VGG-based mineral resource image classification method

Abstract

The invention discloses a mineral resource image classification method based on VGG, and the method comprises the steps: an image preprocessing stage: segmenting a remote sensing image into small blocks according to the coordinates of a mineral point and a classification result, making labels, and carrying out the upsampling, sharpening, filtering and other operations of the image; and a model construction stage: designing a VGG-based classification model, then dividing the sorted images into a training set and a test set, training the model by using the training set, and evaluating the model effect by using the test set. Then model parameters are adjusted, and an optimal parameter combination is found; and a mineral resource prediction stage: carrying out containerization deployment on the trained model, inputting a predicted picture, and obtaining an output result. According to the method, the VGG algorithm is used, a good effect is achieved in mineral resource image classification under real geological data, and the method has wide application value and use prospects in the geological mineral field.

Description

Technical Field

[0001] The present invention relates to a VGG-based mineral resource image classification method, and belongs to the fields of geographical science, deep learning, and image processing. Background Art

[0002] In recent years, the widespread use of GIS technology has driven the development of geological resource image classification. Currently, GIS-based MPM models can generally be divided into knowledge-driven and data-driven models. However, these methods require certain assumptions, including specific data distributions and / or conditional independence between input features, to generate unbiased models. These assumptions are rarely met in mineral imagery. Therefore, to address these issues, the VGG algorithm from deep learning is considered. Deep learning methods have the advantage of not requiring assumptions about data distribution and being able to handle nonlinear relationships between mineral deposits and evidence features. Furthermore, they are powerful pattern recognition engines capable of processing a large number of spatial features to obtain reliable results.

[0003] Many researchers have proposed methods for classifying geological resource images in different scenarios, but these methods all have certain limitations. Pan et al. [1] used a convolutional neural network model based on AlexNet to process fusion data of remote sensing images and geochemical survey data. Tao et al. [2] used a classification model combining the VanillaNet network and GhostNet to classify rock images. Chen et al. [3] constructed a two-layer convolutional neural network to identify nine lithologies in the ocean basin study area. However, the experimental data of the above research results are relatively ideal and cannot accurately classify more complex actual images.

[0004] The VGG model is an architecture based on deep convolutional neural networks that efficiently processes and understands image data. Through a multi-layered feature extraction process, the model gradually transforms input images into high-level semantic features, thereby supporting subsequent classification or regression tasks. The core of the VGG model lies in its layered feature extraction mechanism. The output features of each layer completely determine the input features of the next layer. This layer-by-layer transfer allows the model to gradually abstract the semantic information of the image.

[0005] [1]T.Pan, R.Zuo and Z.Wang, "Geological Mapping via ConvolutionalNeural Network Based on Remote Sensing and Geochemical Survey Data inVegetation Coverage Areas," in IEEE Journal of Selected Topics in AppliedEarth Observations and Remote Sensing, vol.16, pp.3485-3494, 2023.

[0006] [2] L.Tao,

[0007] [3] L.Tao, Summary of the Invention

[0008] Purpose of the invention: In order to overcome the deficiencies in the prior art, the present invention provides a mineral image classification method using the VGG algorithm, targeting geological data and remote sensing images, taking into account the complex relationship between geological data in real scenarios.

[0009] Technical solution: To achieve the above purpose, the technical solution adopted by the present invention is:

[0010] A VGG-based mineral resource image classification method includes the following stages:

[0011] A. Image preprocessing: Using geological data and remote sensing images, principal component analysis is used to fuse remote sensing image features. The images are then divided into small blocks based on the mine coordinates and classification results and labeled. The images are then upsampled, sharpened, and filtered.

[0012] B. Model building phase: Design a VGG-based classification model, then divide the organized image set into a training set and a test set. Use the training set to train the model, and use the test set to evaluate the model's performance. Then adjust the model parameters to find the optimal parameter combination.

[0013] C. Mineral resource prediction stage: The trained model is deployed in a container and exposed to the outside world. The predicted image is input through the interface to obtain the output result.

[0014] The specific steps of the image preprocessing stage are as follows:

[0015] A1. Assume that the matrix of the K remote sensing source images to be fused, which are derived from radiometric values, is {X1, X2,…, X k}, the size of each image is H × L. First, each image is flattened into a column vector Construct joint data matrix D = [x1-μx2-μ…x k -μ], where is the global mean vector;

[0016] A2. Calculate the covariance matrix of the joint data matrix D

[0017] A3. Solve the characteristic equation Cv i =λ i v i (i=1,2,…,K)where λ1≥λ2≥…≥λ K is the eigenvalue, v i is the corresponding orthogonal eigenvector;

[0018] A4. Each image is passed through p i =Dv i (i=1,2,…,K) is projected into the principal component space, and the fusion strategy adopted is: for the first principal component p1, it reflects the commonality of multiple images and is directly retained; for j≥2 principal components, the absolute value maximization criterion is used to select the detail component, In the formula Represents the projection coefficient of the k-th image on the j-th principal component;

[0019] A5. The fused principal component Reconstructed by inverse PCA transform, Finally, y is reshaped into an H×L fused image matrix Y;

[0020] A6. Split the fused feature image into p×p pixel blocks based on the coordinates of the mineral point. The center of the image is the coordinate of the mineral point. Each mineral point is split into a p×p remote sensing image and labeled with the mineral type.

[0021] A7. Upsample the image using bilinear interpolation from p×p size to P×P size.

[0022] A8. Sharpen and filter the image. Use the mild SHARPEN sharpener, Gaussian filter, and a filter parameter of 5.

[0023] The model building phase includes the following steps:

[0024] B1. Assume that the input image size is (H×L×C). Where H and L are the height and width of the image, C is the number of channels, and there are Y categories in total. Divide an image into several image blocks of the same size, P×P. Let the input image be I. In the convolution layer, the first stage passes through two convolution layers and one pooling layer in sequence, and the output is O1=MaxPool(ReLU(W2(ReLU(W1I+b1))+b2)). Similarly, the output of the second stage is O2=MaxPool(ReLU(W4(ReLU(W3O1+b3))+b4)). The output of the third stage is O3=MaxPool(ReLU((W7(ReLU(W6(ReLU(W5O2+b5))+b6))+b7))). The output of the fourth stage is O4=MaxPool(ReLU((W 10 (ReLU(W9(ReLU(W8O3+b8))+b9))+b 10 ))), the output of the fifth stage is O5=MaxPool(ReLU((W 13 (ReLU(W 12 (ReLU(W 11 O4+b 11 ))+b 12 ))+b 13 ))), and finally, a global average pooling O6 = AdaptiveAvgPool2d(O5) is performed, which is the output of the convolutional layer. Here, MaxPool() is the pooler, the pooling method is maximum pooling, the pooling kernel size is 2×2, the stride is 2, and the activation function is ReLU(). All convolution kernels are 3×3, the stride is 1, and the padding is 1. W and b are the convolution kernel and bias of the convolutional layer, respectively.

[0025] B2. The fully connected layer consists of three intermediate layers, with the number of layers being 4096, 4096, and the number of categories. The output of the first layer is O7=ReLU(W 14 Flatten(O6)+b 14 ), the second layer output is O8=ReLU(W 15 Flatten(O7)+b 15 ), the third layer output is O9=ReLU(W 16 Flatteen(O8)+b 16 ). ReLU() is the activation function, W and b are the weight matrix and bias of the fully connected layer respectively. The final output is a Y-dimensional vector.

[0026] B3. The input sequence after processing by the softmax function is Y is the dimension of the output layer, o i represents the value of the i-th dimension of the output layer, represents the probability of category i;

[0027] B4. Integrate the above components into a VGG model, divide the entire image into several blocks, perform several rounds of convolution, pooling, and fully connected layer processing, and finally use the softmax function for classification;

[0028] B5. Split the training set and test set into an 80%-20% ratio. The ratio of test set to training set for each category should be evenly distributed.

[0029] B6. Train the model for 100 rounds, using classification accuracy as the metric. The highest classification accuracy among the 100 rounds of training is selected to evaluate the training effect. The initial learning rate is 0.001, and the batch size is 4.

[0030] B7. Use grid search to adjust the learning rate and batch size. Test the learning rate from 0.001 to 0.1 with a step size of 0.001. Test the batch size from 4 to 128 in powers of 2, for a total of 60 combinations. Select each parameter combination and proceed to step B6.

[0031] B8. Organize the experimental results and find the optimal model;

[0032] The mineral deposit prediction stage includes the following steps:

[0033] C1. Deploy the trained model in Docker format, using python:3.8-slim as the base image, install relevant third-party libraries on top of it, and use the Flask framework to expose the interface.

[0034] C2. Call the interface via HTTP, input the remote sensing image to be predicted, and obtain the prediction result.

[0035] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor. When the processor executes the program, the VGG-based mineral resource image classification method is implemented.

[0036] A computer-readable storage medium stores computer instructions, which, when executed by a processor, implement the VGG-based mineral resource image classification method.

[0037] The VGG-based mineral resource image classification method provided by the present invention has the following beneficial effects compared with the existing technology:

[0038] (1) This invention is specifically designed for the characteristics of geological and mineral remote sensing images. In the preprocessing stage, the quality and consistency of the image blocks are effectively improved through a combination of precise mineral point coordinate segmentation, bilinear interpolation upsampling, mild SHARPEN sharpening, and Gaussian filtering (parameter 5). Upsampling ensures that the input image meets the resolution required by the model; sharpening enhances geological edge and texture information, which is beneficial for the model to recognize subtle features; Gaussian filtering effectively suppresses the common noise in remote sensing images, significantly reducing the negative impact of noise on subsequent classification accuracy, and providing more "clean" and more prominent input data for model training.

[0039] (2) A VGG-based model architecture is used to effectively extract deep and abstract features of remote sensing images by stacking multiple layers of small convolutional kernels (3x3) (5 convolutional blocks). Max pooling (2x2, stride 2) gradually reduces the spatial dimension while retaining significant features, improving the model's translation invariance and computational efficiency. The application of global average pooling (GAP) replaces the traditional fully connected layer (in the spatial dimension), effectively reducing the number of model parameters, reducing the risk of overfitting, and directly mapping the high-level feature maps extracted by the convolutional layer to the category space.

[0040] (3) During the model construction phase, the training and test sets were evenly distributed (80%-20%, uniformly distributed across categories) to ensure objectivity in model evaluation. A grid search method was used to systematically and comprehensively tune key hyperparameters (learning rate, batch size) (learning rate 0.001-0.1 with a step size of 0.001, batch size 4-128 in powers of 2), significantly improving the model's generalization ability and predictive stability on unknown data, effectively avoiding poor performance due to inappropriate parameter selection.

[0041] (4) In the mineral resource prediction phase, Docker containerization technology is used for model deployment. This is built on the python:3.8-slim base image, ensuring high consistency and portability of the model's operating environment. Exposing the HTTP interface through the Flask framework greatly simplifies the model's practical application process. Users or systems only need to submit the remote sensing image to be predicted through a standard HTTP request to quickly obtain the prediction results, significantly lowering the barrier to use and facilitating integration into the actual workflow or information system of mineral exploration, enabling efficient and automated mineral resource identification. BRIEF DESCRIPTION OF THE DRAWINGS

[0042] Figure 1 It is the remote sensing image in the example of the present invention and the remote sensing image describing the mining point information after splitting,

[0043] Figure 2 is a schematic diagram of the VGG algorithm in the method of the present invention,

[0044] Figure 3 It is a flowchart of a specific implementation algorithm of the method of the present invention. DETAILED DESCRIPTION

[0045] The present invention is further illustrated below with reference to the accompanying drawings and specific implementation examples. It should be understood that these examples are only used to illustrate the present invention and are not used to limit the scope of the present invention. After reading the present invention, modifications of various equivalent forms of the present invention made by those skilled in the art all fall within the scope defined by the claims attached to this application.

[0046] Example: A method for classifying mineral resource images based on VGG, comprising the following stages:

[0047] Assume that at this time, Figure 1 The satellite image on the left shows the following stages:

[0048] A. Preprocessing stage: A1. Assume that the matrix of the K remote sensing source images to be fused, which are derived from radiometric values, is {X1, X2,…, X k}, the size of each image is H × L. First, each image is flattened into a column vector Construct joint data matrix D = [x1-μx2-μ…x k -μ], where is the global mean vector; A2. Calculate the covariance matrix of the joint data matrix D A3. Solve the characteristic equation Cv i =λ i v i (i=1,2,…,K)where λ1≥λ2≥…≥λ K is the eigenvalue, vi is the corresponding orthogonal eigenvector; A4. Each image is passed through p i=Dv i (i=1,2,…,K) is projected into the principal component space, and the fusion strategy adopted is: for the first principal component p1, it reflects the commonality of multiple images and is directly retained; for j≥2 principal components, the absolute value maximization criterion is used to select the detail component, In the formula Represents the projection coefficient of the kth image on the jth principal component; A5. The fused principal component Reconstructed by inverse PCA transform, Finally, y is reshaped into an H×L fused image matrix Y; A6. Based on the coordinates of the mineral points, the original remote sensing image is divided into several p×p pixel image blocks, and the center point of each image block corresponds to a mineral point coordinate. For each divided image block, the type of mineral deposit is labeled according to the data of the corresponding mineral point. A7. For all p×p pixel image blocks obtained in step A1, the bilinear interpolation method is used for upsampling operation, and their resolution is uniformly increased to P×P pixels. A8. The upsampled P×P pixel image blocks are subjected to image enhancement processing, including: applying SHARPEN sharpener for sharpening processing, and using Gaussian filtering (parameter is 5) for filtering processing, and finally the following is obtained. Figure 1 The preprocessed image block shown on the right. This preprocessing stage corresponds to Figure 3 The "latitude and longitude gridding" and "image segmentation" processing steps shown in .

[0049] B. Model construction phase: B1. Use the Torch library to build the required convolutional layer and pooling layer; B2. Use the Torch library to build the required fully connected layer B3. Use the softmax function to implement the probabilistic classifier. B4. Integrate the above components into a VGG model. The model structure is as follows Figure 2 As shown; B5. Split the training set and test set at a ratio of 80%-20%. B6. Train the model for 100 rounds, using classification accuracy as the indicator. Select the highest classification accuracy value in 100 rounds of training to evaluate the training effect. The initial learning rate is 0.001 and the batch size is 4. B7. Use the grid search method to adjust the learning rate and batch parameters. The learning rate is tested from 0.001 to 0.1 with a step size of 0.001. The batch size is tested from 4 to 128 according to the power of 2, for a total of 60 combinations. After selecting each parameter combination, go to step B6. B8. Organize the experimental results and find the optimal model. The model construction and training phase corresponds to Figure 3 The "convolutional layer", "fully connected layer", and "classification layer" processing flow shown in the figure finally outputs the classification results.

[0050] C. Mineral Deposit Prediction: C1. Deploy the trained model in Docker format and expose the interface using the Flask framework. C2. Call the interface via HTTP, input the remote sensing image to be predicted, and obtain the prediction results.

[0051] It should be noted that the above embodiments are not intended to limit the scope of protection of the present invention, and equivalent changes or substitutions made on the basis of the above technical solutions fall within the scope of protection of the claims of the present invention.

Claims

1. A VGG-based mineral resource image classification method, characterized by: The method comprises the following steps, A. Image preprocessing stage: Using geological data and remote sensing images, the remote sensing images are divided into small blocks according to the coordinates of the mining points and the classification results, and labeled. The images are then subjected to image upsampling, sharpening, and filtering operations. B. Model building phase: Design a VGG-based classification model, then divide the organized image installation ratio into a training set and a test set. Use the training set to train the model, and use the test set to evaluate the model effect. Then adjust the model parameters to find the optimal parameter combination. C. Mineral resource prediction stage: The trained model is deployed in a container and exposed to the outside world. The predicted image is input through the interface to obtain the output result.

2. The VGG-based mineral resource image classification method according to claim 1, characterized in that: The specific steps of the image preprocessing stage are as follows: A1. Constructing a joint data matrix of the remote sensing images to be fused based on radiometric values; A2. Calculate the covariance matrix of the matrix; A3. Solve the characteristic equation to obtain the eigenvalue and its corresponding eigenvector; A4. Project the image into the principal component space; A5. Perform inverse PCA transformation on the fused principal components and reshape them into an image of the original size to obtain a feature-fused image. A6. Divide the fused feature image into small blocks based on the coordinates of the mineral points and label the mineral types. A7. Upsample the feature image. A8. Sharpen and filter the feature image.

3. The VGG-based mineral resource image classification method according to claim 2, characterized in that: The model building phase includes the following steps: B1. Construct the convolutional layer, which includes five stages. The first two stages consist of two convolutional layers and one pooling layer, and the last three stages consist of three convolutional layers and one pooling layer. B2. Construct a fully connected layer. There are three intermediate layers with 4096 neurons, 4096 neurons, and the number of categories, respectively. B3. Construct a classification layer to output the probability that an image belongs to each category; B4. Integrate the above components into a VGG model; B5. Split the training set and test set; B6. Model training and evaluation of training results; B7. Adjust the parameter combination and go to step B6; B8. Organize the experimental results and find the optimal model.

4. The VGG-based mineral resource image classification method according to claim 3, characterized in that: The mineral deposit prediction stage includes the following steps: C1. Deploy the trained model in Docker format and expose the interface externally; C2. Call the interface, input the remote sensing image to be predicted, and obtain the prediction result.

5. The VGG-based mineral resource image classification method according to claim 4, characterized in that: The remote sensing image is an image obtained by sorting radioactive source data. Each mining site has multiple different types of minerals that need to be trained separately. Each mineral has only six types of mineral deposits.

6. The VGG-based mineral resource image classification method according to claim 1, characterized in that: The model building phase includes the following steps: B1. Assume that the input image size is (H×L×C), where H and L are the height and width of the image, C is the number of channels, and there are Y categories in total. Divide an image into several P×P image blocks of equal size. Assume that the input image is I. In the convolutional layer, the first stage passes through two convolutional layers and one pooling layer in sequence. The output is O1=MaxPool(ReLU(W2(ReLU(W1I+b1))+b2)), the output of the second stage is O2=MaxPool(ReLU(W4(ReLU(W3O1+b3))+b4)), the output of the third stage is O3=MaxPool(ReLU((W7(ReLU(W6(ReLU(W5O2+b5))+b6))+b7))), and the output of the fourth stage is O4=MaxPool(v((W 10 (ReLU(W9(ReLU(W8O3+b8))+b9))+b 10 ))), the output of the fifth stage is O5=MaxPool(ReLU((W 13 (ReLU(W 12 (ReLU(W 11 O4+b 11 ))+b 12 ))+b 13 ))), and finally perform a global average pooling O6 = AdaptiveAvgPool2d (O5), which is the output result of the convolution layer, where MaxPool() is the pooler, the pooling method is maximum pooling, the pooling kernel size is 2 × 2, the step size is 2, ReLU() is the activation function, all convolution kernels are 3 × 3, the step size is 1, and the padding is 1. W, b are the convolution kernel and bias of the convolution layer respectively; B2. The fully connected layer consists of three intermediate layers, with the number of layers being 4096, 4096, and the number of categories. The output of the first layer is O7=ReLU(W 14 Flatten(O6)+b 14 ), the second layer output is O8=ReLU(W 15 Flatten(O7)+b 15 ), the third layer output is O9=ReLU(W 16 Flatten(O8)+b 16 ), where ReLU() is the activation function, W and b are the weight matrix and bias of the fully connected layer, and the final output is a Y-dimensional vector; B3. The input sequence after processing by the softmax function is Y is the dimension of the output layer, o i represents the value of the i-th dimension of the output layer, represents the probability of category i; B4. Integrate the above components into a VGG model, divide the entire image into several blocks, perform several rounds of convolution, pooling, and fully connected layer processing, and finally use the softmax function for classification; B5. Split the training set and test set into an 80%-20% ratio. The ratio of test set to training set for each category should be evenly distributed. B6. Train the model for 100 rounds, using classification accuracy as the metric. The highest classification accuracy among the 100 rounds of training is selected to evaluate the training effect. The initial learning rate is 0.001, and the batch size is 4. B7. Use grid search to adjust the learning rate and batch size. Test the learning rate from 0.001 to 0.1 with a step size of 0.

001. Test the batch size from 4 to 128 in powers of 2, for a total of 60 combinations. Select each parameter combination and proceed to step B6. B8. Organize the experimental results and find the optimal model.

7. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein: When the processor executes the program, the VGG-based mineral resource image classification method as described in any one of claims 1 to 6 above is implemented.

8. A computer-readable storage medium having computer instructions stored thereon, characterized in that: When the computer instructions are executed by a processor, the VGG-based mineral resource image classification method as described in any one of claims 1 to 6 is implemented.

Images