Rock particle segmentation method based on PANet and related equipment
The rock particle segmentation method based on PANet integrates a bottom-up feature fusion layer of a feature pyramid network and a path aggregation network, as well as an adaptive feature pooling layer. This solves the problems of low efficiency and low accuracy in traditional methods, and achieves efficient and accurate rock particle segmentation.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- PETROCHINA CO LTD
- Filing Date
- 2024-10-23
- Publication Date
- 2026-04-24
AI Technical Summary
Traditional rock particle segmentation methods rely on manual operation, which is inefficient and inaccurate. Traditional convolutional neural networks tend to ignore features between different layers during feature extraction, and cannot guarantee the integrity and diversity of features.
A rock particle segmentation method based on PANet is adopted, which combines a bottom-up feature fusion layer, an adaptive feature pooling layer, and a fully connected fusion layer of a feature pyramid network and a path aggregation network. The model is optimized through transfer learning to improve the completeness and diversity of feature extraction.
It significantly improves the accuracy and efficiency of rock particle segmentation, enabling rapid and accurate segmentation of rock particles and reducing waste of manpower and resources.
Smart Images
Figure CN121921501A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the field of rock particle segmentation technology, specifically a rock particle segmentation method and related equipment based on PANet. Background Technology
[0002] Grain segmentation is a prerequisite for automated analysis and identification of rock thin sections. Its purpose is to extract mineral grains from rock thin section images and divide them into individual grain target regions. Therefore, the quality of segmentation directly affects the accuracy of the analysis and identification process.
[0003] Traditional visual estimation methods require a high level of operational experience from researchers, and the work of particle labeling is quite arduous. Moreover, it is easy for some samples to be labeled incorrectly due to the subjective factors of the labelers. This method requires a lot of manpower and time, and is inefficient and inaccurate.
[0004] Compared to traditional manual screening methods, image segmentation technology is simpler, faster, and less costly. There are many traditional image segmentation algorithms, each with its own advantages and disadvantages. Threshold-based segmentation algorithms are relatively simple to implement, but require the segmented object to have significant differences from the background features in spatial and color dimensions, placing high demands on the dataset. Edge detection-based segmentation algorithms can improve the problem of discontinuous edges caused by noise disturbances, and their advantage lies in accurate edge localization. However, when the image information is complex, there are many edge breaks, and detail processing becomes blurred. Clustering-based segmentation methods are also commonly used in image segmentation, classifying different categories based on a certain similarity between sample pixels. The difficulty lies in choosing a suitable similarity metric function, and it usually requires manually specifying the final number of categories, which is difficult to implement in rock particle segmentation.
[0005] In recent years, deep learning technology has developed rapidly. It extracts random and salient features from images using neural networks, and then samples these random features multiple times to obtain multiple sets of different random features. Finally, it combines the salient features with these multiple sets of different random features to obtain the optimal segmentation result. The most common method is to use convolutional neural networks for granular segmentation. However, traditional convolutional neural networks, such as CNNs, tend to ignore features between different layers when extracting features, failing to guarantee the completeness and diversity of features. Summary of the Invention
[0006] This invention provides a rock particle segmentation method and related equipment based on PANet, which solves the problem that traditional convolutional neural networks, such as CNN, tend to ignore features between different layers when extracting features, and cannot guarantee the integrity and diversity of features.
[0007] To achieve the above objectives, the present invention provides the following technical solution:
[0008] A rock grain segmentation method based on PANet includes:
[0009] Obtain images of rock particles;
[0010] Input the rock particle image into the trained rock particle segmentation model to obtain the rock particle segmentation result;
[0011] The trained rock particle segmentation model includes a feature pyramid network, a bottom-up feature fusion layer added to the path aggregation network, an adaptive feature pooling layer, a bounding box prediction head of the path aggregation network, and a fully connected fusion layer.
[0012] Preferably, the feature pyramid network comprises:
[0013] Bottom-up on one side, top-down on the other side, and the lateral connection when connecting the bottom-up and top-down;
[0014] The bottom-up process involves inputting an image into a backbone network to extract features. The size of the feature map output by the backbone network remains unchanged or is reduced by a factor of 2. Layers with unchanged output size are grouped into a stage, and the features output by the last layer of each stage are extracted.
[0015] The top-down process involves upsampling the feature map obtained from the higher layers and then passing it down.
[0016] The process of lateral connection is as follows: first, a 1*1 convolution is performed on the feature map output by each stage to reduce the dimensionality. Then, the obtained feature map is added to the feature map sampled from the previous layer. After the addition, a 3*3 convolution is performed to obtain the feature output of this layer.
[0017] Preferably, the bottom-up feature fusion layer added to the path aggregation network includes shallow features. The shallow features are laterally connected in the original path aggregation network and then added to the top layer from the bottom-up feature fusion layer along the path aggregation network. The bottom-up feature fusion layer is formed by passing a convolution with a size of 3*3 and a stride of 2, reducing the feature map size to half of its original size, and then adding it to the feature map. The result is then passed through a convolutional layer with a kernel size of 3*3 and a stride of 1 to obtain the final feature map.
[0018] Preferably, the adaptive feature pooling layer includes multiple feature layers. Each region of interest obtained by the path aggregation network is aligned with multiple feature layers to obtain multiple different feature maps. Then, the multiple different feature maps are fused together to obtain the final feature.
[0019] Preferably, the training method for the path aggregation network model is as follows:
[0020] Construct a rock grain segmentation dataset;
[0021] A rock particle segmentation model based on a path aggregation network was constructed using a rock particle segmentation dataset.
[0022] The rock particle segmentation model is iteratively trained to obtain a well-trained rock particle segmentation model.
[0023] Preferably, the steps for constructing the rock particle segmentation dataset are as follows:
[0024] Images were acquired using a thin section scanner to obtain full-view, fully stitched rock thin section scan images. The scanned images were then sliced into 2178 views, with approximately 60-80 instances of different types of images selected for annotation in each view. Finally, the training set, validation set, and test set were divided in a 6:2:2 ratio.
[0025] Preferably, the steps for iteratively training the rock particle segmentation model are as follows:
[0026] Each sample image in the rock particle segmentation dataset is input into the feature pyramid network structure to obtain multiple fused feature images at different scales.
[0027] Feature images of different scales are input into a bottom-up feature fusion layer to obtain multiple fused feature images of different scales.
[0028] Features are further fused through an adaptive feature pooling layer;
[0029] The fused features are input into the fully connected fusion layer of the prediction mask to obtain the mask prediction result;
[0030] The fused features are input into the bounding box prediction head to predict the category and detection box;
[0031] Based on the predicted segmented image and the labeled image corresponding to the sample image, the gradient descent algorithm is used to optimize the initial path aggregation network model.
[0032] Repeat the above process until the preset conditions are met, then stop training to obtain a rock particle segmentation model based on path aggregation network.
[0033] A PANet-based rock particle segmentation system includes:
[0034] Data acquisition module: used to acquire images of rock particles;
[0035] Segmentation module: Used to input rock particle images into a trained rock particle segmentation model to obtain rock particle segmentation results;
[0036] The trained rock particle segmentation model includes FPN, a bottom-up feature fusion layer added to PANet, an adaptive feature pooling layer, a PANet bounding box prediction head, and a fully connected fusion layer.
[0037] A computer device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the steps of a PANet-based rock particle segmentation method.
[0038] A computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of a PANet-based rock particle segmentation method.
[0039] Compared with existing technologies, this invention has the following advantages: This invention proposes a rock particle segmentation method based on PANet. The proposed rock particle segmentation model integrates a Feature Pyramid Network (FPN), a bottom-up feature fusion layer of a Path Aggregation Network (PANet), an adaptive feature pooling layer, PANet's bounding box prediction head, and a fully connected fusion layer. This model enhances the ability to recognize rock particles of different scales and shapes by fusing multi-layer features through FPN; while PANet's bottom-up fusion layer further improves the utilization of low-level information, enhancing the model's adaptability to complex scenes. The adaptive feature pooling layer improves computational efficiency and robustness, enabling the model to flexibly handle various particles. Combined with PANet's bounding box prediction head, the model achieves accurate particle localization. The fully connected fusion layer enhances feature representation capabilities by integrating multi-view predictions. Overall, this model significantly improves the accuracy and efficiency of rock particle segmentation. Attached Figure Description
[0040] Figure 1 This is a flowchart of a rock particle segmentation method based on PANet according to the present invention;
[0041] Figure 2 These are scanned images of rock thin sections according to embodiments of the present invention;
[0042] Figure 3 This is a structural diagram of a rock particle segmentation model according to an embodiment of the present invention;
[0043] Figure 4 This is a diagram of the adaptive feature pooling layer structure in an embodiment of the present invention;
[0044] Figure 5 This is a diagram of the fully connected fusion layer in an embodiment of the present invention;
[0045] Figure 6 This is a segmentation result diagram of an embodiment of the present invention. Detailed Implementation
[0046] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.
[0047] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.
[0048] It should be noted that similar labels and letters in the following figures indicate similar items. Therefore, once an item is defined in one figure, it does not need to be further defined and explained in subsequent figures.
[0049] To enable those skilled in the art to better understand the technical solution of the present invention, the present invention will be further described in detail below with reference to the accompanying drawings.
[0050] like Figure 1 As shown, one embodiment of the present invention provides a rock particle segmentation method based on PANet, including: acquiring rock particle images;
[0051] Input the rock particle image into the trained rock particle segmentation model to obtain the rock particle segmentation result;
[0052] The trained rock particle segmentation model includes a feature pyramid network, a bottom-up feature fusion layer added to the path aggregation network, an adaptive feature pooling layer, a bounding box prediction head of the path aggregation network, and a fully connected fusion layer.
[0053] A model suitable for rock grain segmentation is constructed. The specific process of this invention is as follows: Figure 1 As shown.
[0054] The detailed steps of this invention are as follows:
[0055] Step 1: Establish a rock grain segmentation dataset
[0056] Images were acquired using the Axio Scan.Z7 thin section scanner to obtain full-view, fully stitched rock thin section scan images. Because the full-view images are too large, and the model input is not that large, information loss would occur. Therefore, we performed a slice operation on the scanned images, dividing them into 2178 views, with approximately 60-80 instances per view. Figure 2 The images are the segmented data. Then, images of different types were selected for annotation, and finally, the training, validation, and test sets were divided in a 6:2:2 ratio.
[0057] Step 2: Construct a PANet-based rock particle segmentation model
[0058] This invention constructs a rock particle segmentation model based on PANet, and the model structure is as follows: Figure 3 As shown, it mainly consists of 5 core modules. Among them, (a) is an FPN, (b) is a bottom-up feature fusion layer added to PANet, (c) is an adaptive feature pooling layer, (d) is the bounding box prediction head of PANet, and (e) is a fully connected fusion layer used for predicting the mask.
[0059] (a) The architecture of this part is mainly a feature pyramid, including a bottom-up on the left, a top-down on the right, and a lateral connection when connecting the left and right parts. The bottom-up process involves inputting the image into the backbone network to extract features. The size of the feature maps output by the backbone network varies, some having a constant size and others reduced by a factor of two. Layers with constant output sizes are grouped into a stage, and the features output from the last layer of each stage are extracted. The top-down process on the right involves upsampling the feature maps obtained from higher layers and then passing them down. This allows the rich semantic information contained in the higher-level features to propagate to the lower-level features, making the lower-level features also rich in semantic information. The lateral connection when connecting the left and right parts mainly involves: first, performing a 1*1 convolution to reduce the dimensionality of the feature map output from each stage; then, combining the resulting features with the feature map sampled from the previous layer. Pn+1 The features are added together, and then passed through a 3x3 convolution to obtain the feature output of this layer. Pn The red arrow in the diagram indicates the bottom-up process in FPN. Features from shallow layers need to go through many layers to reach the top layer, which can cause the loss of a lot of detailed information.
[0060] (b) describes the bottom-up feature fusion layer optimized by PANet. The green arrows in the diagram represent the bottom-up structure added by PANet. In this structure, shallow features are connected laterally in the original FPN network and then fed to the top layer via the bottom-up feature fusion layer added along PANet, effectively preserving shallow feature information. The N3, N4, and N5 in the diagram represent the same feature map, but N3, N4, and N5 are different from P3, P4, and P5; N3, N4, and N5 are the result of fusing P3, P4, and P5. The bottom-up feature fusion layer involves a 3x3 convolution with a stride of 2, reducing the feature map size to half its original size, and then adding it to the Pi+1 feature map. The result is then passed through a 3x3 convolutional layer with a stride of 1 to obtain the final N. i+1 .
[0061] (c) The adaptive feature pooling layer still performs feature fusion. In past RPN networks, the ROIs obtained were based on single-layer features, and FPN was similarly based on single-layer features. In PANet's adaptive feature pooling layer, features from different layers are extracted for each ROI and fused, which is beneficial for improving model performance. The detailed structure of the adaptive feature pooling layer is as follows: Figure 4 As shown, each ROI obtained by the RPN network needs to undergo ROI Align operation with the N2, N3, N4, and N5 feature layers respectively. In this way, four different feature maps are extracted for each ROI. Then, the four different feature maps are fused together to obtain the final feature. Subsequent classification and regression are based on this final feature.
[0062] The final (d) part is a fully connected fusion layer for the prediction mask, structured as follows: Figure 5 This approach introduces a fully connected branch for foreground binary classification into the original segmentation branch, fusing the outputs of these two branches to obtain a more accurate segmentation result. The upper part of the diagram represents the original mask branch, while the lower part shows the fully connected fusion layer added to the PANet network. This is because the fully connected layer can extract different information than the convolutional layers. The main structure adds a conv4_fc-to-mask branch to the original mask branch for fusion. The added branch contains two 3x3 convolutional layers to reduce computation by halving the output channels of the second convolution, followed by a fully connected layer. Here, the fc is used to predict either foreground or background, and then a reshape operation is performed to obtain foreground or background with the same dimensions as the branch above. Finally, the outputs of these two branches are fused using the mask to obtain a more refined final result.
[0063] Step 3: Iterative training of the model
[0064] Loss function settings;
[0065] During model training, the classification loss function of RPN in PANet is set to cross-entropy, and the regression loss function of the bounding box is set to SmoothL1; the classification loss function of the prediction header of the detection box in PANet is set to cross-entropy, the regression loss function of the bounding box is set to SmoothL1, and the segmentation header loss function is set to mask binary cross-entropy.
[0066] Transfer learning based on a general scenario instance segmentation model
[0067] The model construction process utilizes the concept of transfer learning, which has the advantage of finding similarities between existing and new knowledge, helping to learn more general features. At the same time, the initial performance of the model is higher before fine-tuning; the model improves faster during training; and the model converges better after training.
[0068] This invention uses a model trained on the COCO dataset as a foundation, and fine-tunes the model based on the dataset described in this patent through transfer learning to construct a rock particle segmentation model.
[0069] Step 4: Obtain rock particle segmentation results
[0070] By inputting the rock particle image into the PANet network model trained according to the present invention, the result of rock particle instance segmentation can be obtained.
[0071] The PANet network proposed in this invention achieves good results in rock particle segmentation, enabling rapid and efficient segmentation of rock particles. This solves the problems of low accuracy and wasted manpower and resources associated with traditional methods for this task. The effectiveness of this invention in accomplishing this task is as follows: Figure 6 As shown in the figure, mAP was selected as the evaluation metric, and the final evaluation metric was the accuracy on the validation set. Under this evaluation metric, the experimental results are shown in Table 1.
[0072] Table 1 Quantitative Evaluation of Rock Particle Segmentation Effect
[0073]
[0074] Another embodiment of the present invention provides a method for obtaining rock particle segmentation results based on PANet, the method comprising the following steps:
[0075] S101: Data collection and acquisition.
[0076] S102: Data processing.
[0077] S103: Input the acquired field-of-view image of the sandstone casting to be identified into the PANet-based rock grain segmentation model to obtain the grain segmentation results.
[0078] In step S101 above, data acquisition utilizes an optical microscope or an Axio Scan.Z7 thin section scanner to acquire single-polarized images of sandstone castings, adjusts the image magnification, and focuses the microscope to obtain clear microscopic images. Data collection involves downloading thin section microscopic image data of the region of interest and typical regions from a cloud database.
[0079] In step S102 above, data processing refers to dividing the acquired microscopic images into segments of approximately 60-80 instances per field of view.
[0080] In step S103 above, the obtained field-of-view image of the sandstone casting thin section to be identified is input into the PANet-based rock grain segmentation model to obtain the rock grain segmentation result output by the PANet-based rock grain segmentation model.
[0081] In this embodiment of the application, the rock grain segmentation model based on PANet is obtained in the following manner:
[0082] First, a rock grain segmentation dataset was established by collecting and acquiring single-polarized light microscopic images of thin sections of rock cores. If the images are too large, and the input size of the model is also too large, information loss will occur, and the number of target instances the model can detect will be limited. Therefore, the images were cropped to approximately 60-80 instances per view. Then, images from typical view areas were selected and annotated by rock and mineral identification experts, and the training, validation, and test sets were divided in a 6:2:2 ratio.
[0083] Then, the initial PANet model is constructed. The structure of the initial PANet model is as follows: Figure 2 As shown, it includes 5 core modules, namely the Feature Pyramid (FPN), the bottom-up feature fusion layer, the adaptive feature pooling layer, the bounding box prediction head, and the fully connected fusion layer for prediction masks.
[0084] The training steps for constructing a PANet-based rock particle segmentation model based on a rock particle segmentation dataset include at least the following steps.
[0085] S1031: Input each sample image in the rock particle segmentation dataset into the FPN structure to obtain multiple feature images P at different scales after fusion.
[0086] The feature pyramid consists of two stages: Bottom-up and Top-down. The Bottom-up process inputs the image into the backbone network, outputting feature maps with different pooling levels. The Top-down process upsamples the feature maps from higher layers and then propagates them down, allowing the rich semantic information contained in the higher-level features to be transmitted to the lower-level features. Specifically, for each layer's output feature map, a 1x1 convolution is first performed to reduce its dimensionality. Then, the resulting features are combined with the feature map P sampled from the previous layer. n+1 The features are added together, and then subjected to a 3x3 convolution to obtain the feature output P of this layer. n .
[0087] S1032: Input feature images P of different scales into the bottom-up feature fusion layer to obtain multiple feature images N of different scales after fusion.
[0088] The bottom-up feature fusion layer connects shallow features laterally to P in the original FPN network, and then from P to the top layer along the bottom-up feature fusion layer added to PANet, which can better preserve shallow feature information.
[0089] S1033: Features are further fused through an adaptive feature pooling layer.
[0090] The detailed structure of the adaptive feature pooling layer is as follows: Figure 3 As shown, each ROI obtained by the RPN network is subjected to ROI Align operation with the N2, N3, N4, and N5 feature layers respectively, extracting four different feature maps, which are then fused together to obtain the final feature map.
[0091] S1034: Input the fused features into the fully connected fusion layer of the prediction mask to obtain the mask prediction result.
[0092] The fully connected fusion layer structure of the prediction mask is as follows Figure 4 It introduces a fully connected branch for foreground binary classification into the original segmentation branch, and obtains a more accurate segmentation result by fusing the outputs of these two branches.
[0093] S1035: Input the fused features into the bounding box prediction head to predict the category and detection box.
[0094] S1036: Based on the predicted segmented image and the labeled image corresponding to the sample image, optimize the initial PANet model using the gradient descent algorithm.
[0095] S1037: Repeat the iterative training process from S1031 to S1036 until the preset conditions are met, then stop training to obtain the rock particle segmentation model based on PANet.
[0096] The preset conditions are not specifically limited, but are generally set to reach a certain number of iterations, or the accuracy does not change within a certain number of iterations. The preset PANet model can be pre-trained on a large segmentation dataset such as COCO, or it can be pre-trained.
[0097] Another embodiment of the present invention provides a PANet-based rock particle segmentation system, comprising:
[0098] Data acquisition module: used to acquire images of rock particles;
[0099] Segmentation module: Used to input rock particle images into a trained rock particle segmentation model to obtain rock particle segmentation results;
[0100] The trained rock particle segmentation model includes FPN, a bottom-up feature fusion layer added to PANet, an adaptive feature pooling layer, a PANet bounding box prediction head, and a fully connected fusion layer.
[0101] Another embodiment of the present invention provides a rock grain segmentation device for microscopic images of sandstone castings based on PANet, which specifically comprises the following four modules:
[0102] Image input module 101 is used to input collected microscopic images of sandstone casting thin sections;
[0103] Image cropping module 102 is used to crop images based on manual settings;
[0104] Particle segmentation module 103 uses a PANet-based rock particle segmentation model, and the rock particle segmentation results are shown below.
[0105] The result output module 104 outputs and displays the results of rock particle segmentation on the monitor.
[0106] An embodiment of the present invention provides a terminal device. This terminal device includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the computer program, it implements the steps in the various method embodiments described above. Alternatively, when the processor executes the computer program, it implements the functions of each module / unit in the various device embodiments described above.
[0107] The computer program can be divided into one or more modules / units, which are stored in the memory and executed by the processor to complete the present invention.
[0108] The terminal device may be a desktop computer, laptop, handheld computer, or cloud server, etc. The terminal device may include, but is not limited to, a processor and a memory.
[0109] The processor may be a central processing unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc.
[0110] The memory can be used to store the computer program and / or module. The processor implements various functions of the terminal device by running or executing the computer program and / or module stored in the memory and calling the data stored in the memory.
[0111] If the modules / units integrated into the terminal device are implemented as software functional units and sold or used as independent products, they can be stored in a computer-readable storage medium. Based on this understanding, all or part of the processes in the methods of the above embodiments can also be implemented by a computer program instructing related hardware. The computer program can be stored in a computer-readable storage medium, and when executed by a processor, it can implement the steps of the various method embodiments described above. The computer program includes computer program code, which can be in the form of source code, object code, executable files, or certain intermediate forms. The computer-readable medium can include: any entity or device capable of carrying the computer program code, recording media, USB flash drives, portable hard drives, magnetic disks, optical disks, computer memory, read-only memory (ROM), random access memory (RAM), electrical carrier signals, telecommunication signals, and software distribution media, etc. It should be noted that the content included in the computer-readable medium can be appropriately added or removed according to the requirements of legislation and patent practice in the jurisdiction. For example, in some jurisdictions, according to legislation and patent practice, computer-readable media do not include electrical carrier signals and telecommunication signals.
[0112] Although embodiments of the present invention have been described above in conjunction with the accompanying drawings, the present invention is not limited to the specific embodiments and application fields described above. The specific embodiments described above are merely illustrative and instructive, and not restrictive. Those skilled in the art, guided by the specification, can make many other modifications without departing from the scope of the claims of the present invention, and all of these modifications are within the scope of protection of the present invention.
Claims
1. A rock particle segmentation method based on PANet, characterized in that, include: Obtain images of rock particles; Input the rock particle image into the trained rock particle segmentation model to obtain the rock particle segmentation result; The trained rock particle segmentation model includes a feature pyramid network, a bottom-up feature fusion layer added to the path aggregation network, an adaptive feature pooling layer, a bounding box prediction head of the path aggregation network, and a fully connected fusion layer.
2. The rock particle segmentation method based on PANet according to claim 1, characterized in that, The feature pyramid network includes: Bottom-up on one side, top-down on the other side, and the lateral connection when connecting the bottom-up and top-down; The bottom-up process involves inputting an image into a backbone network to extract features. The size of the feature map output by the backbone network remains unchanged or is reduced by a factor of 2. Layers with unchanged output size are grouped into a stage, and the features output by the last layer of each stage are extracted. The top-down process involves upsampling the feature map obtained from the higher layers and then passing it down. The process of lateral connection is as follows: first, a 1*1 convolution is performed on the feature map output by each stage to reduce the dimensionality. Then, the obtained feature map is added to the feature map sampled from the previous layer. After the addition, a 3*3 convolution is performed to obtain the feature output of this layer.
3. The rock particle segmentation method based on PANet according to claim 1, characterized in that, The bottom-up feature fusion layer added to the path aggregation network includes shallow features. The shallow features are connected laterally in the original path aggregation network and then added to the top layer from the bottom-up feature fusion layer along the path aggregation network. The bottom-up feature fusion layer is formed by a 3*3 convolution with a stride of 2, which reduces the feature map size to half of its original size and then adds it to the feature map. The result is then passed through a convolutional layer with a kernel size of 3*3 and a stride of 1 to obtain the final feature map.
4. The rock particle segmentation method based on PANet according to claim 1, characterized in that, The adaptive feature pooling layer consists of multiple feature layers. Each region of interest obtained by the path aggregation network is aligned with multiple feature layers to obtain multiple different feature maps. Then, the multiple different feature maps are fused together to obtain the final feature.
5. The rock particle segmentation method based on PANet according to claim 1, characterized in that, The training method for the path aggregation network model is as follows: Construct a rock grain segmentation dataset; A rock particle segmentation model based on a path aggregation network was constructed using a rock particle segmentation dataset. The rock particle segmentation model is iteratively trained to obtain a well-trained rock particle segmentation model.
6. The rock particle segmentation method based on PANet according to claim 5, characterized in that, The specific steps for constructing a rock grain segmentation dataset are as follows: Images were acquired using a thin section scanner to obtain full-view, fully stitched rock thin section scan images. The scanned images were sliced and divided into 2178 views. Each view was further divided into 60-80 instances of different types of images for annotation. Finally, the training set, validation set, and test set were divided in a 6:2:2 ratio.
7. The rock particle segmentation method based on PANet according to claim 5, characterized in that, The specific steps for iteratively training the rock particle segmentation model are as follows: Each sample image in the rock particle segmentation dataset is input into the feature pyramid network structure to obtain multiple fused feature images at different scales. Feature images of different scales are input into a bottom-up feature fusion layer to obtain multiple fused feature images of different scales. Features are further fused through an adaptive feature pooling layer; The fused features are input into the fully connected fusion layer of the prediction mask to obtain the mask prediction result; The fused features are input into the bounding box prediction head to predict the category and detection box; Based on the predicted segmented image and the labeled image corresponding to the sample image, the gradient descent algorithm is used to optimize the initial path aggregation network model. Repeat the above process until the preset conditions are met, then stop training to obtain a rock particle segmentation model based on path aggregation network.
8. A rock particle segmentation system based on PANet, characterized in that, include: Data acquisition module: used to acquire images of rock particles; Segmentation module: Used to input rock particle images into a trained rock particle segmentation model to obtain rock particle segmentation results; The trained rock particle segmentation model includes FPN, a bottom-up feature fusion layer added to PANet, an adaptive feature pooling layer, a PANet bounding box prediction head, and a fully connected fusion layer.
9. A computer device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the computer program, it implements the steps of the PANet-based rock particle segmentation method as described in any one of claims 1 to 7.
10. A computer-readable storage medium storing a computer program, characterized in that, When the computer program is executed by the processor, it implements the steps of the PANet-based rock particle segmentation method as described in any one of claims 1 to 7.