Identification method of organic matter microscopic components based on cross-modal fusion and attention mechanism

Abstract

The present invention discloses an organic matter microscopic component identification method based on cross-modal fusion and attention mechanism, which belongs to the field of artificial intelligence and geology. Traditional organic matter microscopic component identification methods have the problems of high labor intensity, high subjectivity, low efficiency, etc., which limit their application in the industrial field. The automated identification method proposed in this application converts the component identification task in organic matter microscopic images into a pixel-level classification problem. Its core technologies include: proposing a CFA-Unet model, fusing oil-immersion reflected light and fluorescence images, and realizing pixel-level classification of organic matter microscopic components. A cross-modal fusion method is introduced to strengthen the interaction of feature information between modalities, and a weighted attention mechanism is adopted to reduce component misidentification. Experiments have shown that the CFA-Unet model in this application has achieved excellent performance, with stronger segmentation performance, anti-interference ability and generalization performance. The application of this technology can effectively improve the efficiency and accuracy of organic matter resource quality assessment.

Description

Technical Field

[0001] The present application relates to the fields of artificial intelligence and geology, and specifically to a semantic segmentation method of organic matter microscopic components with cross-modal fusion and attention mechanism. Background Art

[0002] The microscopic components of organic matter play a key role in determining its origin, physicochemical properties (including reactivity and coking), process characteristics (such as coking and mechanical strength), etc., and directly affect the combustion efficiency and industrial application of organic matter. Traditional methods for analyzing the microscopic components of organic matter mainly rely on visual observation under a microscope, requiring testers to have rich knowledge of organic matter and proficient microscope operation skills. However, these methods are time-consuming and labor-intensive, and are easily affected by the subjective experience and resolution of the testers, resulting in highly subjective analysis results and lack of strong inter-laboratory comparability, which limits the industrial application of organic matter microscopic component identification and parameter analysis.

[0003] In addition, existing image-based methods for identifying organic matter microscopic components mainly rely on analyzing common features such as grayscale, color, and texture. However, the microscopic components of organic matter are complex and diverse, making some components difficult to distinguish under common features. In addition, these methods usually require manual feature selection, which may affect the recognition results. In terms of extracting shape features, specific components still need to be manually segmented, and seamless pixel-level analysis and fully automatic recognition cannot be achieved.

[0004] The Chinese patent application, publication number CN114581907A, describes a method and system for classifying inert group microscopic images based on improved MobileNetV3. The invention improves the MobileNetV3 network to obtain a lighter and more accurate network model, which can reduce the number of model parameters and improve the accuracy of inert group classification. However, the invention only classifies a single microscopic image belonging to the inert group, and the output of the model is the category to which the entire image belongs. When there are other microscopic components in an image except the inert group, the invention loses its function.

[0005] The Chinese patent application, publication number CN113177965A, describes a method for extracting all components of organic matter based on an improved U-net network and its application. The invention achieves good results by segmenting all components of organic matter through an improved U-net network. However, the invention only uses data from a single modality when training the model, and cannot segment some components with similar features under a single modality, resulting in a low segmentation accuracy. Summary of the invention

[0006] 1. Technical issues to be solved

[0007] In response to the problems existing in the prior art, such as low accuracy of single-modal microscopic image recognition and lack of end-to-end semantic segmentation and recognition solutions, this application provides an organic matter microscopic component identification method based on cross-modal fusion and attention mechanism, using an improved CFA-Unet network structure, treating the identification of organic matter microscopic components as a pixel-level classification task, and achieving comprehensive identification of organic matter microscopic components through fusion analysis of organic matter microscopic images under oil-immersion reflected light and fluorescence, solving the restrictive problems of traditional methods. In addition, this application provides a unique cross-modal fusion and attention weighting method, which deeply explores the complementarity and correlation of component features between images of different light sources, enhances relevant information, reduces irrelevant details, realizes cross-modal interaction of feature information, improves component identification accuracy, and reduces the problem of misidentification.

[0008] 2. Technical solution

[0009] The purpose of the present invention is achieved by the following technical solutions. The specific method steps of this application are as follows:

[0010] (1) Obtain dual-modal microscopic images under oil immersion reflected light and fluorescence.

[0011] (2) The regions with more organic microscopic components were cropped from the dual-modal microscopic images, and the regions with more irrelevant components such as minerals and background resin were excluded. The cropped dual-modal microscopic images were annotated with four types of organic microscopic components, namely background resin, vitrinite, inertinite and exinite, at the pixel level. A corresponding category label was generated for each pixel of the microscopic image, and the category label was one of background resin, vitrinite, inertinite or exinite. The dataset was constructed and divided into a training set and a test set.

[0012] (3) Preprocess the cropped dual-modal microscopy images to ensure the consistency of image size and feature distribution.

[0013] (4) Design a deep learning model CFA-Unet based on the U-shaped semantic segmentation network (Unet). Two independent encoders are constructed in the CFA-Unet network to deeply mine the microscopic component feature information of oil-immersion reflectance and fluorescence organic matter microscopic images. The two independent encoders fuse the feature outputs obtained from different modal learning with the feature outputs in the decoding process through jump connections to form a more complete and rich feature representation, and finally perform dense prediction to achieve semantic segmentation of organic matter microscopic components. The CFA-Unet network also includes an attention gating module.

[0014] (5) Use the jointly optimized Focal Loss and Dice Loss as the loss function to train the model, and the weighted ratio of the two is 1:1; use the training set to train the CFA-Unet model.

[0015] (6) Save the optimal model after training, use the trained CFA-Unet model to perform semantic segmentation of organic matter microscopic components, input the dual-modal organic matter microscopic image under oil immersion reflected light and fluorescence, and the model can output the category to which each pixel of the microscopic image belongs, that is, which category the pixel belongs to among background resin, vitrinite, inertinite and exinite. Accurate identification of components such as background resin, vitrinite, inertinite and exinite is achieved.

[0016] Preferably, the dual-modal organic matter microscopic image described in steps 1-3 contains four microscopic components: background resin, vitrinite, inertinite and exinite. The microscopic image is selected and annotated at the pixel level by professionals to become a dual-modal microscopic image dataset that is conducive to network training and has a consistent size and feature distribution.

[0017] Preferably, a cross-modal feature fusion method based on skip connection and efficient channel attention (ECA) is adopted in two independent encoders. This method first introduces the microscopic component feature information of the fluorescence modality into the feature extraction network of the oil-immersion reflected light organic matter microscopy image through skip connection, and then uses the ECA module to weight the obtained fused feature maps of different scales to ensure cross-modal information interaction in the feature extraction process.

[0018] Preferably, an attention gating module is introduced into the CFA-Unet network to improve the network's attention to useful feature information. The input of this module is the feature map output by the encoder and decoder, respectively. and To represent, the input is subjected to feature fusion through the convolution layer, and then through nonlinear transformation and The activation function is weighted, and finally the residual connection is used to focus on the beneficial feature areas to improve the recognition accuracy of organic matter microscopic components. The specific formula of the attention gate is as follows:

[0019]

[0020] In the formula: express ReLU Activation function; express Sigmoid Activation function; Indicates application Sigmoid The attention parameter matrix before the activation function; represents the final attention parameter matrix; Wx and Represents convolution operation; T represents matrix transpose; and represents bias; represents the model parameters; Represents the feature map after attention weighting.

[0021] Preferably, the network loss function described in step 5 adopts a joint optimization strategy of Focal Loss and Dice Loss to balance the loss contribution of different components and improve the robustness of the model. The specific formulas of Focal Loss and Dice Loss loss functions are as follows:

[0022]

[0023] in: It represents the predicted probability that the sample is class 1, ranging from 0 to 1. Represents the true value label, which is 0 or 1. and is a hyperparameter used to address the class imbalance problem.

[0024]

[0025] In the formula, and The number of elements in the intersection, and and Respectively represent sets and The number of elements in . Multiplying the numerator by 2 eliminates the denominator. and The impact of repeated calculations caused by common elements. This ensures that the value of the loss function is in the range of 0 to 1.

[0026] Preferably, during the training of the network in step 5, the best model is selected based on the performance of the model in the test set. The performance evaluation indicators used in this application include pixel accuracy (PA), mean pixel accuracy (MPA), intersection over union (IOU) and mean intersection over union (MIOU), which are used to evaluate the performance of the network in the task of identifying organic microscopic components. The formulas for each evaluation indicator are as follows:

[0027]

[0028]

[0029]

[0030]

[0031] In the above formula, we assume that there is kind. Indicates that it belongs to the class and is predicted to be class The number of pixels, Indicates that it belongs to the class But it is predicted to be The number of pixels.

[0032] Preferably, the training of the network in step 5 adopts a five-fold cross-validation strategy to ensure the independence of the data set and provide a more comprehensive and objective network performance evaluation; during the training of the network, the AdamW optimizer is used, and the initial learning rate is 2×10 -4 .

[0033] 3. Beneficial effects

[0034] Compared with the prior art, the advantages of the present invention are:

[0035] (1) This application is based on a bimodal feature fusion network and is trained using bimodal data. By combining the information of the two modalities, it is possible to easily distinguish components in reflected light microscopy images that are difficult to distinguish using existing technologies.

[0036] (2) Based on semantic segmentation technology, this application can classify all pixels of a microscopic image. The classification range includes four categories: background resin, vitrinite, inertinite and exinite, which can handle situations that cannot be handled by existing technologies.

[0037] (3) The CFA-Unet network constructs two independent encoders to deeply mine the feature information of the microscopic components of oil-immersed reflectance and fluorescence organic matter microscopic images. The two independent encoders fuse the feature outputs obtained from different modal learning with the feature outputs in the decoding process through jump connections to form a more complete and rich feature representation, and finally perform dense prediction to achieve semantic segmentation of organic matter microscopic components.

[0038] Through this application, the accuracy and robustness of the identification of organic matter microscopic components have been significantly improved, greatly reducing the difficulty of the tester's work, and realizing the fully automated, efficient, and accurate identification of organic matter microscopic components. In addition, this application has strong inter-laboratory comparability, provides efficient and accurate tools and technical support for organic matter analysis and utilization in the industrial field, and has significant application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] Figure 1 The overall structure of the method of the present invention;

[0040] Figure 2 The specific structure of the ECA module;

[0041] Figure 3 Schematic diagram of the cross-modal interaction method based on ECA;

[0042] Figure 4 It is the attention gating module structure;

[0043] Figure 5 The results of organic matter microscopic component recognition based on different semantic segmentation methods. DETAILED DESCRIPTION

[0044] Example 1

[0045] The technical solution of the present invention is described in detail below in conjunction with the accompanying drawings, but the content of the present invention is not limited thereto.

[0046] The present invention provides an organic matter microscopic component identification method based on cross-modal fusion and attention mechanism, which can quickly identify and classify organic matter microscopic components. The specific steps are as follows:

[0047] Step 1: Data collection and preprocessing.

[0048] In the present invention, a polarizing microscope is used to observe organic matter samples to obtain the required oil-immersion reflection and fluorescence microscopic images. It should be noted that the resolution of the acquired microscopic images is inconsistent, and some images contain mineral impurities that are irrelevant to the organic matter microscopic components. These problems may have an adverse effect on the quality and reliability of the image. Therefore, the present invention first excludes areas with too many irrelevant microscopic components such as minerals. Then, the remaining suitable areas are cropped and preprocessed. Therefore, a total of 70 pairs of organic matter microscopic images of the same size of 672x672 pixels are obtained. This processing method ensures the consistency and reliability of the image and provides consistent input data for subsequent segmentation tasks.

[0049] In addition, in order to perform semantic segmentation of organic matter microscopic components, corresponding category labels need to be generated for microscopic images. This helps to annotate background resin, vitrinite, inertinite, and exinite at the pixel level. Figure 5 Examples of microscopic images under oil-immersion reflected light and fluorescence and their corresponding annotation results are shown. In order to ensure the professionalism and accuracy of the generated labels, researchers with rich experience in the analysis of organic matter microscopic components were invited to provide professional guidance in the annotation process. In this embodiment, the researchers are a highly professional team of coal petrologists who have excellent professional backgrounds and rich experience in pixel-level identification of coal and rock microscopic components. The members of this team all have doctoral degrees in the field of coal petrology and have demonstrated excellent technical capabilities in multiple coal mining projects. They use advanced microscopes and sophisticated image processing algorithms to ensure that the microscopic components of coal and rock samples can be highly accurately identified when performing pixel-level identification. In order to ensure the consistency and credibility of the results, the team adopts strict quality control measures, including multi-person cross-validation and calibration using standard samples. They will provide detailed reports and documents to demonstrate the identification results, analysis methods, and technical parameters.

[0050] Step 2: CFA-Unet network structure design.

[0051] This application uses the Unet network as the basic structure, which is a U-shaped semantic segmentation network consisting of an encoder and a decoder. The encoder extracts features from the input image, while the decoder obtains compressed features in the encoding process through jump connections and deeply mines feature information. To maintain symmetry, the encoder and decoder are combined through jump connections, which effectively fuses shallow and deep feature information, makes up for the information lost during the downsampling process, and makes the segmentation output retain more details, which is suitable for the identification of organic microscopic components.

[0052] Since Unet was originally designed for medical image segmentation, and the complex composition and diverse morphological features of organic microscopic images are very different from medical images, when processing these complex data, Unet often finds it difficult to fully extract important features of interest, limiting the further improvement of segmentation performance. To solve this problem, this application proposes a semantic segmentation method based on deep learning and multimodal learning, and introduces the Cross-Fusion-Attention-Unet (CFA-Unet) model. Specifically, the model constructs two independent encoders for in-depth mining of microscopic component feature information of oil-immersed reflectance and fluorescence organic microscopic images. After the features learned in different modalities are fused, they are connected to the feature maps in the decoding process through jump connections to form a more complete and richer feature representation. Finally, dense prediction is performed to achieve semantic segmentation of organic microscopic components. In order to further improve the recognition accuracy of components such as exoskeleton microscopic components and background resin, this application introduces improvements in feature fusion and key information mining, and proposes cross-modal feature fusion and attention weighting methods.

[0053] Step 3: Cross-modal feature fusion.

[0054] Traditional feature fusion strategies effectively process oil-immersion reflectance and fluorescent organic matter microscopy images, but there are some limitations: each modality-specific feature extraction network is trained independently without considering the participation of feature information from other modalities. Feature information of different modalities often exhibits interdependent and complementary characteristics, while independent training strategies often ignore the complementary effects of feature information of different modalities, limiting further improvement of model performance. To solve this problem, this application proposes a cross-modal fusion strategy based on skip connections and efficient channel attention (ECA), which makes full use of the interdependence and complementarity between feature information of different modalities and enhances the expression ability of the network. The implementation details of ECA are as follows: Figure 2 As shown, the feature map undergoes GAP (global average pooling), convolutional layers and After the activation function, the attention weight of the corresponding channel is obtained and weighted with the original feature map. The ECA module assigns a larger attention weight to the beneficial feature information and a smaller attention weight to the irrelevant interference information. The overall cross-modal feature fusion method is as follows Figure 3 As shown. First, the feature extraction network of the fluorescence modality (extracting the characteristic information of the microscopic components of the fluorescence modality) is introduced into the feature extraction network of the oil-immersed reflectance organic microscopic image through a jump connection, and the characteristic information of the microscopic components of the fluorescence modality is used as a priori knowledge to guide the feature extraction process of the oil-immersed reflectance organic microscopic image. Secondly, the fused feature maps of different scales are ECA-weighted, and in the subsequent layers of the feature extraction network of the oil-immersed reflectance organic microscopic image, the weighted feature maps of different scales are used as input to ensure cross-modal information interaction in the entire feature extraction process.

[0055] Step 4: Attention weighting.

[0056] In order to increase the network's attention to feature information that is beneficial to improving component recognition, this application introduces an attention weighting method, which is achieved by introducing an attention gating module in the segmentation model. This module can fully learn and understand key semantic information. This method autonomously learns and retains network parameters that are beneficial to specific tasks during network training. Its characteristics are that it is plug-and-play, simple and efficient. Figure 4 is a schematic diagram of the structure of the attention gating module. The input of this module is the feature map output by the encoder and decoder, respectively. and To represent, the input is subjected to feature fusion through the convolution layer, and then through nonlinear transformation and The activation function is weighted, and finally the residual connection is used to focus on the beneficial feature areas to improve the recognition accuracy of organic matter microscopic components. The specific formula of the attention gate is as follows:

[0057]

[0058] In the formula: express ReLU Activation function; express Sigmoid Activation function; Indicates application Sigmoid The attention parameter matrix before the activation function; represents the final attention parameter matrix; Wx and Represents convolution operation; T represents matrix transpose; and represents bias; represents the model parameters; Represents the feature map after attention weighting.

[0059] Step 5: Training and optimization.

[0060] The data obtained in step 1 and its label group are formed into a data set, and the data set is divided into a training set and a test set according to an 8:2 ratio. This application uses the divided training set data to train the network, adopts a joint optimization strategy, and defines the loss function in combination with Focal Loss and Dice Loss to solve the problem of unbalanced proportions of different microscopic components. Focal Loss is a loss function designed to solve the problem of class imbalance. The core idea is to adjust the weight of the sample so that the model can better adapt to the characteristics of the sample distribution. It assigns greater weights to challenging samples, so that the model pays more attention to these samples and improves the learning effect on minority classes. The calculation formula is as follows:

[0061]

[0062] in: It represents the predicted probability that the sample is class 1, ranging from 0 to 1. Represents the true value label, which is 0 or 1. and It is a hyperparameter used to solve the class imbalance problem. Dice Loss is a loss function used for image segmentation tasks. It guides training by measuring the overlap between the segmentation results generated by the model and the true label. Its calculation formula is as follows:

[0063]

[0064] In the formula, and The number of elements in the intersection, and and Respectively represent sets and The number of elements in . Multiplying the numerator by 2 eliminates the denominator. and The impact of repeated calculations caused by common elements. This ensures that the value of the loss function is in the range of 0 to 1.

[0065] The model was trained using the AdamW optimizer and five-fold cross validation was used to ensure the stability and reliability of the model. During the training of the network, the initial learning rate was 2×10 -4 , achieving efficient feature learning and recognition.

[0066] Step 6: Performance evaluation of the model.

[0067] This application selects the best model based on the performance of the model in the test set. The performance evaluation indicators used in this application include pixel accuracy (PA), mean pixel accuracy (MPA), intersection over union (IOU) and mean intersection over union (MIOU), which are used to evaluate the performance of the network in the task of identifying organic matter microscopic components. The formulas for each evaluation indicator are as follows:

[0068]

[0069]

[0070]

[0071]

[0072] In the above formula, we assume that there is kind. Indicates that it belongs to the class and is predicted to be class The number of pixels, Indicates that it belongs to the class But it is predicted to be The number of pixels.

[0073] Step 7: Model component identification. Save the optimal model after training, input the dual-modal organic matter microscopic image under oil immersion reflected light and fluorescence to be predicted, and the model can output the category to which each pixel of the microscopic image belongs, that is, which category the pixel belongs to: background resin, vitrinite, inertinite, or exinite.

[0074] In order to verify the feasibility and effectiveness of the proposed organic matter microscopic component identification method based on the fusion of oil-immersion reflected light and fluorescent organic matter microscopic images, Table 1 shows the performance comparison of several mainstream semantic segmentation networks used in the identification of organic matter microscopic components. The selected semantic segmentation networks include Pyramid Scene Parsing Network (PSPNet), DeepLabV3+ and Unet. It is worth noting that in order to make a fair comparison, the three networks were trained using a multimodal dataset similar to CFA-Unet in the experiment. This means that the image input to the network is a concatenation of microscopic images taken under two different conditions. It can be seen that using multimodal data for network training can significantly improve the segmentation performance compared to using single modality data.

[0075] As can be seen from Table 1, compared with other mainstream semantic segmentation methods, the proposed CFA-Unet network achieved the best performance in all evaluation indicators (PA: pixel accuracy, MPA: mean pixel accuracy, MIOU: mean intersection over union), reaching 94.84%, 79.79% and 73.06% respectively.

[0076] Table 1 Identification performance of organic matter microscopic components under different methods

[0077]

[0078] The CFA-Unet network and its multimodal learning method proposed in this application have achieved remarkable results in the field of organic matter microscopic component identification. Compared with other mainstream semantic segmentation networks, CFA-Unet not only excels in pixel accuracy, but also achieves excellent results in average pixel accuracy and average intersection-over-union ratio. Its smooth and detailed component identification results show that this method can more accurately distinguish organic matter microscopic components, providing efficient and accurate tools and technical support for organic matter analysis, resource exploration and organic matter utilization in the industrial field. By improving the accuracy and robustness of component identification, this application provides a reliable intelligent means for resource production and utilization, effectively improves production efficiency, and reduces labor costs. It can be widely used in the field of automated identification of organic matter microscopic components, and makes an important contribution to the sustainable development of energy.

[0079] The above schematically describes the invention and its implementation methods, which is not restrictive. Without departing from the spirit or basic features of the invention, the invention can be implemented in other specific forms. What is shown in the accompanying drawings is only one of the implementation methods of the invention. The actual structure is not limited to this, and any figure mark in the claims should not limit the claims involved. Therefore, if a person of ordinary skill in the art is inspired by it, without departing from the purpose of the invention, a structural method and an embodiment similar to the technical solution are designed without creativity, which should all fall within the scope of protection of this patent. In addition, the word "including" does not exclude other elements or steps, and the word "one" before the element does not exclude the inclusion of "multiple" elements. The multiple elements stated in the product claim can also be implemented by one element through software or hardware. The words first, second, etc. are used to indicate names, and do not indicate any specific order.

Claims

1. A method for identifying organic matter microscopic components based on cross-modal fusion and attention mechanism, comprising the following steps: (1) Obtain dual-modal microscopic images under oil immersion reflected light and fluorescence; (2) Crop out the areas with more organic microscopic components from the dual-modal microscopic images, exclude the areas with a large proportion of irrelevant components such as minerals and background resins, and annotate the cropped dual-modal microscopic images at the pixel level according to the categories of organic microscopic components; construct a dataset and divide it into a training set and a test set; (3) Preprocess the cropped dual-modal microscopy images to ensure the consistency of image size and feature distribution; (4) Designing a deep learning model CFA-Unet based on the U-shaped semantic segmentation network Unet, two independent encoders are constructed in CFA-Unet, which are used to deeply mine the microscopic component feature information of oil-immersion reflectance and fluorescent organic matter microscopic images respectively; the two independent encoders fuse the feature outputs obtained from different modal learning with the feature outputs in the decoding process through jump connections to form a complete feature representation; the CFA-Unet network also includes an attention gating module; the input of the attention gating module is the feature map output by the encoder and the decoder; (5) Use the jointly optimized Focal Loss and Dice Loss as the loss function to train the model, with the weighted ratio of the two being 1:1; use the training set to train the CFA-Unet model; (6) The optimal model after training is saved, and the trained CFA-Unet model is used to perform semantic segmentation of organic matter microscopic components. The dual-modal organic matter microscopic images under oil-immersion reflected light and fluorescence are input, and the model outputs the category of organic matter microscopic components to which each pixel of the microscopic image belongs.

2. According to claim 1, a method for identifying organic matter microscopic components based on cross-modal fusion and attention mechanism is characterized in that: The categories of organic matter microscopic components include background resin, vitrinite, inertinite and exinite.

3. The method for identifying organic matter microscopic components based on cross-modal fusion and attention mechanism according to claim 1, characterized in that: The image preprocessing method is as follows: first, histogram equalization or bilateral filtering algorithm is used to improve the image quality and enhance the image features, and then the image size is changed to a fixed 672×672 pixels.

4. The method for identifying organic matter microscopic components based on cross-modal fusion and attention mechanism according to claim 1, characterized in that: The method for pixel-level annotation of dual-modal microscopic images is: generating a corresponding category label for each pixel of the microscopic image, wherein the category label is one of background resin, vitrinite, inertinite or exinite.

5. The method for identifying organic matter microscopic components based on cross-modal fusion and attention mechanism according to claim 1, characterized in that: CFA-Unet first introduces the microscopic component feature information of the fluorescence modality into the feature extraction network of the oil-immersion reflected light organic matter microscopy image through skip connections, and then uses the ECA module to weight the obtained fusion feature maps of different scales.

6. The method for identifying organic matter microscopic components based on cross-modal fusion and attention mechanism according to claim 1, characterized in that: The input of the attention gating module is the feature map output by the encoder and decoder. The input is fused through the convolution layer, weighted by nonlinear transformation and activation function, and finally the attention is paid to the beneficial feature areas through residual connection.

7. The method for identifying organic matter microscopic components based on cross-modal fusion and attention mechanism according to claim 6, characterized in that: The input of the attention gating module is the feature map output by the encoder and decoder, represented by x and g respectively; the specific formula is as follows: α=σ2(θ att (x, g, Θ)) x * =x·a In the formula: σ1 represents the ReLU activation function; σ2 represents the Sigmoid activation function; θ att represents the attention parameter matrix before applying the Sigmoid activation function; α represents the final attention parameter matrix; Wx and W g Represents the convolution operation; T represents matrix transpose; b g and b ψ represents bias; Θ represents model parameters; x * Represents the feature map after attention weighting.

8. The method for identifying organic matter microscopic components based on cross-modal fusion and attention mechanism according to claim 1, characterized in that: The specific formula of the Focal Loss loss function is as follows: Where: p represents the predicted probability that the sample is class 1, ranging from 0 to 1, y represents the true value label, the value is 0 or 1, α and γ are hyperparameters to solve the class imbalance problem.

9. The method for identifying organic matter microscopic components based on cross-modal fusion and attention mechanism according to claim 1, characterized in that: The calculation formula of Dice Loss loss function is as follows: Here, |X∩Y| represents the number of elements in the intersection of sets X and Y, while |X| and |Y| represent the number of elements in sets X and Y, respectively.

10. A method for identifying organic matter microscopic components based on cross-modal fusion and attention mechanism according to any one of claims 1 to 9, characterized in that: The network was trained using a five-fold cross-validation strategy. During the network training, the AdamW optimizer was used with an initial learning rate of 2×10 -4 .

Images