Zircon classification method based on multi-scale deep convolutional neural network

Abstract

The application discloses a zircon classification method based on a multi-scale deep convolutional neural network, which comprises the following steps: (1) pre-processing and data enhancement are performed on a zircon CL image; (2) a classification feature parameter of the zircon CL image is obtained by using transfer learning; (3) a bimodal neural network is constructed, wherein the bimodal neural network comprises a double-feature extraction module, a channel attention module, a fusion information module and an output judgment module; (4) an input zircon CL image is recognized, pre-training weights are assigned to the bimodal neural network, then feature extraction is performed on the zircon CL image, finally, according to the specific features of the zircon CL image, reverse propagation is performed to obtain appropriate weights; and (5) the bimodal neural network is used to classify the CL image of the zircon. The bimodal attention deep convolutional neural network is adopted, so that the zircon classification precision and prediction stability of the CL image are improved, and in particular, the classification precision of ordinary magmatic zircon and metamorphic zircon is significantly improved.

Description

Technical Field

[0001] This invention relates to the field of geology, and in particular to a zircon classification method based on a multi-scale deep convolutional neural network. Background Technology

[0002] Zircon is an excellent geochronometer because it can incorporate uranium while excluding lead during crystallization, and it exhibits high resistance to chemical and physical weathering, maintaining the stability of these isotopic ratios. Therefore, zircon is of great significance in determining geological age, the formation of igneous and metamorphic rocks, and the source regions and maximum depositional ages of sedimentary rocks.

[0003] Igneous, metamorphic, and hydrothermal zircons are common zircon types, and classifying these types is a prerequisite for further geochronological and geochemical analysis. This is because the ages obtained from zircon grains in igneous and metamorphic rocks reflect the timing of magma crystallization and metamorphism, respectively. Accurate classification of zircon types is necessary to obtain the precise age of specific geological activities. Furthermore, distinguishing hydrothermal zircon grains from those in igneous and metamorphic rocks is also important. First, the relative abundance of hydrothermal zircons reflects the activity of hydrothermal events reported in relation to different types of economically valuable deposits. Second, removing hydrothermal zircons is sometimes important for many studies because hydrothermal zircon grains may contain non-radiative Pb and have undergone Pb loss, both of which reduce the usefulness of these zircon grains.

[0004] Current methods for zircon classification include: 1. Sorting zircons under a microscope and classifying them according to the Th / U ratio. However, during manual sorting, zircon geochronology may increase bias. 2. Cathodoluminescence (CL) or backscattered electron (BSE) imaging methods, which can fully study the internal structure of zircons, thus classifying zircon types before geological dating and isotopic analysis. Corfu et al. (2003) found that CL is more useful than BSE in distinguishing zircon grains because CL images can better detect the internal structure, fractures, and other mineral inclusions of zircon grains. Therefore, CL imaging is a more widely used imaging technique for zircon classification and selection analysis.

[0005] Traditional methods of classifying zircons in logarithmic images (CL images) are manual processes based on prior knowledge of zircon grain morphology and the geological background of the rock sample. This procedure is labor-intensive, requires specialized knowledge, and its accuracy depends on the interpreter's expertise and caution. Manual target selection is common in geoscience missions, and these manual procedures inevitably introduce biases due to sample complexity and limitations of individual knowledge. Zircon classification in CL images is no exception. Therefore, an automated, unbiased method for classifying zircons in CL images is necessary, and deep learning technology is an optimal solution.

[0006] Based on the above, the inventors proposed a deep learning-based zircon classification method from cathodoluminescence images (Dongyu Zheng et al., *Geoscience Frontiers*, Vol. 13, No. 6, 2022, hereinafter referred to as "Prior Art") in 2022. This method uses deep convolutional neural networks (DCNNs) and transfer learning to classify common zircon types (i.e., magmatic zircon, metamorphic zircon, and hydrothermal zircon). According to the "Conclusion" section of the Prior Art, the DCNN model can classify most hydrothermal zircons, but distinguishing between magmatic and metamorphic zircons remains challenging. Specifically, DCNNs have difficulty differentiating between magmatic and metamorphic zircons with similar characteristics, such as metamorphic zircons with inherited magmatic zircon cores and ordinary magmatic zircons.

[0007] Through in-depth analysis, the inventors discovered that this is because DCNN is limited by the size of the convolution kernel, and can only focus on small areas of the image, making it unable to accurately extract the feature structures of different zircon categories. For example, DCNN only focuses on local features of the zircon CL image, which can easily lead to misclassification by the model; for another example, for metamorphic zircons with inherited magmatic zircon cores, the model often only extracts information about the magmatic zircon core, ignoring the entire metamorphic zircon grain, thus classifying these metamorphic zircons as ordinary magmatic zircons.

[0008] Therefore, based on the above factors, the prediction accuracy and stability of existing technologies for zircon classification have a certain degree of uncertainty, and further improvements are needed. Summary of the Invention

[0009] To address the shortcomings of the existing technologies, this invention provides a zircon classification method based on a multi-scale deep convolutional neural network, which further improves the accuracy and stability of zircon classification prediction by employing a new training network.

[0010] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0011] The zircon classification method based on multi-scale deep convolutional neural networks includes the following steps:

[0012] (1) Preprocessing and data augmentation of known zircon CL images;

[0013] (2) Use transfer learning to obtain the classification feature parameters of zircon CL images and obtain the pre-training weights of the neural network.

[0014] (3) Construct a bimodal neural network, which includes a dual feature extraction module, a channel attention module, a fusion information module, and an output decision module. The dual feature extraction module consists of two parallel convolutional neural networks (CNNs), each of which has two residual convolutional blocks and a global pooling layer. The channel attention module includes a compression block and an activation block. The output decision module consists of a global pooling layer, a fully connected layer, and a softmax layer.

[0015] The dual feature extraction module is used to receive the original zircon CL image and a cropped image of significant structural features within the original zircon image, and use them as input;

[0016] The compression block in the channel attention module is used to convert the input N×N×C matrix into a 1×1×C matrix and output a vector containing the input feature map information to the activation block in the channel attention module, where N and C are both non-zero natural numbers.

[0017] The excitation block is used to take the output of the compression block as input, then generate a set of channel weights for each feature, and then output information of the input feature map with weighted feature importance;

[0018] The fusion information module performs a Hadamard product based on the output of the excitation block, so that the neural network contains global and local information from the input zircon CL image, and then outputs it to the output judgment module.

[0019] The output judgment module is used to calculate the predicted probability of each category and output the category with the highest probability as the final prediction;

[0020] (4) Identify the input zircon CL image and assign the pre-trained weights to the bimodal neural network as the weights of the bimodal neural network. Then, extract features from the input zircon CL image. Finally, backpropagate based on the specific features of the zircon CL image to obtain the most suitable weights and use them as the basis for classification.

[0021] (5) Repeat step (4) to complete the training of the bimodal neural network;

[0022] (6) Classify the CL images of zircon using a trained bimodal neural network.

[0023] Step (2) specifically involves using transfer learning to obtain the pre-trained weights of the neural network from the ImageNet-1K image dataset.

[0024] Furthermore, in step (3), the dual feature extraction module is activated by the linear rectifier unit.

[0025] Furthermore, in step (4), the process of generating a set of channel weights for each feature by the activation block is as follows: the 1×1×C matrix is ​​processed through two fully connected layers, and the Sigmoid activation function is applied to restrict the output value to between 0 and 1. Then, these values ​​are multiplied by each channel of the original N×N×C matrix to obtain a set of channel weights.

[0026] Furthermore, in step (3), the process of the fusion information module performing the Hadamard product is as follows: receiving two matrices of the same size and generating a new matrix, wherein each element is the product of the corresponding elements of the original matrix.

[0027] Specifically, the process of obtaining the most suitable weight in step (4) is as follows:

[0028] 1) Set the initial learning rate to 0.0001 and the decay learning rate to a decay factor of 0.5 every 20 periods;

[0029] 2) Select class cross-entropy loss and stochastic gradient descent as the loss function and optimization algorithm, respectively;

[0030] 3) Use the pre-trained weights obtained from transfer learning as the initial weights for training the bimodal neural network. Then input the training dataset and backpropagate according to the actual features of the training dataset to change the initial weights of the bimodal neural network. As data is continuously input, the weights are adjusted accordingly until the most suitable weights are obtained, which are then used as the basis for classification.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] This invention constructs a novel neural network—a bimodal attention-based deep convolutional neural network. This bimodal neural network integrates an attention mechanism and bimodal (i.e., global and local feature) input, enabling it to acquire global and feature structure information of zircons. Then, it employs global pooling to output a vector containing information from the input feature map. Next, it generates a set of channel weights for each feature and outputs information from the input feature map containing weighted feature importance. Finally, it performs a Hadamard product (also known as inter-element product), thus enabling the neural network to incorporate both global and local information from the input zircon CL image. After training with transfer learning methods, the resulting bimodal neural network can better classify zircon CL images. Experiments show that the bimodal neural network designed in this invention, by expanding its focus beyond small regions of the image to include both global and local information, further improves the overall prediction accuracy compared to the DCNN model, especially significantly improving the accuracy for common magmatic zircons and metamorphic zircons with similar features. Attached Figure Description

[0033] Figure 1 This is a flowchart illustrating an embodiment of the present invention.

[0034] Figure 2 This is a logical schematic diagram of a dual-modal neural network in an embodiment of the present invention. Detailed Implementation

[0035] The present invention will be further described below with reference to the accompanying drawings and embodiments. The embodiments of the present invention include, but are not limited to, the following embodiments.

[0036] Example

[0037] This embodiment provides a zircon classification method based on a multi-scale deep convolutional neural network, which is a further improvement on the inventor's original scheme. The aim is to enhance the accuracy and stability of zircon classification, and to improve the accuracy of distinguishing between magmatic and metamorphic zircons with similar characteristics. Figure 1 As shown, the main process of this embodiment is as follows:

[0038] I. Preprocessing and data augmentation of known zircon CL images

[0039] After collecting rock samples from the field, thin sections were ground to extract the zircon grains to be studied, which were then observed under a scanning electron microscope. Cathodoluminescence (CL) is light or electromagnetic radiation in the ultraviolet (UV) to near-infrared (NIR) range of electromagnetic waves, produced by the rapid electrons (cathode rays) of an electron beam; it reveals the internal structure of zircon grains, cracks, and other mineral inclusions.

[0040] Data preprocessing and augmentation: Since the original zircon CL images have variable widths and heights, the original images needed to be resized to have constant widths and heights. Furthermore, data augmentation was performed to expand the original dataset and avoid overfitting, including horizontal and vertical flipping, random cropping, and rotation.

[0041] II. Using transfer learning to obtain classification feature parameters from zircon CL images, and obtaining pre-trained weights for the neural network.

[0042] ImageNet-1K is one of the benchmarks for computer vision, consisting of 10 million images across 1000 categories. It can be used to extract rich features (e.g., shape, brightness, size, etc.) useful for new classification tasks. This embodiment obtains the pre-trained weights of the neural network from the ImageNet-1K image dataset using transfer learning, thereby extracting classification feature information and parameters of identification significance from zircon CL images.

[0043] III. Constructing a Dual-Mode Neural Network

[0044] The main innovation of this embodiment lies in the use of a novel bimodal attention-based deep convolutional neural network for training, thereby improving the prediction accuracy and stability of zircon classification, and enhancing the accuracy in distinguishing between magmatic and metamorphic zircons that are difficult to differentiate due to similar characteristics. The neural network used in this embodiment is a bimodal neural network, integrating an attention mechanism and bimodal (global and local feature) inputs, such as... Figure 2 As shown, it specifically includes four modules: dual feature extraction module, channel attention module, information fusion module, and output judgment module.

[0045] The following section describes the functions and operations of each module.

[0046] 1. Dual Feature Extraction Module: The dual feature extraction module consists of two parallel convolutional neural networks (CNNs), each with two residual convolutional blocks and a global pooling layer activated by a linear rectified unit (ReLU). To capture both global and local features in rock classification, the dual feature extraction module receives a cropped image of significant skeletal grains from both the rock optical micrograph and the original optical micrograph as input. For example, for a feldspathic sandstone image, one input would be the original thin section image, while the other input would be representative spinel-subcrystal feldspar grains with twinning.

[0047] 2. Channel Attention Module: The channel attention module includes a compression block and an activation block. In this embodiment, the compression operation is a global pooling operation, which converts the input N×N×C matrix into a 1×1×C matrix and outputs a vector containing the input feature map information to the activation block in the channel attention module, as shown in formula (1):

[0048]

[0049] Among them, Z c F is the matrix after 1×1×C compression; sq Represents compression operation; 'c' represents the c-th channel; U c It is N×N×C, which is the original input matrix.

[0050] The output of the compression operation (Z) c ) is a vector containing information about the input feature map.

[0051] The activation block is used to take the output of the compression block as input and then generate a set of channel weights for each feature. Specifically, the 1×1×C matrix is ​​processed through two fully connected layers, and the Sigmoid activation function is applied to restrict the output values ​​to between 0 and 1. These values ​​are then multiplied by each channel of the original N×N×C matrix to obtain a set of channel weights, as shown in formulas (2) and (3).

[0052] α=F ex (Z,W)=σ(g(Z,W))=σ(W2δ(W1Z)), (2)

[0053] X c =F scale (U c ,α c )=α c U c (3)

[0054] In the formula, α represents the vector of weights for each channel; F ex δ represents the activation operation; Z is the information matrix after compression in the previous step; g represents the dimensionality descent operation; W, W1, and W2 represent the learnable weight matrices; δ represents the ReLU activation function; σ represents the sigmoid activation function; X c The output of the channel attention module; α c F represents the weight of the C-th channel; scale This represents weighting operations.

[0055] Finally, the output is information (α) of the input feature map with weighted feature importance.

[0056] 3. Feature Fusion Module. The information fusion module performs a Hadamard product based on the output of the activation block, enabling the neural network to incorporate global and local information from the input rock image, which is then output by the output judgment module. The process of performing the Hadamard product is as follows: two matrices of the same size are received and a new matrix is ​​generated, where each element is the product of the corresponding elements of the original matrix. For example, given matrices A and B of dimension m×n from the attention block of the previous channel, the fusion feature extraction is the Hadamard product of A and B, as shown in formula (4):

[0057] (A⊙B) mn =(A) mn *(B) mn (4)

[0058] Here, ⊙ and * represent the Hadamard product operation, that is, the dot product operation.

[0059] 4. Output Decision Module: The output decision module consists of a global pooling layer, a fully connected layer, and a softmax layer. It is used to calculate the predicted probability of each category and output the category with the highest probability as the final prediction.

[0060] IV. Training a bimodal attention-based deep convolutional neural network and classifying CL images of zircon.

[0061] First, the input zircon CL image is identified, and pre-trained weights are assigned to the bimodal neural network as its weights. Then, features are extracted from the input zircon CL image. Finally, backpropagation is performed based on the specific features of the zircon CL image to obtain the most suitable weights (i.e., achieving the best match between the input and output classes). The specific training process is as follows:

[0062] 1) Set the initial learning rate to 0.0001 and the decay learning rate to a decay factor of 0.5 every 20 periods;

[0063] 2) Select class cross-entropy loss and stochastic gradient descent as the loss function and optimization algorithm, respectively;

[0064] 3) Use the pre-trained weights obtained from transfer learning as the initial weights for training the bimodal neural network. Then input the training dataset and backpropagate according to the actual features of the training dataset to change the initial weights of the bimodal neural network. As data is continuously input, the weights are adjusted accordingly until the most suitable weights are obtained, which are then used as the basis for classification.

[0065] The zircon CL images are repeatedly input for recognition to complete the training of the bimodal neural network. Then, the trained bimodal neural network is used to classify the zircon CL images. That is, the zircon CL images are automatically classified based on the identification significance of the feature information (weights) obtained from the classification system through transfer learning.

[0066] V. Selection of the Optimal Bimodal Neural Network

[0067] The constructed bimodal neural network was evaluated using conventional evaluation metrics for image classification tasks, including accuracy, recall, and F1-score, and the optimal bimodal neural network was ultimately selected.

[0068] The following metrics were used to evaluate the bimodal neural network:

[0069] Evaluation metrics accuracy = (TP + TN) / (TP + FN + FP + TN)

[0070] Precision = TP / (TP + FP)

[0071] Recall = TP / (TP + FN)

[0072] F1 Score = 2 * Precision * Recall / (Precision + Recall)

[0073] in:

[0074] TP—True positive, both label and predictive are positive;

[0075] TN—True negative, both label and predictive are negative;

[0076] FP—a false positive, labeled as negative but predicted as positive;

[0077] FN—False positive, labeled as positive, predicted as negative.

[0078] The evaluation results compared with existing technologies are shown in Table 1-3 (the evaluation results of existing technologies are based on data recorded in the inventor's articles):

[0079] Table 1

[0080]

[0081] Table 2

[0082]

[0083] Table 3

[0084]

[0085] In summary, by employing a novel neural network and combining it with transfer learning methods, this invention effectively improves the prediction accuracy and stability of zircon classification, particularly enhancing the accuracy and stability of distinguishing between common magmatic zircons and metamorphic zircons with similar characteristics. Therefore, compared with existing technologies, this invention possesses outstanding substantive features and significant advancements.

[0086] The above embodiments are merely one of the preferred embodiments of the present invention and should not be used to limit the scope of protection of the present invention. Any modifications or refinements made to the main design concept and spirit of the present invention that are not of substantial significance, but solve the same technical problem as the present invention, should be included within the scope of protection of the present invention.

Claims

1. A zircon classification method based on multi-scale deep convolutional neural networks, characterized in that, Includes the following steps: (1) Preprocessing and data augmentation of known zircon CL images; (2) Use transfer learning to obtain the classification feature parameters of zircon CL images and obtain the pre-training weights of the neural network; (3) Construct a bimodal neural network, which includes a dual feature extraction module, a channel attention module, a fusion information module and an output judgment module. The dual feature extraction module consists of two parallel convolutional neural networks, each of which has two residual convolutional blocks and a global pooling layer. The channel attention module includes a compression block and an activation block. The output judgment module consists of a global pooling layer, a fully connected layer and a softmax layer. The dual feature extraction module is used to receive the zircon CL image processed in step (1) and the region image containing significant structural features cropped from the zircon CL image, and use it as input; The compression block in the channel attention module is used to convert the input N×N×C matrix into a 1×1×C matrix and output a vector containing the input feature map information to the activation block in the channel attention module, where N and C are both non-zero natural numbers. The excitation block is used to take the output of the compression block as input, then generate a set of channel weights for each feature, and then output information of the input feature map with weighted feature importance; The fusion information module performs a Hadamard product based on the output of the excitation block, so that the neural network contains global and local information from the input zircon CL image, and then outputs it to the output judgment module. The output judgment module is used to calculate the predicted probability of each category and output the category with the highest probability as the final prediction; (4) Read the input zircon CL image and assign the pre-trained weights to the bimodal neural network as the weights of the bimodal neural network. Then, extract features from the input zircon CL image. Finally, backpropagate based on the specific features of the zircon CL image to obtain the most suitable weights and use them as the basis for classification. (5) Repeat step (4) to complete the training of the bimodal neural network; (6) Classify the CL images of zircon using a trained bimodal neural network.

2. The zircon classification method based on a multi-scale deep convolutional neural network according to claim 1, characterized in that, Step (2) specifically involves using transfer learning to obtain the pre-trained weights of the neural network from the ImageNet-1K image dataset.

3. The zircon classification method based on a multi-scale deep convolutional neural network according to claim 1 or 2, characterized in that, In step (3), the dual feature extraction module is activated by the linear rectifier unit.

4. The zircon classification method based on a multi-scale deep convolutional neural network according to claim 3, characterized in that, In step (3), the process of generating a set of channel weights for each feature by the activation block is as follows: the 1×1×C matrix is ​​processed through two fully connected layers, and the Sigmoid activation function is applied to restrict the output value to between 0 and 1. Then, these values ​​are multiplied by each channel of the original N×N×C matrix to obtain the feature map after channel weighting.

5. The zircon classification method based on a multi-scale deep convolutional neural network according to claim 4, characterized in that, In step (3), the process of the fusion information module performing the Hadamard product is as follows: receiving two matrices of the same size and generating a new matrix, where each element is the product of the corresponding elements of the original matrix.

6. The zircon classification method based on a multi-scale deep convolutional neural network according to claim 5, characterized in that, The specific process for obtaining the most suitable weight in step (4) is as follows: 1) Set the initial learning rate to 0.0001 and the decay learning rate to a decay factor of 0.5 every 20 periods; 2) Select class cross-entropy loss and stochastic gradient descent as the loss function and optimization algorithm, respectively; 3) Use the pre-trained weights obtained from transfer learning as the initial weights for training the bimodal neural network. Then input the training dataset and backpropagate according to the actual features of the training dataset to change the initial weights of the bimodal neural network. As data is continuously input, the weights are adjusted accordingly until the most suitable weights are obtained, which are then used as the basis for classification.

Images