Method and device for zircon classification, electronic equipment and storage medium

Abstract

The invention provides a zircon classification method and device, electronic equipment and a storage medium. The method comprises the following steps: collecting and sorting a cathode luminescence CL image of zircon to be classified, wherein zircon feature information in the CL image at least comprises a zircon growth girdle, a crack and an inclusion; identifying the to-be-classified zircon by using the target zircon classification model to obtain a zircon CL image classification result; the target zircon classification model comprises a feature extraction module, a multi-scale feature fusion module, a graph attention module and a classification module, the feature extraction module is used for extracting feature information of zircon, and the multi-scale feature fusion module is used for capturing and fusing multi-scale feature graphs; the image attention module processes the fused feature image to obtain high-dimensional features, and the classification module processes the high-dimensional features to obtain a zircon classification result.

Description

Technical Field

[0001] The present invention belongs to the field of artificial intelligence and geoscience big data, and specifically relates to a method, device, electronic equipment and storage medium for zircon classification. Background Art

[0002] Image classification is an important task in the field of computer vision. In today's era of continuous progress in sedimentary geology, advanced image classification continues to emerge. A large amount of geological data information has always been an important material in geological research. Specifically, in zircon analysis, scholars have been exploring how to quickly and accurately obtain zircon categories. Among them, zircon CL image data sets can include characteristics such as growth rings, cracks and inclusions in the image. In related technologies, the current traditional zircon classification method relies on manual microscope observation or classifies zircon types based on Th / U ratios, but often due to human and growth reasons, there are large deviations in classification. Later, it was discovered that zircon cathode luminescence (CL) images can be used as an important analysis method. The CL images of zircons can reveal their internal bedding structures, inclusions and microcracks, and can provide key information on the formation and evolution of zircons. However, most of the current classification methods manually classify zircon crystal forms based on CL images. Because the internal structure of zircon is very different and diverse due to the growth environment of zircon, manual classification requires professional geological background knowledge, which makes it highly subjective and requires a lot of time and labor. Therefore, many geologists are actively exploring the introduction of automation technology to automatically classify images using features in the images to improve efficiency and reduce duplication of work.

[0003] In mineralogy and petrology research, deep learning models have begun to be used to automatically identify and analyze the microscopic features of mineral images, which can improve the efficiency and accuracy of data processing, but there are still many challenges in zircon classification. However, due to the large differences in zircon image data from different sources and the diversity within zircon images, the generalization ability of the migration model may be affected. Secondly, the deep learning model is not interpretable enough and it is difficult to provide sufficient geological basis in practical applications.

[0004] Therefore, the present application contemplates a new method for zircon classification. Summary of the invention

[0005] In view of the problems existing in the prior art, the present invention provides a method, device, electronic device and storage medium for zircon classification, aiming to improve the accuracy and efficiency of zircon classification.

[0006] In a first aspect, an embodiment of the present invention provides a method for classifying zircon, comprising:

[0007] Collecting and arranging cathode luminescence (CL) images of zircons to be classified, wherein characteristic information of zircons in the CL images at least includes zircon growth zones, cracks and inclusions;

[0008] The target zircon classification model is used to identify the zircon to be classified, so as to obtain the zircon classification result;

[0009] The target zircon classification model includes a feature extraction module, a multi-scale feature fusion module, a graph attention module and a classification module. The feature extraction module is used to extract feature information of zircon, the multi-scale feature fusion module is used to capture and fuse multi-scale feature maps, the graph attention module processes the fused feature maps to obtain high-dimensional features, and the classification module processes the high-dimensional features to obtain zircon classification results.

[0010] Furthermore, before using the target zircon classification model to identify the zircon to be classified to obtain the classification result of the zircon, it also includes training and obtaining the target zircon classification model, and the training and obtaining the target zircon classification model includes:

[0011] Collecting and arranging the CL images of the zircon to be tested, and obtaining the category labels corresponding to the CL images of the zircon to be tested;

[0012] Using the CL images of the zircon to be tested and the category labels to create a training data set, a verification data set and a test data set;

[0013] Construct an initial zircon classification model;

[0014] Based on the training data set and the validation data set, the initial zircon classification model is optimized using a stochastic gradient descent method to obtain an optimized zircon classification model;

[0015] Performing test segmentation on the zircon CL images in the test data set based on the optimized zircon classification model to obtain a predicted category result;

[0016] The predicted category result is compared and analyzed with the category label corresponding to the test data set to obtain the target zircon classification model.

[0017] Further, obtaining the optimized zircon classification model includes:

[0018] Training the initial zircon classification model using a stochastic gradient descent optimizer and a cross entropy loss function; and

[0019] A data enhancement strategy was used to optimize the initial zircon classification model.

[0020] Further, the CL image is input into a target zircon classification model according to a preset size;

[0021] In the feature extraction module, a 3×3 convolution operation is performed to capture local fine-grained information in the input image to obtain a feature map, the feature map is normalized, and an activation function and an inverted residual block including an extended convolution, a depth convolution, and a projection convolution are used to perform feature transformation and dimension adjustment on the feature map to obtain a multi-scale feature map;

[0022] In the multi-scale feature fusion module, the multi-scale feature map is processed using dilated convolutions with different dilation rates to capture multi-scale information, the feature maps are spliced, each feature map is globally averaged pooled to generate global features, the spliced ​​feature maps are 1×1 convolved to integrate information of different scales;

[0023] In the graph attention module, a linear transformation is performed through a weight matrix, and a new feature representation is obtained based on the integration of information at different scales. The attention mechanism is used to construct a combined feature for each pair of nodes. The neighborhood of each node is normalized using the Softmax function, and then feature aggregation is performed. The features of neighboring nodes are aggregated using the normalized attention coefficient to obtain high-dimensional features.

[0024] In the classification module, high-dimensional features are mapped to the category space through fully connected layers, activation functions and average pooling operations.

[0025] Furthermore, the inverted residual block including the dilated convolution, the depth convolution and the projection convolution performs feature transformation and dimension adjustment on the feature map to obtain a multi-scale feature map in the following process:

[0026] x exp =Conv2d(x, W exp )

[0027] x dw =DepthwiseConv2d(x exp , W dw )

[0028] x proj =Conv2d(x dw , W proj )

[0029] y=x+x proj

[0030] Where x is the input tensor, W exp is the weight matrix of the dilated convolution, x exp is the expanded feature map, W dw is the weight matrix of the depthwise convolution, x dw is the feature map after deep convolution, W proj is the weight matrix of the projected convolution, x proj is the feature map after projection, xproj is the feature map after projection, and y is the output tensor after residual connection.

[0031] Furthermore, obtaining the multi-scale feature map includes:

[0032] y atrous =Conv2d(x, W, dilation=d), d=1, 6, 12, 18

[0033] y concat =[y d＝1 ;y d＝6 ;y d＝12 ;y d＝18 ]

[0034] y gap =GlobalAvgPool(x)

[0035] y=Conv2d(y concat , W 1×1 )

[0036] Where x is the feature map obtained by the feature extraction module, W is the weight matrix of the convolution kernel, d is the expansion rate, and y is atrous is the output feature map of the dilated convolution, y concat is the concatenated feature map, y gap is the feature map after global average pooling, W 1×1 is the weight matrix of 1×1 convolution, and y is the feature map after multi-scale feature fusion.

[0037] Furthermore, obtaining the high-dimensional features includes:

[0038] H′=HW

[0039] e ij =LeakyReLU(a t [H′ i ||H′ J ])

[0040]

[0041] H″ i =σ(∑ k∈N(i) α ij H′ j

[0042] Where H is the feature map after multi-scale feature fusion, W is the weight matrix of feature transformation, H′ is the transformed feature matrix, and H′ i , H′ j is the transformed feature vector, a is the weight vector of the attention mechanism, and e ijis the attention score between nodes i and j, N(i) is the neighboring node of node i, α ij is the normalized attention coefficient, H′ j is the feature vector of the neighbor node, σ is the activation function, H″ j is the aggregated feature vector.

[0043] In a second aspect, an embodiment of the present disclosure provides a device for classifying zircons, the device comprising:

[0044] An image acquisition unit is configured to: acquire and organize zircon cathode luminescence (CL) images to be classified, wherein the CL images include characteristic information such as zircon growth zones, cracks, and inclusions;

[0045] A zircon classification unit is configured to use a target zircon classification model to identify zircons to be classified to obtain a classification result of the zircon; wherein the target zircon classification model includes a feature extraction module, a multi-scale feature fusion module, a graph attention module and a classification module, the feature extraction module is used to extract feature information of zircon, the multi-scale feature fusion module is used to capture and fuse multi-scale feature maps, the graph attention module processes the fused feature maps to obtain high-dimensional features, and the classification module processes the high-dimensional features to obtain a zircon classification result.

[0046] In a third aspect, an embodiment of the present disclosure provides an electronic device, the electronic device comprising:

[0047] at least one processor; and,

[0048] a memory communicatively connected to the at least one processor; wherein,

[0049] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the above-mentioned method for zircon classification.

[0050] In a fourth aspect, an embodiment of the present disclosure provides a non-transitory computer-readable storage medium, which stores computer instructions for causing the computer to execute the above-mentioned method for zircon classification.

[0051] Other optional features and technical effects of the embodiments of the present invention are partially described below, and partially can be understood by reading this document.

[0052] Compared with the prior art, the present invention has the following beneficial technical effects:

[0053] The present invention provides a method for zircon classification, including: collecting and collating cathode luminescence CL images of zircons to be classified, wherein the zircon feature information in the CL images at least includes zircon growth rings, cracks and inclusions; using a target zircon classification model to identify the zircons to be classified to obtain zircon classification results; the target zircon classification model includes a feature extraction module, a multi-scale feature fusion module, a graph attention module and a classification module, wherein the feature extraction module is used to extract feature information of zircons, the multi-scale feature fusion module is used to capture and fuse multi-scale feature maps, the graph attention module processes the fused feature maps to obtain high-dimensional features, and the classification module processes the high-dimensional features to obtain zircon classification results. The present application can automatically identify the zircon category by sending the zircon CL images to be processed into the target zircon classification model, and has efficient classification capabilities. Compared with the classification results in other classification models, the present application only needs to perform simple resolution adaptation to classify zircons. Therefore, the present application has a wide range of applications in geological image classification. BRIEF DESCRIPTION OF THE DRAWINGS

[0054] Figure 1 A schematic flow chart of a method for zircon classification according to an embodiment of the present disclosure is shown;

[0055] Figure 2 A zircon CL image dataset according to an embodiment of the present disclosure is shown;

[0056] Figure 3 A schematic diagram of a zircon classification model according to an embodiment of the present disclosure is shown;

[0057] Figure 4 A schematic diagram of the optimization of the zircon classification model according to the present disclosure is shown;

[0058] Figure 5 The confusion matrix results of the zircon classification model of the embodiment of the present disclosure and other models are shown;

[0059] Figure 6 The visualization result of the zircon classification model of the embodiment of the present disclosure is shown;

[0060] Figure 7 A diagram showing a device for classifying zircon according to an embodiment of the present disclosure is shown. DETAILED DESCRIPTION

[0061] The embodiments of the present disclosure are described in detail below with reference to the accompanying drawings.

[0062] The following describes the embodiments of the present disclosure through specific examples, and those skilled in the art can easily understand other advantages and effects of the present disclosure from the contents disclosed in this specification. Obviously, the described embodiments are only a part of the embodiments of the present disclosure, rather than all of the embodiments. The present disclosure can also be implemented or applied through other different specific embodiments, and the details in this specification can also be modified or changed in various ways based on different viewpoints and applications without departing from the spirit of the present disclosure. It should be noted that the following embodiments and features in the embodiments can be combined with each other without conflict. Based on the embodiments in the present disclosure, all other embodiments obtained by ordinary technicians in the field without making creative work are within the scope of protection of the present disclosure.

[0063] It should be noted that various aspects of the embodiments within the scope of the appended claims are described below. It should be apparent that the aspects described herein may be embodied in a wide variety of forms, and any specific structure and / or function described herein is merely illustrative. Based on the present disclosure, it should be understood by those skilled in the art that an aspect described herein may be implemented independently of any other aspect, and two or more of these aspects may be combined in various ways. For example, any number of aspects described herein may be used to implement the device and / or practice the method. In addition, other structures and / or functionalities other than one or more of the aspects described herein may be used to implement this device and / or practice this method.

[0064] It should also be noted that the illustrations provided in the following embodiments are only schematic illustrations of the basic concept of the present disclosure. The drawings only show components related to the present disclosure rather than being drawn according to the number, shape and size of components in actual implementation. In actual implementation, the type, quantity and proportion of each component may be changed arbitrarily, and the component layout may also be more complicated.

[0065] Additionally, in the following description, specific details are provided to facilitate a thorough understanding of the examples. However, it will be understood by those skilled in the art that the aspects described may be practiced without these specific details.

[0066] Figure 1 A flow chart 100 of a method for classifying zircon according to an embodiment of the present disclosure is shown. Figure 1 As shown, including:

[0067] In step S101, cathode luminescence (CL) images of zircons to be classified are collected and sorted, wherein the zircon characteristic information in the CL images at least includes zircon growth zones, cracks and inclusions. Figure 2 The zircon CL image dataset of the embodiment of the present disclosure is shown, wherein ai respectively show different zircon CL image datasets, and the scale bars thereof are all 50 μm.

[0068] Next, go to step S102.

[0069] In step S102, the target zircon classification model is used to identify the zircon to be classified to obtain the zircon classification result. Specifically, Figure 3 The target zircon classification model of the embodiment of the present disclosure is shown, and the target zircon classification model includes a feature extraction module (Feature Extraction Module, FEM), a multi-scale feature fusion module (Feature Fusion Module, FFM), a graph attention module (Graph Attention Module, GAT) and a classification module (or classifier). The feature extraction module is used to extract the feature information of zircon, the multi-scale feature fusion module is used to capture and fuse multi-scale feature maps, the graph attention module processes the fused feature maps to obtain high-dimensional features, and the classification module processes the high-dimensional features to obtain the zircon classification results.

[0070] More specifically, if Figure 3 As shown, in an embodiment of the present disclosure, the zircon CL image (Image) is input into the target zircon classification model according to a preset size, that is, according to an input image of a given H×W size.

[0071] In the Feature Extraction Module (FEM), a 3×3 convolution operation is first performed on the input image to capture local fine-grained information in the image, reduce the size of the feature map, and extract important image detail information. The feature map is normalized (Batch Normalization, BN), and the Hardswish (CBH for short: Conv+BatchNorm+Hardswish) activation function is used to increase the nonlinear expression ability of the model. By using an inverted residual block (InvertedResidual block) containing extended convolution (Depthwise Conv), deep convolution (Conv) and projection convolution (Projection Conv), the feature map is transformed and dimensionally adjusted, and multi-scale features are effectively extracted and fused. At the same time, the residual connection (Residual Connection, Res) helps to alleviate the gradient vanishing problem, improve the training efficiency of the model, and obtain multi-scale feature maps.

[0072] In the multi-scale feature fusion module (FFM), atrous convolution with different dilation rates (dilation rate = 1, 6, 12, 18) is used to process the multi-scale feature map, capture multi-scale information, and splice the feature map. Global average pooling is performed on the processed multi-scale feature map to generate global features, and then 1×1 convolution (1×1Conv) is performed on the spliced ​​feature map to integrate information of different scales. In this way, information of different scales is effectively fused, and finally a fusion map with global and local features is obtained.

[0073] In the Graph Attention Module (GAT), the features are first matrix multiplied to achieve linear transformation of the features and obtain preliminary feature representation. Then, the feature concatenation operation is used to combine features from different sources to enhance the information expression capability. The attention mechanism is used to construct combined features for each pair of nodes, and the neighborhood of each node is normalized (Batch Normalization, BN) using the Softmax function to generate attention weights. Subsequently, the neighborhood node features are aggregated using the normalized (Batch Normalization, BN) attention coefficient through a weighted sum operation to obtain a high-dimensional feature representation. In this process, the interaction between the LeakyReLU activation function and the weight matrix contributes to the nonlinear transformation of features.

[0074] In the classification module (Classifier), the high-dimensional features are mapped to the category space through the fully connected layer (Fully Connected Layer), and the classification expression ability is enhanced by using the Hardswish activation function. Average pooling is applied to the high-dimensional features for dimensionality reduction. Finally, the image classification task is completed through the Softmax function.

[0075] Specifically, in some embodiments, the inverted residual block (InvertedResidual block) including the extended convolution (Depthwise Conv), the depth convolution (Conv) and the projection convolution (Projection Conv) performs feature transformation and dimension adjustment on the feature map to obtain a multi-scale feature map, and the process of obtaining the multi-scale feature map by residual connection (Residual Connection, Res) is as follows:

[0076] x exp =Conv2d(x, W exp )

[0077] x dw =DepthwiseConv2d(x exp , W dw )

[0078] x proj =Conv2d(x dw , W proj )

[0079] y=x+x proj

[0080] Where x is the input tensor, W exp is the weight matrix of the dilated convolution, x exp is the expanded feature map, W dw is the weight matrix of the depthwise convolution, x dw is the feature map after deep convolution, W proj is the weight matrix of the projected convolution, x proj is the feature map after projection, x proj is the feature map after projection, and y is the output tensor after residual connection.

[0081] Specifically, in some embodiments, atrous convolution with different dilation rates (dilation rate = 1, 6, 12, 18) is used to process the multi-scale feature map, capture multi-scale information, perform feature map splicing, perform global average pooling on the processed multi-scale feature map, generate global features, and then perform 1×1 convolution (1×1Conv) on the spliced ​​feature map. The multi-scale feature map obtained by integrating information of different scales includes:

[0082] y atrous =Conv2d(x, W, dilation=d), d=1, 6, 12, 18

[0083] y concat=[y d＝1 ;y d＝6 ;y d＝12 ;y d＝18 ]

[0084] y gap =GlobalAvgPool(x)

[0085] y=Conv2d(y concat , W 1×1 )

[0086] Where x is the feature map obtained by the feature extraction module, W is the weight matrix of the convolution kernel, d is the expansion rate, and y is atrous is the output feature map of the dilated convolution, y concat is the concatenated feature map, y gap is the feature map after global average pooling, W 1×1 is the weight matrix of 1×1 convolution, and y is the feature map after multi-scale feature fusion.

[0087] Specifically, in some embodiments, matrix multiplication is performed on the features to achieve linear transformation of the features and obtain preliminary feature representation. The feature concatenation operation is used to combine features from different sources to enhance the ability to express information. The attention mechanism is used to construct combined features for each pair of nodes, and the Softmax function is used to normalize the neighborhood of each node (BatchNormalization, BN) to generate attention weights. Subsequently, the neighborhood node features are aggregated using the normalized (BatchNormalization, BN) attention coefficient through a weighted sum operation to obtain a high-dimensional feature representation. In this process, the interaction between the LeakyReLU activation function and the weight matrix contributes to the nonlinear transformation of the features, and the high-dimensional features obtained include:

[0088] H′=HW

[0089] e ij =LeakyReLU(a T [H′ i ||H′ J ])

[0090]

[0091] H″ i =σ(∑ k∈N(i) α ij H′ j

[0092] Where H is the feature map after multi-scale feature fusion, W is the weight matrix of feature transformation, H′ is the transformed feature matrix, and H′ i , H′ j is the transformed feature vector, a is the weight vector of the attention mechanism, and e ij is the attention score between nodes i and j, N(i) is the neighboring node of node i, α ij is the normalized attention coefficient, H′ j is the feature vector of the neighbor node, σ is the activation function, H″ i is the aggregated feature vector.

[0093] Specifically, in some embodiments, the input features are mapped to a high-dimensional space through a fully connected layer, a Hardswish activation function is used, an average pooling is used, and finally a Softmax function is used to map the features to a category space, where:

[0094] y fc =W fc x+b fc

[0095] y act =Hardswish(y fc )

[0096] y pool =AdaptiveAvgPool2d(y act , (1, 1))

[0097]

[0098] Where x is the feature vector calculated by graph attention, W fc is the weight matrix of the fully connected layer, b fc is the bias term of the fully connected layer, y fc is the feature after the fully connected layer, y act is the activated feature, y pool is the feature after average pooling, W out is the weight matrix of the output layer, b out The bias term of the output layer, is the final classification result.

[0099] In some embodiments, before using the target zircon classification model to identify the zircon to be classified to obtain the classification result of the zircon, it also includes training and obtaining the target zircon classification model, and the training and obtaining the target zircon classification model includes:

[0100] Collecting and arranging the CL images of the zircon to be tested, and obtaining the category labels corresponding to the CL images of the zircon to be tested;

[0101] Using the CL images of the zircon to be tested and the category labels to create a training data set, a verification data set and a test data set;

[0102] Construct an initial zircon classification model;

[0103] Based on the training data set and the validation data set, the initial zircon classification model is optimized using a stochastic gradient descent method to obtain an optimized zircon classification model;

[0104] Performing test segmentation on the zircon CL images in the test data set based on the optimized zircon classification model to obtain a predicted category result;

[0105] The predicted category result is compared and analyzed with the category label corresponding to the test data set to obtain the target zircon classification model.

[0106] In some embodiments, the CL image to be tested needs to be preprocessed to exclude some non-zircon, overexposed and blurred images to reduce the interference of random noise and improve the clarity of the image; specifically, Gaussian filtering and median filtering methods are used for denoising preprocessing, and the image is scaled to a preset resolution to obtain a preprocessed CL image; then, the preprocessed CL image is divided into a training data set, a verification data set and a test data set according to a preset ratio.

[0107] In an embodiment of the present disclosure, obtaining the optimized zircon classification model includes:

[0108] The initial zircon classification model is trained using a stochastic gradient descent (SGD) optimizer and a cross entropy loss function; wherein an initial value of the learning rate of the initial zircon classification model is set to 0.005; and

[0109] A data enhancement strategy was used to optimize the initial zircon classification model.

[0110] like Figure 4 As shown, by performing data augmentation (Augment) after inputting the image (Image). In some embodiments, the data augmentation strategy includes random cropping, resizing, horizontal flipping, random rotation, color adjustment and automatic augmentation strategy (Auto Augment).

[0111] In some embodiments, the first prediction result of the zircon classification model is compared with the first prediction results of other classification models to analyze the improvement of accuracy, recall, precision and F1 score, where the F1 score is the harmonic mean of precision and recall. Figure 5 The confusion matrix results of the optimized zircon classification model of the disclosed embodiment and other models are shown, wherein the other models include model Resnet50, model VGG16, model Efficientnet, model MobilenetV2 and model MobilenetV3, as well as the optimized zircon classification model Ours in this embodiment, and the evaluation indicators include data augmentation Data_aug, accuracy Accuary, recall Recall, precision Precision and F1 score F1score.

[0112] Table 1 shows the evaluation indicators of the optimized zircon classification model of the embodiment of the present disclosure and other models.

[0113] Table 1. Evaluation indicators of the optimized zircon classification model and other models

[0114]

[0115] The visualization results include three types: Hydrothermal, Igneous and Metamorphic. Hydrothermal means that these zircons may be igneous rocks, and Metamorphic means that these zircons may be crystalline structures. The criteria for judging these types are shown in Table 2:

[0116] Table 2. Evaluation indexes of three types of zircon

[0117]

[0118]

[0119] The second embodiment of the present invention further provides a device for classifying zircons, the device comprising:

[0120] An image acquisition unit is configured to acquire and organize zircon cathode luminescence (CL) images to be classified, wherein the zircon characteristic information in the CL images at least includes zircon growth zones, cracks and inclusions;

[0121] A zircon classification unit is configured to use a target zircon classification model to identify zircons to be classified to obtain a classification result of the zircon; wherein the target zircon classification model includes a feature extraction module, a multi-scale feature fusion module, a graph attention module and a classification module, the feature extraction module is used to extract feature information of zircon, the multi-scale feature fusion module is used to capture and fuse multi-scale feature maps, the graph attention module processes the fused feature maps to obtain high-dimensional features, and the classification module processes the high-dimensional features to obtain a zircon classification result.

[0122] A third embodiment of the present invention further provides an electronic device, the electronic device comprising:

[0123] at least one processor; and,

[0124] a memory communicatively connected to the at least one processor; wherein,

[0125] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to perform the method for zircon classification of any one of the aforementioned embodiments.

[0126] The fourth embodiment of the present invention further provides a non-transitory computer-readable storage medium, wherein the non-transitory computer-readable storage medium stores computer instructions, and the computer instructions are used to enable the computer to execute the method for zircon classification described in any of the aforementioned embodiments.

[0127] The fifth embodiment of the present invention also provides a computer program product, which includes a computer program stored on a non-transitory computer-readable storage medium, and the computer program includes program instructions. When the program instructions are executed by a computer, the computer executes the method for zircon classification of any of the aforementioned embodiments.

[0128] Figure 6 A schematic diagram of a method or device 1000 that can implement an embodiment of the present invention is shown, which may include more or fewer devices than shown in the figure in some embodiments. In some embodiments, it can be implemented using a single or multiple devices. In some embodiments, it can be implemented using cloud or distributed devices.

[0129] like Figure 6As shown, the device 1000 includes a processor 1001, which can perform various appropriate operations and processes according to the program and / or data stored in the read-only memory (ROM) 1002 or the program and / or data loaded from the storage part 1008 to the random access memory (RAM) 1003. Processor 1001 can be a multi-core processor or can include multiple processors. In some embodiments, processor 1001 can include a general main processor and one or more special coprocessors, such as a central processing unit (CPU), a graphics processing unit (GPU), a neural network processor (NPU), a digital signal processor (DSP), etc. In RAM 1003, various programs and data required for the operation of device 1000 are also stored. Processor 1001, ROM 1002 and RAM 1003 are connected to each other via bus 1004. Input / output (I / O) interface 1005 is also connected to bus 1004.

[0130] The processor and the memory are used together to execute the program stored in the memory. When the program is executed by the computer, the methods, steps or functions described in the above embodiments can be implemented.

[0131] The following components are connected to the I / O interface 1005: an input section 1006 including a keyboard, a mouse, a touch screen, etc.; an output section 1007 including a cathode ray tube (CRT), a liquid crystal display (LCD), etc., and a speaker, etc.; a storage section 1008 including a hard disk, etc.; and a communication section 1009 including a network interface card such as a LAN card, a modem, etc. The communication section 1009 performs communication processing via a network such as the Internet. A drive 1010 is also connected to the I / O interface 1005 as needed. A removable medium 1011, such as a magnetic disk, an optical disk, a magneto-optical disk, a semiconductor memory, etc., is installed on the drive 1010 as needed, so that a computer program read therefrom is installed into the storage section 1008 as needed. Figure 6 Only some components are schematically shown in the figure, which does not mean that the device 1000 only includes Figure 6 Components shown.

[0132] The systems, devices, modules or units described in the above embodiments may be implemented by a computer or its associated components. The computer may be, for example, a mobile terminal, a smart phone, a personal computer, a laptop computer, a vehicle-mounted human-computer interaction device, a personal digital assistant, a media player, a navigation device, a game console, a tablet computer, a wearable device, a smart TV, an Internet of Things system, a smart home, an industrial computer, a server or a combination thereof.

[0133] Although not shown, in an embodiment of the present invention, a computer-readable storage medium is provided, on which a computer program / instruction is stored, and when the computer program / instruction is executed by a processor, the method for classifying zircon described in the embodiment is implemented.

[0134] Storage media in embodiments of the present invention include permanent and non-permanent, removable and non-removable items that can be used to store information by any method or technology. Examples of storage media include, but are not limited to, phase change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technology, compact disk read-only memory (CD-ROM), digital versatile disk (DVD) or other optical storage, magnetic cassettes, magnetic tape magnetic disk storage or other magnetic storage devices or any other non-transmission medium that can be used to store information that can be accessed by a computing device.

[0135] Although not shown, an embodiment of the present invention further provides a computer program product, including: a computer program / instruction, which, when executed by a processor, implements the method for classifying zircon described in the embodiment.

[0136] The methods, programs, systems, devices, etc. of the embodiments of the present invention may be executed or implemented in a single or multiple networked computers, or may be practiced in a distributed computing environment. In the embodiments of this specification, in these distributed computing environments, tasks may be performed by remote processing devices connected via a communication network.

[0137] Those skilled in the art should understand that the embodiments of the present specification may be provided as methods, systems or computer program products. Therefore, those skilled in the art may imagine that the functional modules / units or controllers and related method steps described in the above embodiments may be implemented in software, hardware or a combination of software / hardware.

[0138] Unless explicitly stated, the actions or steps of the methods, programs, and embodiments of the present invention do not have to be performed in a specific order and can still achieve the desired results. In some implementations, multitasking and parallel processing are also possible or may be advantageous.

[0139] In this article, multiple embodiments of the present invention are described, but for the sake of brevity, the description of each embodiment is not exhaustive, and the same or similar features or parts between the embodiments may be omitted. In this article, "one embodiment", "some embodiments", "examples", "specific examples", or "some examples" are meant to be applicable to at least one embodiment or example according to the present invention, but not all embodiments. The above terms do not necessarily mean to refer to the same embodiment or example. In the absence of contradiction, those skilled in the art can combine and combine the different embodiments or examples described in this specification and the features of the different embodiments or examples.

[0140] The exemplary systems and methods of the present invention have been specifically shown and described with reference to the above embodiments, which are merely examples of the best modes for implementing the present systems and methods. It will be appreciated by those skilled in the art that various changes may be made to the embodiments of the systems and methods described herein when implementing the present systems and / or methods without departing from the spirit and scope of the present invention as defined in the appended claims.

[0141] In addition, the method, system, device, and medium for zircon classification according to the present invention can also be implemented in the following manner:

[0142] (1) A method for zircon classification, comprising:

[0143] Collecting and collating cathode luminescence (CL) images of zircons to be classified, wherein the CL images include characteristic information such as zircon growth zones, cracks and inclusions;

[0144] The target zircon classification model is used to identify the zircon to be classified, so as to obtain the zircon classification result;

[0145] The target zircon classification model includes a feature extraction module, a multi-scale feature fusion module, a graph attention module and a classification module. The feature extraction module is used to extract feature information of zircon, the multi-scale feature fusion module is used to capture and fuse multi-scale feature maps, the graph attention module processes the fused feature maps to obtain high-dimensional features, and the classification module processes the high-dimensional features to obtain zircon classification results.

[0146] (2) The method according to (1) is characterized in that before using the target zircon classification model to identify the zircon to be classified to obtain the zircon classification result, it also includes training and obtaining the target zircon classification model, and the training and obtaining the target zircon classification model includes:

[0147] Collecting and arranging the CL images of the zircon to be tested, and obtaining the category labels corresponding to the CL images of the zircon to be tested;

[0148] Using the CL images of the zircon to be tested and the category labels to create a training data set, a verification data set and a test data set;

[0149] Construct an initial zircon classification model;

[0150] Based on the training data set and the validation data set, the initial zircon classification model is optimized using a stochastic gradient descent method to obtain an optimized zircon classification model;

[0151] Performing test segmentation on the zircon CL images in the test data set based on the optimized zircon classification model to obtain a predicted category result;

[0152] The predicted category result is compared and analyzed with the category label corresponding to the test data set to obtain the target zircon classification model.

[0153] (3) The method according to (2) is characterized in that obtaining the optimized zircon classification model comprises:

[0154] Training the initial zircon classification model using a stochastic gradient descent optimizer and a cross entropy loss function; and

[0155] A data enhancement strategy was used to optimize the initial zircon classification model.

[0156] (4) The method according to (1), characterized in that

[0157] Inputting the CL image into a target zircon classification model according to a preset size;

[0158] In the Feature Extraction Module (FEM), a 3×3 convolution operation is first performed on the input image to capture local fine-grained information in the image, reduce the size of the feature map, and extract important image detail information. The feature map is normalized (Batch Normalization, BN), and the Hardswish (CBH for short: Conv+BatchNorm+Hardswish) activation function is used to increase the nonlinear expression ability of the model. By using an inverted residual block (InvertedResidual block) containing extended convolution (Depthwise Conv), deep convolution (Conv) and projection convolution (Projection Conv), the feature map is transformed and dimensionally adjusted, and multi-scale features are effectively extracted and fused. At the same time, the residual connection (Residual Connection, Res) helps to alleviate the gradient vanishing problem, improve the training efficiency of the model, and obtain multi-scale feature maps.

[0159] In the multi-scale feature fusion module (FFM), atrous convolution with different dilation rates (dilation rate = 1, 6, 12, 18) is used to process the multi-scale feature map, capture multi-scale information, splice the feature map, perform global average pooling on the processed multi-scale feature map, generate global features, and then perform 1×1 convolution (1×1Conv) on the spliced ​​feature map to integrate information of different scales. In this way, information of different scales is effectively fused, and finally a fusion map with global and local features is obtained.

[0160] In the Graph Attention Module (GAT), the features are first matrix multiplied to achieve linear transformation of the features and obtain preliminary feature representation. Then, the feature concatenation operation is used to combine features from different sources to enhance the ability to express information. The attention mechanism is used to construct combined features for each pair of nodes, and the neighborhood of each node is normalized (Batch Normalization, BN) using the Softmax function to generate attention weights. Subsequently, the neighborhood node features are aggregated using the normalized (Batch Normalization, BN) attention coefficient through a weighted sum operation to obtain a high-dimensional feature representation. In this process, the interaction between the LeakyReLU activation function and the weight matrix contributes to the nonlinear transformation of features.

[0161] In the classification module (Classifier), the high-dimensional features are mapped to the category space through the fully connected layer (Fully Connected Layer), and the classification expression ability is enhanced by using the Hardswish activation function. Average pooling is applied to the high-dimensional features for dimensionality reduction. Finally, the image classification task is completed through the Softmax function.

[0162] (5) The method according to (4) is characterized in that the inverted residual block (InvertedResidual block) including the depthwise convolution (Depthwise Conv), the depthwise convolution (Conv) and the projection convolution (Projection Conv) performs feature transformation and dimension adjustment on the feature map to obtain a multi-scale feature map, and the process of obtaining the multi-scale feature map by residual connection (ResidualConnection, Res) is as follows:

[0163] x exp =Conv2d(x, W exp )

[0164] x dw =DepthwiseConv2d(x exp , W dw )

[0165] x proj =Conv2d(x dw , W proj )

[0166] y=x+x proj

[0167] Where x is the input tensor, W exp is the weight matrix of the dilated convolution, x exp is the expanded feature map, W dw is the weight matrix of the depthwise convolution, x dw is the feature map after deep convolution, W proj is the weight matrix of the projected convolution, x proj is the feature map after projection, x proj is the feature map after projection, and y is the output tensor after residual connection.

[0168] (6) The method according to (4) is characterized in that atrous convolution with different dilation rates (dilation rate = 1, 6, 12, 18) is used to process the multi-scale feature map, capture multi-scale information, concatenate the feature maps, perform global average pooling on the processed multi-scale feature map to generate global features, and then perform 1×1 convolution (1×1Conv) on the concatenated feature map to integrate information of different scales to obtain the multi-scale feature map, including:

[0169] y atrous =Conv2d(x, W, dilation=d), d=1, 6, 12, 18

[0170] y concat =[y d＝1 ;y d＝6 ;y d＝12 ;y d＝18 ]

[0171] y gap =GlobalAvgPool(x)

[0172] y=Conv2d(y concat , W 1×1 )

[0173] Where x is the feature map obtained by the feature extraction module, W is the weight matrix of the convolution kernel, d is the expansion rate, and y is atrous is the output feature map of the dilated convolution, y concat is the concatenated feature map, y gap is the feature map after global average pooling, W 1×1 is the weight matrix of 1×1 convolution, and y is the feature map after multi-scale feature fusion.

[0174] (7) The method according to (4) is characterized in that matrix multiplication (Matrix Multiplication) is performed on the features to achieve feature linear transformation and obtain preliminary feature representation. Feature concatenation (Feature Concatenation) operation is used to combine features from different sources to enhance information expression capabilities. An attention mechanism (Attention Mechanism) is used to construct combined features for each pair of nodes, and the neighborhood of each node is normalized (Batch Normalization, BN) using the Softmax function to generate attention weights. Subsequently, a weighted sum operation is performed, that is, the normalized (Batch Normalization, BN) attention coefficient is used to aggregate the neighborhood node features to obtain a high-dimensional feature representation. In this process, the interaction between the LeakyReLU activation function and the weight matrix contributes to the nonlinear transformation of the features, and the high-dimensional features obtained include:

[0175] H′=HW

[0176] e ij =LeakyReLU(a T [H′ i ||H′ J ])

[0177]

[0178] Where H is the feature map after multi-scale feature fusion, W is the weight matrix of feature transformation, H′ is the transformed feature matrix, and H′i , H′ j is the transformed feature vector, a is the weight vector of the attention mechanism, and e ij is the attention score between nodes i and j, N(i) is the neighboring node of node i, α ij is the normalized attention coefficient, H′ j is the feature vector of the neighbor node, σ is the activation function, H″ i is the aggregated feature vector.

[0179] (8) A device for classifying zircons, based on the method for classifying zircons according to any one of (1) to (7), comprising:

[0180] An image acquisition unit is configured to acquire and organize cathode luminescence (CL) images of zircons to be classified, wherein the CL images include characteristic information such as zircon growth zones, cracks, and inclusions;

[0181] The zircon classification unit is configured to use a target zircon classification model to identify the zircon to be classified to obtain a classification result of the zircon; wherein the target zircon classification model includes a feature extraction module (Feature Extraction Module, FEM), a multi-scale feature fusion module (Feature Fusion Module, FFM), a graph attention module (Graph Attention Module, GAT) and a classification module (or classifier) ​​(Classifier), the feature extraction module is used to extract the feature information of the zircon, the multi-scale feature fusion module is used to capture and fuse multi-scale feature maps, the graph attention module processes the fused feature map to obtain high-dimensional features, and the classification module processes the high-dimensional features to obtain the zircon classification result.

[0182] (9) According to the device described in (8), the zircon classification unit is further configured to, before using the target zircon classification model to identify the zircon to be classified to obtain the zircon classification result, also include training and obtaining the target zircon classification model, the training and obtaining the target zircon classification model includes:

[0183] Collecting and arranging the CL images of the zircon to be tested, and obtaining the category labels corresponding to the CL images of the zircon to be tested;

[0184] Using the CL images of the zircon to be tested and the category labels to create a training data set, a verification data set and a test data set;

[0185] Construct an initial zircon classification model;

[0186] Based on the training data set and the validation data set, the initial zircon classification model is optimized using a stochastic gradient descent method to obtain an optimized zircon classification model;

[0187] Performing test segmentation on the zircon CL images in the test data set based on the optimized zircon classification model to obtain a predicted category result;

[0188] The predicted category result is compared and analyzed with the category label corresponding to the test data set to obtain the target zircon classification model.

[0189] (10) According to the apparatus described in (8), the zircon classification unit is further configured to obtain the optimized zircon classification model including:

[0190] Training the initial zircon classification model using a stochastic gradient descent optimizer and a cross entropy loss function; and

[0191] A data enhancement strategy was used to optimize the initial zircon classification model.

[0192] (11) According to the device described in (8), the zircon classification unit is further configured to input the zircon CL image into a target zircon classification model according to a preset size;

[0193] In the Feature Extraction Module (FEM), a 3×3 convolution operation is first performed on the input image to capture local fine-grained information in the image, reduce the size of the feature map while extracting important image detail information, normalize the feature map (Batch Normalization, BN), and use the Hardswish (CBH for short: Conv+BatchNorm+Hardswish) activation function to increase the nonlinear expression ability of the model. By using an inverted residual block (InvertedResidual block) containing extended convolution (Depthwise Conv), depth convolution (Conv) and projection convolution (Projection Conv), the feature map is transformed and dimensionally adjusted to effectively extract and fuse multi-scale features. At the same time, the residual connection (Residual Connection, Res) helps to alleviate the gradient vanishing problem, improve the training efficiency of the model, and obtain multi-scale feature maps;

[0194] In the multi-scale feature fusion module (FFM), atrous convolution with different dilation rates (dilation rate = 1, 6, 12, 18) is used to process the multi-scale feature map, capture multi-scale information, and splice the feature map. Global average pooling is performed on the processed multi-scale feature map to generate global features, and then 1×1 convolution (1×1Conv) is performed on the spliced ​​feature map to integrate information of different scales. In this way, information of different scales is effectively fused, and finally a fusion map with global and local features is obtained;

[0195] In the Graph Attention Module (GAT), the features are first matrix multiplied to achieve linear transformation of the features and obtain preliminary feature representation. Then, the feature concatenation operation is used to combine features from different sources to enhance the ability to express information. The attention mechanism is used to construct combined features for each pair of nodes, and the neighborhood of each node is normalized using the Softmax function to generate attention weights. Subsequently, a weighted sum operation is performed, that is, the normalized attention coefficients are used to aggregate the neighborhood node features to obtain a high-dimensional feature representation. In this process, the interaction between the LeakyReLU activation function and the weight matrix contributes to the nonlinear transformation of features;

[0196] In the classification module (Classifier), the high-dimensional features are mapped to the category space through the fully connected layer (Fully Connected Layer), and the classification expression ability is enhanced by using the Hardswish activation function. Average pooling is applied to the high-dimensional features for dimensionality reduction. Finally, the image classification task is completed through the Softmax function.

[0197] (12) According to the device described in (11), the zircon classification unit is further configured as follows:

[0198] The inverted residual block (InvertedResidual block) including the extended convolution (Depthwise Conv), the depth convolution (Conv) and the projection convolution (ProjectionConv) performs feature transformation and dimension adjustment on the feature map to obtain a multi-scale feature map, and the process of obtaining the multi-scale feature map by residual connection (Residual Connection, Res) is as follows:

[0199] x exp =Conv2d(x, W exp )

[0200] x dw =DepthwiseConv2d(x exp , W dw )

[0201] x proj =Conv2d(x dw , W proj )

[0202] y=x+x proj

[0203] Where x is the input tensor, W exp is the weight matrix of the dilated convolution, x exp is the expanded feature map, W dw is the weight matrix of the depthwise convolution, x dw is the feature map after deep convolution, W proj is the weight matrix of the projected convolution, x proj is the feature map after projection, x proj is the feature map after projection, and y is the output tensor after residual connection.

[0204] (13) According to the device described in (11), the zircon classification unit is further configured as follows:

[0205] Atrous Convolution with different dilation rates (dilation rate = 1, 6, 12, 18) is used to process the multi-scale feature map, capture multi-scale information, perform feature map splicing, perform global average pooling on the processed multi-scale feature map, generate global features, and then perform 1×1 convolution (1×1Conv) on the spliced ​​feature map to integrate information of different scales to obtain the multi-scale feature map, including:

[0206] y atrous =Conv2d(x, W, dilation=d), d=1, 6, 12, 18

[0207] y concat =[y d＝1 ;y d＝6 ;y d＝12 ;y d＝18 ]

[0208] y gap =GlobalAvgPool(x)

[0209] y=Conv2d(y concat , W 1×1 )

[0210] Where x is the feature map obtained by the feature extraction module, W is the weight matrix of the convolution kernel, d is the expansion rate, and y is atrous is the output feature map of the dilated convolution, y concat is the concatenated feature map, y gap is the feature map after global average pooling, W 1×1 is the weight matrix of 1×1 convolution, and y is the feature map after multi-scale feature fusion.

[0211] (13) According to the device described in (11), the zircon classification unit is further configured as follows:

[0212] Perform matrix multiplication on the features to achieve linear transformation of the features and obtain preliminary feature representation. Use feature concatenation to combine features from different sources to enhance information expression capabilities. Use the attention mechanism to construct combined features for each pair of nodes, and use the Softmax function to normalize the neighborhood of each node (Batch Normalization, BN) to generate attention weights. Subsequently, a weighted sum operation is performed, that is, the normalized (Batch Normalization, BN) attention coefficient is used to aggregate the neighborhood node features to obtain a high-dimensional feature representation. In this process, the interaction between the LeakyReLU activation function and the weight matrix contributes to the nonlinear transformation of the features, and the high-dimensional features obtained include:

[0213] H′=HW

[0214] e ij =LeakyReLU(a T [H′ i ||H′ J ])

[0215]

[0216] Where H is the feature map after multi-scale feature fusion, W is the weight matrix of feature transformation, H′ is the transformed feature matrix, and H′ i , H′ j is the transformed feature vector, a is the weight vector of the attention mechanism, and e ij is the attention score between nodes i and j, N(i) is the neighboring node of node i, αij is the normalized attention coefficient, H′ j is the feature vector of the neighbor node, σ is the activation function, H″ i is the aggregated feature vector.

[0217] (14) An electronic device, characterized in that the electronic device comprises:

[0218] at least one processor; and,

[0219] a memory communicatively connected to the at least one processor; wherein,

[0220] The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor to enable the at least one processor to execute the method for zircon classification described in any one of (1) to (13).

[0221] (15) A non-transitory computer-readable storage medium, characterized in that the non-transitory computer-readable storage medium stores computer instructions, and the computer instructions are used to enable the computer to execute the method for zircon classification described in any one of (1) to (13).

[0222] (16) A computer program product, comprising a computer program stored on a non-transitory computer-readable storage medium, the computer program comprising program instructions which, when executed by a computer, cause the computer to execute the method for zircon classification described in any one of (1) to (13).

Claims

1. A method for classifying zircon, characterized in that: include: Collecting and arranging cathode luminescence (CL) images of zircons to be classified, wherein characteristic information of zircons in the CL images at least includes zircon growth zones, cracks and inclusions; The target zircon classification model is used to identify the zircon to be classified, so as to obtain the zircon classification result; The target zircon classification model includes a feature extraction module, a multi-scale feature fusion module, a graph attention module and a classification module. The feature extraction module is used to extract feature information of zircon, the multi-scale feature fusion module is used to capture and fuse multi-scale feature maps, the graph attention module processes the fused feature maps to obtain high-dimensional features, and the classification module processes the high-dimensional features to obtain zircon classification results.

2. The method according to claim 1, characterized in that Before using the target zircon classification model to identify the zircon to be classified to obtain the classification result of the zircon, it also includes training and obtaining the target zircon classification model, and the training and obtaining the target zircon classification model includes: Collecting and arranging the CL images of the zircon to be tested, and obtaining the category labels corresponding to the CL images of the zircon to be tested; Using the CL images of the zircon to be tested and the category labels to create a training data set, a verification data set and a test data set; Construct an initial zircon classification model; Based on the training data set and the validation data set, the initial zircon classification model is optimized using a stochastic gradient descent method to obtain an optimized zircon classification model; Performing test segmentation on the zircon CL images in the test data set based on the optimized zircon classification model to obtain a predicted category result; The predicted category result is compared and analyzed with the category label corresponding to the test data set to obtain the target zircon classification model.

3. The method according to claim 2, characterized in that Obtaining the optimized zircon classification model includes: Training the initial zircon classification model using a stochastic gradient descent optimizer and a cross entropy loss function; and A data enhancement strategy was used to optimize the initial zircon classification model.

4. The method according to claim 1, characterized in that: Inputting the CL image into a target zircon classification model according to a preset size; In the feature extraction module, a 3×3 convolution operation is performed to capture local fine-grained information in the input image to obtain a feature map, the feature map is normalized, and an activation function and an inverted residual block including an extended convolution, a depth convolution, and a projection convolution are used to perform feature transformation and dimension adjustment on the feature map to obtain a multi-scale feature map; In the multi-scale feature fusion module, the multi-scale feature map is processed using dilated convolutions with different dilation rates to capture multi-scale information, the feature maps are spliced, each feature map is globally averaged pooled to generate global features, the spliced ​​feature maps are 1×1 convolved to integrate information of different scales; In the graph attention module, a linear transformation is performed through a weight matrix, and a new feature representation is obtained based on the integration of information at different scales. The attention mechanism is used to construct a combined feature for each pair of nodes. The neighborhood of each node is normalized using the Softmax function, and then feature aggregation is performed. The features of neighboring nodes are aggregated using the normalized attention coefficient to obtain high-dimensional features. In the classification module, high-dimensional features are mapped to the category space through fully connected layers, activation functions and average pooling operations.

5. The method according to claim 4, characterized in that The inverted residual block including the dilated convolution, the depth convolution and the projection convolution performs feature transformation and dimension adjustment on the feature map to obtain a multi-scale feature map in the following process: x exp =Conv2d(x,W exp ) x dw =DepthwiseConv2d(x exp ,W dw ) x proj =Conv2d(x dw ,W proj ) y=x+x proj Where x is the input tensor, W exp is the weight matrix of the dilated convolution, x exp is the expanded feature map, W dw is the weight matrix of the depthwise convolution, x dw is the feature map after deep convolution, W proj is the weight matrix of the projected convolution, x proj is the feature map after projection, x proj is the feature map after projection, and y is the output tensor after residual connection.

6. The method according to claim 4, characterized in that Obtaining the multi-scale feature map includes: y atrous =Conv2d(x,W,dilation=d),d=1,6,12,18 and concat =[and d＝1 ;and d＝6 ;and d＝12 ;and d＝18 ] y gap =GlobalAvgPool(x) and=Conv2d ( and concat ,W 1×1 ) In the formula, x is the feature map obtained by the feature extraction module, W is the weight matrix of the convolution kernel, d is the expansion rate, and y is atrous is the output feature map of the dilated convolution, y concat is the concatenated feature map, y gap is the feature map after global average pooling, W 1×1 is the weight matrix of 1×1 convolution, and y is the feature map after multi-scale feature fusion.

7. The method according to claim 4, characterized in that Obtaining the high-dimensional features includes: H'=HW have been ij =LeakyReLU(a T [H' i ||H' J ] Where H is the feature map after multi-scale feature fusion, W is the weight matrix of feature transformation, H' is the transformed feature matrix, and H' i , H' j is the transformed feature vector, a is the weight vector of the attention mechanism, and e ij is the attention score between nodes i and j, N(i) is the neighboring node of node i, α ij is the normalized attention coefficient, H' j is the feature vector of the neighbor node, σ is the activation function, H” i is the aggregated feature vector.

8. A device for classifying zircon, characterized in that: Based on the method for zircon classification according to any one of claims 1 to 7, the device comprises: An image acquisition unit is configured to acquire and organize zircon cathode luminescence (CL) images to be classified, wherein the zircon characteristic information in the CL images at least includes zircon growth zones, cracks and inclusions; A zircon classification unit is configured to use a target zircon classification model to identify zircons to be classified to obtain a classification result of the zircon; wherein the target zircon classification model includes a feature extraction module, a multi-scale feature fusion module, a graph attention module and a classification module, the feature extraction module is used to extract feature information of zircon, the multi-scale feature fusion module is used to capture and fuse multi-scale feature maps, the graph attention module processes the fused feature maps to obtain high-dimensional features, and the classification module processes the high-dimensional features to obtain a zircon classification result.

9. An electronic device, characterized in that: The electronic device includes: at least one processor; and, a memory communicatively connected to the at least one processor; wherein, The memory stores instructions that can be executed by the at least one processor, and the instructions are executed by the at least one processor so that the at least one processor can execute the method for zircon classification as described in any one of claims 1 to 7.

10. A non-transitory computer-readable storage medium, characterized in that: The non-transitory computer-readable storage medium stores computer instructions, which are used to cause the computer to execute the method for zircon classification as described in any one of claims 1 to 7.

Images