Power transmission and transformation project tower footing recessive disease GPR image recognition method based on improved Faster R-CNN

Abstract

The invention belongs to the field of image processing and convolutional neural networks, and particularly provides a power transmission and transformation project tower footing recessive disease GPR image recognition method based on an improved Faster R-CNN, and the method comprises the steps: carrying out a series of forward simulation on the recessive disease of a tower footing based on the actual measurement data of a rigid straight column type foundation; processing the GPR image through a generative adversarial network to expand the data set; a traditional Faster R-CNN model is optimized, ResNet-50 is adopted as a feature extraction backbone network, an attention mechanism is combined, and a soft-NMS algorithm is adopted to improve the detection capability of closely connected targets; and carrying out deep learning training on underground hidden defect samples by using the preprocessed data, and testing the recognition capability of the optimized Faster R-CNN model on hidden diseases. According to the method, potential concrete diseases can be more effectively identified and positioned, the detection efficiency and accuracy of recessive disease characteristics are improved, and the method has practical application significance.

Description

Technical Field

[0001] The present invention belongs to the fields of image processing and convolutional neural networks, and specifically relates to a method for identifying hidden diseases in GPR images of transmission and transformation project tower bases based on improved Faster R-CNN. Background Art

[0002] Overhead lines are important facilities for long-distance power transmission in the power system, with the advantages of rapid construction, wide applicability, low cost, strong bearing capacity, and easy fault location. However, due to long-term exposure, they are vulnerable to natural erosion, which may cause damage to components such as conductors and towers, increasing the risk of faults. Especially for reinforced concrete structures, due to problems such as non-compliance of the mix ratio and incomplete mixing during construction, the strength of the concrete drops severely. At the same time, tower bases will also gradually deteriorate under the action of environmental factors such as crustal movement and erosion in the soil, affecting their performance. Therefore, regular inspection and maintenance work is particularly necessary.

[0003] Reinforced concrete tower bases are widely used in transmission lines due to their high load-bearing capacity and rigidity. To ensure the functionality and reliability of transmission lines, tower bases need to prevent problems such as settlement or cracking during normal use. In addition, tower bases also need to have sufficient stability to resist external forces and avoid abnormal working states of the towers. Therefore, tower bases are not only the structural foundation for supporting towers but also the key elements to ensure the safe and stable operation of transmission lines.

[0004] Traditional detection methods such as mechanical drilling, electromagnetic method, and acoustic wave method not only cause structural damage but also have problems such as low efficiency and low accuracy of image recognition. Ground Penetrating Radar (GPR) is a non-destructive geophysical detection technology that uses electromagnetic waves to perform real-time imaging of underground structures, providing a solution for underground shallow health monitoring. However, GPR signals are not intuitive and are blurred in terms of recognition, and its performance is also easily restricted by the environment. In addition, in addition to relying on manual analysis of disease characteristics in GPR images, this method is also limited by the subjective evaluation of analysts, which may lead to low analysis speed and accuracy. To improve the efficiency of disease recognition and reduce the consumption of human resources, many scholars have devoted themselves to developing and improving the application errors of GPR in detecting underground hidden diseases. These studies cover many aspects from algorithm optimization to data processing, aiming to improve the automation level and recognition accuracy of radar equipment.

[0005] In the field of image processing, traditional methods for detecting diseases usually use machine vision to analyze the B-Scan images of GPR, and identify characteristic information such as hyperbolic reflections in the images. However, these methods rely on manually labeled image data, cannot meet the requirements of real-time processing, and are easily affected by factors such as noise and light, which affect the recognition effect. In addition, the modern machine vision field is gradually integrating cutting-edge technologies such as deep learning. By using Convolutional Neural Network (CNN) and other advanced algorithms, image data can be better processed, the ability of feature extraction can be improved, and the generalization and real-time processing capabilities of the system can be enhanced. However, the ability of deep learning models in object recognition depends to a large extent on the training dataset. Given the complexity and cumbersome nature of using GPR equipment to verify detection results, as well as the high cost involved in obtaining prior information about detection targets, there are significant challenges in constructing a dataset that meets the requirements. This problem has led to the lack of a comprehensive dataset for underground diseases in current open-source data platforms. Summary of the Invention

[0006] The technical problem to be solved by the present invention is to provide a method for identifying GPR images of hidden diseases in the tower bases of power transmission and transformation projects based on improved Faster R-CNN, which can more effectively identify and locate potential concrete diseases, and improve the detection efficiency and accuracy of hidden disease features.

[0007] To solve the above technical problems, the technical solution adopted by the present invention is: a method for identifying GPR images of hidden diseases in the tower bases of power transmission and transformation projects based on improved Faster R-CNN, comprising the following steps: Step 1: Data acquisition: Based on the actual measurement data of rigid straight-column foundations, a series of forward simulation simulations of hidden diseases in tower bases are carried out to obtain GPR images; real GPR image data of rigid straight-column foundations are collected through on-site ground penetrating radar equipment; Step 2: Data preprocessing is used to expand the dataset.

[0008] Step 3: Optimize the traditional Faster R-CNN model; Step 4: Use the preprocessed data in Step 2 to perform deep learning training on underground hidden defect samples, and test the recognition ability of the optimized Faster R-CNN model for hidden diseases.

[0009] In a preferred solution, in the second step, the GPR images obtained by simulation are processed through a generative adversarial network, and real GPR image data are collected through on-site ground penetrating radar equipment. The adversarial loss between the generator and the discriminator is minimized to optimize the network parameters.

[0010] In the preferred solution, in step three, optimizing the traditional Faster R-CNN model includes: 1) Improvement of the feature extraction network: Select ResNet-50 as the basic network model and combine it with the attention mechanism; 2) Improvement of soft non-maximum suppression: Use the Soft-NMS algorithm to evaluate the scores of candidate boxes in the Faster R-CNN framework and select the candidate box with the highest score. The expression is as follows: (1); In the above formula, represents the candidate box with the highest score; is the candidate box to be evaluated; if and the overlap degree is less than the threshold then retain 's detection score. Otherwise, decay the score of ; σ represents the parameter controlling the decay speed; D represents the set storing the final detection box results; represents the candidate box score.

[0011] In the improvement of the feature extraction network in the preferred solution, embed the SE-Net module after each convolutional layer of ResNet-50, add the SE module at each level of the network and the SE module is located in front of the residual block.

[0012] In the preferred solution, in step four, the training steps are as follows: 1) Initialize the RPN network: Independently train the region proposal network, use the ImageNet pre-trained model as the starting point of the RPN network, and fine-tune it to generate candidate bounding boxes; 2) Train the Faster R-CNN network: Independently train the Faster R-CNN network, use the ImageNet pre-trained model, and use the candidate boxes of the RPN to train the detection network; 3) Fine-tune the RPN parameters: Use the parameters of the Faster R-CNN network to initialize the RPN, keep the shared convolutional layer unchanged, and only adjust the parameters of the RPN; 4) Optimize the detection network: Keep the shared CNN network unchanged, use the candidate boxes optimized by the RPN, and perform separate fine-tuning on the detection network.

[0013] In the preferred solution, in step four, use the mean average precision Quantitatively judge the model performance by or / and mean average precision mAP or / and precision P or / and recall rate R, and combine the number of frames transmitted per second to reflect the comprehensive performance of the network model.

[0014] In a preferred solution, the average accuracy The calculation formula is: (2); In the formula, represents the precision-recall rate R curve, which shows the precision of the model at different recall rate levels; represents the integral of from r = 0 to r = 1.

[0015] In a preferred solution, the calculation formula of the mean average precision mAP is: (3); In the formula, represents the average accuracy value of the nth category , N is the total number of categories, represents the sum of the average accuracy values of the 1st to Nth categories sum.

[0016] In a preferred solution, the calculation formula of the precision P is: (4); In the formula, represents the number of samples correctly predicted as positive by the model; represents the number of samples wrongly predicted as positive by the model.

[0017] In a preferred solution, the calculation formula of the recall rate R is: (5); In the formula, represents the number of samples correctly predicted as positive by the model; represents the number of positive category samples wrongly predicted as negative by the model.

[0018] A method for identifying hidden diseases in GPR images of transmission and transformation project tower bases based on improved Faster R-CNN provided by the present invention has the following beneficial effects: 1. An improved Faster R-CNN algorithm model is proposed, which can be used to quickly identify hidden diseases in GPR images of transmission and transformation project tower bases.

[0019] 2. To meet the identification requirements of hidden diseases in the tower bases of power transmission and transformation projects, the existing algorithm is improved by introducing Soft-NMS and channel attention mechanism, and the data set is expanded, so as to more accurately locate and classify disease features when processing GPR images of tower bases.

[0020] 3. The optimized Faster R-CNN model has mAP and F-Score of 88.59% and 83.38% respectively in disease identification, and can effectively distinguish main disease types such as cracks and cavities. Even when disease features overlap or compound diseases occur, the model can still accurately identify them. Brief Description of the Drawings

[0021] The present invention will be further described below in conjunction with the drawings and embodiments: Figure 1 is the Faster R-CNN object detection framework; Figure 2 is the RPN structure diagram; Figure 3 is the forward model of cavity disease; Figure 4 is the forward model of crack disease; Figure 5 is the GAN structure diagram; Figure 6 is the expanded GPR image; Figure 7 is the SE module structure; Figure 8 is the comparison result of error rates when the SE module is added at different levels; Figure 9 is the comparison result of error rates when the SE module is placed in different positions; Figure 10 is the detection result of hidden diseases in actual application; Figure 11 is the test result of three feature extraction networks; Figure 12 is the comparison of object detection models. Detailed Embodiments

[0022] In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0023] A method for identifying GPR images of hidden diseases in tower bases of power transmission and transformation projects based on improved Faster R-CNN includes the following steps: Step 1: Based on the actual measurement data of the rigid straight-column foundation, mainly dimension data, a series of forward simulation and emulation are carried out on the hidden diseases of the tower foundation.

[0024] In this embodiment, the gprMax software is used for emulation.

[0025] The hidden diseases of the concrete of the tower foundation in the power transmission and transformation project generally refer to the internal damages or defects that are not easily directly observable in the concrete structure, such as cracks and cavities, etc. The gprMax software is used to simulate a forward model consistent with the actual tower foundation of the power transmission and transformation tower, and the GPR echo characteristic curves of different diseases are obtained. The model takes into account the actual size and dielectric constant distribution, and simulates the common diseases of the tower foundation. Figure 3 Modeling is carried out for the two different morphological cavity defect diseases of circular and rectangular shapes. Since the models established by the built-in geometric commands of gprMax are regular, such as spheres, cubes, cylinders, triangular prisms, etc., Matlab is combined to generate arbitrary irregular geometric shapes for modeling. Figure 4 Modeling is carried out for the hidden disease of cracks.

[0026] The relevant technical parameters of the ground penetrating radar equipment are shown in Table 1.

[0027]

[0028] A series of forward simulation and emulation are carried out on the hidden diseases of the tower foundation using the gprMax simulation software. First, emulation is carried out based on a single disease model to ensure in-depth analysis of each disease type. On this basis, in this embodiment, a combined model containing multiple diseases is created, and 800 simulation echo images are generated. At the same time, 200 real echo images are collected using the field ground penetrating radar equipment.

[0029] Step 2: Process the GPR images through the Generative Adversarial Networks (GAN) to expand the dataset.

[0030] Enhance the forward simulation B-scan dataset through the GAN network, and its structure is shown in Figure 5 . Figure 5 A system including two core networks, the generator G and the discriminator D, is shown. In the figure, both G and D are based on the CNN network architecture. G receives a random noise vector as input and generates a simulated GPR target image . Relatively, D receives these synthetic two-dimensional images and the real GPR two-dimensional images simulated using the gprMax software as input, and its task is to distinguish the synthetic data from the real data , and evaluate the probability that the synthetic data matches the real data. The core goal of model training is to make the authenticity of the synthetic data as close to the real situation as possible. Among them, the discriminator labels the real data as 1 and the forged data as 0. The objective function used in the training process aims to optimize this process. The objective function of model training is: ; In the above formula: is the average probability of the real data; is the average probability that the synthetic data is judged as real. The goal of training the model is to increase the probability of identifying real data and reduce the misjudgment probability, so as to achieve the maximum recognition rate , minimize the rate , and conform to the images in the actual detection situation as much as possible to obtain different disease detection data sets.

[0031] Use the trained GAN to expand the data set and combine it with the actually measured GPR map. Before feeding the target map into the network, crop the GPR map to ensure that the side length does not exceed 1000, and label all GPR maps. The ground penetrating radar B-Scan images are divided into three parts: training, validation and testing, with a ratio of 8:1:1. For the specific allocation, see Table 2.

[0032]

[0033] Step 3: Optimize the traditional Faster R-CNN model.

[0034] The Faster R-CNN object detection framework takes the ground penetrating radar B-Scan map of hidden diseases as input and then performs convolutional processing. Along with the increase of the network levels, the model gradually extracts the feature information of the hidden diseases in the concrete of the transmission and transformation project tower foundation. The whole network consists of four main parts, working together to achieve the effective recognition of disease features.

[0035] The optimization of the traditional Faster R-CNN model includes: 1) Improvement of the feature extraction network: Select ResNet-50 as the basic network model and combine it with the attention mechanism.

[0036] Deep neural networks play a role in feature extraction in image processing architectures, and their recognition accuracy often improves with the increase in the number of network layers. However, there is no linear positive correlation between network depth and recognition accuracy. When the network depth exceeds a certain threshold, it may lead to a decrease in the efficiency of model training and affect the detection speed of the model. In the current research, VGG19, ResNet-50, and InceptionNet-V3 are several models widely used for image feature extraction. The number of network layers and the number of pre-trained parameters of the three network models are shown in Table 3 below.

[0037]

[0038] Compared with VGG19 and InceptionNet-V3, the ResNet-50 model uses its residual structure to enhance the convergence rate of the network. Based on this advantage, the present invention selects ResNet-50 as the basic network model. This selection is to utilize its fast convergence characteristics to improve the research efficiency and performance.

[0039] To improve the detection ability of diseases, a channel attention mechanism, abbreviated as SE-Net, is integrated into the feature extraction network. The purpose of this integration is to enhance the efficiency of the model in identifying disease features. As Figure 7 shown, the SE-Net evaluates the importance of each channel feature by weighting the feature channels. This structural design helps the model to more accurately obtain key disease features.

[0040] SE-Net incorporates an attention mechanism by assigning corresponding weights to each channel to reflect their key roles in subsequent processing. At the same time, it is made into an insertable module, which is very convenient for combination with various basic networks. Using the SE-Net module, not only can the interaction connections between channels be recognized, but also the significance of each channel feature can be automatically learned and evaluated. This mechanism helps to suppress those channels with less information and enhance the model's attention to information-rich channels, thereby improving the quality of feature expression.

[0041] By conducting comparative experiments on adding the SE module at different levels of the network and placing the SE module in different positions, the optimal position of the SE module in the model is determined, so that the model can reduce the redundant computational burden on the basis of extracting key information in the feature map.

[0042] Conduct experiments on adding the SE module at different levels of the network and placing the SE module in different positions, and compare the evaluation metrics of Top-1 error rate and Top-5 error rate, as Figure 8 and Figure 9As shown in the figure, taking ResNet as the experimental object, it is found that adding SE modules at each layer and placing the SE module in front of the residual block gives the best performance. After comprehensive consideration, this study selects to embed the SE-Net module after each convolutional layer of ResNet-50. In addition, since the SE-Net is integrated before the residual block and will be trained end-to-end with the entire CNN, the performance cost is negligible.

[0043] 2) Improvement of Soft-Non-Maximum Suppression: In the Faster R-CNN framework, the non-maximum suppression algorithm (NMS) is used to evaluate the scores of candidate boxes and select the candidate box with the highest score. However, in the disease detection of GPR B-Scan images, due to the possible close proximity of targets, the traditional NMS algorithm may lead to missed detections.

[0044] To solve this problem, the present invention adopts the Soft-NMS method. Different from the traditional NMS that directly removes overlapping candidate boxes, Soft-NMS deals with it by reducing the scores of overlapping boxes. This method reduces false positives while ensuring the effective identification of disease targets in close positions.

[0045] Generally, Soft-NMS uses Gaussian and linear functions to adjust the box scores. In this embodiment, a linear function is selected for the experiment to improve the model's ability to identify adjacent disease targets. The expression is as follows: ; In the above formula, represents the candidate box with the highest score; is the candidate box to be evaluated; if and the overlap degree between is less than the threshold , then the detection score of is retained. Otherwise, the score of is attenuated; σ represents a parameter that controls the attenuation speed, D represents the set storing the final detection box results, represents the candidate box score.

[0046] Step 4: Use the data preprocessed in Step 2 to perform deep learning training on the underground hidden defect samples, and test the recognition ability of the optimized Faster R-CNN model for invisible diseases.

[0047] Based on frameworks such as TensorFlow and Keras, Windows 10 Professional Edition is selected as the experimental operating system to perform deep learning training on the underground hidden defect samples. The initial learning rate of the experiment is 0.001.

[0048] Meanwhile, a learning rate decay strategy is used. During the training process, the learning rate is multiplied by 0.005 every 2000 iterations to gradually reduce the learning rate, thereby improving the convergence speed and performance of the model. On this basis, the threshold is set to 0.45 to determine whether the bounding box contains disease information. If the IoU value of a certain bounding box exceeds 0.45, it is considered that it contains underground disease-related information; if not, the area is regarded as not containing relevant information and should be excluded.

[0049] The training steps are as follows: 1) Initialize the RPN network: Independently train the region proposal network. Use the ImageNet pre-trained model as the starting point of the RPN network and fine-tune it to generate candidate bounding boxes; 2) Train the Faster R-CNN network: Independently train the Faster R-CNN network. Use the ImageNet pre-trained model and use the candidate boxes of the RPN to train the detection network; 3) Fine-tune the RPN parameters: Use the parameters of the Faster R-CNN network to initialize the RPN, keep the shared convolutional layer unchanged, and only adjust the parameters of the RPN; 4) Optimize the detection network: Keep the shared CNN network unchanged, use the candidate boxes optimized by the RPN, and perform separate fine-tuning on the detection network.

[0050] The average accuracy or / and the mean average precision mAP or / and the precision P or / and the recall rate R are used to quantitatively judge the performance of the model, and the frames per second (FPS) is combined to reflect the comprehensive performance of the network model.

[0051] The average accuracy The calculation formula is: ; In the formula, represents the precision P-recall rate R curve, which shows the precision of the model at different recall rate levels; represents the integral of from r = 0 to r = 1.

[0052] The calculation formula of the mean average precision mAP is: ; In the formula, represents the average accuracy value of the nth category , N is the total number of categories, represents the average accuracy value from the 1st to the Nth category The sum.

[0053] The calculation formula for the precision P is as follows: ; In the formula, represents the number of samples correctly predicted as positive by the model; represents the number of samples incorrectly predicted as positive by the model.

[0054] The calculation formula for the recall rate R is as follows: ; In the formula, represents the number of samples correctly predicted as positive by the model; represents the number of positive-class samples incorrectly predicted as negative by the model.

[0055] F-Score is the harmonic mean of precision and recall, used to measure the performance of a classification model. It is particularly suitable for application scenarios that require a balance between precision and recall, such as information retrieval, text classification, and disease diagnosis. It is a comprehensive indicator for evaluating the accuracy of the model in disease recognition, and the expression is as follows: ; In the formula, is a parameter used to adjust the weights of P and R. P represents the precision value, that is, the proportion of actually positive samples among all samples predicted as positive. R represents the recall rate value, that is, the proportion of samples predicted as positive among all actually positive samples.

[0056] In the improved Faster R-CNN object detection algorithm, the loss function mainly includes two parts: classification loss and bounding box regression loss. The classification loss is used to measure the matching degree between the category predicted by the model and the true label, and the bounding box regression loss is used to measure the difference between the predicted bounding box and the true bounding box. The formula is as follows: ; ; ; ; In the above formula, represents the difference between the predicted value of the model and the true value , is the probability that the model predicts the true category , is the classification loss function, usually represents the category probability distribution predicted by the model, represents the true category, represents the regression loss function, represents the regression value predicted by the model, represents the true value, represents the logarithmic function, represents the smooth transition of the loss function from L2 Loss to L1 Loss, represents the difference between the regression value predicted by the model and the true value.

[0057] The above formula is the loss function, which is a function used in machine learning and deep learning to measure the difference between the predicted value of the model and the true value. It is used to guide the optimization of the model during training. By minimizing the value of the loss function, the parameters of the model are adjusted to improve the performance of the model.

[0058] To verify the effectiveness of the proposed model in identifying hidden diseases of transmission and transformation project tower bases, this embodiment uses typical disease samples in GPR images and applies the improved Faster R-CNN algorithm for identification. The research is based on the model pre-trained on the simulation dataset for further training, and 2000 iterations are set to ensure that the model can fully learn and optimize.

[0059] The trained model shows good feature extraction ability in the detection of actual disease GPR images, such as Figure 10 shown. It can be seen from the detection results that although the measured image is closer to the real situation and the recognition difficulty increases compared with the simulation results, this model can still accurately identify different disease features.

[0060] To evaluate the performance of the model of the present invention, ablation experiments are carried out, comparing the traditional model with the models integrated with different improvement methods such as Soft-NMS, SE-Net attention mechanism, and GAN, as well as the comparison of different permutations and combinations of the above 3 improvement methods.

[0061] To comprehensively evaluate the performance of the proposed model, the traditional model is compared with the models integrated with different improvement methods such as Soft-NMS, SE-Net attention mechanism, and GAN, as well as the comparison of different permutations and combinations of the above 3 improvement methods. Table 4 shows the comparison results of the Faster R-CNN structures with different improvement combinations. It can be obtained from Table 4 that the improved Fater R-CNN has a high accuracy, and its evaluation indexes mAP and F-Score are 88.59% and 83.38% respectively.

[0062]

[0063] The detection performance comparison test of the improved model and SSD300 and Yolo V5 was carried out on the same data. When the IoU value was set to 0.45, the recognition efficiency of different models was compared and analyzed, and the results were summarized in Table 5. Figure 12 The results in Figure 12 show that the improved model performs better in terms of both mAP and F-Score, and exhibits good performance stability in multiple tests.

[0064]

[0065] After verification, the model can accurately identify the underground disease characteristics in the image, and the GPR target detection results are actually verified, which also proves the robustness of the model. At the same time, it shows a relatively fast recognition speed during the disease recognition process, with an average time of 0.081 seconds. Different from traditional machine learning algorithms, this model can effectively detect multiple overlapping targets and provide reliable recognition conclusions.

[0066] It is easy for those skilled in the art to understand that the above are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements, and improvements made within the spirit and principle of the present invention should be included within the protection scope of the present invention.

Claims

1. A GPR image recognition method for hidden defects of tower foundations of power transmission and transformation projects based on improved Faster R-CNN, characterized in that: The following steps are involved: Step 1: Data acquisition: Based on the actual measurement data of the rigid column foundation, a series of forward simulations are performed on the hidden diseases of the tower foundation to obtain GPR images; the real GPR image data of the rigid column foundation is collected through the field ground penetrating radar equipment; Step 2: Data preprocessing to expand the data set. Step 3: Optimize the traditional Faster R-CNN model; Step 4: Use the preprocessed data from step 2 to conduct deep learning training on underground hidden defect samples, and test the ability of the optimized Faster R-CNN model to identify invisible diseases.

2. According to claim 1, a GPR image recognition method for hidden defects of tower foundation of power transmission and transformation engineering based on improved Faster R-CNN is characterized in that: In the step 2, the GPR image obtained by simulation is processed by a generative adversarial network, and the real GPR image data is collected by a field ground penetrating radar device, and the network parameters are optimized by minimizing the adversarial loss between the generator and the discriminator.

3. According to claim 1, a GPR image recognition method for hidden defects of tower foundation of power transmission and transformation engineering based on improved Faster R-CNN is characterized in that: In step 3, optimizing the traditional Faster R-CNN model includes: 1) Feature extraction network improvement: ResNet-50 is selected as the basic network model and combined with the attention mechanism; 2) Soft non-maximum suppression improvement: The Soft-NMS algorithm is used to evaluate the scores of the candidate boxes in the Faster R-CNN framework and filter out the candidate boxes with the highest scores. The expression is as follows: （1）； In the above formula, Represents the candidate box with the highest score; is the candidate box to be evaluated; if and The overlap between Less than threshold When The detection score of The fraction is attenuated; σ Represents the parameter that controls the decay speed; D represents the set that stores the final detection box results; Represents the candidate box score.

4. According to claim 3, a GPR image recognition method for hidden defects of tower foundation of power transmission and transformation engineering based on improved Faster R-CNN is characterized in that: In the feature extraction network improvement, the SE-Net module is embedded after each convolutional layer of ResNet-50, and the SE module is added to each level of the network and is located before the residual block.

5. According to claim 1, a GPR image recognition method for hidden defects of tower foundation of power transmission and transformation engineering based on improved Faster R-CNN is characterized in that: In step 4, the training steps are as follows: 1) Initialize the RPN network: Train the region generation network independently, use the ImageNet pre-trained model as the starting point of the RPN network, and fine-tune it to generate candidate bounding boxes; 2) Train the Faster R-CNN network: Train the Faster R-CNN network independently, use the ImageNet pre-trained model, and use the RPN candidate box to train the detection network; 3) Fine-tune RPN parameters: Use the parameters of the Faster R-CNN network to initialize the RPN, keep the shared convolutional layer unchanged, and only adjust the parameters of the RPN; 4) Optimize the detection network: Keep the shared CNN network unchanged, use the candidate boxes optimized by RPN, and make separate fine adjustments to the detection network.

6. According to claim 1, a GPR image recognition method for hidden defects of tower foundation of power transmission and transformation engineering based on improved Faster R-CNN is characterized in that: In step 4, the average accuracy is used Or / and mean average precision (mAP) or / and precision (P) or / and recall (R) are used to quantify the model performance and are combined with the number of frames transmitted per second to reflect the comprehensive performance of the network model.

7. The method for GPR image recognition of hidden defects of tower foundation of power transmission and transformation project based on improved Faster R-CNN according to claim 6 is characterized in that: The average accuracy The calculation formula is: （2）； In the formula, It represents the precision P-recall R curve, which shows the accuracy of the model at different recall levels; Express The integration is performed from r=0 to r=1.

8. The method for GPR image recognition of hidden defects of tower foundation of power transmission and transformation project based on improved Faster R-CNN according to claim 6 is characterized in that: The calculation formula of the mean average precision mAP is: （3）； In the formula, Represents the average accuracy value of the nth category , N is the total number of categories, Represents the average accuracy of the 1st to N categories sum.

9. The method for GPR image recognition of hidden defects of tower foundation of power transmission and transformation project based on improved Faster R-CNN according to claim 6, characterized in that: The calculation formula of the precision P is: （4）； In the formula, Indicates the number of samples correctly predicted by the model as positive categories; Indicates the number of samples that the model incorrectly predicts as the positive class.

10. The method for GPR image recognition of hidden defects of tower foundation of power transmission and transformation project based on improved Faster R-CNN according to claim 6, characterized in that: The calculation formula of the recall rate R is: （5）； In the formula, Indicates the number of samples correctly predicted by the model as positive categories; Indicates the number of positive class samples that the model incorrectly predicts as negative classes.

Images