Biological particle segmentation and identification method, system, device and storage medium

Abstract

This invention discloses a method, system, device, and storage medium for bioparticle segmentation and identification. The method includes collecting a sequence of images of thin rock sections and preprocessing the sequence of images; performing feature fusion on the preprocessed sequence of images to obtain a fused image; performing edge detection on the fused image to extract initial edge information of the target bioparticles and obtain an edge feature map; and using an improved PSPNet semantic segmentation model based on the edge feature map to accurately identify and segment the target bioparticles in the sequence of images. This invention overcomes the shortcomings of traditional carbonate rock bioparticle identification methods, achieving efficient and accurate identification and segmentation of bioparticles.

Description

Technical Field

[0001] This invention relates to the field of geological petrology, and more particularly to methods, systems, devices, and storage media for the segmentation and identification of biological particles. Background Technology

[0002] Approximately 60% of China's oil and gas resources are found in carbonate rocks. Carbonate rocks are also major metallurgical solvents, chemical raw materials, refractory raw materials, and raw materials for refining metallic magnesium. The structure and composition of carbonate rocks are closely related to their sedimentary environment and post-diagenetic processes, reflecting their genetic characteristics to a certain extent. This structure is not only an important identification marker but also a key basis for the classification and naming of carbonate rocks. Therefore, strengthening fundamental theoretical research in petrology and lithofacies paleogeography of carbonate rocks, and using this knowledge to guide the exploration and development of oil and gas, groundwater, and various metallic and non-metallic minerals or industrial raw materials, has significant theoretical and practical implications.

[0003] In existing technologies, carbonate rock samples are typically cut into thin sections and observed using a polarizing microscope. The particles in the optical microscopic images are then identified and segmented. However, traditional methods are time-consuming, inefficient, inaccurate, and dependent on the subjectivity of the identification personnel.

[0004] Therefore, there is an urgent need for a biological particle segmentation and identification method that can overcome the shortcomings of traditional carbonate rock biological particle identification methods and achieve efficient and accurate identification and segmentation of biological particles. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, device and storage medium for the segmentation and identification of biological particles, to overcome the shortcomings of traditional methods for identifying biological particles in carbonate rocks, and to achieve efficient and accurate identification and segmentation of biological particles.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] In a first aspect, embodiments of the present invention provide a method for biological particle segmentation and identification, comprising:

[0008] Collect a sequence of images of rock thin sections and preprocess the sequence of images;

[0009] Feature fusion is performed on the preprocessed sequence images to obtain a fused image;

[0010] Edge detection is performed on the fused image to extract the initial edge information of the target biological particles and obtain an edge feature map;

[0011] Based on edge feature maps, the improved PSPNet semantic segmentation model is used to accurately identify and segment target biological particles in the sequence images.

[0012] In a second aspect, embodiments of the present invention provide a biological particle segmentation and identification system, comprising:

[0013] A collection and processing unit is used to collect sequential images of rock thin sections and preprocess the sequential images;

[0014] The fusion unit is used to perform feature fusion on the preprocessed sequence images to obtain a fused image;

[0015] An edge detection unit is used to perform edge detection on the fused image, extract the initial edge information of the target biological particles, and obtain an edge feature map;

[0016] The identification and segmentation unit is used to accurately identify and segment target biological particles in the sequence images based on edge feature maps and using the improved PSPNet semantic segmentation model.

[0017] Thirdly, embodiments of the present invention also provide an electronic device, including a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program or instructions to implement the aforementioned biological particle segmentation and identification method.

[0018] Fourthly, embodiments of the present invention also provide a computer storage medium storing a computer program or instructions, which, when executed by a processor, implement the aforementioned biological particle segmentation and identification method.

[0019] Fifthly, embodiments of the present invention also provide a computer program product, including a computer program or instructions, which, when executed by a processor, implement the aforementioned biological particle segmentation and recognition method.

[0020] The technical effects and advantages of this invention are as follows: This invention utilizes feature fusion and edge detection techniques to improve the segmentation and recognition model's ability to perceive biological particle features and highlight boundary information, thereby improving the accuracy of biological particle identification and segmentation. At the same time, it utilizes deep learning technology in image denoising and data augmentation to achieve automatic processing and analysis of a large amount of carbonate rock thin section image data, greatly improving work efficiency.

[0021] Other features and advantages of the invention will be set forth in the description which follows, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures pointed out in the description, claims and drawings. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] Figure 1 This is a flowchart of a biological particle segmentation and recognition method according to an embodiment of the present invention;

[0024] Figure 2(a) is a single-polarized light image of a carbonate rock thin section sample in an embodiment of the present invention;

[0025] Figure 2(b) is a 0-degree orthogonally polarized light image of a carbonate rock thin section sample in an embodiment of the present invention;

[0026] Figure 2(c) is a 15-degree orthogonal polarized light image of a carbonate rock thin section sample in an embodiment of the present invention;

[0027] Figure 2(d) is a 30-degree orthogonally polarized light image of a carbonate rock thin section sample in an embodiment of the present invention;

[0028] Figure 2(e) is a 45-degree orthogonally polarized light image of a carbonate rock thin section sample in an embodiment of the present invention;

[0029] Figure 2(f) is a 60-degree orthogonally polarized light image of a carbonate rock thin section sample in an embodiment of the present invention;

[0030] Figure 2(g) is a 75-degree orthogonally polarized light image of a carbonate rock thin section sample in an embodiment of the present invention;

[0031] Figure 3 This is a sample label illustration from an embodiment of the present invention;

[0032] Figure 4 This refers to a sequence of images in an embodiment of the present invention where the background occupies more than 60% of the image.

[0033] Figure 5(a) is a sequence image before noise removal in an embodiment of the present invention;

[0034] Figure 5(b) shows the sequence image after noise removal in an embodiment of the present invention;

[0035] Figure 6 This is a schematic diagram of the segmentation and recognition model in an embodiment of the present invention;

[0036] Figure 7 This is a schematic diagram of the structure of a biological particle segmentation and recognition system according to an embodiment of the present invention;

[0037] Figure 8 This is a schematic diagram of the structure of an electronic device according to an embodiment of the present invention;

[0038] Figure 9(a) is a schematic diagram of the effect of the FCN semantic segmentation network in an embodiment of the present invention;

[0039] Figure 9(b) is a schematic diagram of the effect of the SegNet semantic segmentation network in an embodiment of the present invention;

[0040] Figure 9(c) is a schematic diagram of the effect of the DeepLabv3+ semantic segmentation network in an embodiment of the present invention;

[0041] Figure 9(d) is a schematic diagram of the effect of the segmentation and recognition model in the embodiment of the present invention. Detailed Implementation

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

[0043] To address the shortcomings of existing technologies, this invention discloses a method for biological particle segmentation and recognition, such as... Figure 1 As shown, it includes the following steps:

[0044] Step S1: Collect a sequence of images of rock thin sections and preprocess the sequence of images;

[0045] Step S2: Perform feature fusion on the preprocessed sequence images to obtain a fused image;

[0046] Step S3: Perform edge detection on the fused image, extract the initial edge information of the target biological particles, and obtain an edge feature map;

[0047] Step S4: Based on the edge feature map, the improved PSPNet semantic segmentation model is used to accurately identify and segment the target biological particles in the sequence images.

[0048] In some specific implementations, step S1: collecting a sequence of images of rock thin sections and preprocessing the sequence of images, including the following steps:

[0049] Step S11: Take a sequence of images of thin sections of carbonate rock using a polarizing microscope;

[0050] One rock thin section sample (i.e., one carbonate rock thin section sample) corresponds to a set of image sequences, such as... Figures 2(a)-2(g) The sequence of images shown includes one single-polarized image and six cross-polarized images;

[0051] Under the same field of view, an orthogonal polarized light image of a rock thin section sample was acquired at 15° intervals. That is, the angles of 6 orthogonal polarized light images in a series of images were 0 degrees, 15 degrees, 30 degrees, 45 degrees, 60 degrees and 75 degrees respectively.

[0052] The specific image acquisition must meet the following standards:

[0053] Acquisition lens magnification: typically 5x, 10x, or 20x;

[0054] Sequence images of rock thin section samples are typically in .jpg or .png format;

[0055] Acquisition channel requirements: Acquire one single-polarized image and six orthogonally polarized images from a single field of view on the rock thin section (i.e., acquire one image every 15 degrees of stage rotation within a 90-degree rotation range).

[0056] Image acquisition without offset: Image acquisition is performed with the stage stationary, and the particle position remains unique and unchanged under multiple channels, that is, the relative position of each particle in the single-polarized light and cross-polarized light images should remain consistent.

[0057] Step S12: By manually annotating the outline information of biological particles and the biological category in the sequence images by experts, sample labels (i.e., the types of biological particles in the sequence images) and labeled particles (i.e., the identified biological particles) are obtained; sample label example image is shown below. Figure 3 As shown.

[0058] Among them, the bioparticle outline information and biological categories include background, unknown bioclasts, bivalves, foraminifera, brachiopods, sponges, algal debris, echinoderms, ostracods, gastropods, and cement.

[0059] The sample label categories (i.e., the categories of sample labels) include bivalves, foraminifera, brachiopods, sponges, algal debris, echinoderms, ostracods, gastropods, and cement.

[0060] Step S13: Preprocess the sequence images using a segmentation and recognition model, including the following steps:

[0061] Step a: Input the sequence of rock thin section images into the segmentation and recognition model. The segmentation and recognition model uniformly crops the sequence of rock thin section images and sample labels into sub-images of the same size (1024×1024) to preserve the integrity of particles in the image, improve the model's computational efficiency and hardware computational adaptability.

[0062] Step b: Use the segmentation and recognition model to perform quality screening and noise reduction on the sequence images to obtain the final sequence images; specifically including:

[0063] The segmentation and recognition model calculates the area ratio of the background and label particles in the sequence images respectively, and deletes sequence images where the background area ratio exceeds a predetermined ratio (e.g., 60%). Figure 4 As shown), to obtain the filtered sequence images; the present invention improves the quality of training samples in the training set by filtering the sequence images.

[0064] The segmentation and recognition model uses a bilateral filter to remove noise from the filtered sequence images based on the spatial relationship between pixels and the similarity of pixels in the grayscale space, thus obtaining the denoised sequence images, which are the final sequence images. Figures 5(a) and 5(b) show the sequence images before and after denoising, respectively.

[0065] The denoised image sequence is obtained using the following formula:

[0066]

[0067] In the formula, p(m,n) represents the denoised image sequence, m and n represent the coordinates of image points, k and l represent the center coordinates of the current image window, f(k,l) represents the pixel value corresponding to the center coordinates of the current image window, and w(m,n,k,l) ​​represents the bilateral filter convolution kernel function;

[0068] The bilateral filter convolution kernel function is calculated using the following formula:

[0069] w(m,n,k,l)=d(m,n,k,l)·r(m,n,k,l),

[0070] In the formula, d(m,n,k,l) ​​represents the spatial kernel function, and r(m,n,k,l) ​​represents the range kernel function;

[0071] The spatial kernel function and the range kernel function are respectively:

[0072]

[0073] In the formula, f(m,n) represents the pixel value of the current image window, f(k,l) represents the pixel value corresponding to the center coordinate point of the current image window, and σ r This represents the standard deviation of the Gaussian kernel function over the pixel value range.

[0074] Step c: Use the final sequence images as training samples, and construct a dataset based on the training samples and sample labels. According to the number and proportion of samples corresponding to each sample label category in the dataset, and taking the number of unknown biodebris categories as a benchmark, perform sample data augmentation (through image rotation, image flipping, and combination methods, etc.) on sample label categories that exceed the benchmark predetermined number (e.g., 10) to obtain the number of biological particles in the dataset as shown in Table 1. Divide the training set and test set according to a certain ratio (e.g., 7:3). Subsequently, train and validate the segmentation and recognition model using the training set and test set, respectively.

[0075] Table 1. Number of biogranules on thin sections of carbonate rocks

[0076] category Before augmentation After augmentation category Before augmentation After augmentation Unknown raw material 4575 5130 Algae debris 2821 3241 bivalves 122 545 Echinoderms 226 885 Foraminifera 1631 1748 Ostracods 160 940 Brachiopods 80 630 Gastropods 91 625 sponge 182 750 cement 1306 1475

[0077] The present invention improves image quality through the preprocessing process in step S2 in order to establish an accurate and reliable dataset.

[0078] In some specific embodiments, step S2: performing feature fusion on the preprocessed sequence images to obtain a fused image includes the following steps:

[0079] Step S21: The segmentation and recognition model calculates the Euclidean distance between each pixel of each pair of single-polarized and orthogonal-polarized images in each image sequence to reflect the degree of difference between the images, and generates difference images based on the Euclidean distance; that is, it performs a one-to-one correspondence calculation between one single-polarized image and six orthogonal-polarized images in a set of image sequences to obtain six difference images. Each pixel value in the difference image represents the Euclidean distance at the corresponding position.

[0080] Step S22: The segmentation and recognition model determines the weight w of each difference image based on the difference image method. i [0.12,0.15,0.18,0.2,0.25,0.1], the pixel values ​​of each difference image are multiplied by the corresponding weight to obtain a weighted difference image, and each weighted difference image is accumulated one by one to obtain a fused image; the accumulation process here can preserve the key features in the image and reduce noise.

[0081] Step S23: The segmentation and recognition model performs a type conversion on the fused image, that is, converts the fused image from a floating-point type to an 8-bit integer type; and ensures that the pixel values ​​of the converted image are between 0 and 255. If the pixel values ​​exceed this range, truncation is performed. This embodiment of the invention, by converting the type of the fused image, can help reduce storage space and improve the efficiency of subsequent image processing by the segmentation and recognition model.

[0082] Among them, steps S22 and S23 in the feature fusion process can be expressed by the following formulas:

[0083]

[0084] In the formula, F represents, clip(x,0.255) represents the pixel value constraint function, and w i I represents the weight of the i-th difference image. si Let I represent the i-th single-polarized image. oi Let I represent the i-th orthogonally polarized image. Since there is only one single-polarized image in the sequence, each indexed single-polarized image is the same single-polarized image, i.e., Ii. si -I oi Indicates the first For single-polarized light images and cross-polarized light images.

[0085] In some specific embodiments, step S3: performing edge detection on the fused image to extract the initial edge information of the target biological particles and obtain an edge feature map includes the following steps:

[0086] like Figure 6 As shown, the segmentation and recognition model includes an edge detection module and a semantic segmentation module based on dense upsampling;

[0087] Based on the theory of convolutional neural networks, a lightweight edge detection network is constructed and trained using a training set. The trained edge detection network is then used as the edge detection module.

[0088] The lightweight edge detection network comprises five stages, each consisting of 3×3, 1×1-2, and 1×1-1 convolutional layers, with each adjacent stage connected by a 2×2 pooling layer. During the training of the lightweight edge detection network, the cross-entropy loss function and the Sigmoid layer are calculated, and the network parameters are updated.

[0089] The segmentation and recognition model inputs the fused image and the sequence image into the edge detection module. The edge detection module upsamples the fused image through deconvolution to restore the fused image to the same size as the input image (i.e., the sequence image).

[0090] The segmentation and recognition model then uses an edge detection module to perform edge detection on the fused image to obtain the initial contour information of the target biological particles, including:

[0091] During edge detection, the outputs of each stage in the edge detection module are superimposed, and multi-channel features are merged through a 1×1-1 convolutional layer to achieve the purpose of fusing the feature information extracted from all stages (i.e., 5 stages). Finally, the edge feature map is output, which is the output result of the edge detection module.

[0092] In some specific embodiments, step S4: Based on the edge feature map, the improved PSPNet semantic segmentation model is used to accurately identify and segment the target biological particles in the sequence image, including the following steps:

[0093] Step S3 mentions that the segmentation and recognition model also includes a semantic segmentation module based on dense upsampling; wherein, the semantic segmentation module includes a feature extraction submodule and a feature processing submodule.

[0094] Step S41: In this invention, ResNet50 is used as a feature extraction submodule. The segmentation and recognition model uses the feature extraction submodule to extract feature information related to the edge of the target biological particle in the input image (i.e., the sequence image) to obtain a feature image.

[0095] Step S42: Introduce dense upsampling technology to improve the PSPNet semantic segmentation model. Use the training set to train the improved PSPNet semantic segmentation model to obtain the trained PSPNet semantic segmentation model, and use the trained PSPNet semantic segmentation model as a feature processing sub-module.

[0096] The segmentation and recognition model uses a feature processing submodule to perform four convolution operations of different sizes (1×1, 2×2, 3×3, and 6×6) on the feature image to extract feature information at different scales.

[0097] Subsequently, the feature processing submodule performs dense upsampling on the feature image after the convolution operation to obtain the output result of the feature processing submodule, so that the feature image after the convolution operation is restored to the same size as the input image (i.e. the sequence image).

[0098] The cross-entropy loss function of the improved PSPNet semantic segmentation model during training is as follows:

[0099]

[0100] In the formula, L represents the cross-entropy loss function of the improved PSPNet semantic segmentation model, N represents the number of samples, j represents the sample index; M represents the number of sample label categories, and c represents the index of the sample label category; y jc This represents the true label of the j-th sample belonging to the label category of the c-th sample, where y jc The value of p is either 0 or 1; jc This represents the predicted probability that the j-th sample belongs to the label category of the c-th sample.

[0101] Step S43: Finally, the segmentation and recognition model uses a convolutional layer to further fuse the output results of the edge detection module (i.e., edge feature map) and the output results of the feature processing submodule (i.e., feature image after dense upsampling) to obtain the output image, thereby achieving accurate identification and segmentation of biological particles in the sequence image.

[0102] In some specific embodiments, after step S4, the trained segmentation and recognition model is tested and evaluated using a test set to obtain the final segmentation and recognition model and the final output image; specifically including:

[0103] The trained segmentation and recognition model was validated and evaluated using pixel principal accuracy (PA) and mean intersection-over-union ratio (mIoU) as evaluation criteria to obtain the final segmentation and recognition model for carbonate rock particles. In this embodiment, the PA and mIoU values ​​on the dataset were 0.841 and 0.659, respectively.

[0104] The evaluation criterion formula for the segmentation and recognition model is as follows:

[0105]

[0106] Where PA represents the pixel main accuracy of the segmentation and recognition model, P i′j′ Let represent the number of samples of class i′ whose label is predicted to be of class j′, k represent the number of sample label classes, mIoU represent the model's average intersection-over-union ratio, and i′ and j′ both represent the index of the sample label class. P i′i′ P represents the number of pixels that were correctly classified. j′i′ This represents the number of samples of type j' whose label category is predicted to be the same as the sample of type i.

[0107] Based on the same inventive concept, embodiments of the present invention disclose a biological particle segmentation and recognition system, such as... Figure 7 As shown, it includes:

[0108] A collection and processing unit is used to collect sequential images of rock thin sections and preprocess the sequential images;

[0109] The fusion unit is used to perform feature fusion on the preprocessed sequence images to obtain a fused image;

[0110] An edge detection unit is used to perform edge detection on the fused image, extract the initial edge information of the target biological particles, and obtain an edge feature map;

[0111] The identification and segmentation unit is used to accurately identify and segment target biological particles in the sequence images based on edge feature maps and using the improved PSPNet semantic segmentation model.

[0112] Regarding the system in the above embodiments, the specific manner in which each unit module performs operations has been described in detail in the embodiments related to the method, and will not be elaborated here.

[0113] Based on the same inventive concept, embodiments of the present invention also provide an electronic device, the structure of which is as follows: Figure 8 As shown, it includes a memory, a processor, and a computer program stored in the memory. The processor executes the computer program or instructions to implement the aforementioned biological particle segmentation and recognition method.

[0114] Based on the same inventive concept, embodiments of the present invention also provide a computer storage medium storing a computer program or instructions, which, when executed by a processor, implements the aforementioned biological particle segmentation and recognition method.

[0115] Based on the same inventive concept, embodiments of the present invention also provide a computer program product, including a computer program or instructions, which, when executed by a processor, implement the aforementioned biological particle segmentation and recognition method.

[0116] An embodiment of the present invention for model validation:

[0117] A series of images of carbonate rock thin sections were captured using a polarizing microscope, and a dataset was created. The dataset contains 200 sets of images of carbonate rock thin sections. Each set of images consists of 6 orthogonally polarized images and 1 single-section image, and each image is 5333×4719 pixels in size.

[0118] As shown in the figure, within the range of 0° to 90°, one orthogonal polarized light image is acquired every 15°, resulting in a sequence of all orthogonal polarized light images. Then, relevant experts annotate the sequence of image data to obtain... Figure 3 The sample label example shown is illustrated in Table 2 below, where the biological label information (i.e., biological particle outline information and biological category) is as follows.

[0119] Table 2 Label information for carbonate rock thin sections

[0120] category RGB Label category RGB Label background (0,0,0) 0 Algae debris (215,215,0) 6 Unknown raw material (255,128,128) 1 Echinoderms (148,0,211) 7 bivalves (0,192,0) 2 Ostracods (255,0,255) 8 Foraminifera (255,255,0) 3 Gastropods (139,0,139) 9 Brachiopods (0,255,255) 4 cement (255,255,240) 10 sponge (139,139,0) 5

[0121] This invention selects three semantic segmentation network models—DeepLabv3+, SegNet, and FCN—for comparative experiments to verify the comprehensive performance of the segmentation and recognition model provided by this invention. A Python 3.9 compilation environment was used, built on PyTorch framework version 1.13.1. For the optimizer, stochastic gradient descent was adopted, with an initial learning rate of 0.001, a weight decay value of 0.0001, and a momentum factor of 0.9. During training, the batch size was set to 8, the maximum number of iterations was 1000, and the cross-entropy loss function was used.

[0122] Experimental results show that the segmentation and recognition model proposed in this invention is superior to commonly used deep learning segmentation and recognition schemes. Furthermore, the results of different algorithms are compared in Table 3, and as shown in Table 3, and... Figures 9(a) to 9(d) The data shows that the segmentation and recognition model of this invention has better segmentation performance than DeepLabv3+, SegNet and FCN semantic segmentation network models, and can accurately identify biological particles in carbonate rock thin sections.

[0123] Table 3. PA and mIoU values ​​in different neural network evaluations.

[0124]

[0125] Finally, it should be noted that the above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for segmenting and identifying biological particles, characterized in that, include: Collect a sequence of images of rock thin sections and preprocess the sequence of images; Feature fusion is performed on the preprocessed sequence images to obtain a fused image; Edge detection is performed on the fused image to extract the initial edge information of the target biological particles and obtain an edge feature map; Based on edge feature maps, the improved PSPNet semantic segmentation model is used to accurately identify and segment target biological particles in the sequence images.

2. The biological particle segmentation and identification method according to claim 1, characterized in that, Collect a sequence of images of rock thin sections, including: A sequence of images of rock thin sections was captured using a polarizing microscope; one rock thin section sample corresponds to a set of images, and a set of images includes single-polarized images and cross-polarized images. Sample labels and labeled particles are obtained by manually annotating the outline information of biological particles and biological categories in the sequence images; Among them, the bioparticle outline information and biological categories include background, unknown bioclasts, bivalves, foraminifera, brachiopods, sponges, algal debris, echinoderms, ostracods, gastropods, and cement. The sample label categories include bivalves, foraminifera, brachiopods, sponges, algal debris, echinoderms, ostracods, gastropods, and cement.

3. A biological particle segmentation and identification method according to claim 1 or 2, characterized in that, Preprocessing the sequence of images includes: The sequence images of rock thin sections and sample labels were uniformly cropped into sub-images of the same size to preserve the integrity of the grains in the images; The sequence images are subjected to quality screening and noise reduction processing to obtain the final sequence images; The final sequence of images is used as training samples, and a dataset is constructed based on the training samples and sample labels; Based on the number and proportion of samples corresponding to each sample label category in the dataset, and taking the number of unknown debris categories as a benchmark, sample data augmentation is performed on sample label categories that exceed the benchmark predetermined number, and training and test sets are divided.

4. The biological particle segmentation and identification method according to claim 3, characterized in that, The quality screening process for the sequence of images includes: Calculate the area ratio of the background and label particles in the sequence images respectively, and delete the sequence images in which the background area ratio exceeds a predetermined ratio to obtain the filtered sequence images.

5. The biological particle segmentation and identification method according to claim 3, characterized in that, The noise reduction process for the image sequence includes: Based on the spatial relationship between pixels in the sequence image and the similarity of pixels in the gray space, a bilateral filter is used to remove noise from the sequence image to obtain a denoised sequence image. The denoised image sequence is obtained using the following formula: In the formula, p(m,n) represents the denoised image sequence, m and n represent the coordinates of image points, f(k,l) represents the image sequence before denoising, k and l represent the center coordinates of the current image window, and w(m,n,k,l) ​​represents the bilateral filter convolution kernel function.

6. The biological particle segmentation and identification method according to claim 1, characterized in that, The preprocessed sequence images are fused using feature fusion to obtain a fused image, including: For each set of sequence images, including single-polarized images and cross-polarized images, calculate the Euclidean distance between each pixel of each pair of single-polarized images and cross-polarized images, and generate a difference image based on the Euclidean distance. The weight of each difference image is determined by the difference image method. The pixel value of each difference image is multiplied by the corresponding weight to obtain a weighted difference image. Then, each weighted difference image is accumulated one by one to obtain a fused image.

7. The biological particle segmentation and identification method according to claim 1, characterized in that, Edge detection is performed on the fused image to extract the initial edge information of the target biological particles and obtain an edge feature map, including: Based on the theory of convolutional neural networks, a lightweight edge detection network is constructed and trained to obtain the trained edge detection network. The fused image and the sequence image are input into the trained edge detection network. The trained edge detection network upsamples the fused image through deconvolution operation, so that the fused image is restored to the same size as the sequence image. The trained edge detection network performs edge detection on the fused image to obtain the initial contour information of the target biological particles. During the edge detection process, the outputs of each stage in the trained edge detection network are superimposed, and multi-channel feature merging is performed through a 1×1-1 convolutional layer to fuse the feature information extracted from all stages, and the output is an edge feature map. The lightweight edge detection network consists of five stages, each stage including multiple convolutional layers, and adjacent stages are connected by a pooling layer.

8. The biological particle segmentation and identification method according to claim 1, characterized in that, Based on edge feature maps, the improved PSPNet semantic segmentation model is used to accurately identify and segment target biological particles in the sequence images, including: ResNet50 is used as a feature extraction submodule, and ResNet50 is used to extract the feature information related to the edge of the target biological particles in the sequence image to obtain the feature image; Dense upsampling technique is introduced to improve the PSPNet semantic segmentation model, and the improved PSPNet semantic segmentation model is used as a feature processing submodule. The feature processing submodule performs convolution operations of different sizes on the feature images to extract feature information at different scales. The feature processing submodule performs dense upsampling on the feature image after the convolution operation, so that the feature image after the convolution operation is restored to the same size as the sequence image. The convolutional layer is used to further fuse the densely upsampled feature image and the edge feature map to obtain the output image.

9. A biological particle segmentation and recognition system, characterized in that, include: A collection and processing unit is used to collect sequential images of rock thin sections and preprocess the sequential images; The fusion unit is used to perform feature fusion on the preprocessed sequence images to obtain a fused image; An edge detection unit is used to perform edge detection on the fused image, extract the initial edge information of the target biological particles, and obtain an edge feature map; The identification and segmentation unit is used to accurately identify and segment target biological particles in the sequence images based on edge feature maps and using the improved PSPNet semantic segmentation model.

10. An electronic device, characterized in that, The device includes a memory, a processor, and a computer program stored in the memory, wherein the processor executes the computer program or instructions to implement a biological particle segmentation and identification method according to any one of claims 1-8.

11. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program or instructions, which, when executed by a processor, implement the biological particle segmentation and identification method according to any one of claims 1-8.

Images