A variable-scale deep learning underground target imaging method and imaging system

Abstract

The present invention discloses an underground target imaging method and imaging system based on variable-scale deep learning, which belongs to the field of geophysical exploration technology and includes the following steps: S1: establishing a multi-scale magnetic survey data set; S2: constructing a variable-scale deep learning network model based on UNet++, and the dynamic information fusion layer performs feature fusion on the information after pooling and split convolution, which can adapt to input data of any size and compensate for the information loss caused by pooling; S3: designing a multi-constraint loss function consisting of mean square error and loss weighting; S4: adopting transfer learning, taking the model parameters trained with a fixed size as the initial weights, and realizing progressive training of multi-size data by freezing and fine-tuning the decoder parameters, and finally verifying the effectiveness of the entire network model according to the prediction effect of the test set. This not only improves the versatility and flexibility of the network when processing data of different scales, but also can quickly adapt to the large-size data characteristics of a specific area.

Description

Technical Field

[0001] The present invention belongs to the field of geophysical exploration technology, and in particular relates to a variable-scale deep learning underground target imaging method and imaging system. Background Art

[0002] Geophysical exploration plays a vital role in resource exploration and geological disaster early warning. In the field of geophysical exploration, the accurate detection and location of underground anomalies (including but not limited to ore bodies, cavities, geological structural zones, etc.) has always been a research focus and difficulty.

[0003] As targets of geophysical exploration, the precise depth and horizontal position information of underground anomalies is of inestimable value for geological structure analysis, mineral resource assessment, and engineering geological stability evaluation. Currently, many scholars are committed to studying how to more effectively detect these underground anomalies, from early methods based on empirical formulas to modern numerical simulation techniques, especially deep learning-based underground target imaging methods, which have also developed with the continuous development of artificial intelligence technology. However, existing deep learning-based underground target imaging methods still have certain limitations in processing fixed-size data. When responding to complex and changing geological exploration needs, they often need to rely on a large number of preprocessing steps, which reduces the efficiency and universality of the model. In actual data processing, it is still impossible to process and image data of different scales.

[0004] Therefore, there is an urgent need for a variable-scale deep learning underground target imaging method that reduces preprocessing steps and can fully utilize diverse information in multi-scale scenarios. Summary of the Invention

[0005] To address these issues, this proposal proposes a method and system for underground target imaging based on scale-dependent deep learning. This approach establishes a scale-dependent deep learning network model based on UNet++ and designs a dynamic information fusion layer that automatically adjusts internal parameters and structure based on the size of the input data. Based on the trained UNet++ scale-dependent deep learning network, a transfer learning approach is employed to significantly improve the prediction performance of the scale-dependent deep learning network when processing large-scale data. This effectively addresses the high training cost and poor performance of deep learning models when processing large-scale data, and further addresses the scale-dependent issue.

[0006] A method for underground target imaging based on variable scale deep learning includes the following steps:

[0007] S1: Establish a multi-scale magnetic survey dataset, including underground target magnetic anomaly data of different sizes. The multi-scale magnetic survey dataset is divided into a training set, a validation set, and a test set. The data of the multi-scale magnetic survey dataset is obtained based on the magnetic anomaly forward calculation formula;

[0008] S2: Build a scale-varying deep learning network model based on UNet++, including a dynamic information fusion layer. The dynamic information fusion layer includes an adaptive adjustment module and a cross-scale feature alignment module. The dynamic information fusion layer can adapt to input data of any size and compensate for information loss caused by pooling.

[0009] S3: A multi-constrained loss function is used to train the scale-varying deep learning network. The multi-constrained loss function is composed of a weighted numerical error term and a spatial consistency assessment term.

[0010] S4: Using transfer learning, the trained training set parameters are used as the initial weights. By freezing the encoder parameters and fine-tuning the decoder parameters, progressive training of multi-scale data is achieved. After training, the validation set is used to verify the effect. Finally, the test set is used to predict the effect to verify the effectiveness of the entire variable-scale deep learning network model.

[0011] Preferably, the multi-scale magnetic survey dataset includes data of sizes of 16×16, 32×32, 64×64, and 100×100, and initially 30,000 datasets of size 32×32 are used, with a division ratio of training set, test set, and validation set = 6:2:2;

[0012] The forward calculation formula for magnetic anomaly is:

[0013] (1);

[0014] Where μ0 is the magnetic permeability in vacuum, μ0 unit is H / m, k is the magnetic susceptibility, H is the magnetic field strength of the magnetized medium, H unit is A / m, ξ, η, ζ are the coordinates of the endpoints of the cube in the X, Y, and Z directions respectively, and the coordinate unit is meter. The meaning of is the weight coefficient of forward calculation, which expresses the influence of the field source at the underground (ξ, η, ζ) position on the magnetic field intensity at the ground (x, y, z) observation point. The meaning of is the triple integral with respect to the weight coefficients.

[0015] Preferably, the multi-constraint loss function is a weighted combination of the mean square error and the Dice loss function. The multi-constraint loss function expression is:

[0016] (2);

[0017] The mean square error is used to measure the numerical difference between the predicted value and the true value and is defined as:

[0018] (3);

[0019] The Dice loss function is used to evaluate the overlap consistency between the predicted area and the true area. The Dice loss function coefficient is defined as:

[0020] (4);

[0021] T is the overall predicted value, P is the overall true value, is the predicted value of a single data point, is the true value of a single data, is the balance coefficient used to control the weight of the mean square error and loss function.

[0022] Preferably, the adaptive adjustment module is an adaptive pooling layer, which is responsible for dynamically adjusting the input feature map to the target size without relying on the size of a fixed pooling window;

[0023] The cross-scale feature alignment module includes a split convolution module, which includes at least two convolution kernels of different sizes set in parallel;

[0024] The dynamic adjustment formula of the input feature map of the adaptive pooling layer is:

[0025] (5);

[0026] Among them, the adaptive pooling layer inputs the feature map Convert to output feature map , is the output of the cth channel, is the input of the cth channel, and are the coordinates of the output feature map, and their value ranges are 1 and 1 , and are the dimensions of the input area in height and width, respectively. Output feature map for the cth channel, Input feature map for the c-th channel.

[0027] Preferably, the balance coefficient is 0.5.

[0028] Preferably, in step S2:

[0029] The scale-varying deep learning network was trained using the Adam optimizer on the training set. The initial learning rate was set to 0.001, the batch size was 16, and the learning rate was decayed by 50% every 20 training cycles.

[0030] The validation set loss is monitored through the early stopping mechanism, and training is terminated if the loss does not decrease for 5 consecutive cycles.

[0031] Preferably, step S4 includes the following steps:

[0032] S41: Using the model parameters pre-trained on a 32x32 fixed-size dataset as the initial weights, transfer learning is used to fine-tune the scaled deep learning network. The fine-tuning dataset includes images of 64x64 and 100x100 sizes.

[0033] S42: During the fine-tuning phase, the convolutional layer parameters of the encoder are frozen, and only the decoder and dynamic information fusion layer parameters are updated. This allows the scale-varying deep learning model to adaptively learn the spatial context information of large-scale data while retaining the underlying feature extraction capabilities, thereby achieving progressive training of multi-scale data.

[0034] A subsurface target imaging system based on variable-scale deep learning uses a subsurface target imaging method based on variable-scale deep learning.

[0035] Compared with the prior art, the present invention can achieve the following beneficial effects:

[0036] 1. Build a scale-variable deep learning network model based on UNet++ and design a dynamic information fusion layer that automatically adjusts internal parameters and structure based on the size of the input data. The core principle of this improvement lies in the design of a dynamically adjustable pooling operation and the fusion of pooled information with the information after the split convolution. This allows the model to adapt to input data of any size and compensate for the information loss caused by pooling. This mechanism enables the scale-variable deep learning network to maintain good performance when processing geophysical data of different scales, improving the network's versatility and flexibility.

[0037] 2. Based on the trained UNet++ scale-variable deep learning network, a transfer learning approach is employed to enable it to learn basic geophysical characteristics and data patterns. The pre-trained model is fine-tuned using a small amount of large-scale data from a specific region. During the fine-tuning process, only parameters of a subset of the scale-variable deep learning network layers are updated. This allows the scale-variable deep learning network to quickly adapt to the large-scale data characteristics of a specific region while retaining the general characteristics of the pre-trained model. This approach significantly improves the predictive performance of the scale-variable deep learning network when processing large-scale data, effectively addressing the high training cost and poor performance of deep learning models when processing large-scale data, and better addressing the scale-variable problem. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 This is a method flow chart of an underground target imaging method and system based on variable scale deep learning provided by the present invention;

[0039] Figure 2A method for underground target imaging based on variable scale deep learning provided by the present invention and a feature fusion graph of the dynamic information fusion layer in the system;

[0040] Figure 3 It is a predicted original image of an underground target imaging method and system based on variable scale deep learning provided by the present invention;

[0041] Figure 4 It is the predicted result image after the original image provided by the present invention has been subjected to scale-varying deep learning and transfer learning;

[0042] Figure 5 It is the intersection-over-union ratio of the underground target imaging method based on variable-scale deep learning provided by the present invention and the training set and verification set of the system. DETAILED DESCRIPTION

[0043] Hereinafter, embodiments of the present invention will be described with reference to the accompanying drawings. In the following description, identical modules are denoted by identical reference numerals. In the case of identical reference numerals, their names and functions are also identical. Therefore, their detailed description will not be repeated.

[0044] In order to make the purpose, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and do not constitute a limitation of the present invention.

[0045] Figure 1 A method flow chart of an underground target imaging method and system based on variable-scale deep learning provided by the present invention is shown.

[0046] like Figure 1 As shown, a method for underground target imaging based on variable scale deep learning includes the following steps:

[0047] S1: Establish a multi-scale magnetic survey dataset, including magnetic anomaly data of underground targets of different sizes, detect underground targets, and draw corresponding boundaries on the ground according to the boundary shape of the underground magnetic anomaly body. Use this as a label for the magnetic anomaly data, annotate the target data, and divide the dataset into training set, validation set, and test set. Based on the magnetic anomaly forward calculation formula, perform forward calculations on magnetic anomalies caused by arbitrary shapes in the underground space.

[0048] The four datasets are of different sizes: 16×16, 32×32, 64×64, and 100×100. The data volume of each size is 30,000. The forward calculation formula for the magnetic anomaly is:

[0049] (1);

[0050] Where μ0 is the magnetic permeability in vacuum, μ0 unit is H / m, k is the magnetic susceptibility, H is the magnetic field strength of the magnetized medium, H unit is A / m, ξ, η, ζ are the coordinates of the endpoints of the cube in the X, Y, and Z directions respectively, and the coordinate unit is meter. The meaning of is the weight coefficient of forward calculation, which expresses the influence of the field source at the underground (ξ, η, ζ) position on the magnetic field intensity at the ground (x, y, z) observation point. The meaning of is the triple integral of the weight coefficients.

[0051] The dataset of each size is 30,000, and the division ratio is 6:2:2 for training set, test set, and validation set. Ensure that the dataset is divided reasonably, the training set is used for model training, the validation set is used to monitor model performance, and the test set is used for final evaluation.

[0052] S2: Build a variable-scale deep learning network model based on UNet++ and design a dynamic information fusion layer to process data of different scales.

[0053] The scale-varying deep learning network includes a dynamic information fusion layer, which includes an adaptive adjustment module and a cross-scale feature alignment module;

[0054] The core innovation of the dynamic information fusion layer lies in its modular design:

[0055] Adaptive adjustment module: includes the adaptive pooling layer, which is responsible for dynamically mapping the input feature map to the target size without relying on the fixed pooling window size.

[0056] The formula for dynamically adjusting the size of the input feature map of the adaptive pooling layer is:

[0057] (5);

[0058] Among them, the adaptive pooling layer inputs the feature map Convert to output feature map , the value of each output unit is obtained as above, is the output of the cth channel, is the input of the cth channel, and are the coordinates of the output feature map, and their value ranges are 1 and 1 , and are the dimensions of the input area in height and width, respectively. Output feature map for the cth channel, Input feature map for the c-th channel, The following coordinates are and , that is, the coordinates of the output feature map are behind the input of the c-th channel, because the coordinates of the output feature map are used as the index basis, which is in turn mapped to the input feature map for pooling calculation.

[0059] The adaptive adjustment module is used to eliminate the input data size limitation, which can be achieved through adaptive pooling, deformable convolution and other technologies. The feature map size is dynamically adjusted by formula (5), so that the scale-varying deep learning network can be compatible with input data from 16×16 to 100×100 without modifying the structure.

[0060] Cross-scale feature alignment module: includes a split convolution module. The split convolution module contains two convolution kernels of different sizes set in parallel, using parallel 3×3 and 5×5 convolution kernels, and finally a split convolution structure with a 1×1 convolution kernel. The cross-scale feature alignment module is used to solve the problem of multi-resolution feature fusion, and can choose split convolution, attention mechanism and other solutions.

[0061] Figure 2 A feature fusion diagram of an underground target imaging method based on variable-scale deep learning and a dynamic information fusion layer in the system provided by the present invention is shown.

[0062] like Figure 2 As shown in the figure, the upper branch is the dynamic information fusion layer that performs adaptive pooling on the feature map, and the lower branch is the dynamic information fusion layer that performs split convolution on the feature map. Split convolution is divided into two steps: the first step is depthwise separable convolution, and the second step is pointwise separable convolution. Depthwise separable convolution is an algorithm that improves the standard convolution calculation in convolutional neural networks. By splitting the correlation between the spatial dimension and the channel depth dimension, the number of parameters required for convolution calculation is reduced. Pointwise separable convolution is usually used for interaction between feature channels. It does not involve the integration of neighborhood information, but only performs a weighted sum with all input channels at each position.

[0063] Finally, the feature maps processed by the upper and lower branches are fused, that is, the dynamic information fusion layer performs a feature fusion operation on the pooled information and the information after the split convolution, so that it can adapt to input data of any size and make up for the information loss caused by pooling.

[0064] The Adam optimizer is used to train the 32×32 training set in a variable-scale deep learning network. The initial learning rate (lr) is set to 0.001, the batch size is 16, and the learning rate decay cycle is 20 training rounds (epochs). The decay ratio is 50% of the original learning rate to improve the convergence speed.

[0065] Early stopping on the validation set is used to prevent overfitting. When the loss function value does not decrease for N consecutive training cycles, the training termination condition is triggered to prevent overfitting, thereby preventing the training process from falling into an infinite loop and improving training efficiency. It has been verified that N=5 can achieve the optimal balance between training efficiency and model performance. The training process is carried out for a total of 50 training rounds (epochs), and the adaptive optimization algorithm is used in the initial training stage.

[0066] When a fixed learning rate of 0.001 is used, the validation set loss of the model is reduced from 0.92 to 0.11 after 50 training rounds, and the training set loss is reduced from 1.03 to 0.08.

[0067] The early stopping technology is a "training termination condition based on the validation set" that is compatible with various stopping strategies, such as loss plateau monitoring and accuracy threshold judgment.

[0068] S3: Design a multi-constraint loss function, which is composed of a weighted numerical error term and a spatial consistency evaluation term.

[0069] In this scheme, the numerical error term is the mean square error, the spatial consistency evaluation term is the Dice loss function, and the weighted expression is:

[0070] (2);

[0071] The mean square error is used to measure the numerical difference between the predicted value and the true value. The mean square error formula is:

[0072] (3);

[0073] The Dice loss function is used to evaluate the overlap consistency between the predicted area and the true area. The Dice loss function coefficient formula is:

[0074] (4);

[0075] In the formula, P is the overall predicted value, T is the overall true value, is the predicted value of a single data point, is the true value of a single data, The balance coefficient is used to control the weight of mean square error and loss. MSE is mean square error. Dice in Dice loss function is an English word and is usually kept as is without translation.

[0076] Dice comes from the Dice Similarity Coefficient, a metric proposed by Lee R. Dice to measure the similarity between two sets. The Dice loss function is primarily used in image segmentation tasks. By maximizing the overlap between the predicted segmentation result and the ground-truth annotation, it effectively addresses class imbalance, improving segmentation accuracy and connectivity of the target region. In literature and engineering practice, the term "Dice Loss" is often used to refer to the problem.

[0077] The mean square error is used to constrain energy matching, accurately restore the field value magnitude, constrain loss shape similarity, sharpen boundaries, and suppress artifacts.

[0078] Figure 3 The figure shows the predicted original image of an underground target imaging method and system based on variable-scale deep learning provided by the present invention.

[0079] like Figure 3 As shown, Figure 3 The sizes of the magnetic anomaly map on the middle left from top to bottom are as follows: Figure 3 The size of (a) is 32, Figure 3 The size of (b) is 25, Figure 3 The size of (c) is 64, and the trained variable scale deep learning network is put into it for prediction.

[0080] Figure 4 The figure shows the predicted result image of the original image provided by the present invention after scale-varying deep learning and transfer learning.

[0081] like Figure 4 The right side shows the prediction results, where Figure 4 (a') is the image result of size 32, Figure 4 The image result of size 25 is shown in (b'). Figure 4 (a') and Figure 4 The image in (b') becomes significantly clearer. Figure 4 The size of (c') is 64. The prediction effect is poor when it is directly put into the variable scale deep learning network. It is still fuzzy. However, after transfer learning, the prediction is made again, such as Figure 4 As shown in (c''), the image result becomes significantly clearer, which shows the effectiveness of transfer learning of this scheme.

[0082] Figure 5 The intersection-over-union ratio of the training set and validation set of the underground target imaging method based on variable-scale deep learning provided by the present invention is shown.

[0083] like Figure 5As shown in the figure, when selecting the optimal α, ten different α values ​​are tested, namely α = 0.1, 0.2, 0.3, 0.4, 0.5, 0.6, 0.7, 0.8, 0.9, and 1. Each test uses a multi-constraint loss function with different weight combinations for training to obtain the optimal α value, and calculates the average IOU value on the validation set to measure the performance of the variable-scale deep learning network model.

[0084] The IOU value is the intersection-over-union ratio. IOU is a standard for measuring the accuracy of detecting corresponding objects in a specific data set. It is used to measure the degree of overlap between the predicted area and the true area. The calculation formula is: IOU = the intersection area of ​​the predicted area and the true area / the union area of ​​the predicted area and the true area. The value range is [0,1]. The higher the value, the closer the prediction is to the truth.

[0085] like Figure 5 As shown in the figure, the experimental results show that when α=0.5, the average IOU of the validation set reaches the maximum value, indicating that the α weight at this time can achieve the best balance between the mean square error (MSE) and the loss function (Dice), so the optimal weight is 0.5.

[0086] S4: Through transfer learning, the parameters of the variable-scale deep learning network model pre-trained on a 32x32 fixed-size dataset are used as the initial weights to accelerate the convergence of deep learning training to the target volume imaging accuracy. By freezing the encoder parameters and fine-tuning the decoder parameters, progressive training of multi-scale data is achieved. Finally, the effectiveness of the entire variable-scale deep learning network model is verified based on the prediction results of the test set.

[0087] Transfer learning is used to fine-tune the scaled deep learning network, and the fine-tuning dataset includes images of size 64x64 and 100x100.

[0088] During the fine-tuning stage, the convolutional layer parameters of the encoder are frozen, and only the parameters of the decoder and dynamic information fusion layer are updated. This allows the variable-scale deep learning model to adaptively learn the spatial context information of large-scale data while retaining the underlying feature extraction capabilities, thereby achieving progressive training of multi-scale data. The purpose of this fine-tuning process is to allow the variable-scale deep learning network to adapt to larger-sized input images and improve its performance on large images without retraining the entire variable-scale deep learning network. Finally, the effectiveness of the entire variable-scale deep learning network model is verified based on the prediction results of the test set in 20% of the dataset.

[0089] Initially adopting a small 32×32 size can reduce the computational complexity of initial training and accelerate the convergence of variable-scale deep learning network models. Direct training on large-scale data requires reinitializing network parameters, which increases computing resource consumption by 3-5 times and can easily lead to overfitting due to insufficient sample size. Therefore, this paper uses a parameter migration strategy to transfer the feature extraction capabilities gained from small-scale data training to large-scale data processing, achieving a gradual expansion of model capabilities rather than repeated training.

[0090] 60% of the dataset is used as a training set for training, 20% as a validation set for validation, and 20% as a test set for prediction.

[0091] It includes the following three stages:

[0092] 1. Training phase: Use the training set to train the model and minimize the loss function by continuously adjusting the parameters.

[0093] Data preparation: Extract training, validation, and test sets from the multi-scale magnetic survey dataset. Ensure that the dataset is properly divided: the training set is used for model training, the validation set is used to monitor model performance, and the test set is used for final evaluation.

[0094] Training loop: Train the training set batch by batch. After each training cycle, perform the following operations.

[0095] 2. Validation phase: During the training process, the validation set is used to evaluate the performance of the scaled deep learning network model and select the optimal model parameters. Through methods such as cross-validation, the model can be further optimized in the combination of multiple training sets and validation sets.

[0096] Verification process:

[0097] Data input: Input the validation set data into the currently trained model.

[0098] Loss calculation: Calculate the multi-constraint loss value on the validation set, including numerical error terms and spatial consistency evaluation terms.

[0099] Numerical error terms: such as mean squared error, which measures the difference between the model's predicted value and the true value.

[0100] Spatial consistency evaluation items: such as the structural similarity index, which measures the smoothness and consistency of the model prediction results in space.

[0101] Performance indicator calculation: Calculate other performance indicators on the validation set, such as accuracy, precision, recall, F1 value, etc.

[0102] Monitor performance: Monitor the performance of the model by comparing the validation set loss and performance metrics across different training epochs. If the validation set loss stops decreasing or starts increasing over multiple consecutive epochs, you can stop training early.

[0103] Save the best model: Based on the performance indicators on the validation set, save the model parameters that perform best. Usually, the model parameters with the lowest validation set loss value can be selected as the final model.

[0104] 3. Testing Phase: After the training and validation of the scale-variable deep learning network model is completed, the test set is used to evaluate the final performance of the scale-variable deep learning network model. The results of the test set are used to report the generalization ability of the model and ensure the performance of the scale-variable deep learning network model on new data.

[0105] Testing process:

[0106] Data input: After training, use the test set to evaluate the final saved model. Input the test set data into the model.

[0107] Loss calculation: Calculate the multi-constraint loss value on the test set, including numerical error terms and spatial consistency evaluation terms.

[0108] Numerical error terms: such as mean squared error, which measures the difference between the model's predicted value and the true value.

[0109] Spatial consistency evaluation items: such as the structural similarity index, which measures the smoothness and consistency of the model prediction results in space.

[0110] Performance indicator calculation: Calculate other performance indicators on the test set, such as accuracy, precision, recall, F1 value, etc.

[0111] Evaluate the model: Verify the effectiveness of the entire scale-varying deep learning network model based on the evaluation results on the test set. If the performance indicators on the test set meet the project requirements, it means that the model has good generalization ability and can be used in practical applications.

[0112] A system for underground target imaging based on variable-scale deep learning uses a method for underground target imaging based on variable-scale deep learning. An electronic device includes a memory, a processor, and a computer program stored in the memory. When the processor executes the program, it implements the method for underground target imaging based on variable-scale deep learning.

[0113] Table 1 The time taken to train the network is as follows:

[0114] Task Type Time consuming to train a new network Transfer learning takes time 100x100 183 minutes and 8 seconds (3 hours, 3 minutes and 8 seconds) 15 minutes and 6 seconds + 1 hour, 33 minutes and 40 seconds (1 hour, 48 minutes and 46 seconds) 64x64 93 minutes and 40 seconds (1 hour, 33 minutes and 40 seconds) 8 minutes 27 seconds + 1 hour 7 minutes 5 seconds (1 hour 15 minutes 32 seconds) 32x32 1 hour, 7 minutes, 5 seconds 28 minutes 16 seconds + 3 minutes 12 seconds (31 minutes 28 seconds)

[0115] As shown in Table 1, the first column shows the time required to train the new network, while the second column shows the time required to transfer learning, which is the combined time required to generate the dataset and retrain the network. This time depends on the computer's hardware. Compared to existing technologies, this solution improves computational efficiency and mitigates information loss.

[0116] The dynamic information fusion network of the present invention can be implemented based on an encoder-decoder architecture (such as UNet++), but does not exclude the application of other network frameworks (such as DeepLabV3+ and HRNet).

[0117] It should be understood that the various forms of the processes shown above can be used to reorder, add, or delete steps. For example, the steps described in the present disclosure can be performed in parallel, sequentially, or in a different order, as long as the desired results of the technical solutions disclosed in the present disclosure can be achieved. This is not limited herein.

[0118] The above specific embodiments do not limit the scope of protection of the present invention. Those skilled in the art will appreciate that various modifications, combinations, sub-combinations, and substitutions may be made based on design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention are intended to be included within the scope of protection of the present invention.

Claims

1. A method for underground target imaging based on variable scale deep learning, characterized in that: The steps include: S1: Establish a multi-scale magnetic survey dataset, including underground target magnetic anomaly data of different sizes. The multi-scale magnetic survey dataset is divided into a training set, a validation set, and a test set. The data of the multi-scale magnetic survey dataset is obtained based on the magnetic anomaly forward calculation formula; S2: Build a scale-varying deep learning network model based on UNet++, including a dynamic information fusion layer. The dynamic information fusion layer includes an adaptive adjustment module and a cross-scale feature alignment module. The dynamic information fusion layer can adapt to input data of any size and compensate for information loss caused by pooling. S3: A multi-constrained loss function is used to train the scale-varying deep learning network. The multi-constrained loss function is composed of a weighted numerical error term and a spatial consistency assessment term. S4: Using transfer learning, the trained training set parameters are used as the initial weights. By freezing the encoder parameters and fine-tuning the decoder parameters, progressive training of multi-scale data is achieved. After training, the validation set is used to verify the effect. Finally, the test set is used to predict the effect to verify the effectiveness of the entire variable-scale deep learning network model. Adaptive adjustment module: includes adaptive pooling layer, which is responsible for dynamically mapping the input feature map to the target size without relying on the fixed pooling window size; The formula for dynamically adjusting the size of the input feature map of the adaptive pooling layer is: （5） Among them, the adaptive pooling layer inputs the feature map Convert to output feature map , the value of each output unit is obtained as above, is the output of the cth channel, is the input of the cth channel, and are the coordinates of the output feature map, and their value ranges are 1 and 1 , and are the dimensions of the input area in height and width, respectively. Output feature map for the cth channel, Input feature map for the c-th channel, The following coordinates are and , that is, the input of the cth channel is followed by the coordinates of the output feature map, which is used as the index basis and in turn mapped to the input feature map for pooling calculation; Cross-scale feature alignment module: includes a split convolution module. The split convolution module contains two convolution kernels of different sizes set in parallel, using parallel 3×3 and 5×5 convolution kernels, and finally a split convolution structure with a 1×1 convolution kernel. The cross-scale feature alignment module is used to solve the problem of multi-resolution feature fusion.

2. The underground target imaging method based on variable scale deep learning according to claim 1, characterized in that: The multi-scale magnetic survey dataset contains data of sizes 16×16, 32×32, 64×64, and 100×100. Initially, 30,000 datasets of size 32×32 were used, and the ratio of training set, test set, and validation set was 6:2:

2. The forward calculation formula for magnetic anomaly is: （1）； Where μ0 is the magnetic permeability in vacuum, μ0 unit is H / m, k is the magnetic susceptibility, H is the magnetic field strength of the magnetized medium, H unit is A / m, ξ, η, ζ are the coordinates of the endpoints of the cube in the X, Y, and Z directions respectively, and the coordinate unit is meter. The meaning of is the weight coefficient of forward calculation, which expresses the influence of the field source at the underground (ξ, η, ζ) position on the magnetic field intensity at the ground (x, y, z) observation point. The meaning of is the triple integral with respect to the weight coefficients.

3. The underground target imaging method based on variable scale deep learning according to claim 1, characterized in that: The multi-constraint loss function is composed of the weighted mean square error and Dice loss function. The expression of the multi-constraint loss function is: （2）； The mean square error is used to measure the numerical difference between the predicted value and the true value and is defined as: （3）； The Dice loss function is used to evaluate the overlap consistency between the predicted area and the true area. The Dice loss function coefficient is defined as: （4）； T is the overall predicted value, P is the overall true value, is the predicted value of a single data point, is the true value of a single data, is the balance coefficient used to control the weight of the mean square error and loss function.

4. The underground target imaging method based on variable scale deep learning according to claim 1, characterized in that: Balance coefficient is 0.

5.

5. The underground target imaging method based on variable scale deep learning according to claim 1, characterized in that: It is characterized by: In step S2: The scale-varying deep learning network was trained using the Adam optimizer on the training set. The initial learning rate was set to 0.001, the batch size was 16, and the learning rate was decayed by 50% every 20 training cycles. The validation set loss is monitored through the early stopping mechanism, and training is terminated if the loss does not decrease for 5 consecutive cycles.

6. The underground target imaging method based on variable scale deep learning according to claim 1, characterized in that: Step S4 includes the following steps: S41: Using the model parameters pre-trained on a 32x32 fixed-size dataset as the initial weights, transfer learning is used to fine-tune the scaled deep learning network. The fine-tuning dataset includes images of 64x64 and 100x100 sizes. S42: During the fine-tuning phase, the convolutional layer parameters of the encoder are frozen, and only the decoder and dynamic information fusion layer parameters are updated. This allows the scale-varying deep learning model to adaptively learn the spatial context information of large-scale data while retaining the underlying feature extraction capabilities, thereby achieving progressive training of multi-scale data.

7. An underground target imaging system based on variable scale deep learning, using an underground target imaging method based on variable scale deep learning as described in claims 1-6.

Images