Method for detecting pores among lattice fringes of aromatic hydrocarbon in coal based on HRTEM image

Abstract

The invention provides a coal aromatic hydrocarbon lattice fringe pore detection method based on an HRTEM image, and belongs to the technical field of coal aromatic hydrocarbon lattice fringe pore detection. The technical problem to be solved is to provide the method for detecting the pores between the lattice fringes of the aromatic hydrocarbon of the coal based on the HRTEM image. According to the technical scheme, the method comprises the steps that an image of the surface of a coal body to be detected is collected to serve as an original image, and gray processing, shearing, Fourier-inverse Fourier transformation, image binarization processing and blocking are sequentially conducted on the original image; carrying out normalization processing on the image after the binarization processing; marking stripe pores in the normalized image by adopting marking software to obtain main pore types and positions; processing the obtained pore image by adopting a target detection algorithm, generating a pore feature map from the pore image, and inputting the pore feature map into a detection module; the method is used for extracting the pores between the coal aromatic hydrocarbon lattice fringes.

Description

Technical Field

[0001] The present invention provides a method for detecting pores between coal aromatic hydrocarbon lattice fringes based on HRTEM images, belonging to the technical field of pore detection between coal aromatic hydrocarbon lattice fringes. Background Technique

[0002] Coalbed methane is an important part of unconventional natural gas and has become an important alternative energy type in the process of energy transformation and development. However, at present, the single-well production of coalbed methane wells is low, the production decline is fast, and the development effect is poor. The transformation and drainage of coalbed methane are urgent problems to be solved at present.

[0003] From a microscopic perspective, the micropores in coal are important factors affecting the storage, migration, penetration, adsorption and desorption, and production efficiency of coalbed methane. Their characteristics directly affect the permeability of coal seams and the recoverability of coalbed methane. At the same time, it is also of great significance for coal mine safety production and coalbed methane resource assessment. Therefore, studying the micropore structure of coal can provide new ideas for coalbed methane drainage to improve development efficiency and production.

[0004] At present, experimental methods such as low-temperature carbon dioxide adsorption and small-angle X-ray diffraction can be used to study the structural composition and characteristics of coal substances, but the research scale is limited to the micron level and the micro-nano level; the existing patent CN118587164A has carried out quantitative research on the morphology of coal HRTEM lattice fringes, improving the extraction accuracy and efficiency of coal HRTEM lattice fringes. There is also a patent CN113888409B that obtains the distribution, length, angle, and aspect ratio information of lattice fringes by using the separation and merging method of aromatic hydrocarbon lattice fringe intersections; the above-mentioned publicly disclosed technical solutions have carried out precise quantitative research on each lattice fringe, but there is less research on the micropores between coal HRTEM lattice fringes. At present, corresponding pore extraction and detection schemes need to be proposed to improve the research. Summary of the Invention

[0005] In order to overcome the deficiencies in the prior art, the technical problem to be solved by the present invention is to provide a method for detecting pores between coal aromatic hydrocarbon lattice fringes based on HRTEM images.

[0006] In order to solve the above technical problem, the technical solution adopted by the present invention is: a method for detecting pores between coal aromatic hydrocarbon lattice fringes based on HRTEM images, including the following detection steps: S100: Collect an image of the surface of the coal body to be detected as the original image, and sequentially perform gray processing, shearing, Fourier-inverse Fourier transform, image binarization processing, and block division on the original image; S200: Perform normalization processing on the binarized image, and adjust the image size to the size required by the model; S300: Use a labeling software to label the striped pores in the normalized image to obtain the main pore categories and positions; S400: Process the obtained pore images using an object detection algorithm. After optimizing the data of the pore images through the CSC module in the Backbone network and the newly added TA layer, input them into the Neck network. After multi-scale fusion of the newly added P2 layer and the P3-P5 layers in the Neck network to obtain feature maps of corresponding sizes, generate the pore images into pore feature maps; S500: Input the pore feature maps into the Head detection module. The Head detection module generates an information-containing tensor. After screening by the confidence threshold and non-maximum suppression, output the object category labels and the coordinates of the label boxes. Screen out the prediction results with corresponding scores on the pore categories, and return the category labels, positions, and confidences of the prediction results as the final detection results.

[0007] The specific method for collecting and processing the coal body image in step S100 is as follows: S101: Grind the coal body sample with a grinder to less than 200 mesh. Use tweezers to pick up a small amount of coal powder and place it in 5 ml of anhydrous ethanol. Through ultrasonic oscillation for 15 min, make the coal powder fully dispersed in the anhydrous ethanol and let it stand for 5 min. Use a capillary glass tube to suck the uniform mixture on the upper layer of the anhydrous ethanol and drop it on a 200-mesh copper mesh. After the anhydrous ethanol dries, place the copper mesh under an electron microscope for observation and photography; S102: Select the edge area of the dried mixture under the electron microscope for analysis. Import the original HRTEM image into the Digital Micrograph software and frame the target area; S103: On the premise of retaining the microcrystalline structure information in the coal, perform Fourier and inverse Fourier transform processing on the image to eliminate the noise in the image; S104: Use Adobe photoshop software to adjust the threshold of the filtered image to obtain a HRTEM image with clear lattice fringes.

[0008] The specific method for normalizing the image in step S200 is as follows: Adjust the image size to 640 * 640 pixels, and normalize the pixel values to the 0-1 interval by dividing each pixel value by 255, so that the image meets the model processing requirements.

[0009] The specific method for labeling the striped pores in the image in step S300 is as follows: Use the labelimg software for labeling, and save the labeled labels in txt format. The data with bounding box annotation information and the HRTEM image together constitute the training dataset.

[0010] The specific method of using the object detection algorithm to process the pore image in step S400 is as follows: S401: Use the object detection algorithm to extract features from the pore image: Transfer the pore image data to the Backbone network. The CSC module uses the split layer to split the feature map along the channel dimension into feature maps X1 and X2, where: After the feature map X1 is convolved with different kernel sizes, it passes through the S-Bottleneck module and the BN batch normalization layer respectively, and then is fused by the Concat module and passed through the ReLU activation function to obtain the first output feature; The feature map X2 passes through two sequentially connected S-Bottleneck modules to obtain the second output feature. After the two are fused, a 1*1 convolution is performed to generate the output feature of the CSC module. A new TA layer is added between the CSC module and the Neck network. After the TA layer optimizes and integrates the features, they are passed into the Neck network; Perform multi-scale downsampling on the image through a convolutional neural network to generate a series of feature maps of the pores between the stripes; S402: In the Neck network, the output of the newly added P2 layer is used as the small target detection layer. In the feature enhancement stage, it performs multi-scale feature fusion with the features output by the P3-P5 layers, and uses upsampling, downsampling, feature splicing, and convolution operations to fully integrate the feature advantages of each layer to generate an enhanced feature map with a size of 160*160*64, which is transmitted to the Head detection head; S403: Use predicted anchor boxes to predict a fixed number of bounding boxes for each anchor box at a preset scale. Each bounding box contains the center point coordinates, aspect ratio, probability of the corresponding category, and confidence. The non-maximum suppression algorithm is used to remove the bounding boxes with high similarity but close positions to eliminate duplicate predictions; S404: During the training process, use the cross-entropy loss function to detect the difference between the prediction result and the true label. The loss function quantifies the deviation in category determination by calculating the cross threshold between the predicted category probability distribution and the true category label; At the same time, in combination with the CIoU loss calculation method, the category loss and the bounding box loss are integrated according to the corresponding weights to form the final loss; Based on the total loss, use the backpropagation algorithm to calculate the gradient information of the model parameters with respect to the loss. Use the Adam optimizer to calculate the gradient, learning rate, and momentum parameter rules, and gradually update and optimize the weight and bias parameters in the model. Perform multiple rounds of iteration, evaluate the model performance on the validation set, calculate the gap between the prediction result and the true label, and optimize the model parameters.

[0011] The specific method of step S500 is as follows: The Head detection head converts the received feature map into an output tensor containing the dimensions of the object category probability channels, the dimensions of the bounding box coordinate information, and the object confidence value using a convolutional layer. The output tensor is processed by first filtering out low-confidence prediction boxes according to the confidence threshold and then removing redundant prediction boxes through the non-maximum suppression algorithm. Finally, the class labels and bounding box coordinate information of the objects in the image are output.

[0012] The beneficial effects of the present invention compared with the prior art are as follows: The coal aromatic hydrocarbon lattice fringe pore extraction scheme proposed by the present invention can provide a new identification research method for studying nano-scale micropores, and provides a new way for studying the relationship between the micropore distribution of coal and the basic units of the coal macromolecular structure; this scheme approximately fits the pores of the coal body into rectangles and uses a specific image recognition algorithm to intuitively reproduce the distribution of the pores between the coal aromatic lattice fringes, which plays a good guiding role in studying the relationship between the micropore distribution and the basic units of the coal macromolecular structure; the detection scheme provided by the present invention can complete the whole process from image input to accurate target detection results output. Through the improvement and coordinated operation of each module, the performance of the model in small target detection is effectively improved. BRIEF DESCRIPTION OF THE DRAWINGS

[0013] The present invention will be further described below with reference to the drawings: Figure 1 It is a flowchart of the steps for extracting the pores between the coal aromatic hydrocarbon lattice fringes of the present invention; Figure 2 It is an effect diagram of the processing process of the coal HRTEM image in the embodiment of the present invention; Figure 3 It is the effect of pore detection on the coal HRTEM image in the embodiment of the present invention Figure 1 ; Figure 4 It is the effect of pore detection on the coal HRTEM image in the embodiment of the present invention Figure 2 . DETAILED DESCRIPTION OF THE EMBODIMENTS

[0014] The high-resolution transmission electron microscope (HRTEM) adopted by the present invention can quantitatively evaluate the characteristics such as the size of aromatic units and the orientation of aromatic structures from the nano-scale, can directly display the micropores in coal and conduct morphological analysis, and is an attractive supplementary technology in the field of direct observation of coal structure, providing a basis for the accurate detection of micropores in the present invention.

[0015] Such as Figure 1As shown in the figure, the method for detecting pores between coal aromatic hydrocarbon lattice fringes based on HRTEM images provided by the present invention mainly includes the following image detection steps: First, perform gray-scale processing, shearing, Fourier-inverse Fourier transform, image binarization processing, and block division on the original image; then label the stripe pores of the preprocessed image using labelimg, and save the labeled tags as a text file in txt format; then use the yolo semantic segmentation algorithm to extract features, segment grids, predict anchor boxes, and perform non-maximum suppression on the most representative target pores; finally, screen out the predictions with higher scores in the pore category, and return the category label, position, and confidence as the final detection result.

[0016] Furthermore, the method for detecting pores between coal aromatic hydrocarbon lattice fringes adopted above specifically includes the following detection steps: S100: Perform gray-scale processing, shearing, Fourier-inverse Fourier transform, image binarization processing, and block division on the original image; S200: Normalize the image after binarization processing and adjust its size to make the data meet the model requirements.

[0017] S300: Use a labeling software to label the stripe pores of the normalized image to obtain the main pore categories and positions; S400: Use an object detection algorithm to process the image pores. In the Backbone network, use the CSC module (including feature map segmentation, different processing, and fusion) and add a TA layer to optimize the input to the Neck network. The Neck network uses the newly added P2 layer and the P3-P5 layers for multi-scale fusion to obtain a feature map of a specific size, and generate a pore feature map from the image data; S500: Input the feature map into the Head detection. The Head detection generates an information-containing tensor. After screening by the confidence threshold and non-maximum suppression, output the object category label and the label box coordinates, screen out the predictions with higher scores in the pore category, and return the category label, position, and confidence as the final detection result.

[0018] Specifically, the above step S100 adopts the following method to collect and process the image: S101: Grind the sample with a grinder to less than 200 mesh, pick up a small amount of coal powder with tweezers and place it in 5 ml of anhydrous ethanol. Oscillate it ultrasonically for 15 min to fully disperse the coal powder in the anhydrous ethanol, let it stand for 5 min, and finally suck the upper homogeneous mixture with a capillary glass tube and drop it on a 200-mesh copper mesh (with a diameter of about 3 mm). After the anhydrous ethanol dries, place the copper mesh under a high-resolution transmission electron microscope for observation and photography; S102: Select a thin edge area for analysis under a high-resolution transmission electron microscope, import the original HRTEM image into the Digital Micrograph software, and frame the area of interest; S103: On the premise of retaining the microcrystalline structure information in the coal, perform Fourier and inverse Fourier transforms on the image to eliminate noise; S104: Use Adobe photoshop software to adjust the threshold of the filtered image to obtain a HRTEM image with clear lattice fringes.

[0019] Specifically, the above step S200 adopts the following method to normalize the image: S201: Adjust the image size to 640*640 pixels, and normalize the pixel values to the 0-1 interval by dividing each pixel value by 255, so that the image meets the requirements of model processing.

[0020] Specifically, the above step S300 adopts the following method to label the stripe pores in the image: S301: Perform image annotation on the preprocessed HRTEM image to provide accurate supervision information for model training. Determine the pore category of each target, accurately outline the bounding box of the target pore, and strictly follow the standards and specifications of the pores between the stripes during the annotation process to ensure the consistency and accuracy of the annotation. Use the labelimg software for annotation, save the labeled tags in txt format, and form a training data set together with the HRTEM image with bounding box annotation information.

[0021] Specifically, the above step S400 adopts the following method to process the image pores: S401: Use the object detection algorithm to extract features from the high-resolution image, and transfer the data to the Backbone network. This network uses the CSC module to replace the original C2f module, splits the feature map along the channel dimension into feature maps X1 and X2 through the split layer. After X1 is convolved with different kernel sizes, it passes through the S-Bottleneck module and the BN batch normalization layer respectively, and then is fused by the Concat module and activated by the ReLU activation function to obtain the first output feature. X2 passes through two sequentially connected S-Bottleneck modules to obtain the second output feature. After the two are fused, a 1*1 convolution is performed to generate the output feature of the CSC module, and a new TA layer is added between the CSC module and the Neck network. The TA layer optimizes and integrates the features and then passes them into the Neck network. Perform multi-scale downsampling on the image through a convolutional neural network (CNN) to generate a series of feature maps of the pores between the stripes. The feature maps contain different levels of visual information from low-level to high-level.

[0022] S402: In the Neck network, the output of the newly added P2 layer serves as the small object detection layer. During the feature enhancement stage, multi-scale feature fusion is carried out with the features output by the P3 - P5 layers. By means of operations such as upsampling, downsampling, feature splicing, and convolution, the advantages of features from each layer are fully integrated to generate an enhanced feature map with a size of 160*160*64, which is then transmitted to the Head detection head.

[0023] S403: Using the predicted anchor boxes, at a pre-set scale, a fixed number of bounding boxes are predicted for each anchor box. Each bounding box contains 5 elements: the coordinates of the center point, the aspect ratio, and the probabilities and confidences corresponding to the categories. The non-maximum suppression algorithm is applied to remove the bounding boxes that are highly similar but close in position, eliminating duplicate predictions.

[0024] During the training process, the cross-entropy loss function is used to accurately measure the degree of difference between the prediction results and the true labels. The loss function quantifies the deviation in category determination by calculating the cross-threshold between the predicted category probability distribution and the true category labels. At the same time, in combination with loss calculation methods such as CIoU, the category loss and the bounding box loss are integrated according to specific weights to form the final loss. Based on the total loss, the gradient information of the model parameters with respect to the loss is calculated using the backpropagation algorithm. Using the Adam optimizer, the gradients, as well as parameters such as the learning rate and momentum, are calculated to gradually update and optimize the weight and bias parameters in the model, improving the accuracy and generalization ability of pore detection in the task. Multiple rounds of iteration are carried out throughout the process, and the performance of the model is evaluated on the validation set during development, calculating the gap between the prediction results and the true labels to optimize the model parameters.

[0025] Specifically, the above step S500 adopts the following image detection method: S501: Based on the received feature map, the Head detection head uses the convolutional layer to convert it into a specific format output tensor containing the dimension of the object category probability channel, the dimension of the bounding box coordinate information, and the object confidence value. Finally, this tensor undergoes post-processing. First, the low-confidence prediction boxes are filtered out according to the confidence threshold, and then the redundant prediction boxes are removed through the non-maximum suppression algorithm. Finally, the category labels and the bounding box coordinate information of the objects in the image are output, completing the whole process from image input to accurate target detection results output. Through the improvement and coordinated operation of each module, the performance of the model in small object detection is effectively improved.

[0026] In the embodiment of the present invention, taking the No. 8 coal in a certain mine in a coalfield as an example, the method for accurately extracting the pores between lattice fringes in the HRTEM image proposed by the present invention is verified. A thinner and clearly defined area in the high-resolution image of the No. 8 coal in a certain mine is selected for analysis. The detection process for extracting the pores between the HRTEM lattice fringes of the No. 8 coal is as follows Figure 1As shown, the processing effect of the acquired images during the detection process is as follows Figure 2 As shown, the original acquired images are respectively subjected to Fourier transform infrared (FTIR), noise reduction, and binarization processing, and finally the aromatic framework of the HRTEM image is obtained. Among the processing effects of the HRTEM images, a is the ROI, b is the Fourier transform infrared, c is the noise reduction, d is the binarization, and e is the aromatic framework; the recognition effect of the pores in the HRTEM image is as follows Figure 3 and Figure 4 As shown, a1, b1, c1, and d1 in the figure are the segmented HRTEM images of the embodiments of the present invention. a2, b2, c2, and d2 in the figure are the txt files of the striped pore labels of the embodiments of the present invention. a3, b3, c3, and d3 in the figure are the pore feature maps predicted by the algorithm of the embodiments of the present invention, including pore labels, positions, and confidence levels; the specific experimental steps are as follows Perform pretreatment on coal No. 8. Grind the sample with a grinder to less than 200 mesh. Use tweezers to pick up a small amount of coal powder and place it in 5 ml of anhydrous ethanol. Oscillate it ultrasonically for 15 min to fully disperse the coal powder in the anhydrous ethanol, and let it stand for 5 min. Finally, use a capillary glass tube to suck the upper-layer uniform mixture and drop it on a 200-mesh copper grid. After the anhydrous ethanol dries, place the copper grid under a high-resolution transmission electron microscope for observation and photography (as shown in a in Figure 2 ); Import the high-resolution image of coal No. 8 into the Digital Micrograph software, select a thin area for lattice fringe analysis, and frame out the area of interest; then perform Fourier transform (as shown in b in Figure 2 ) and inverse Fourier transform (as shown in c in Figure 2 ) on the image to eliminate noise; finally, use an annular filter to retain a more ordered microcrystalline structure (as shown in d in Figure 2 ); Import the filtered image into the Adobe Photoshop software for threshold adjustment to obtain a HRTEM image with clear lattice fringes (as shown in e in Figure 2 ); Perform block processing on the binarized image to more accurately predict the target pores (as shown in a1, b1 in Figure 3 , Figure 4 c1, d1 in ); Figure 3 Import the clear HRTEM image after pretreatment into the software labelimg for striped pore annotation, and the label format is a txt file (as shown in a2, b2 in Figure 4 , The high-resolution image of No. 8 coal is subjected to feature extraction using a Backbone, and the HRTEM image is downsampled at multiple scales through a Convolutional Neural Network (CNN) to generate a series of pore feature maps. As the network layer deepens, the size of the feature map gradually decreases, and the number of channels gradually increases. The feature map contains different levels of visual information from low-level to high-level; The Neck module is used to fuse the features at different levels extracted by the backbone network to improve the accuracy of object detection. The segmentation grid is used to divide the feature map into square grids (cells), and each grid is responsible for detecting the pores that may exist in the area.

[0027] Finally, the Head is used to convert the features extracted by the backbone network into prediction results, including the category and location of the object, as well as the location of the key points. The size of the feature map is doubled using the nearest neighbor interpolation method; the concatenation layer concatenates the output and the output of the 6th layer in the backbone network in the channel dimension. The upsampling layer and the concatenation layer gradually enlarge the size of the feature map and fuse the features from different levels to better detect pores and key points; the pose module is defined to receive the outputs of the 16th, 19th, and 22nd layers and use these features to predict the category, location, and key points of the object; the predicted anchor boxes are used to predict a fixed number of bounding boxes at a preset scale for each anchor box. Each bounding box contains 5 elements: the center point coordinates, aspect ratio, and the probability and confidence of the corresponding category; The non-maximum suppression algorithm is applied to find the box with the highest confidence from all the detection boxes, calculate its IOU with the remaining boxes, and if the value is greater than a certain threshold (too high overlap), the box will be removed to eliminate duplicate predictions.

[0028] The predictions with higher scores in the stripe pore category are selected, and the category label, location, and confidence are returned as the final detection results (such as Figure 3 a3, b3 in Figure 4 shown as c3, d3 in

[0029] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit it; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some or all of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for detecting pores between lattice fringes of coal aromatic hydrocarbons based on HRTEM images, characterized in that: The detection steps include the following: S100: collecting an image of the surface of the coal body to be detected as an original image, and sequentially performing grayscale processing, shearing, Fourier-inverse Fourier transform, image binarization processing, and block division on the original image; S200: performing normalization processing on the binarized image and adjusting the image size to the size required by the model; S300: using annotation software to annotate the stripe pores in the normalized image to obtain the main pore categories and positions; S400: The pore image obtained by the target detection algorithm is passed to the Neck network after data optimization by the CSC module and the newly added TA layer in the Backbone network. The feature map of the corresponding size is obtained by multi-scale fusion of the newly added P2 layer and the P3-P5 layers in the Neck network, and the pore image is generated into a pore feature map; S500: The pore feature map is input into the Head detection module, which generates an information tensor. After confidence threshold screening and non-maximum suppression, the object category label and label box coordinates are output, and the prediction results of the corresponding scores on the pore category are screened out. The category label, position, and confidence of the prediction result are returned as the final detection result.

2. The method for detecting pores between lattice fringes of coal aromatic hydrocarbons based on HRTEM images according to claim 1, characterized in that: The specific method for collecting and processing the coal body image in step S100 is: S101: Grind the coal sample with a grinder to less than 200 mesh, take a small amount of coal powder with tweezers and place it in 5 ml of anhydrous ethanol, vibrate it with ultrasound for 15 minutes, make the coal powder fully dispersed in the anhydrous ethanol and let it stand for 5 minutes, use a capillary glass tube to absorb the uniform mixture on the upper layer of anhydrous ethanol, and drop it on a 200-mesh copper mesh. After the anhydrous ethanol is dried, place the copper mesh under an electron microscope for observation and photography; S102: Select the edge area of ​​the dried mixed solution for analysis under an electron microscope, import the original HRTEM image into Digital Micrograph software, and select the target area; S103: Under the premise of retaining the microcrystalline structure information in the coal, the image is processed by Fourier and inverse Fourier transform to eliminate the noise in the image; S104: Using Adobe Photoshop software, threshold adjustment is performed on the filtered image to obtain a HRTEM image with clear lattice fringes.

3. The method for detecting pores between lattice fringes of coal aromatic hydrocarbons based on HRTEM images according to claim 2, characterized in that: The specific method of normalizing the image in step S200 is: The image size is resized to 640*640 pixels, and each pixel value is normalized to the range of 0-1 by dividing it by 255, so that the image meets the model processing requirements.

4. The method for detecting pores between lattice fringes of coal aromatic hydrocarbons based on HRTEM images according to claim 3, characterized in that: The specific method of marking the stripe pores in the image in step S300 is: Use labelimg software for labeling, save the labeled labels in txt format, and use the data with bounding box annotation information and HRTEM images to form a training data set.

5. The method for detecting pores between lattice fringes of coal aromatic hydrocarbons based on HRTEM images according to claim 4, characterized in that: The specific method of using the target detection algorithm to process the pore image in step S400 is: S401: Extract features from pore images using target detection algorithms: The pore image data is transmitted to the Backbone network, and the CSC module is used to split the feature map into feature maps X1 and X2 along the channel dimension through the split layer, where: After the feature map X1 is convolved with different kernel sizes, it passes through the S-Bottleneck module and the BN batch normalization layer respectively, and then is fused by the Concat module and activated by the ReLU function to obtain the first output feature; The feature map X2 is passed through two S-Bottleneck modules connected in sequence to obtain the second input feature. After the two are fused, a 1*1 convolution is performed to generate the output feature of the CSC module. A TA layer is added between the CSC module and the Neck network. The TA layer optimizes and integrates the features and then passes them into the Neck network. The image is downsampled at multiple scales through a convolutional neural network to generate a series of feature maps of the pores between stripes; S402: In the Neck network, the newly added P2 layer output is used as the small target detection layer. In the feature enhancement stage, multi-scale feature fusion is performed with the features output by the P3-P5 layers. Upsampling, downsampling, feature concatenation, and convolution operations are used to fully integrate the advantages of features in each layer, generate an enhanced feature map of 160*160*64 size, and transmit it to the Head detection head; S403: using the predicted anchor boxes, predicting a fixed number of bounding boxes for each anchor box at a preset scale, each bounding box includes the center point coordinates, aspect ratio, probability and confidence of the corresponding category, and using the non-maximum suppression algorithm to remove highly similar but close-positioned bounding boxes to eliminate duplicate predictions; S404: During the training process, a cross entropy loss function is used to detect the difference between the predicted result and the true label. The loss function quantifies the deviation in category determination by calculating the cross threshold between the predicted category probability distribution and the true category label. At the same time, combined with the CIoU loss calculation method, the category loss and bounding box loss are integrated according to the corresponding weights to form the final loss; Based on the total loss, the back-propagation algorithm is used to calculate the gradient information of the model parameters relative to the loss. The Adam optimizer is used to calculate the gradient, learning rate, and momentum parameter rules. The weights and bias parameters in the model are gradually updated and optimized. Multiple rounds of iterations are performed to evaluate the model performance on the validation set, calculate the gap between the predicted results and the true labels, and optimize the model parameters.

6. The method for detecting pores between lattice fringes of coal aromatic hydrocarbons based on HRTEM images according to claim 5, characterized in that: The specific method of step S500 is: The Head detection head uses a convolutional layer to convert the received feature map into an output tensor containing the object category probability channel dimension, the bounding box coordinate information dimension, and the object confidence value. The output tensor is processed by first filtering out low-confidence prediction boxes based on the confidence threshold, and then removing redundant prediction boxes through the non-maximum suppression algorithm. Finally, the category label and bounding box coordinate information of the object in the image are output.

Images