Underground target imaging method and imaging system for variable-scale deep learning

Abstract

The invention discloses an underground target imaging method and imaging system based on variable scale deep learning, and belongs to the technical field of geophysical exploration, and the method comprises the following steps: S1, building a multi-scale magnetic survey data set; s2, constructing a UNet + +-based variable-scale deep learning network model, and performing feature fusion on pooled information and information subjected to split convolution by a dynamic information fusion layer so as to adapt to input data of any size and make up for information loss caused by pooling; s3, designing a multi-constraint loss function formed by mean square error and loss weighting; s4, migration learning is adopted, model parameters trained in a fixed size serve as initial weights, progressive training of multi-size data is achieved by freezing and fine tuning of decoder parameters, and finally the effectiveness of the whole network model is verified according to a test set prediction effect, so that the universality and flexibility of the network are improved when data of different sizes are processed, and the network performance is improved. And large-size data features in a specific area can be quickly adapted.

Description

Technical Field

[0001] The present invention belongs to the technical field of geophysical exploration, and particularly relates to an underground target imaging method and imaging system based on variable-scale deep learning. Background Art

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

[0003] As the target body of geophysical exploration, the accurate depth and horizontal position information of underground anomalies are of inestimable value for geological structure analysis, mineral resource assessment, engineering geological stability evaluation, etc. At present, many scholars are committed to studying how to more effectively detect these underground anomalies. From the early methods based on empirical formulas to modern numerical simulation technologies, especially the underground target imaging methods based on deep learning, they have also developed with the continuous development of artificial intelligence technology. However, the existing underground target imaging methods based on deep learning still have certain limitations in processing fixed-size data, making them often rely on a large number of preprocessing steps when dealing with complex and variable geological exploration requirements, thus reducing the efficiency and universality of the model, and unable to realize the processing and imaging of different-scale data in actual data processing.

[0004] Therefore, there is an urgent need for an underground target imaging method based on variable-scale deep learning that reduces preprocessing steps and can make full use of diversity information in multi-scale scenarios. Summary of the Invention

[0005] To solve the above problems, the present invention provides an underground target imaging method and imaging system based on variable-scale deep learning, establishes a variable-scale deep learning network model based on UNet++, designs a dynamic information fusion layer, and automatically adjusts internal parameters and structures according to the size of the input data. On the basis of the trained UNet++ variable-scale deep learning network, the transfer learning method is adopted to significantly improve the prediction effect of the variable-scale deep learning network when processing large-size data, effectively solve the problems of high training cost and poor performance of the deep learning model when processing large-size data, and better solve the variable-scale problem.

[0006] An underground target imaging method based on variable-scale deep learning includes the following steps: S1: Establish a multi-scale magnetic measurement data set, including underground target magnetic anomaly data of different sizes. The multi-scale magnetic measurement data set is divided into a training set, a validation set, and a test set, and the data of the multi-scale magnetic measurement data set is obtained based on the magnetic anomaly forward calculation formula; S2: Construct a variable-scale 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 the information loss caused by pooling. S3: Use a multi-constraint loss function to train the variable-scale deep learning network. The multi-constraint loss function is composed of a weighted numerical error term and a spatial consistency evaluation term. S4: Adopt transfer learning. Use the parameters of the trained training set as the initial weights. By freezing the encoder parameters and fine-tuning the decoder parameters, progressive training of multi-size data is achieved. After the training is completed, use the validation set to verify the effect, and finally predict the effect according to the test set to verify the effectiveness of the entire variable-scale deep learning network model.

[0007] Preferably, the multi-scale magnetic measurement data set contains data with sizes of 16×16, 32×32, 64×64, and 100×100 respectively. Initially, 30,000 data sets of size 32×32 are used, and the division ratio is training set: test set: validation set = 6:2:2. The forward calculation formula for magnetic anomaly is: (1); In the formula, μ0 is the magnetic permeability in vacuum, the unit of μ0 is H / m, k is the magnetic susceptibility, H is the magnetic field strength of the magnetized medium, the unit of H is A / m, ξ, η, ζ are the end point coordinates of the cube in the X, Y, Z directions respectively, and the coordinate unit is meter. The meaning of is the weight coefficient of the forward calculation. The weight coefficient expresses the influence of the field source at the underground position (ξ, η, ζ) on the magnetic field strength of the observation point on the ground (x, y, z). The meaning of is the triple integral for the weight coefficient.

[0008] Preferably, 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 region and the true region. The coefficient of the Dice loss function is defined as: (4); T is the overall predicted value, P is the overall true value, is the predicted value of a single data, is the true value of a single data, The balance coefficient is used to control the weights of the mean square error and the loss function.

[0009] 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. The cross-scale feature alignment module includes a split convolution module, which contains at least two convolution kernels of different sizes set in parallel. The dynamic adjustment formula for the input feature map of the adaptive pooling layer is: (5); Among them, the adaptive pooling layer converts the input feature map into the output feature map , is the output of the c-th channel, is the input of the c-th channel, and are the coordinates of the output feature map, and the value ranges are 1 and 1 , and are the sizes of the input area in height and width respectively, is the output feature map of the c-th channel, is the input feature map of the c-th channel.

[0010] Preferably, the balance coefficient is 0.5.

[0011] Preferably, in step S2: The variable-scale deep learning network uses the Adam optimizer to train the training set. The initial learning rate is set to 0.001, the batch size is 16, and the learning rate decays by 50% every 20 training epochs. Monitor the validation set loss through the early stopping mechanism. If the loss does not decrease for 5 consecutive epochs, terminate the training.

[0012] Preferably, step S4 includes the following steps: S41: Use the model parameters pre-trained on the 32x32 fixed-size dataset as the initial weights, and use transfer learning to fine-tune the variable-scale deep learning network. The fine-tuning dataset includes images of sizes 64x64 and 100x100. S42: In the fine-tuning stage, freeze the convolution layer parameters of the encoder part, and only update the parameters of the decoder and the dynamic information fusion layer, so that the variable-scale deep learning model can retain the underlying feature extraction ability while adaptively learning the spatial context information of large-size data, and achieve progressive training of multi-size data.

[0013] An underground target imaging system based on variable-scale deep learning uses an underground target imaging method based on variable-scale deep learning.

[0014] Compared with the prior art, the present invention can achieve the following beneficial effects: 1. Establish a variable-scale deep learning network model based on UNet++, design a dynamic information fusion layer, and automatically adjust internal parameters and structures according to the size of input data. The core principle of this improvement lies in designing a dynamically adjustable pooling operation, and performing feature fusion operations on the pooled information and the information after split convolution, enabling it to adapt to input data of any size and making up for the information loss caused by pooling. This mechanism enables the variable-scale deep learning network to maintain good performance when processing geophysical data of different scales, improving the versatility and flexibility of the network.

[0015] 2. Based on the trained UNet++ variable-scale deep learning network, adopt the transfer learning method to enable it to learn basic geophysical features and data patterns. Use a small amount of large-size data in a specific area to fine-tune the pre-trained model. During the fine-tuning process, only update the parameters of some layers of the variable-scale deep learning network. In this way, on the basis of retaining the general features in the pre-trained model, it can quickly adapt to the large-size data features in a specific area. Through this method, the prediction effect of the variable-scale deep learning network when processing large-size data can be significantly improved, effectively solving the problems of high training cost and poor performance of deep learning models when processing large-size data, and better solving the variable-scale problem. BRIEF DESCRIPTION OF THE DRAWINGS

[0016] Figure 1 is a method flow chart in an underground target imaging method and system based on variable-scale deep learning provided by the present invention; Figure 2 is a feature fusion diagram of the dynamic information fusion layer in an underground target imaging method and system based on variable-scale deep learning provided by the present invention; Figure 3 is the predicted original image in an underground target imaging method and system based on variable-scale deep learning provided by the present invention; Figure 4 is the predicted result image after the predicted original image provided by the present invention undergoes variable-scale deep learning and transfer learning; Figure 5 is the intersection-over-union ratio of the training set and the validation set in an underground target imaging method and system based on variable-scale deep learning provided by the present invention. DETAILED DESCRIPTION OF THE INVENTION

[0017] In the following, embodiments of the present invention will be described with reference to the accompanying drawings. In the following description, the same modules are denoted by the same reference numerals. In the case of the same reference numerals, their names and functions are also the same. Therefore, their detailed descriptions will not be repeated.

[0018] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention will be further described in detail below with reference to 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 to the present invention.

[0019] Figure 1 The flowchart of the method in a method and system for underground target imaging based on variable-scale deep learning provided by the present invention is shown.

[0020] As Figure 1 shown, a method for underground target imaging based on variable-scale deep learning includes the following steps: S1: Establish a multi-scale magnetic measurement data set, including magnetic anomaly data of underground targets with different sizes, detect underground targets, and draw the corresponding boundary on the ground according to the boundary shape of the underground magnetic anomaly body, which is used as the label of the magnetic anomaly data, label the target data, divide the data set into a training set, a validation set, and a test set, and perform forward calculation of the magnetic anomaly caused by any shape in the underground space based on the magnetic anomaly forward calculation formula.

[0021] The sizes of the four different data sets are 16×16, 32×32, 64×64, and 100×100 respectively, and the data volume of each size is 30,000. The forward calculation formula for the magnetic anomaly is: (1); In the formula, μ0 is the magnetic permeability in vacuum, the unit of μ0 is H / m, k is the magnetic susceptibility, H is the magnetic field strength of the magnetized medium, the unit of H is A / m, ξ, η, ζ are the end point coordinates 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 the forward calculation. The weight coefficient expresses the influence of the field source at the underground (ξ, η, ζ) position on the magnetic field strength at the ground (x, y, z) observation point. The meaning of is the triple integral of the weight coefficient.

[0022] The data volume of each size of the data set is 30,000, and the division ratio is training set: test set: validation set = 6:2:2; ensure that the division of the data set is reasonable. The training set is used for model training, the validation set is used for monitoring the model performance, and the test set is used for the final evaluation.

[0023] S2: Construct a variable-scale deep learning network model based on UNet++, and design a dynamic information fusion layer to process data of different scales.

[0024] The variable-scale deep learning network includes a dynamic information fusion layer, and the dynamic information fusion layer includes an adaptive adjustment module and a cross-scale feature alignment module; The core innovation of the dynamic information fusion layer lies in the modular design: Adaptive adjustment module: It includes an adaptive pooling layer, and the adaptive pooling layer is responsible for dynamically mapping the input feature map to the target size without relying on the size of a fixed pooling window.

[0025] The dynamic adjustment formula for the size of the input feature map of the adaptive pooling layer is: (5); Among them, the adaptive pooling layer maps the input feature map to the output feature map , and the value of each output unit is obtained as above. is the output of the c-th channel, is the input of the c-th channel, and are the coordinates of the output feature map, and the value ranges are 1 and 1 , and are the sizes of the input area in height and width respectively, is the output feature map of the c-th channel, is the input feature map of the c-th channel, The following coordinates are and , that is, the coordinates of the output feature map follow the input of the c-th channel because the coordinates of the output feature map are used as the indexing basis and are mapped back to the input feature map for pooling calculation.

[0026] The adaptive adjustment module is used to eliminate the limitation of the input data size and can be realized by technologies such as adaptive pooling and deformable convolution; the size of the feature map is dynamically adjusted through formula (5), enabling the variable-scale deep learning network to be compatible with input data from 16×16 to 100×100 without modifying the structure.

[0027] Cross-scale feature alignment module: It includes a split convolution module, and the split convolution module contains two convolution kernels with different sizes set in parallel, adopting a split convolution structure with parallel 3×3 and 5×5 convolution kernels and finally a 1×1 convolution kernel. The cross-scale feature alignment module is used to solve the problem of multi-resolution feature fusion and can select solutions such as split convolution and attention mechanism.

[0028] Figure 2 shows the feature fusion diagram of the dynamic information fusion layer in an underground target imaging method and system based on variable-scale deep learning provided by the present invention.

[0029] As Figure 2 shown, the upper branch is the dynamic information fusion layer performing adaptive pooling on the feature map, and the lower branch is the dynamic information fusion layer performing split convolution on the feature map. The split convolution is divided into two steps. The first step is to perform depthwise separable convolution, and the second step is to perform pointwise separable convolution. Depthwise separable convolution is an algorithm obtained by improving the standard convolution calculation in a convolutional neural network. By splitting the correlation between the spatial dimension and the channel depth dimension, it reduces the number of parameters required for convolution calculation. Pointwise separable convolution is usually used for the interaction between feature channels, without involving the integration of neighborhood information, but only performing weighted summation with all input channels at each position.

[0030] 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 split convolution, enabling it to adapt to input data of any size and compensate for the information loss caused by pooling.

[0031] Using the Adam optimizer, a 32×32 size training set is placed in the variable-scale deep learning network for training. The initial learning rate (lr) is set to 0.001, the batch size is 16, the learning rate decay period is 20 training epochs, and the decay ratio is 50% of the original learning rate to improve the convergence speed.

[0032] The early stopping technique 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 avoiding getting stuck in an infinite loop during the training process and improving the 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 50 training epochs, and the adaptive optimization algorithm is adopted in the initial training stage.

[0033] When using a fixed learning rate of 0.001, the loss value on the validation set of the model decreases from 0.92 to 0.11 after 50 training epochs, and the loss value on the training set decreases from 1.03 to 0.08.

[0034] The early stopping technique, "training termination conditions based on the validation set", is compatible with multiple stopping strategies, such as loss plateau monitoring and accuracy threshold judgment.

[0035] S3: Design a multi-constraint loss function, which is composed of a weighted sum of a numerical error term and a spatial consistency evaluation term.

[0036] In this scheme, the numerical error term is the mean squared error, and the spatial consistency evaluation term is the Dice loss function. The weighted composition expression is: (2); The mean squared error is used to measure the numerical difference between the predicted value and the true value. The formula for the mean squared error is: (3); The Dice loss function is used to evaluate the overlap consistency between the predicted region and the true region. The formula for the coefficient of the Dice loss function is: (4); In the formula, P is the overall predicted value, T is the overall true value, is the predicted value of a single data, is the true value of a single data, is the balance coefficient used to control the weights of the mean squared error and the loss. MSE is the mean squared error. In the Dice loss function, Dice is an English word and is usually kept unchanged without translation. Dice is derived from the Dice coefficient (Dice Similarity Coefficient), which is an index to measure the similarity between two sets and was proposed by Lee R. Dice. The Dice loss function is mainly used in image segmentation tasks. By maximizing the overlap between the predicted segmentation result and the true annotation, it can effectively handle the problem of class imbalance, improve the segmentation accuracy and the connectivity of the target region. In literature and engineering practice, the English "Dice Loss" is usually directly used to represent it.

[0037] The mean squared error is used for energy matching constraints, accurately restoring the magnitude of the field value, loss form similarity constraints, sharpening the boundary, and suppressing artifacts.

[0038] Figure 3 Shows the original prediction image of an underground target imaging method and system based on variable-scale deep learning provided by the present invention.

[0039] As Figure 3 shown, Figure 3 In the left magnetic anomaly map, the sizes from top to bottom are as follows: Figure 3 In (a) of it, the size is 32, Figure 3 In (b) of it, the size is 25, Figure 3 In (c) of it, the size is 64, and it is put into the trained variable-scale deep learning network for prediction.

[0040] Figure 4 Shows the predicted result image after variable-scale deep learning and transfer learning of the original prediction image provided by the present invention.

[0041] As Figure 4 shown on the right, the predicted result is as follows. Among them, Figure 4 In (a') of it, the image result is of size 32, Figure 4 In (b') of it, the image result is of size 25. Thus, it can be seen that Figure 4In (a') and Figure 4 the image results in (b') are significantly clearer. Figure 4 For (c') with a size of 64 directly input into the variable-scale deep learning network, the prediction effect is not good and it is still blurry. However, after performing transfer learning and then predicting, as Figure 4 shown in (c''), the image results are significantly clearer. Thus, the effectiveness of the transfer learning in this solution can be seen.

[0042] Figure 5 Shows the intersection over union of the training set and the validation set of a method and system for underground target imaging based on variable-scale deep learning provided by the present invention.

[0043] As Figure 5 shown, 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, 1. Each test uses a multi-constraint loss function with different weight combinations for training to obtain the optimal α value, and the average IOU value is calculated on the validation set to measure the performance of the variable-scale deep learning network model.

[0044] The IOU value, that is, the intersection over union, is a standard for measuring the accuracy of detecting corresponding objects in a specific dataset, used to measure the overlapping degree between the predicted region and the true region. The calculation formula is: IOU = the intersection area of the predicted region and the true region / the union area of the predicted region and the true region, and the value range is [0, 1]. The higher the value, the closer the prediction is to the truth.

[0045] As Figure 5 shown, 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). Therefore, the optimal weight is 0.5.

[0046] S4: Through transfer learning, use the parameters of the variable-scale deep learning network model pre-trained on a 32x32 fixed-size dataset as the initial weights to accelerate the imaging accuracy of the target object with good convergence in deep learning training. By freezing the encoder parameters and fine-tuning the decoder parameters, progressive training of multi-size data is achieved. Finally, the effectiveness of the entire variable-scale deep learning network model is verified according to the prediction effect on the test set.

[0047] Use transfer learning to fine-tune the variable-scale deep learning network. The fine-tuning dataset includes images with sizes of 64x64 and 100x100.

[0048] During the fine-tuning stage, the convolutional layer parameters of the encoder part are frozen, and only the parameters of the decoder and the dynamic information fusion layer are updated. This enables the variable-scale deep learning model to adaptively learn the spatial context information of large-size data while retaining the underlying feature extraction ability, achieving progressive training of multi-size data. The purpose of this fine-tuning process is to enable the variable-scale deep learning network to adapt to larger-size input images and improve its performance on large images without having to retrain 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.

[0049] Initially using a small size of 32×32 can reduce the computational complexity of initial training and accelerate the convergence of the variable-scale deep learning network model. If directly training large-size data, it is necessary to re-initialize the network parameters, which will lead to a 3 - 5-fold increase in computational resource consumption and is prone to overfitting due to insufficient sample size. Therefore, through the parameter transfer strategy of the present invention, the feature extraction ability obtained from training small-size data is transferred to the processing of large-size data to achieve progressive expansion of the model's capabilities rather than repeated training.

[0050] 60% of the dataset is used as the training set for training, 20% is used as the validation set for validation, and 20% is used as the test set for prediction.

[0051] Specifically, it includes the following three stages: 1. Training stage: Use the training set to train the model and minimize the loss function by continuously adjusting the parameters.

[0052] Data preparation: Extract the training set, validation set, and test set from the multi-scale magnetic measurement dataset. Ensure a reasonable division of the dataset, where the training set is used for model training, the validation set is used to monitor the model performance, and the test set is used for final evaluation.

[0053] Training loop: Train the training set batch by batch. After each training cycle ends, perform the following operations.

[0054] 2. Validation stage: During the training process, use the validation set to evaluate the performance of the variable-scale deep learning network model, select the best model parameters, and through methods such as cross-validation, the model can be further optimized among multiple combinations of training sets and validation sets.

[0055] Validation process: Data input: Input the data of the validation set into the currently trained model.

[0056] Loss calculation: Calculate the multi-constraint loss value on the validation set, including the numerical error term and the spatial consistency evaluation term.

[0057] Numerical error term: Such as the mean square error, which measures the difference between the model prediction value and the true value.

[0058] Spatial consistency evaluation item: Such as the structural similarity index, which measures the smoothness and consistency of the model prediction results in space.

[0059] Performance metric calculation: Calculate other performance metrics on the validation set, such as accuracy, precision, recall, F1 score, etc.

[0060] Monitor performance: Monitor the performance of the model by comparing the validation set loss values and performance metrics for different training epochs. If the validation set loss value no longer decreases or starts to increase over multiple consecutive epochs, training can be stopped early.

[0061] Save the best model: Save the model parameters with the best performance according to the performance metrics on the validation set. Usually, the model parameters with the lowest validation set loss value can be selected as the final model.

[0062] 3. Testing phase: After the variable-scale deep learning network model is trained and validated, use the test set to evaluate the final performance of the variable-scale 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 variable-scale deep learning network model on new data.

[0063] Testing process: Data input: After training, use the test set to evaluate the finally saved model. Input the data of the test set into the model.

[0064] Loss calculation: Calculate the multi-constraint loss value on the test set, including the numerical error term and the spatial consistency evaluation item.

[0065] Numerical error term: Such as the mean squared error, which measures the difference between the model prediction value and the true value.

[0066] Spatial consistency evaluation item: Such as the structural similarity index, which measures the smoothness and consistency of the model prediction results in space.

[0067] Performance metric calculation: Calculate other performance metrics on the test set, such as accuracy, precision, recall, F1 score, etc.

[0068] Evaluate the model: Verify the effectiveness of the entire variable-scale deep learning network model based on the evaluation results on the test set. If the performance metrics on the test set meet the project requirements, it indicates that the model has good generalization ability and can be used in practical applications.

[0069] An underground target imaging system based on variable-scale deep learning uses an underground target imaging method based on variable-scale deep learning. An electronic device includes a memory, a processor, and a computer program stored on the memory. When the processor executes the program, it implements the underground target imaging method based on variable-scale deep learning.

[0070] Table 1 The time consumption of the training network is as follows: Task type Time taken for new training network Time taken for transfer learning 100x100 183 minutes 8 seconds (3 hours 3 minutes 8 seconds) 15 minutes 6 seconds + 1 hour 33 minutes 40 seconds (1 hour 48 minutes 46 seconds) 64x64 93 minutes 40 seconds (1 hour 33 minutes 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) As shown in Table 1, the time consumption of the new training network in the first column is the time consumption of the original network, and the time consumption of transfer learning in the second column is the sum of the time consumption of generating the dataset and retraining. This time consumption is determined according to the hardware conditions of the computer. Compared with the prior art, the computing efficiency of this solution is improved, and the problem of information loss is compensated for.

[0071] The dynamic information fusion network of the present invention can be implemented based on an encoder-decoder architecture (such as UNet++), but the application of other network frameworks (such as DeepLabV3+ and HRNet) is not excluded.

[0072] It should be understood that various forms of the processes shown above can be used, steps can be reordered, added or deleted. For example, the steps described in the disclosure of the present invention can be executed in parallel, sequentially or in a different order, as long as the desired results of the technical solution disclosed in the present invention can be achieved. No limitation is made herein.

[0073] The above specific embodiments do not constitute a limitation to the protection scope of the present invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions and improvements made within the spirit and principle of the present invention shall be included within the protection scope of the present invention.

Claims

1. An underground target imaging method based on variable-scale deep learning, characterized in that It includes the following steps: S1: Establish a multi-scale magnetic measurement data set, including magnetic anomaly data of underground targets with different sizes. The multi-scale magnetic measurement data set is divided into a training set, a validation set, and a test set. The data of the multi-scale magnetic measurement data set is obtained based on the magnetic anomaly forward calculation formula; S2: Construct a variable-scale deep learning network model based on UNet++. The model includes a dynamic information fusion layer, which 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 the information loss caused by pooling; S3: Use a multi-constraint loss function to train the variable-scale deep learning network. The multi-constraint loss function is composed of a weighted numerical error term and a spatial consistency evaluation term; S4: Adopt transfer learning. Use the parameters of the trained training set as the initial weights. By freezing the encoder parameters and fine-tuning the decoder parameters, progressive training of multi-size data is achieved. After the training is completed, use the validation set to verify the effect. Finally, according to the prediction effect of the test set, verify the effectiveness of the entire variable-scale deep learning network model.

2. The underground target imaging method based on variable-scale deep learning according to claim 1, characterized in that The multi-scale magnetic measurement data set contains data with sizes of 16×16, 32×32, 64×64, and 100×100 respectively. Initially, 30,000 data sets of size 32×32 are used, and the division ratio is training set:test set:validation set = 6:2:2; The magnetic anomaly forward calculation formula is: （1）； In the formula, μ0 is the magnetic permeability in vacuum, the unit of μ0 is H / m, k is the magnetic susceptibility, H is the magnetic field strength of the magnetized medium, the unit of H is A / m, ξ, η, and ζ are the endpoint coordinates of the cube in the X, Y, and Z directions respectively, and the unit of the coordinates is meters. The meaning of is the weighting coefficient of the forward calculation. The weighting coefficient expresses the influence of the field source at the underground position (ξ, η, ζ) on the magnetic field strength at the observation point (x, y, z) on the ground. The meaning of is the triple integral for the weighting coefficient.

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 a weighted mean square error and a 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 region and the true region. The coefficient of the Dice loss function is defined as: （4）； T is the overall predicted value, and P is the overall true value. is the predicted value of a single data point. is the true value of a single data point. is the balance coefficient used to control the weights of the mean squared error and the loss function.

4. The underground target imaging method based on variable-scale deep learning according to claim 1, characterized in that, The adaptive adjustment module is an adaptive pooling layer. The adaptive pooling layer is responsible for dynamically adjusting the input feature map to the target size without relying on the size of a fixed pooling window. The cross-scale feature alignment module includes a split convolution module; The dynamic adjustment formula of the input feature map of the adaptive pooling layer is: （5）； Among them, the adaptive pooling layer converts the input feature map into an output feature map , is the output of the c-th channel, is the input of the c-th channel, and are the coordinates of the output feature map, and their value ranges are 1 and 1 , and are the sizes of the input area in height and width respectively, is the output feature map of the c-th channel, is the input feature map of the c-th channel.

5. A method for underground target imaging based on variable-scale deep learning according to claim 1, characterized in that Balance coefficient is 0.

5.

6. A method for underground target imaging based on variable-scale deep learning according to claim 1, characterized in that It is characterized in that In step S2: The variable-scale deep learning network uses the Adam optimizer to train the training set. The initial learning rate is set to 0.001, the batch size is 16, and the learning rate decays by 50% every 20 training epochs; Monitor the validation set loss through the early stopping mechanism. If the loss does not decrease for 5 consecutive epochs, the training is terminated.

7. A method for underground target imaging based on variable-scale deep learning according to claim 1, characterized in that Step S4 includes the following steps: S41: Use the model parameters pre-trained on the 32x32 fixed-size data set as the initial weights, and use transfer learning to fine-tune the variable-scale deep learning network. The fine-tuning data set includes images of sizes 64x64 and 100x100; S42: In the fine-tuning stage, freeze the convolutional layer parameters of the encoder part, and only update the parameters of the decoder and the dynamic information fusion layer, so that the variable-scale deep learning model can retain the underlying feature extraction ability while adaptively learning the spatial context information of large-size data, and achieve progressive training of multi-size data.

8. 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 1-7.

Images