Intelligent identification system and intelligent identification method for sandstone slice

Abstract

The invention provides an intelligent sandstone slice identification system and an intelligent sandstone slice identification method, and belongs to the technical field of geological exploration. In the system, a data analysis module integrates three core algorithms: a data enhancement module adopts an improved conditional cycle consistency generative adversarial network to solve the problem of scarcity of rock samples; the multi-modal feature understanding module fuses image visual features and petrology text features by using a cross-modal attention mechanism; and the intelligent identification module completes recognition and classification of mineral composition, structure construction and diagenesis in parallel based on a multi-task learning architecture. The method comprises a system training stage and a sample identification stage. According to the method, the model generalization ability is improved by generating a large number of synthetic samples, the geological rationality of the identification result is enhanced through multi-modal semantic fusion, efficient, accurate and objective automatic intelligent identification of the sandstone slice is realized, and a key technical support is provided for reservoir evaluation in the field of petroleum geology.

Description

Technical Field

[0001] This invention relates to the field of geological exploration technology, and in particular to an intelligent identification system and method for sandstone thin sections. Background Technology

[0002] Sandstone thin section identification is a fundamental and crucial task in petroleum geological research. Precise analysis of the mineral composition, structure, and diagenesis of sandstone provides direct geological evidence for reservoir evaluation, sedimentary environment reconstruction, and the study of hydrocarbon accumulation patterns. Traditional identification methods heavily rely on manual observation and description by geologists under a polarizing microscope. This process is not only time-consuming and labor-intensive, but the results are also easily influenced by personal experience, subjective judgment, and physical condition, making standardization and quantification difficult. This has become a technical bottleneck restricting large-scale, high-efficiency core analysis.

[0003] In recent years, with the development of computer vision technology, deep learning-based image recognition methods have been introduced into the field of rock thin section identification. These methods typically rely on convolutional neural networks to extract features and classify thin section images. However, such methods face two prominent challenges in practical applications: First, acquiring a large number of thin section image samples precisely annotated by geological experts is extremely costly and time-consuming, resulting in a limited scale of high-quality datasets available for training. This makes the models prone to weak generalization ability and unsatisfactory recognition accuracy due to insufficient training. Second, existing methods often focus on single-image visual information, failing to effectively integrate multimodal semantic information such as petrological knowledge, well logging data, and geological reports. This results in the identification system lacking geological logical reasoning ability, making it difficult to simulate the comprehensive judgment process of geological experts, and limiting its analytical depth and reliability. Summary of the Invention

[0004] In view of this, the present invention aims to propose an intelligent identification system and method for sandstone thin sections that can overcome the scarcity of samples and deeply integrate multi-source geological information, so as to achieve efficient, accurate and objective automated identification and analysis.

[0005] To achieve the above objectives, the technical solution of the present invention is as follows: an intelligent identification system for sandstone thin sections, comprising: The image acquisition module is used to acquire digital images of sandstone thin sections; The data processing module is communicatively connected to the image acquisition module and is used to receive and process the digital images acquired by the image acquisition module. The data storage module, which is connected in communication with the data processing module, is used to store the original image, the generated image, the model parameters, and the identification results. The data analysis module includes a data augmentation module, a multimodal feature understanding module, and an intelligent identification module; The data augmentation module employs an improved conditional cyclic consistency generative adversarial network to generate training samples based on the original sandstone thin section images. The multimodal feature understanding module uses a cross-modal attention mechanism to fuse the visual features of the sandstone thin section images with petrological knowledge text features to generate multimodal semantic representation vectors. The intelligent identification module adopts a multi-task learning architecture to perform parallel identification and classification of the mineral composition, structure, and diagenesis of the sandstone thin sections based on the multimodal semantic representation vectors. The results output module communicates with the data analysis module to visualize and output identification reports.

[0006] Furthermore, the image acquisition module includes a polarizing microscope unit, a digital imaging unit, and an illumination control unit; The polarizing microscope unit is equipped with a stage capable of 360° rotation and multiple achromatic objectives with magnification ranging from 4× to 100× and a maximum numerical aperture of 0.95. The digital imaging unit is coupled to the eyepiece port of the polarizing microscope unit and is equipped with a CMOS image sensor with 20 million effective pixels and 16-bit dynamic range. The lighting control unit includes a full-spectrum LED transmission light source and a control circuit. The color temperature of the full-spectrum LED transmission light source is continuously adjustable in the range of 2800K to 7000K, and it can work in conjunction with the polarizer and analyzer of the polarizing microscope unit. Through programmed control, it can automatically switch between multiple observation modes such as single polarization, cross polarization and conophthalmology. The image acquisition module, under the control of the data processing module, can perform fully automated operations from stage movement, focusing, illumination mode switching, polarizing filter rotation to image capture.

[0007] Furthermore, the data processing module executes image preprocessing algorithms, employing an algorithm based on SIFT feature detection and FLANN matcher to spatially align sandstone thin section images acquired under different optical modes within the same field of view. It applies a multi-scale Retinex algorithm to perform color constancy processing on the images to eliminate color deviations caused by changes in illumination conditions. A contrast-limited adaptive histogram equalization method is used to enhance the microscopic contrast of the rock structure. A non-local mean denoising algorithm is used to eliminate random noise introduced during image acquisition. Image pixel values ​​are normalized to the [0,1] interval and uniformly scaled to a fixed size, generating a preprocessed image that conforms to the neural network input specifications.

[0008] Furthermore, the data augmentation module is implemented based on an improved conditional cycle-consistent generative adversarial network, which includes a generator G, a discriminator D, and a geological feature extractor F. Generator G is used to generate an image X from source domain A.A And random noise vector z, under the condition y in the target domain B Under the guidance of [the relevant authority], generate images in the style of the target domain. ; The discriminator D adopts a multi-scale PatchGAN architecture. Its input is a real image or a generated image and conditional labels, and its output is the probability distribution of each local region of the image being real. The geological feature extractor F is a pre-trained deep convolutional network used to extract high-dimensional geological semantic features from images. The optimization objective of the data augmentation module is composed of the following loss functions: The adversarial loss is expressed as a least squares loss function, which is shown below: ; Where E represents the mathematical expectation, Meaning: Calculate the expectation of a sample drawn from the real data distribution; Represents: Calculate the expectation of the joint distribution of x and y drawn from the real data distribution and the noise z drawn from the random noise distribution; D(x|y) is the output probability of the discriminator D judging the real image x as true under given condition y; G(x,z|y) is the synthetic image generated by the generator G with the source image x, noise z and condition y as input; D(G(x,z|y)|y) is the output probability of the discriminator judging the generated image G(x,z|y) as true. The cycle consistency loss is defined as: ; G(x,z|y A To transform image x from the original domain to conditional y A Defined style; G(G(x,z|y) A ),z|y B To convert the image generated in the previous step back to the original condition y B ; Geological feature matching loss is defined as: ; F(x) is the high-dimensional semantic feature vector obtained by the feature extractor F from the real image x; F(G(x,z|y)) is the high-dimensional semantic feature vector obtained by the generated image G(x,z|y) through the same feature extractor F. The total loss function is: ; in , , These are the weighting coefficients for each loss term.

[0009] Furthermore, the multimodal feature understanding module adopts a deep neural network architecture based on a cross-modal attention mechanism, including a visual feature extraction network, a text feature extraction network, and a cross-modal fusion network; The visual feature extraction network adopts a pre-trained ResNet-101 architecture. It takes a pre-processed sandstone thin section image I as input and extracts multi-scale visual feature maps through multi-layer convolution operations. ,in , indicating visual features; The text feature extraction network employs the BERT model, taking a textual description of petrology knowledge T as input; and extracts text feature sequences through a multi-layer Transformer encoder. ,in , representing text features; The cross-modal fusion network employs a dual-path cross-attention mechanism, where the attention weights from image to text are calculated as follows: The attention weights for text-to-image processing are calculated as follows: ; in, , These are the parameter matrices; T represents the matrix transpose; α ij β represents the relevance weight between the i-th image region and the j-th text term, indicating the degree to which text term j should be considered when interpreting image region i; ji Let be the relevance weight between the j-th text term and the i-th image region. It represents the degree to which image region i should be considered when interpreting text term j. The optimization objective of the multimodal feature understanding module includes cross-modal contrastive loss, which is calculated as follows: ; in, τ represents the cosine similarity of positive sample pairs in the common space; f and g are the feature projection functions; τ is the temperature hyperparameter; and N represents the total number of samples in a training batch.

[0010] Furthermore, the intelligent identification module adopts a deep neural network architecture based on multi-task learning, including a feature input layer, a multi-task learning layer, and a result output layer; The feature input layer receives the fused semantic vector output from the multimodal feature understanding module. ;in, The dimension of the fusion feature; The multi-task learning layer contains three parallel sub-network branches: The first sub-network branch is the mineral composition identification branch, which uses a softmax classifier and outputs the mineral category probability distribution as follows: in, This is the weight matrix. Number of mineral categories; This is a fused semantic representation vector from the multimodal feature understanding module. It is the bias vector; The second sub-network branch is the structure construction analysis branch, which uses a multilayer perceptron and outputs a structure feature vector. The structure feature vector is: ; Where σ is the sigmoid activation function; As weight, It is the bias vector; The third sub-network branch is the diagenesis identification branch, which uses a multi-label classifier to output the probability of diagenesis type. The probability of diagenesis type is: ; in, For bias vectors, As weight; The total loss function of the intelligent identification module is the weighted sum of the losses of each task, and the total loss function is shown in the following formula: ; Where α, β, γ are task weight coefficients, λ is a regularization coefficient, and Θ is a model parameter.

[0011] Furthermore, the result output module includes a visualization unit and a report generation unit; The visualization unit is used to display the identification results in a multi-dimensional way, including a mineral composition distribution map, a structural annotation map, and a diagenetic analysis map; The report generation unit is used to automatically generate a standardized identification report, which includes basic sample information, image display, data tables, and geological analysis conclusions. The result output module also provides a data interface that supports outputting identification results in image, structured data, and document formats.

[0012] The intelligent identification method employing any one of the above-mentioned intelligent identification systems for sandstone thin sections includes a system training phase and a sample identification phase: The system training phase includes the following steps: Step S1: Collect sandstone thin section samples containing various known mineral compositions, structures, and diagenetic processes to form the original training sample set; Step S2: Obtain the digital image sequence of each sample in the original training sample set under multimodal optical conditions through the image acquisition module, and label the corresponding petrological knowledge text descriptions of the digital images; Step S3: The image obtained in step S2 is preprocessed using the data processing module to generate standardized image data; Step S4: Invoke the data augmentation module, and use the normalized image data obtained in step S3 as a basis to train the improved conditional cycle consistency generative adversarial network to generate synthetic sandstone thin section images, which together with the original images constitute the augmentation training set. Step S5: Using the enhanced training set and the corresponding text description, train the multimodal feature understanding module and the intelligent identification module until the model converges, and obtain the final intelligent identification model for sandstone thin sections. The sample identification stage includes the following steps: Step U1: Acquire digital images of the unknown sandstone thin section to be identified through the image acquisition module; Step U2: Use the data processing module to perform the same preprocessing on the image obtained in step U1 as in the training phase; Step U3: Input the preprocessed image into the trained multimodal feature understanding module and intelligent identification module, and automatically output the identification results of the mineral composition, structure and diagenesis of the unknown sandstone thin section.

[0013] Furthermore, step S4 includes: S4.1 Initialize the network parameters of generator G, discriminator D and geological feature extractor F; S4.2 Input the original training images and their conditional labels into the network, fix the generator G, and update the parameters of the discriminator D to maximize the ability to distinguish between real images and generated images. S4.3, with the discriminator D fixed, update the parameters of the generator G to minimize the total loss function L. total This prompts the generator to produce composite images that are both realistic and consistent with geological patterns; S4.4, repeat steps S4.2 and S4.3, and use geological features to match the loss L. geo Constraint generation process until network convergence; S4.5, save the trained generator G, which is used to generate synthetic images in batches to expand the training set.

[0014] Furthermore, the process of training the multimodal feature understanding module and the intelligent identification module in step S5 includes the following steps: S5.1, Multimodal Feature Understanding Module Pre-training: Using an enhanced training set, the weights of the visual feature extraction network and the text feature extraction network are fixed, and the cross-modal fusion network is trained separately. The optimization objective is to minimize the cross-modal contrast loss L. cmc This brings image features and text features from the same sample closer together in the semantic space; S5.2, Intelligent Identification Module Pre-training: Freeze the parameters of the trained multimodal feature understanding module, and convert the multimodal semantic representation vector h... fusion As input, the multi-task learning layer of the intelligent identification module is trained separately, with the optimization objective being to minimize the total loss function L. total ; S5.3, End-to-End Joint Fine-Tuning: Unfreeze the parameters of the multimodal feature understanding module, connect the multimodal feature understanding module and the intelligent identification module into a whole network, using the total loss function L... total As a global optimization target, an end-to-end approach is used to jointly fine-tune all network parameters; S5.4, Model Validation and Selection: Evaluate the overall performance of the jointly fine-tuned model on an independent validation set. When the overall performance index of the model on the validation set no longer improves for several consecutive training cycles, stop training and save the parameters of the model with the best performance as the final evaluation model.

[0015] Compared with existing technologies, the intelligent identification system and method for sandstone thin sections described in this invention have the following advantages: (1) Effectively solves the problem of small sample learning and significantly improves the model's generalization ability and identification accuracy: This invention introduces an improved conditional cycle consistency generative adversarial network, which can generate a large number of synthetic samples rich in geological regularities and visually realistic using limited original thin section images. This data augmentation process fundamentally expands the scale and diversity of the training set, enabling the subsequent deep neural network model to be fully trained, effectively overcoming the model overfitting problem caused by sample scarcity, thereby significantly improving the accuracy of common mineral identification while also enhancing the identification ability of rare minerals and special structures; (2) Deep fusion of multimodal geological information is achieved, making the identification results more geologically logical and interpretable: This invention breaks through the limitations of traditional methods that rely solely on image information. Through a cross-modal attention mechanism, the visual features of thin section images are aligned and fused with the textual features of professional petrological knowledge at the semantic level. This fusion enables the system to make inferences and judgments by combining visual evidence and theoretical knowledge, just like a geological expert. It can not only give the conclusion of "what" but also provide semantic support of "why" through feature association, which significantly improves the professionalism and credibility of the identification results. (3) The entire process is automated and intelligent, greatly improving the efficiency and objectivity of identification: This invention integrates the entire process from automatic image acquisition, preprocessing, data enhancement to intelligent identification and report generation. This system can replace a large amount of repetitive manual observation and recording work, freeing geological experts from heavy labor and shortening the identification time from several hours to minutes. At the same time, the identification process based on a unified algorithm model avoids the bias caused by human subjective factors, ensuring the objectivity, standardization and repeatability of the identification results, and providing efficient and reliable technical support for petroleum geological research; Attached Figure Description

[0016] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings: Figure 1 This is a schematic diagram of the overall architecture of the system described in this invention; Figure 2 This is a flowchart of the training phase of the system of the present invention; Figure 3 This is a flowchart of the sample identification stage of the present invention; Figure 4 For acquiring multi-angle polarized single-particle images; Figure 5 Synthesized multiple sandstone thin section images generated for the data augmentation module. Detailed Implementation

[0017] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0018] The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0019] like Figures 1-5 As shown, this invention is an intelligent identification system for sandstone thin sections, comprising: Image acquisition module, used to acquire high-resolution digital images of sandstone thin sections; The data processing module is communicatively connected to the image acquisition module and is used to receive and process the digital images acquired by the image acquisition module. The data storage module, which is connected in communication with the data processing module, is used to store the original image, the generated image, the model parameters, and the identification results. The data analysis module includes a data augmentation module, a multimodal feature understanding module, and an intelligent identification module; The data augmentation module employs an improved conditional cyclic consistency generative adversarial network to generate training samples based on a limited set of original sandstone thin section images. The multimodal feature understanding module uses a cross-modal attention mechanism to fuse visual features of sandstone thin section images with petrological knowledge textual features to generate multimodal semantic representation vectors. The intelligent identification module adopts a multi-task learning architecture to perform parallel identification and classification of the mineral composition, structure, and diagenesis of sandstone thin sections based on the multimodal semantic representation vectors. The results output module communicates with the data analysis module to visualize and output identification reports.

[0020] The image acquisition module includes a polarizing microscope unit, a digital imaging unit, and an illumination control unit; The polarizing microscope unit is equipped with a stage capable of 360° rotation and multiple achromatic objectives with magnifications ranging from 4× to 100× and numerical apertures up to 0.95. In the description of this invention, unless otherwise stated, "multiple" means two or more.

[0021] The digital imaging unit is coupled to the eyepiece port of the polarizing microscope unit and is equipped with a CMOS image sensor with 20 million effective pixels and 16-bit dynamic range. The lighting control unit includes a full-spectrum LED transmission light source and a control circuit. The color temperature of the full-spectrum LED transmission light source is continuously adjustable in the range of 2800K to 7000K, and it can work in conjunction with the polarizer and analyzer of the polarizing microscope unit. Through programmed control, it can automatically switch between various observation modes such as single polarization, cross polarization and conophthalmology. The image acquisition module, under the control of the data processing module, can perform fully automated operations from stage movement, focusing, illumination mode switching, polarizing filter rotation to image capture.

[0022] The data processing module executes image preprocessing algorithms, employing an algorithm based on SIFT feature detection and FLANN matcher to accurately align sandstone thin section images of the same field of view acquired under different optical modes. It applies a multi-scale Retinex algorithm to perform color constancy processing on the images to eliminate color deviations caused by changes in lighting conditions. A contrast-limited adaptive histogram equalization method is used to enhance the microscopic contrast of the rock structure. A non-local mean denoising algorithm is used to eliminate random noise introduced during image acquisition. Image pixel values ​​are normalized to the [0,1] interval and uniformly scaled to a fixed size, generating a preprocessed image that conforms to the neural network input specifications.

[0023] The data augmentation module is implemented based on an improved conditional cyclic consistency generative adversarial network, which includes a generator G, a discriminator D, and a geological feature extractor F. In the binary framework of Cyclic Consistency Generative Adversarial Networks (CGNs), a pre-trained geological feature extractor is introduced as a fixed semantic evaluator, and a "geological feature matching loss" is added accordingly. This loss, together with the original adversarial loss and cycle consistency loss, constitutes a new joint optimization objective for geological image generation. This improvement ensures that the large-scale synthetic samples generated by the data augmentation module are not only realistic at the pixel level, but also have high fidelity at the level of geological semantic features, which are crucial for intelligent identification, thus providing a high-quality data foundation for solving the problem of few-shot learning.

[0024] Generator G is used to generate an image X from source domain A. A And random noise vector z, under the condition y in the target domain B Under the guidance of [the relevant authority], generate images in the style of the target domain. ; The discriminator D adopts a multi-scale PatchGAN architecture. Its input is a real image or a generated image and conditional labels, and its output is the probability distribution of each local region of the image being real. The geological feature extractor F is a pre-trained deep convolutional network used to extract high-dimensional geological semantic features from images. The optimization objective of the data augmentation module is composed of the following loss functions: The adversarial loss is expressed as a least squares loss function, which is shown below: ; Where E represents the mathematical expectation, Meaning: Calculate the expectation of a sample drawn from the real data distribution; Represents: Calculate the expectation of the joint distribution of x and y drawn from the real data distribution and the noise z drawn from the random noise distribution; D(x|y) is the output probability of the discriminator D judging the real image x as true under given condition y; G(x,z|y) is the synthetic image generated by the generator G with the source image x, noise z and condition y as input; D(G(x,z|y)|y) is the output probability of the discriminator judging the generated image G(x,z|y) as true. The cycle consistency loss is defined as: ; G(x,z|y A To transform image x from the original domain to conditional y A Defined style; G(G(x,z|y) A ),z|y B To convert the image generated in the previous step back to the original condition y B ; Geological feature matching loss is defined as: ; F(x) is the high-dimensional semantic feature vector obtained by the feature extractor F from the real image x; F(G(x,z|y)) is the high-dimensional semantic feature vector obtained by the generated image G(x,z|y) through the same feature extractor F. The total loss function is: ; in , , These are the weighting coefficients for each loss term.

[0025] The multimodal feature understanding module adopts a deep neural network architecture based on a cross-modal attention mechanism, including a visual feature extraction network, a text feature extraction network, and a cross-modal fusion network; The visual feature extraction network adopts a pre-trained ResNet-101 architecture. It takes a pre-processed sandstone thin section image I as input and extracts multi-scale visual feature maps through multi-layer convolution operations. ,in , indicating visual features; The text feature extraction network employs the BERT model, taking a textual description of petrology knowledge T as input; and extracts text feature sequences through a multi-layer Transformer encoder. ,in , representing text features; The cross-modal fusion network employs a dual-path cross-attention mechanism, where the attention weights from image to text are calculated as follows: The attention weights for text-to-image processing are calculated as follows: ; in, , These are the parameter matrices; T represents the matrix transpose; α ij β represents the relevance weight between the i-th image region and the j-th text term, indicating the degree to which text term j should be considered when interpreting image region i; ji Let be the relevance weight between the j-th text term and the i-th image region. It represents the degree to which image region i should be considered when interpreting text term j. The optimization objective of the multimodal feature understanding module includes cross-modal contrastive loss, which is calculated as follows: in, τ represents the cosine similarity of positive sample pairs in the common space; f and g are the feature projection functions; τ is the temperature hyperparameter; and N represents the total number of samples in a training batch.

[0026] The intelligent identification module adopts a deep neural network architecture based on multi-task learning, including a feature input layer, a multi-task learning layer, and a result output layer. The feature input layer receives the fused semantic vector output from the multimodal feature understanding module. ;in, The dimension of the fusion feature; The multi-task learning layer contains three parallel sub-network branches: The first sub-network branch is the mineral composition identification branch, which uses a softmax classifier and outputs the mineral category probability distribution as follows: in, This is the weight matrix. Number of mineral categories; This is a fused semantic representation vector from the multimodal feature understanding module. It is the bias vector; The second sub-network branch is the structure construction analysis branch, which uses a multilayer perceptron and outputs a structure feature vector. The structure feature vector is: ; Where σ is the sigmoid activation function; As weight, It is the bias vector; The third sub-network branch is the diagenesis identification branch, which uses a multi-label classifier to output the probability of diagenesis type. The probability of diagenesis type is: ; in, For bias vectors, As weight; The total loss function of the intelligent identification module is the weighted sum of the losses of each task, and the total loss function is shown in the following formula: ; Where α, β, γ are task weight coefficients, λ is a regularization coefficient, and Θ is a model parameter.

[0027] The result output module includes a visualization unit and a report generation unit; The visualization unit is used to display the identification results in a multi-dimensional way, including a mineral composition distribution map, a structural annotation map, and a diagenetic analysis map; The report generation unit is used to automatically generate a standardized identification report, which includes basic sample information, image display, data tables, and geological analysis conclusions. The result output module also provides a data interface that supports outputting identification results in image, structured data, and document formats.

[0028] This invention also provides an intelligent identification method for sandstone thin sections, including a system training phase and a sample identification phase: The system training phase includes the following steps: Step S1: Collect sandstone thin section samples containing various known mineral compositions, structures, and diagenetic processes to form the original training sample set; Step S2: Obtain a high-resolution digital image sequence of each sample in the original training sample set under multimodal optical conditions through the image acquisition module, and label the corresponding petrological knowledge text descriptions of the digital images. Step S3: The image obtained in step S2 is preprocessed using the data processing module to generate standardized image data; Step S4: Invoke the data augmentation module. Based on the normalized image data obtained in step S3, train the improved conditional cycle consistency generative adversarial network to generate large-scale, high-quality synthetic sandstone thin section images, which together with the original images constitute the augmentation training set. Step S5: Using the enhanced training set and the corresponding text description, train the multimodal feature understanding module and the intelligent identification module until the model converges, and obtain the final intelligent identification model for sandstone thin sections. The sample identification stage includes the following steps: Step U1: Acquire a high-resolution digital image of the unknown sandstone thin section to be identified through the image acquisition module; Step U2: Use the data processing module to perform the same preprocessing on the image obtained in step U1 as in the training phase; Step U3: Input the preprocessed image into the trained multimodal feature understanding module and intelligent identification module, and automatically output the identification results of the mineral composition, structure and diagenesis of the unknown sandstone thin section.

[0029] Specifically, step S4 includes: S4.1 Initialize the network parameters of generator G, discriminator D and geological feature extractor F; S4.2 Input the original training images and their conditional labels into the network, fix the generator G, and update the parameters of the discriminator D to maximize the ability to distinguish between real images and generated images. S4.3, with the discriminator D fixed, update the parameters of the generator G to minimize the total loss function L. totalThis prompts the generator to produce composite images that are both realistic and consistent with geological patterns; S4.4, repeat steps S4.2 and S4.3, and use geological features to match the loss L. geo Constraint generation process until network convergence; S4.5, save the trained generator G, which is used to generate synthetic images in batches to expand the training set.

[0030] The process of training the multimodal feature understanding module and the intelligent identification module in step S5 includes the following steps: S5.1, Multimodal Feature Understanding Module Pre-training: Using an enhanced training set, the weights of the visual feature extraction network and the text feature extraction network are fixed, and the cross-modal fusion network is trained separately. The optimization objective is to minimize the cross-modal contrast loss L. cmc This brings image features and text features from the same sample closer together in the semantic space; S5.2, Intelligent Identification Module Pre-training: Freeze the parameters of the trained multimodal feature understanding module, and convert the multimodal semantic representation vector h... fusion As input, the multi-task learning layer of the intelligent identification module is trained separately, with the optimization objective being to minimize the total loss function L. total ; S5.3, End-to-End Joint Fine-Tuning: Unfreeze the parameters of the multimodal feature understanding module, connect the multimodal feature understanding module and the intelligent identification module into a whole network, using the total loss function L... total As a global optimization target, an end-to-end approach is used to jointly fine-tune all network parameters; S5.4, Model Validation and Selection: Evaluate the overall performance of the jointly fine-tuned model on an independent validation set. When the overall performance index of the model on the validation set no longer improves for several consecutive training cycles, stop training and save the parameters of the model with the best performance as the final evaluation model.

[0031] Example 1 See Figures 1 to 5 This invention provides an intelligent identification system for sandstone thin sections, including an image acquisition module, a data processing module, a data storage module, a data analysis module, and a result output module; Image acquisition module, used to acquire high-resolution digital images of sandstone thin sections; The data processing module is communicatively connected to the image acquisition module and is used to receive and process the digital images acquired by the image acquisition module. The data storage module, which is connected in communication with the data processing module, is used to store the original image, the generated image, the model parameters, and the identification results. The data analysis module includes a data augmentation module, a multimodal feature understanding module, and an intelligent identification module; The data augmentation module employs an improved conditional cyclic consistency generative adversarial network to generate training samples based on a limited set of original sandstone thin section images. The multimodal feature understanding module uses a cross-modal attention mechanism to fuse visual features of sandstone thin section images with petrological knowledge textual features to generate multimodal semantic representation vectors. The intelligent identification module adopts a multi-task learning architecture to perform parallel identification and classification of the mineral composition, structure, and diagenesis of sandstone thin sections based on the multimodal semantic representation vectors. The results output module communicates with the data analysis module to visualize and output identification reports.

[0032] System construction and sample preparation The hardware environment required for this system implementation includes a workstation equipped with an Intel Xeon Gold 6248R processor, 256GB of RAM, and four NVIDIA RTX A6000 graphics cards (48GB of video memory each). The software environment consists of an Ubuntu 20.04 operating system, the PyTorch 1.12 deep learning framework, and is developed based on Python 3.9.

[0033] Furthermore, the image acquisition module includes a polarizing microscope unit, a digital imaging unit, and an illumination control unit; The polarizing microscope unit is equipped with a stage that can rotate 360° and multiple achromatic objectives, with magnification ranging from 4× to 100× and a numerical aperture of up to 0.95. The digital imaging unit is coupled to the eyepiece port of the polarizing microscope, and uses a CMOS image sensor with 20 million effective pixels and 16-bit dynamic range. The lighting control unit includes a full-spectrum LED transmitted light source and control circuit. The color temperature of the light source is continuously adjustable in the range of 2800K to 7000K, and it can work in conjunction with the microscope's polarizer and analyzer. Through programmed control, it can automatically switch between various observation modes such as single polarization, cross polarization, and conophthalmology. The image acquisition module is connected to the data processing module and can perform fully automated operations from stage movement, focusing, illumination mode switching, polarizing filter rotation to image capture under the control of the data processing module.

[0034] Polarizing microscope unit model: Olympus BX53M industrial-grade research polarizing microscope.

[0035] Objective lens configuration: 4× Plan Achromatic Objectives (Numerical Aperture NA=0.10, Working Distance WD=30mm); 10× plan achromatic objective lens (numerical aperture NA=0.25, working distance WD=10.6mm); 20× plan achromatic objective lens (numerical aperture NA=0.40, working distance WD=3.0mm); 50× plan achromatic objective lens (numerical aperture NA=0.75, working distance WD=0.38mm); 100× oil immersion objective (numerical aperture NA=0.95, working distance WD=0.20mm); Stage: Professional-grade 360° rotating stage, 180mm in diameter, with vernier scale (minimum reading 0.1°). Focusing mechanism: coarse and fine coaxial focusing, minimum reading 1μm, stroke 25mm; Polarizing system: Both the polarizer and analyzer are rotatable with a rotation accuracy of 0.5°; Digital Imaging Unit: Camera Model: Canon EOS R5 Scientific CMOS Camera, Sensor Size: Full-frame 36×24mm, Effective Pixels: 45 million (8192×5464 pixels), Pixel Size: 4.4μm×4.4μm, Dynamic Range: 16-bit ADC, Signal-to-Noise Ratio >65dB, Frame Rate: Up to 20fps at full resolution, Interface: USB 3.2 Gen 2 Type-C, Cooling System: Semiconductor Cooling, Operating Temperature -10℃ to 45℃.

[0036] Lighting control unit: Light source: Lumencor SOLA SE6 full-spectrum LED light source, color temperature range: 2800K-7000K continuously adjustable, color rendering index: CRI>95, brightness stability: <0.5% fluctuation (after 30 minutes of warm-up), control interface: dual interface of RS-232 and Ethernet, polarizer: Glan-Taylor prism, extinction ratio>100,000:1, controller: 32-bit ARM Cortex-M7 processor, response time<1ms.

[0037] Automatic control parameters: Stage movement accuracy: X / Y axis ±1μm, repeatability ±2μm, rotary stage control accuracy: ±0.1°, repeatability ±0.2°, autofocus system: based on contrast detection, focusing accuracy ±0.5μm, image acquisition interval: minimum 100ms, data transmission rate: maximum 5Gbps, operating temperature: 15℃-30℃, humidity <80%RH.

[0038] System integration parameters: Control software: Dedicated control program developed based on Python 3.9; Communication protocol: MODBUS TCP / IP and USB3 Vision; Power requirements: AC 220V±10%, 50 / 60Hz, maximum power consumption 500W; Dimensions: 600mm×500mm×400mm (L×W×H); Weight: Approximately 35kg; Protection rating: IP20; Suitable for laboratory environments. The configuration ensures high-quality digital images in various optical modes, providing a reliable image data foundation for subsequent intelligent identification. All parameters have been verified through actual testing to ensure system stability and repeatability.

[0039] Sandstone samples were collected from typical oil and gas-bearing basins such as the Bohai Bay Basin, Ordos Basin, and Junggar Basin, covering various sedimentary environments including braided river deltas, fan deltas, and littoral-shallow marine facies. A total of 300 sandstone thin sections with known geological properties were prepared as the original training sample set, including: 120 quartz sandstone samples (quartz content >90%) 100 feldspar sandstone samples (feldspar content >25%) 80 rock fragment sandstone samples (rock fragment content >25%) Each thin section sample was imaged using a polarizing microscope following a standard procedure: First, in single-polarization mode, the stage was rotated at 10-degree intervals to acquire image sequences at 36 angles from 0 to 350 degrees; then, in orthogonal polarization mode, the same 36-angle image sequences were acquired; finally, in conical polarization mode, five sets of interferograms at key angles were acquired. Each field of view was acquired separately using 4×, 10×, 20×, and 40× objectives to ensure complete coverage of information from macroscopic structure to microscopic features.

[0040] Furthermore, the data processing module executes image preprocessing algorithms, employing an algorithm based on SIFT feature detection and FLANN matcher to precisely align sandstone thin section images of the same field of view acquired under different optical modes. It applies a multi-scale Retinex algorithm to perform color constancy processing on the images to eliminate color deviations caused by changes in illumination conditions. A contrast-limited adaptive histogram equalization method is used to enhance the microscopic contrast of the rock structure. A non-local mean denoising algorithm is used to eliminate random noise introduced during image acquisition. Image pixel values ​​are normalized to the [0,1] interval and uniformly scaled to a fixed size, generating a preprocessed image that conforms to the neural network input specifications.

[0041] The acquired raw images undergo the following standardized processing flow: Multimodal registration: SIFT feature detection algorithm is used to extract key points of the image, and FLANN matcher is used to achieve spatial alignment of images under different optical modes, with registration error controlled within 2 pixels.

[0042] Color correction: The multi-scale Retinex algorithm is used for color constancy processing, and the Gaussian wrapping scale is set to three scales: 15, 80, and 250 to eliminate color shift caused by uneven lighting.

[0043] Contrast enhancement: The Limiting Contrast Adaptive Histogram Equalization (CLAHE) method is adopted, with the slice size set to 8×8 and the contrast limit set to 2.0, to enhance the microscopic contrast of mineral boundaries.

[0044] Noise suppression: A nonlocal mean denoising algorithm is used, with a search window of 21×21, a similarity window of 7×7, and a filtering parameter of h=10, which effectively suppresses random noise during image acquisition.

[0045] Size normalization: All images were uniformly scaled to 512×512 pixels, and pixel values ​​were normalized to the [0,1] range. The processed image dataset contains approximately 50,000 high-quality labeled images, which serve as the basis for subsequent model training.

[0046] Furthermore, the data augmentation module is implemented based on an improved conditional cyclic consistency generative adversarial network, which includes a generator G, a discriminator D, and a geological feature extractor F; The generator G is used to process the image from source domain A. and random noise vector Under the conditions of the target domain Under the guidance of [the relevant authority], generate images in the style of the target domain. ; The discriminator D adopts a multi-scale PatchGAN architecture. Its input is a real image or a generated image and conditional labels, and its output is the probability distribution of each local region of the image being real. The geological feature extractor F is a pre-trained deep convolutional network used to extract high-dimensional geological semantic features from images. The optimization objective of the data augmentation module is composed of the following loss functions: The adversarial loss is expressed as a least-squares loss function: ; D(x|y) is the output probability of the discriminator D judging the real image x as true under given condition y; G(x,z|y) is the synthetic image generated by the generator G with the source image x, noise z and condition y as input; D(G(x,z|y)|y) is the output probability of the discriminator judging the generated image G(x,z|y) as true. The cycle consistency loss is defined as: ; G(x,z|y A To transform image x from the original domain to conditional y A Defined style; G(G(x,z|y) A ),z|y B To convert the image generated in the previous step back to the original condition y B ; Geological feature matching loss is defined as: ; F(x) is the high-dimensional semantic feature vector obtained by the feature extractor F from the real image x; F(G(x,z|y)) is the high-dimensional semantic feature vector obtained by the generated image G(x,z|y) through the same feature extractor F. The total loss function is: ; in , , These are the weighting coefficients for each loss term.

[0047] E represents the mathematical expectation. This means calculating the expectation of a sample drawn from the real data distribution. In other words, it's the average performance of all possible real images x and their corresponding conditional labels y. This means finding the expectation of the joint distribution of x and y drawn from the real data distribution and the noise z drawn from the random noise distribution.

[0048] The data augmentation module is built upon an improved conditional cycle-consistent generative adversarial network, with the specific network structure as follows: The generator G adopts a U-Net architecture. The encoder part contains 8 convolutional layers with filter counts of 64, 128, 256, 512, 512, 512, 512, and 512 respectively. Each convolutional layer is followed by instance normalization and a LeakyReLU activation function (negative slope 0.2). The decoder part adopts a symmetric structure, using transposed convolutions for upsampling and adding skip connections at corresponding layers. Conditional information is injected into the generator's input and intermediate layers via channel concatenation.

[0049] Discriminator D employs a multi-scale PatchGAN architecture, comprising three independent discriminator networks that process image patches of three scales: 256×256, 128×128, and 64×64, respectively. Each discriminator network consists of five convolutional layers with the following filter counts: 64, 128, 256, 512, and 1. The final layer uses a sigmoid activation function to output the probability of each image patch being true.

[0050] The geological feature extractor F uses a ResNet-50 network pre-trained on ImageNet, removes the last fully connected layer, retains the convolutional layer as the feature extractor, and outputs a 1024-dimensional geological semantic feature vector.

[0051] During training, the batch size was set to 16, the Adam optimizer was used, and the initial learning rate for both the generator and discriminator was 0.0002. The training cycle was 200 epochs. The loss function weights were set to λ. adv =1.0, λ cyc =10.0, λ geo =5.0. After training, the data augmentation module is able to generate synthetic samples that are highly consistent with real thin section images in terms of texture, structure, and mineral assemblage, expanding the training data scale to 200,000 images.

[0052] Furthermore, the multimodal feature understanding module adopts a deep neural network architecture based on a cross-modal attention mechanism, including a visual feature extraction network, a text feature extraction network, and a cross-modal fusion network; The visual feature extraction network adopts a pre-trained ResNet-101 architecture. It takes a pre-processed sandstone thin section image I as input and extracts multi-scale visual feature maps through multi-layer convolution operations. ,in , indicating visual features; The text feature extraction network employs the BERT model, taking a textual description of petrology knowledge T as input; and extracts text feature sequences through a multi-layer Transformer encoder. ,in , representing text features; The cross-modal fusion network employs a dual-path cross-attention mechanism, where the attention weights from image to text are calculated as follows: ; The attention weights for text-to-image processing are calculated as follows: ; in, , These are the parameter matrices; T represents the matrix transpose; α ij β represents the relevance weight between the i-th image region and the j-th text term, indicating the degree to which text term j should be considered when interpreting image region i; ji Let be the relevance weight between the j-th text term and the i-th image region. It represents the degree to which image region i should be considered when interpreting text term j. The optimization objective of the multimodal feature understanding module includes cross-modal contrastive loss: ; in, τ represents the cosine similarity of positive sample pairs in the common space; f and g are the feature projection functions; τ is the temperature hyperparameter; and N represents the total number of samples in a training batch.

[0053] Furthermore, the intelligent identification module adopts a deep neural network architecture based on multi-task learning, including a feature input layer, a multi-task learning layer, and a result output layer; The feature input layer receives the fused semantic vector output from the multimodal feature understanding module. ; in, The dimension of the fusion feature; express A real vector space of dimension 1.

[0054] The multi-task learning layer contains three parallel sub-network branches: The first sub-network branch is the mineral composition identification branch, which uses a softmax classifier and outputs the mineral category probability distribution as follows: in, This is the weight matrix. Number of mineral categories; This is a fused semantic representation vector from the multimodal feature understanding module. It is the bias vector; The second sub-network branch is the structure construction analysis branch, which uses a multilayer perceptron and outputs a structure feature vector. The structure feature vector is: ; Where σ is the sigmoid activation function; As weight, It is the bias vector; The third sub-network branch is the diagenesis identification branch, which uses a multi-label classifier to output the probability of diagenesis type. The probability of diagenesis type is: ; in, For bias vectors, As weight; The total loss function of the intelligent identification module is the weighted sum of the losses of each task, and the total loss function is shown in the following formula: ; Where α, β, γ are task weight coefficients, λ is a regularization coefficient, and Θ is a model parameter.

[0055] Furthermore, the result output module includes a visualization unit and a report generation unit; This module employs a three-level network architecture to achieve multimodal semantic understanding: The visual feature extraction network takes the preprocessed slice image as input and uses a ResNet-101 pre-trained on ImageNet as the backbone network. The last fully connected layer of the original network is removed, and the convolutional layers are retained to extract multi-scale visual features. Features are extracted from 4×, 10×, and 20× objective images respectively, resulting in 512-dimensional feature vectors. Then, multi-scale feature fusion is performed through a Feature Pyramid Network (FPN) to form a unified visual feature representation.

[0056] The text feature extraction network uses a BERT-based Chinese pre-trained model, taking as input textual descriptions of petrology knowledge, including professional descriptions of mineral composition, structural features, and diagenesis. The maximum length of the text sequence is set to 256, and text features are extracted through a 12-layer Transformer encoder, outputting a 768-dimensional text representation vector.

[0057] The cross-modal fusion network employs a two-stream attention mechanism, comprising four attention heads and a hidden layer dimension of 512. The image-to-text attention module calculates the relevance weight between each image region and a text term, while the text-to-image attention module calculates the relevance weight between each term and an image region. The two attention outputs are fused through a gating mechanism to ultimately generate a 1024-dimensional multimodal semantic representation vector.

[0058] During training, a cosine annealing learning rate schedule with hot restart was used, with an initial learning rate of 0.001 and a batch size of 32. The best performance was achieved on the independent validation set after 100 training cycles.

[0059] The intelligent identification module is designed based on a multi-task learning architecture, and its specific implementation is as follows: The feature input layer receives a 1024-dimensional semantic representation vector from the multimodal feature understanding module, which is reduced to 512 dimensions through a fully connected layer. ReLU activation function and Dropout ratio of 0.3 are used to prevent overfitting.

[0060] The multi-task learning layer contains three parallel branches: Mineral composition recognition branch: Contains 3 fully connected layers with 256, 128, and 128 neurons respectively. (Number of mineral categories), using the Softmax activation function to output the probability distribution of 15 common minerals.

[0061] Structural analysis branch: A two-layer perceptron with 256 hidden neurons is used. The output includes regression values ​​of eight structural parameters, such as granularity, sorting, roundness, and cementation type. The Sigmoid function is used to constrain the output to the [0,1] interval.

[0062] Diagenetic process identification branch: Using 3 fully connected layers, output the multi-label classification probability of 6 diagenetic processes, including compaction, cementation, dissolution, and metasomatism.

[0063] During training, a dynamic loss weight adjustment strategy was adopted, with initial weights set to α=1.0 (mineral identification), β=0.8 (structural analysis), and γ=0.6 (diagenesis). Using the AdamW optimizer with a learning rate of 0.0005 and weight decay of 0.01, combined with a linear warm-start training strategy, the system achieved stable convergence within 150 training epochs.

[0064] The visualization unit is used to display the identification results in a multi-dimensional way, including a mineral composition distribution map, a structural annotation map, and a diagenetic analysis map; The report generation unit is used to automatically generate a standardized identification report, which includes basic sample information, image display, data tables, and geological analysis conclusions. The result output module also provides a data interface that supports outputting identification results in image, structured data, and document formats.

[0065] Example 2 See Figures 2 to 3 A smart identification method for sandstone thin sections includes a system training phase and a sample identification phase: The system training phase includes the following steps: Step S1: Collect sandstone thin section samples containing various known mineral compositions, structures, and diagenetic processes to form the original training sample set; Step S2: The image acquisition module acquires a high-resolution digital image sequence of each sample in the original training sample set under multimodal optical conditions, and labels the corresponding petrological knowledge text descriptions of the digital images. Step S3: The image obtained in step S2 is preprocessed using the data processing module to generate standardized image data; Step S4: Invoke the data augmentation module, and based on the normalized image data obtained in step S3, train the improved conditional cycle consistency generative adversarial network to generate large-scale, high-quality synthetic sandstone thin section images, which together with the original images constitute the augmentation training set. Step S5: Using the enhanced training set and the corresponding text description of the enhanced training set, train the multimodal feature understanding module and the intelligent identification module until the model converges, and obtain the final intelligent identification model for sandstone thin sections. The sample identification stage includes the following steps: Step U1: Acquire a high-resolution digital image of the unknown sandstone thin section to be identified through the image acquisition module; Step U2: The data processing module is used to perform preprocessing on the image acquired in step U1, consistent with the training phase. Step U3: Input the preprocessed image into the trained multimodal feature understanding module and the intelligent identification module, and automatically output the identification results of the mineral composition, structure and diagenesis of the unknown sandstone thin section.

[0066] Furthermore, step S4 of the system training phase specifically includes: S4.1, Initialize the network parameters of the generator G, discriminator D and geological feature extractor F; S4.2 Input the original training images and their conditional labels into the network, fix the generator G, and update the parameters of the discriminator D to maximize the ability to distinguish between real images and generated images. S4.3, with the discriminator D fixed, update the parameters of the generator G to minimize the total loss function L. total This prompts the generator to produce composite images that are both realistic and consistent with geological patterns; S4.4, repeat steps S4.2 and S4.3, and use geological features to match the loss L. geo Constraint generation process until network convergence; S4.5, save the trained generator G, which is used to generate synthetic images in batches to expand the training set.

[0067] Furthermore, the process of training the multimodal feature understanding module and the intelligent identification module in step S5 of the system training phase includes the following steps: S5.1, Multimodal Feature Understanding Module Pre-training: Using the enhanced training set, with the weights of the visual feature extraction network and the text feature extraction network fixed, the cross-modal fusion network is trained separately, with the optimization objective being to minimize the cross-modal contrast loss L. cmc This brings image features and text features from the same sample closer together in the semantic space; S5.2, Intelligent identification module pre-training: Freeze the parameters of the trained multimodal feature understanding module, and transfer the multimodal semantic representation vector h... fusion As input, the multi-task learning layer of the intelligent identification module is trained separately, with the optimization objective being to minimize the total loss function L. total; S5.3, End-to-end joint fine-tuning: Unfreeze the parameters of the multimodal feature understanding module, connect the multimodal feature understanding module and the intelligent identification module into a whole network, using the total loss function L. total As a global optimization target, an end-to-end approach is used to jointly fine-tune all network parameters; S5.4, Model Validation and Selection: Evaluate the overall performance of the jointly fine-tuned model on an independent validation set. When the overall performance index of the model on the validation set no longer improves for several consecutive training cycles, stop training and save the parameters of the model with the best performance as the final evaluation model.

[0068] After integrating the various modules, system performance was validated on 50 retained test samples. The test samples covered different sedimentary facies zones and diagenetic stages to ensure the comprehensiveness of the assessment.

[0069] In mineral composition identification, the system achieved an accuracy rate of 96.3% for major minerals such as quartz, feldspar, and rock fragments, and an accuracy rate of 88.5% for rare minerals such as zircon and tourmaline. In structural analysis, the grain size analysis error was less than 0.25Φ, and the sorting accuracy was 92.7%. The overall accuracy rate for diagenesis identification was 94.1%, with the highest accuracy achieved in identifying cementation and dissolution processes, reaching 96.8% and 95.2%, respectively.

[0070] Compared to traditional manual identification methods, this system reduces the identification time for a single sample from 2-3 hours to 5-8 minutes, improving efficiency by approximately 20 times, and providing better repeatability and consistency of results. The system's web interface supports visualized display of identification results and one-click report generation, greatly enhancing the efficiency of geological research.

[0071] Thus far, the description of the above embodiments has been provided for illustrative and descriptive purposes. This is not intended to be exhaustive or limiting of the present disclosure. Individual elements or features of particular embodiments are generally not limited to those particular embodiments, but may be interchanged and used in selected embodiments where applicable, even if not specifically shown or described. In many respects, the same elements or features may also be varied. Such variations are not considered a departure from this disclosure, and all such modifications are intended to be included within the scope of this disclosure.

[0072] Example embodiments are provided so that this disclosure will become thorough and will fully convey the scope to those skilled in the art. Numerous details, such as examples of specific parts, apparatus, and methods, are set forth to provide a thorough understanding of embodiments of this disclosure. It will be apparent to those skilled in the art that the specific details are not required, and the example embodiments may be implemented in many different forms, neither of which should be construed as limiting the scope of this disclosure. In some example embodiments, well-known processes, well-known apparatus structures, and well-known techniques are not described in detail.

[0073] Technical terms are used herein for the purpose of describing specific exemplary embodiments only and are not intended to be limiting. Unless the context clearly indicates otherwise, the singular forms “a” and “the” as used herein may also refer to the plural forms. The terms “comprising” and “having” are inclusive and therefore specify the presence of the stated features, integrals, steps, operations, elements, and / or components, but do not exclude the presence or additional having of one or more other features, integrals, steps, operations, elements, components, and / or combinations thereof. Unless expressly indicated in order of execution, the method steps, processes, and operations described herein are not to be construed as necessarily requiring performance in the specific order discussed and shown. It should also be understood that additional or optional steps may be employed.

Claims

1. A smart identification system for sandstone thin sections, characterized in that, include: The image acquisition module is used to acquire digital images of sandstone thin sections; The data processing module is communicatively connected to the image acquisition module and is used to receive and process the digital images acquired by the image acquisition module. The data storage module, which is connected in communication with the data processing module, is used to store the original image, the generated image, the model parameters, and the identification results. The data analysis module includes a data augmentation module, a multimodal feature understanding module, and an intelligent identification module; The data augmentation module employs an improved conditional cyclic consistency generative adversarial network to generate training samples based on the original sandstone thin section images. The multimodal feature understanding module uses a cross-modal attention mechanism to fuse the visual features of the sandstone thin section images with petrological knowledge text features to generate multimodal semantic representation vectors. The intelligent identification module adopts a multi-task learning architecture to perform parallel identification and classification of the mineral composition, structure, and diagenesis of the sandstone thin sections based on the multimodal semantic representation vectors. The results output module communicates with the data analysis module to visualize and output identification reports.

2. The intelligent identification system for sandstone thin sections according to claim 1, characterized in that: The image acquisition module includes a polarizing microscope unit, a digital imaging unit, and an illumination control unit; The polarizing microscope unit is equipped with a stage capable of 360° rotation and multiple achromatic objectives with magnification ranging from 4× to 100× and a maximum numerical aperture of 0.

95. The digital imaging unit is coupled to the eyepiece port of the polarizing microscope unit and is equipped with a CMOS image sensor with 20 million effective pixels and 16-bit dynamic range. The lighting control unit includes a full-spectrum LED transmission light source and a control circuit. The color temperature of the full-spectrum LED transmission light source is continuously adjustable in the range of 2800K to 7000K, and it can work in conjunction with the polarizer and analyzer of the polarizing microscope unit. Through programmed control, it can automatically switch between multiple observation modes such as single polarization, cross polarization and conophthalmology. The image acquisition module, under the control of the data processing module, can perform fully automated operations from stage movement, focusing, illumination mode switching, polarizing filter rotation to image capture.

3. The intelligent identification system for sandstone thin sections according to claim 1, characterized in that: The data processing module executes image preprocessing algorithms. It employs an algorithm based on SIFT feature detection and FLANN matcher to spatially align sandstone thin section images of the same field of view acquired under different optical modes. It applies the multi-scale Retinex algorithm to perform color constancy processing on the images to eliminate color deviations caused by changes in lighting conditions. It uses a limited contrast adaptive histogram equalization method to enhance the microscopic contrast of the rock structure. It uses a nonlocal mean denoising algorithm to eliminate random noise introduced during image acquisition. It normalizes the image pixel values ​​to the [0,1] interval and uniformly scales them to a fixed size to generate a preprocessed image that conforms to the neural network input specification.

4. The intelligent identification system for sandstone thin sections according to claim 1, characterized in that: The data augmentation module is implemented based on an improved conditional cyclic consistency generative adversarial network, which includes a generator G, a discriminator D, and a geological feature extractor F. Generator G is used to generate an image X from source domain A. A And random noise vector z, under the condition y in the target domain B Under the guidance of [the relevant authority], generate images in the style of the target domain. ; The discriminator D adopts a multi-scale PatchGAN architecture. Its input is a real image or a generated image and conditional labels, and its output is the probability distribution of each local region of the image being real. The geological feature extractor F is a pre-trained deep convolutional network used to extract high-dimensional geological semantic features from images. The optimization objective of the data augmentation module is composed of the following loss functions: The adversarial loss is expressed as a least squares loss function, which is shown below: ； Where E represents the mathematical expectation, Meaning: Calculate the expectation of a sample drawn from the real data distribution; Meaning: Calculate the expectation of the joint distribution of x and y drawn from the real data distribution and the noise z drawn from the random noise distribution; D(x|y) is the output probability of the discriminator D judging the real image x as true under given condition y; G(x,z|y) is the synthetic image generated by the generator G with the source image x, noise z and condition y as input; D(G(x,z|y)|y) is the output probability of the discriminator judging the generated image G(x,z|y) as true. The cycle consistency loss is defined as: ； G(x,z|y A To transform image x from the original domain to conditional y A Defined style; G(G(x,z|y) A ),z|y B To convert the image generated in the previous step back to the original condition y B ; Geological feature matching loss is defined as: ； F(x) is the high-dimensional semantic feature vector obtained by the feature extractor F from the real image x; F(G(x,z|y)) is the high-dimensional semantic feature vector obtained by the generated image G(x,z|y) through the same feature extractor F. The total loss function is: ； in , , These are the weighting coefficients for each loss term.

5. The intelligent identification system for sandstone thin sections according to claim 1, characterized in that: The multimodal feature understanding module adopts a deep neural network architecture based on a cross-modal attention mechanism, including a visual feature extraction network, a text feature extraction network, and a cross-modal fusion network; The visual feature extraction network adopts a pre-trained ResNet-101 architecture. It takes a pre-processed sandstone thin section image I as input and extracts multi-scale visual feature maps through multi-layer convolution operations. ,in , indicating visual features; The text feature extraction network employs the BERT model, taking a textual description of petrology knowledge T as input; and extracts text feature sequences through a multi-layer Transformer encoder. ,in , representing text features; The cross-modal fusion network employs a dual-path cross-attention mechanism, where the attention weights from image to text are calculated as follows: The attention weights for text-to-image processing are calculated as follows: ； in, , These are the parameter matrices; T represents the matrix transpose; α ij β represents the relevance weight between the i-th image region and the j-th text term, indicating the degree to which text term j should be considered when interpreting image region i; ji Let be the relevance weight between the j-th text term and the i-th image region. It represents the degree to which image region i should be considered when interpreting text term j. The optimization objective of the multimodal feature understanding module includes cross-modal contrastive loss, which is calculated as follows: ； in, τ represents the cosine similarity of positive sample pairs in the common space; f and g are the feature projection functions; τ is the temperature hyperparameter; and N represents the total number of samples in a training batch.

6. The intelligent identification system for sandstone thin sections according to claim 1, characterized in that: The intelligent identification module adopts a deep neural network architecture based on multi-task learning, including a feature input layer, a multi-task learning layer, and a result output layer. The feature input layer receives the fused semantic vector output from the multimodal feature understanding module. ;in, The dimension of the fusion feature; The multi-task learning layer contains three parallel sub-network branches: The first sub-network branch is the mineral composition identification branch, which uses a softmax classifier and outputs the mineral category probability distribution as follows: in, This is the weight matrix. Number of mineral categories; This is a fused semantic representation vector from the multimodal feature understanding module. It is the bias vector; The second sub-network branch is the structure construction analysis branch, which uses a multilayer perceptron and outputs a structure feature vector. The structure feature vector is: ； Where σ is the sigmoid activation function; As weight, It is the bias vector; The third sub-network branch is the diagenesis identification branch, which uses a multi-label classifier to output the probability of diagenesis type. The probability of diagenesis type is: ； in, For bias vectors, As weight; The total loss function of the intelligent identification module is the weighted sum of the losses of each task, and the total loss function is shown in the following formula: ； Where α, β, γ are task weight coefficients, λ is a regularization coefficient, and Θ is a model parameter.

7. The intelligent identification system for sandstone thin sections according to claim 1, characterized in that: The result output module includes a visualization unit and a report generation unit; The visualization unit is used to display the identification results in a multi-dimensional way, including a mineral composition distribution map, a structural annotation map, and a diagenetic analysis map; The report generation unit is used to automatically generate a standardized identification report, which includes basic sample information, image display, data tables, and geological analysis conclusions. The result output module also provides a data interface that supports outputting identification results in image, structured data, and document formats.

8. An intelligent identification method using the intelligent identification system for sandstone thin sections according to any one of claims 1-7, characterized in that, Includes the system training phase and the sample identification phase: The system training phase includes the following steps: Step S1: Collect sandstone thin section samples containing various known mineral compositions, structures, and diagenetic processes to form the original training sample set; Step S2: Obtain the digital image sequence of each sample in the original training sample set under multimodal optical conditions through the image acquisition module, and label the corresponding petrological knowledge text descriptions of the digital images; Step S3: The image obtained in step S2 is preprocessed using the data processing module to generate standardized image data; Step S4: Invoke the data augmentation module, and use the normalized image data obtained in step S3 as a basis to train the improved conditional cycle consistency generative adversarial network to generate synthetic sandstone thin section images, which together with the original images constitute the augmentation training set. Step S5: Using the enhanced training set and the corresponding text description, train the multimodal feature understanding module and the intelligent identification module until the model converges, and obtain the final intelligent identification model for sandstone thin sections. The sample identification stage includes the following steps: Step U1: Acquire digital images of the unknown sandstone thin section to be identified through the image acquisition module; Step U2: Use the data processing module to perform the same preprocessing on the image obtained in step U1 as in the training phase; Step U3: Input the preprocessed image into the trained multimodal feature understanding module and intelligent identification module, and automatically output the identification results of the mineral composition, structure and diagenesis of the unknown sandstone thin section.

9. The method according to claim 8, characterized in that, Step S4 includes: S4.1 Initialize the network parameters of generator G, discriminator D and geological feature extractor F; S4.2 Input the original training images and their conditional labels into the network, fix the generator G, and update the parameters of the discriminator D to maximize the ability to distinguish between real images and generated images. S4.3, with the discriminator D fixed, update the parameters of the generator G to minimize the total loss function L. total This prompts the generator to produce composite images that are both realistic and consistent with geological patterns; S4.4, repeat steps S4.2 and S4.3, and use geological features to match the loss L. geo Constraint generation process until network convergence; S4.5, save the trained generator G, which is used to generate synthetic images in batches to expand the training set.

10. The method according to claim 8, characterized in that, The process of training the multimodal feature understanding module and the intelligent identification module in step S5 includes the following steps: S5.1, Multimodal Feature Understanding Module Pre-training: Using an enhanced training set, the weights of the visual feature extraction network and the text feature extraction network are fixed, and the cross-modal fusion network is trained separately. The optimization objective is to minimize the cross-modal contrast loss L. cmc This brings image features and text features from the same sample closer together in the semantic space; S5.2, Intelligent Identification Module Pre-training: Freeze the parameters of the trained multimodal feature understanding module, and convert the multimodal semantic representation vector h... fusion As input, the multi-task learning layer of the intelligent identification module is trained separately, with the optimization objective being to minimize the total loss function L. total ; S5.3, End-to-End Joint Fine-Tuning: Unfreeze the parameters of the multimodal feature understanding module, connect the multimodal feature understanding module and the intelligent identification module into a whole network, using the total loss function L... total As a global optimization target, an end-to-end approach is used to jointly fine-tune all network parameters; S5.4, Model Validation and Selection: Evaluate the overall performance of the jointly fine-tuned model on an independent validation set. When the overall performance index of the model on the validation set no longer improves for several consecutive training cycles, stop training and save the parameters of the model with the best performance as the final evaluation model.

Images