first region adjacency graph image data corresponding to region m _m (i) Representing a normalization result of first region adjacency graph image data corresponding to a region m in the region adjacency graph corresponding to the ith characteristic dimension; h _n Representing a color histogram generated according to second region adjacency graph image data corresponding to the region n in the region adjacency graph; h _n (i) Representing a normalization result of second region adjacency graph image data corresponding to the region n in the region adjacency graph corresponding to the ith characteristic dimension; a region n in the region adjacency graph is an adjacency node of a region m;

step S6, determining the boundary distance h of the first region adjacency graph image data corresponding to the region m and the second region adjacency graph image data corresponding to the region n in the region adjacency graph according to the following formula (2) _e ：

h _e ＝||u _m -u _n || ² ， (2)；

Wherein u is _m Represents the color mean of the region m; u. of _n Represents the color mean of the region n;

step S7, in the orthogonal polarization subimage, determining the distance metric D between the first region adjacency graph image data corresponding to the region m and the second region adjacency graph image data corresponding to the region n in the region adjacency graph according to the following formula (3):

D＝p×hc+(1-p)×he， (3)；

wherein p is a constant between 0 and 1;

s8, determining the gray level similarity h of the single polarization photon image after the hole seams are removed according to the following formula (4) _a ：

Wherein, F _m A gray level histogram representing first region adjacency graph image data corresponding to a region m in the region adjacency graph; f _m (j) A gray level histogram of first region adjacency graph image data corresponding to a region m in the region adjacency graph under a jth intensity value is represented; f _n A gray level histogram representing second region adjacency graph image data corresponding to the region n in the region adjacency graph; f _n (j) Representing a gray level histogram of second region adjacency graph image data corresponding to a region n in the region adjacency graph under a jth intensity value;

step S9, determining the similarity of the region m and the region n in the region adjacency graph and merging the similarity into h according to the following formula (5):

wherein, K represents the shooting angle number of the numbered orthogonal polarization microscopic images at the same position, and D (K) represents the distance measurement value of the image data corresponding to the area m and the image data corresponding to the area n in the orthogonal polarization sub-image area adjacency graph at the kth shooting angle; b is a constant;

step S10, judging whether the similarity is combined to be larger than a preset threshold value or not, if so, combining the region m and the region n into a third region, and combining first region adjacent map image data corresponding to the region m and second region adjacent map image data corresponding to the region n into third image data; determining a color histogram, a color mean value and a gray level histogram corresponding to the third area; updating the region adjacency graph according to the third region;

step S11, repeating the step S5 to the step S10 until no third image data needs to be combined, and obtaining a combined total segmentation area;

s12, under the orthogonal polarization image, carrying out two-class labeling on the combined total image data corresponding to the total segmentation area to obtain a foreground rock particle image and a background other impurity image; taking the foreground rock particle image and the background other impurity images as the input of an automatic rock component recognition model to be trained for training to obtain a trained automatic rock component recognition model;

and S13, classifying the rock slice images to be detected by adopting the trained automatic rock component identification model to obtain the segmentation areas of the rock particles, scratching out the rock particle images of which the corresponding first segmentation area image data contain the rock particles on the orthogonal polarization subimages according to the segmentation areas of the rock particles, and splicing the orthogonal polarization subimages containing the rock particle images according to the numbers to obtain the background images.

5. The automatic rock component recognition method according to claim 1, wherein the automatic rock component recognition model is an EfficientDet model.

6. The method of automatic rock composition identification according to claim 1 wherein the acquiring a qems scan output image of a rock sheet sample using a qems scan method comprises:

and scanning the rock slice sample by a QEMSCAN method to obtain a rock slice backscattering diagram and a mineral component content diagram.

7. The method of claim 6, wherein determining cement constituents in the background image and miscellaneous basis constituents in the background image from the QEMSCAN output image in combination with the alignment results comprises:

acquiring the position of the residual data in the background image;

and according to the alignment result, taking the residual data positions except the positions of the holes and the segmentation result positions in the background image in the mineral component content graph as cement and miscellaneous bases, and counting the number of each pixel value of the residual data positions in the mineral component content graph to obtain the components and the content of the cement and the miscellaneous bases.

8. An automatic rock composition recognition device, comprising:

the device comprises a first acquisition module, a second acquisition module and a third acquisition module, wherein the first acquisition module is used for acquiring single-polarization microscopic images of a rock slice sample at different positions and orthogonal-polarization microscopic images of the rock slice at different positions and different angles; acquiring a QEMSCAN output image of the rock slice sample by adopting a QEMSCAN method;

the preprocessing module is used for respectively preprocessing the single-polarization microscopic image, the orthogonal polarization microscopic image and the QEMSCAN output image to obtain a preprocessed single-polarization global microscopic image, a preprocessed orthogonal polarization global microscopic image and a preprocessed QEMSCAN output image; the preprocessing comprises aligning the single-polarization global microscopic image, the orthogonal-polarization global microscopic image and the QEMSCAN output image;

the slicing module is used for synchronously slicing and numbering the preprocessed single polarization global microscopic image and the preprocessed orthogonal polarization global microscopic image to obtain a plurality of numbered orthogonal polarization sub-images and a plurality of numbered single polarization photon images;

the hole seam information determining module is used for determining hole seam position information and a hole seam ratio in the numbered single-polarization photon image;

the super-pixel segmentation module is used for carrying out super-pixel segmentation on the numbered single-polarization photon images and the numbered orthogonal polarization sub-images by combining with the hole seam position information to obtain rock particle images containing rock particles, and splicing the images except the rock particle images according to the numbers to obtain background images;

the training module is used for taking the rock particle images as a training sample data set so as to train an automatic rock component recognition model to be trained and obtain the trained automatic rock component recognition model;

the second acquisition module is used for acquiring an image of the rock slice to be detected;

the first rock component determining module is used for inputting the rock slice image to be detected into the trained rock component automatic identification model and outputting the rock component of the rock slice to be detected;

and the second rock component determining module is used for determining cement components in the background image and miscellaneous basic components in the background image through the QEMSCAN output image in combination with the alignment result.

Images