Subway tunnel leaky coaxial cable fault image intelligent identification method and system

Abstract

The invention provides a subway tunnel leaky coaxial cable fault image intelligent identification method and system, and the method comprises the steps: carrying out the preprocessing of an original image through an image processing technology, extracting useful feature information, carrying out the feature extraction and classification identification of the preprocessed image through a trained convolutional neural network model, and carrying out the recognition of a fault image. And finally, outputting fault type and position information according to an identification result, thereby realizing a technical effect of intelligently identifying the fault image of the subway tunnel leaky coaxial cable by using the convolutional neural network. According to the invention, the fault condition of the leaky coaxial cable in the subway tunnel can be rapidly and accurately identified.

Description

Technical Field

[0001] The present invention relates to image processing technology, and more particularly to an intelligent recognition method and system for leakage coaxial cable fault images in subway tunnels. Background Art

[0002] The leakage coaxial cable in the subway tunnel is an important device for signal transmission, and its operating state directly affects the communication and signal transmission of subway trains. However, due to the complex environment of the subway tunnel, the cable is vulnerable to the influence of factors such as humidity and corrosion, resulting in frequent failures.

[0003] Traditional fault detection methods mainly rely on manual inspection. This method is not only inefficient but also difficult to ensure the accuracy and timeliness of detection. With the development of image processing technology and deep learning algorithms, automated and intelligent fault recognition methods have gradually become a research hotspot. Summary of the Invention

[0004] The present invention aims to provide an intelligent recognition method and system for leakage coaxial cable fault images in subway tunnels. The method first preprocesses the original image through image processing technology to extract useful feature information, and then uses a trained convolutional neural network model to extract features and classify and recognize the preprocessed image. Finally, the fault type and location information are output according to the recognition result, so as to achieve the technical effect of using a convolutional neural network to intelligently recognize the fault images of leakage coaxial cables in subway tunnels.

[0005] To achieve the above object, the present invention provides an intelligent recognition method for leakage coaxial cable fault images in subway tunnels, including: Step S1: Preprocess the collected image to be detected; Step S2: Construct a convolutional neural network model, collect preprocessed image samples of subway leakage coaxial cables in various types and states, and input them into the model for training to enable it to learn feature extraction to obtain a mature convolutional neural network model, and input the preprocessed image to be detected into the model to output a feature map; Step S3: Design and train a mature softmax classifier including a fully connected layer and a softmax layer, and use it to input the feature map into the softmax classifier to classify and recognize the feature map. The classification and recognition first divide the pixels in the feature map into faulty pixels and non-faulty pixels, mark the pixel value of the faulty pixels as 1, and mark the pixel value of the non-faulty pixels as 0 to obtain a binary image; Step S4: Traverse the pixel values in the binary image, check the adjacent pixels of each faulty pixel, identify the pixels with the pixel value marked as 1 as part of the faulty area, and the pixels with the pixel value marked as 0 as the boundary of the faulty area. Then, mark the boundary of the faulty area in the original image using different colors or lines.

[0006] Further, the preprocessing in step S1 includes: removing noise in the image to be detected using the Gaussian filtering algorithm, grayscale conversion of the image, and binarization of the image.

[0007] Further, the specific steps for training the convolutional neural network model in step S2 include: Enhancing the training data using operations such as rotation, flipping, scaling, and translation; Continuously adjusting the weight parameters using the backpropagation algorithm to minimize the loss function; Evaluating the performance of the model using the recall rate, optimizing the model according to the evaluation results, and using strategies such as early stopping and learning rate decay to prevent overfitting.

[0008] Further, step S3 includes: further subdividing the fault types for the feature map.

[0009] Further, the steps for training the softmax classifier in step S3 specifically include: Calculating the difference between the predicted probability distribution and the true label using the cross-entropy loss function; Calculating the gradient through the backpropagation algorithm and updating the network parameters using the gradient descent algorithm; Evaluating the performance of the softmax classifier using the validation set, including accuracy and recall rate metrics; Adjusting the network structure and learning rate parameters according to the evaluation results to improve the classification performance.

[0010] Further, step S4 includes: processing the boundary by interpolation or smoothing filtering.

[0011] Further, step S4 includes: adopting a method of block processing and parallel computing to improve the computing speed.

[0012] An intelligent recognition system for leakage coaxial cable fault images in subway tunnels, comprising: An image preprocessing module for preprocessing the image to be detected collected. A feature extraction module, which is used to build a convolutional neural network model, collect preprocessed image samples of subway leaky coaxial cables in various types and states, and input them into the model for training, enabling it to learn feature extraction to obtain a mature convolutional neural network model, and input the preprocessed image to be detected into the model to output a feature map; A classification and recognition module, which is used to design and train a mature softmax classifier including a fully connected layer and a softmax layer, and is used to input the feature map into the softmax classifier to classify and recognize the feature map. The classification and recognition first divides the pixels in the feature map into faulty pixels and non-faulty pixels, marks the pixel values of the faulty pixels as 1, and marks the pixel values of the non-faulty pixels as 0 to obtain a binary image; A fault location module, which is used to traverse the pixel values in the binary image, check the adjacent pixels of each faulty pixel, identify the pixels with the pixel value marked as 1 as part of the fault area, identify the pixels with the pixel value marked as 0 as the boundary of the fault area, and then use different colors or lines to mark the boundary of the fault area in the original image.

[0013] In summary, compared with the prior art, the present invention has the following beneficial effects: The present invention uses automated and intelligent means: preprocesses the collected image to be detected, constructs and trains a mature convolutional neural network model capable of feature extraction, inputs the preprocessed image to be detected into the model to output a feature map, then designs and trains a mature softmax classifier including a fully connected layer and a softmax layer, and inputs the feature map into the softmax classifier to classify and recognize the feature map to obtain a binary image. Then, it traverses the pixel values in the binary image, checks the adjacent pixels of each faulty pixel, identifies the pixels with the pixel value marked as 1 as part of the fault area, identifies the pixels with the pixel value marked as 0 as the boundary of the fault area, and finally uses different colors or lines to mark the boundary of the fault area in the original image, thereby being able to quickly and accurately identify the fault condition of the leaky coaxial cable in the subway tunnel. Description of the Drawings

[0014] Figure 1 It is a flowchart of an intelligent recognition method for fault images of leaky coaxial cables in subway tunnels according to the present invention.

[0015] Figure 2 It is a module diagram of an intelligent recognition system for fault images of leaky coaxial cables in subway tunnels. Detailed Embodiments

[0016] The technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those skilled in the art without creative efforts belong to the scope of protection of the present application.

[0017] The following disclosure provides different embodiments or examples for implementing different structures of the present application. To simplify the disclosure of the present application, the components and settings of specific examples are described below. Of course, they are only examples and are not intended to limit the present application. In addition, the present application may repeat reference numerals and / or reference letters in different examples. This repetition is for the purpose of simplification and clarity, and does not itself indicate the relationship between the various embodiments and / or settings discussed. The following provides a detailed description of an intelligent method for identifying images of leakage coaxial cable faults in a subway tunnel provided by the present invention. It should be noted that the description order of the following embodiments does not limit the preferred order of the embodiments.

[0018] Please refer to Figure 1 , an intelligent method for identifying images of leakage coaxial cable faults in a subway tunnel provided by the present invention, includes: Step S1: Preprocess the collected image to be detected; Specifically, the preprocessing in step S1 includes: using the Gaussian filtering algorithm to remove the noise in the image to be detected, grayscale the image, and binarize the image. After denoising and grayscaling, binarizing the grayscale image can facilitate subsequent feature extraction.

[0019] First, the Gaussian filtering algorithm can remove the noise in the image. The Gaussian filtering algorithm replaces the value of each pixel with the weighted average value of the pixels in the surrounding neighborhood of the pixel through weighted averaging, and the weight is determined by the Gaussian function. The value of the Gaussian function is the largest at the central pixel, and as the distance from the central pixel increases, the weight gradually decreases. Therefore, Gaussian filtering can effectively smooth the image while retaining the edge information of the image.

[0020] The specific steps include: setting the Gaussian filter parameters, generating the Gaussian filter kernel, applying the Gaussian filter, and obtaining the denoised image.

[0021] The parameters of the Gaussian filter include the filter size and the standard deviation σ. The size of the filter (also called the kernel) is usually an odd number (such as 3x3, 5x5, 7x7, etc.) because the filter of odd size has a clear central pixel. The standard deviation σ: controls the shape of the Gaussian function and affects the smoothness of the filter. The larger σ is, the smoother the filter is, and the more obvious the image denoising effect is, but it may also cause edge blurring.

[0022] The Gaussian filter kernel is a two-dimensional array, and its element values are calculated by the Gaussian function. The steps to generate the Gaussian filter kernel include: determining the size of the kernel (such as N×N), calculating the center position ((N - 1) / 2, (N - 1) / 2); for each position (i, j) in the kernel, calculating the position difference relative to the center (x = i - (N - 1) / 2, y = j - (N - 1) / 2); substituting the position difference into the Gaussian function to calculate the weight value at that position; normalizing all the weight values so that the sum of all the weight values in the kernel is 1.

[0023] Applying the Gaussian filter means applying the generated Gaussian filter kernel to each pixel in the image, calculating the weighted average value to obtain the denoised image. The specific steps are as follows: For each pixel (x, y) in the image: determine the neighborhood range of this pixel, which is determined by the size of the filter. Cover the filter kernel on the neighborhood, multiply each weight in the kernel by the pixel value at the corresponding position. Calculate the sum of all the products to obtain the weighted average value. Assign the weighted average value to the current pixel (x, y).

[0024] Finally, the obtained denoised image includes: traversing all the pixels in the image, after applying the above steps, the denoised image is obtained.

[0025] Image grayscale conversion is the process of converting a color image into a grayscale image. A grayscale image only contains luminance information and does not contain color information. Grayscale conversion is a common step in image processing, which helps to reduce the data volume and computational complexity.

[0026] Specific grayscale conversion methods include: 1. Component method: Select a certain component (such as one of R, G, B) of the color image as the grayscale value.

[0027] 2. Maximum value method: Select the maximum value among the three components (R, G, B) of the color image as the grayscale value.

[0028] 3. Average value method: Calculate the average value of the three components of the color image as the grayscale value.

[0029] 4. Weighted average value method: According to the sensitivity of the human eye to different colors, assign different weight values to the three components of R, G, B, and calculate the weighted average value as the grayscale value.

[0030] The specific implementation steps of grayscale conversion include: 1. Read the color image and obtain its three components of R, G, B.

[0031] 2. According to the selected grayscale conversion method, calculate the grayscale value of each pixel point.

[0032] 3. Assign the calculated grayscale value to the corresponding pixel points to generate a grayscale image.

[0033] When performing grayscale processing, attention should be paid to the selection of the grayscale method according to the actual image and application requirements. Among them, the weighted average method usually can obtain a more reasonable grayscale image because it takes into account the sensitivity of the human eye to different colors.

[0034] In image binarization, the global threshold method is used for binarization. By setting a global threshold T, all pixel points in the image with grayscale values greater than T are set to white (or a specific high-brightness value, such as 255), while pixel points with grayscale values less than or equal to T are set to black (or a specific low-brightness value, such as 0).

[0035] By performing binarization processing on the grayscale image, it is convenient for subsequent feature extraction.

[0036] The specific framework structure of image binarization is shown in Figure x. It should be noted that: 1. Check the iteration stop condition: Compare the new threshold T_next with the threshold T of the previous iteration. If the absolute value of the difference between them is less than a predefined tolerance ΔT (i.e., |T_next - T| < ΔT), the iteration stops, and the obtained T_next is the global optimal threshold T at this time.

[0037] 2. Update the threshold and repeat: If the iteration does not stop (i.e., |T_next - T| ≥ ΔT), update T to T_next and repeat the steps.

[0038] Step S2: Build a convolutional neural network model, collect preprocessed image samples of subway leaky coaxial cables in various types and states, input them into the model for training, enable it to learn feature extraction to obtain a mature convolutional neural network model, and input the preprocessed image to be detected into the model to output a feature map; Specifically, the specific steps for training the convolutional neural network model in step S2 include: Perform operations such as rotation, flipping, scaling, and translation on the training data for enhancement; Use the backpropagation algorithm to continuously adjust the weight parameters to minimize the loss function; Use the recall rate to evaluate the performance of the model, optimize the model according to the evaluation results, and use the strategies of early stopping and learning rate decay to prevent overfitting.

[0039] Specifically, in the feature extraction step S2, the most important thing is to build and train a CNN model. Considering the particularity of subway leaky coaxial cable feature extraction, it may be necessary to customize the network layer or adjust the network structure.

[0040] By enhancing the training data through operations such as rotation, flipping, scaling, and translation, the generalization ability of the model can be enhanced. By continuously adjusting the weight parameters using the backpropagation algorithm, the loss function can be minimized. The performance of the model is evaluated using the recall rate, and the model is optimized according to the evaluation results. The optimization methods include adjusting the network structure, adding convolutional layers, changing activation functions, etc.

[0041] Among them, before training the model, the steps for building the model include: 1. Design the input layer and select a suitable preprocessed graph.

[0042] 2. Design the convolutional layer and use multiple convolutional kernels to extract local features in the image. As the network depth increases, the convolutional layer can gradually extract higher-level features.

[0043] 3. Design the pooling layer to sample the feature map, reduce the size of the feature map, and retain the most important features.

[0044] 4. Design the fully connected layer to combine the learned features for classification or regression tasks.

[0045] 5. Select the sigmoid activation function to enhance the nonlinear ability of the model.

[0046] After training is completed, a visualization tool can be used to view the feature maps extracted by the convolutional layer to understand how the model extracts and recognizes the features of the subway leaky coaxial cable.

[0047] Step S3: Design and train a mature softmax classifier including a fully connected layer and a softmax layer, and use it to input the feature map into the softmax classifier to classify and recognize the feature map. The classification and recognition first divide the pixels in the feature map into faulty pixels and non-faulty pixels, mark the pixel values of the faulty pixels as 1, and the pixel values of the non-faulty pixels as 0 to obtain a binary image; Specifically, the fully connected layer is used to calculate the class scores, the softmax layer is used to output the probability distribution, and the softmax classifier can set corresponding labels for the non-faulty class and the faulty class.

[0048] Furthermore, step S3 includes: further subdividing the fault types of the feature map. Among them, the fault types include core wire break, break, inner and outer conductor short circuit, etc.

[0049] Specifically, the steps for training the softmax classifier in step S3 specifically include: Use the cross-entropy loss function to calculate the difference between the predicted probability distribution and the true label; Calculate the gradient through the backpropagation algorithm and update the network parameters using the gradient descent algorithm; Evaluate the performance of the softmax classifier using the validation set, including accuracy and recall metrics; Adjust the network structure and learning rate parameters according to the evaluation results to improve the classification performance.

[0050] Among them, cross-entropy refers to a function that measures the difference between two probability distributions. The predicted probability distribution refers to the probability distribution output by the model, which is used to represent the predicted probability of each class. The true label can be regarded as a probability distribution (in binary classification, the true label is 0 or 1; in multi-class classification, the true label is a one-hot encoded vector).

[0051] Among them, the formulas used in the calculation process of the cross-entropy loss function are divided into two types: ① In binary classification problems, assume there are two classes. The true label can be expressed as y ∈ {0, 1}, and the predicted probability of the model is ŷ ∈ [0, 1]. The formula for the cross-entropy loss function is: CrossEntropy = -1 / N ∑_{i=1}^N (y_i log(ŷ_i)) Among them, N represents the number of samples, y_i represents the true label (0 or 1) of the i-th sample, and (ŷ_i) represents the predicted probability of the i-th sample.

[0052] ② In multi-class classification problems, assume there are K classes. The true label can be expressed as y ∈ {0, 1,..., k - 1}, and the predicted probability of the model is ŷ ∈ [0, 1]^k. The formula for the cross-entropy loss function is: CrossEntropy = -1 / N ∑_{i=1}^N ∑_{k=0}^{K - 1} y_{ik} log(ŷ_{ik}) Among them, N represents the number of samples, y_{ik} represents the true label (0 or 1) of the i-th sample belonging to the k-th class, and (ŷ_{ik}) represents the predicted probability of the i-th sample belonging to the k-th class.

[0053] The measurement process is as follows: 1. Calculate the predicted probability: The model makes predictions on the input samples and outputs the predicted probability of each class.

[0054] 2. Determine the true label: Obtain the true class label of the input sample and convert it into the form of a probability distribution (in binary classification, it is directly 0 or 1; in multi-class classification, it is a one-hot encoded vector).

[0055] 3. Calculate the cross-entropy: Substitute the predicted probability distribution and the true label distribution into the above formula to calculate the cross-entropy loss value.

[0056] 4. Evaluate model performance: The smaller the cross - entropy loss value, the smaller the difference between the predicted probability distribution and the true label distribution, that is, the more accurate the model's prediction.

[0057] The specific steps to calculate the gradient through the backpropagation algorithm include: 1. Forward propagation: First, perform forward propagation to calculate the predicted probability distribution and the loss value.

[0058] Input the feature vector x, and after linear transformation, obtain the weighted sum z = Wx + b.

[0059] Apply the Softmax function to convert the weighted sum z into a probability distribution y^ = softmax(z).

[0060] Calculate the cross - entropy loss L = -Σ y_i log(y_i), where y_i is the predicted probability.

[0061] 2. Calculate the gradient of the loss with respect to the predicted probability: The gradient of the cross - entropy loss with respect to the predicted probability y^_i is: ∂L / ∂y_i However, it should be noted that since the outputs y^_i of the Softmax function are mutually related (their sum is 1), we cannot directly use this gradient to update the weights. We need to calculate the gradient of the loss with respect to the weighted sum z, and then obtain the gradient of the loss with respect to the weights and biases through the chain rule.

[0062] 3. Calculate the gradient of the loss with respect to the weighted sum: Using the derivative of the Softmax function, we can obtain the gradient of the loss with respect to the weighted sum z: ∂L / ∂z_i = y^_i - y_i This gradient represents the difference between the predicted probability and the true label.

[0063] 4. Calculate the gradient of the loss with respect to the weights and biases: Finally, through the chain rule, we can obtain the gradient of the loss with respect to the weight W and the bias b.

[0064] Once we have the gradient of the loss function with respect to the parameters, we can use the gradient descent algorithm to update the parameters. The following is the detailed process of updating the parameters using the Stochastic Gradient Descent (SGD) method: 1. Stochastic Gradient Descent (SGD) method: (1) Initialize the learning rate η.

[0065] (2) For each training sample (or batch of samples): Calculate the forward propagation and the loss; Calculate the gradient; Update the weights and biases.

[0066] The Adam algorithm is an adaptive learning rate optimization algorithm that combines the advantages of the momentum method and the RMSProp algorithm. The following are the steps to update the parameters using the Adam algorithm: 1. Initialize the parameters: learning rate η, exponential decay rate β1 for momentum estimation (usually close to 1), exponential decay rate β2 for squared gradient estimation (usually close to 1), and a small constant ε for numerical stability.

[0067] 2. Initialize the first-order momentum vectors m_W and m_b as zero vectors, and the second-order momentum vectors v_W and v_b as zero vectors (or small positive vectors to prevent division by zero).

[0068] 3. For each training sample (or batch of samples): Calculate the forward propagation and loss; Calculate the gradient; Update the first-order momentum; Update the second-order momentum; Calculate the bias-corrected first-order and second-order momenta; Update the weights and biases.

[0069] By repeatedly executing the above process, the parameters of the Softmax classifier will gradually converge to the optimal values, thereby minimizing the loss function and improving the classification accuracy.

[0070] Step S4: Traverse the pixel values in the binary image and check the adjacent pixels of each faulty pixel. Those with a pixel value of 1 are identified as part of the faulty area, and those with a pixel value of 0 are identified as the boundary of the faulty area. Then, use different colors or lines to mark the boundary of the faulty area in the original image.

[0071] Specifically, the accuracy of classification and recognition directly affects the result of fault location. Therefore, the training effect and recognition accuracy of the classifier should be ensured in the previous steps.

[0072] Among them, Step S4 also includes: processing the boundary by interpolation or smoothing filtering methods, and adopting block processing and parallel computing methods to improve the computing speed.

[0073] Specifically, when determining the fault boundary, attention should be paid to the smoothness of the boundary. Processing the boundary by interpolation, smoothing filtering and other methods can improve its smoothness and accuracy, thus avoiding the boundary sawtooth or discontinuity phenomenon caused by pixel-level traversal.

[0074] The method of pixel-level traversal is simple and accurate, but has a large amount of computation. Especially for high-resolution images, traversing the entire image may take a long time. Therefore, in practical applications, the issue of computational efficiency needs to be considered, and the computational speed can be improved by adopting methods such as block processing and parallel computing.

[0075] As Figure 2 shown, the present invention also provides an intelligent recognition system for the fault image of a leaky coaxial cable in a subway tunnel, including: An image preprocessing module 11, configured to preprocess the acquired image to be detected; A feature extraction module 12, configured to build a convolutional neural network model, collect preprocessed image samples of subway leaky coaxial cables of various types and in various states and input them into the model for training, so that it learns feature extraction to obtain a mature convolutional neural network model, and input the preprocessed image to be detected into the model to output a feature map; A classification and recognition module 13, configured to design and train a mature softmax classifier including a fully connected layer and a softmax layer according to the result of feature extraction, and use it to input the feature map into the softmax classifier to classify and recognize the feature map. The classification and recognition first divides the pixels in the feature map into faulty pixels and non-faulty pixels, marks the pixel value of the faulty pixels as 1, and marks the pixel value of the non-faulty pixels as 0 to obtain a binary image; A fault location module 14, configured to traverse the pixel values in the binary image, check the adjacent pixels of each faulty pixel, recognize the pixels with a pixel value of 1 as part of the fault area, and recognize the pixels with a pixel value of 0 as the boundary of the fault area, and then mark the boundary of the fault area in the original image using different colors or lines.

[0076] The above has introduced in detail an intelligent recognition method and system for the fault image of a leaky coaxial cable in a subway tunnel provided by the embodiments of the present application. Specific examples are used in this article to elaborate on the principle and implementation manner of the present application. The description of the above embodiments is only used to help understand the technical solution and its core idea of the present application; those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present application.

Claims

1. An intelligent recognition method for the image of leakage coaxial cable faults in subway tunnels, characterized in that, Including: Step S1: Preprocess the collected image to be detected. Step S2: Construct a convolutional neural network model, collect preprocessed image samples of subway leaky coaxial cables in various types and states, input them into the model for training, enable it to learn feature extraction to obtain a mature convolutional neural network model, and input the preprocessed image to be detected into the model to output a feature map. Step S3: Design and train a mature softmax classifier including a fully connected layer and a softmax layer, and use it to input the feature map into the softmax classifier to classify and identify the feature map. The classification and identification first divide the pixels in the feature map into faulty pixels and non-faulty pixels, mark the pixel value of the faulty pixels as 1, and the pixel value of the non-faulty pixels as 0 to obtain a binary image. Step S4: Traverse the pixel values in the binary image, check the adjacent pixels of each faulty pixel, identify the pixels with the pixel value marked as 1 as part of the faulty area, and the pixels with the pixel value marked as 0 as the boundary of the faulty area. Then use different colors or lines to mark the boundary of the faulty area in the original image.

2. The intelligent recognition method for the leakage coaxial cable fault image in the subway tunnel according to claim 1, characterized in that, The preprocessing in Step S1 includes: using a Gaussian filtering algorithm to remove noise in the image to be detected, grayscale the image, and binarize the image.

3. The intelligent recognition method for the leakage coaxial cable fault image in the subway tunnel according to claim 1, wherein, The specific steps for training the convolutional neural network model in Step S2 include: Enhance the training data by operations such as rotation, flipping, scaling, and translation. Use the backpropagation algorithm to continuously adjust the weight parameters to minimize the loss function. Use the recall rate to evaluate the performance of the model, optimize the model according to the evaluation results, and use the early stopping method and the learning rate decay strategy to prevent overfitting.

4. The intelligent recognition method for leakage coaxial cable fault images in subway tunnels according to claim 1, characterized in that, Step S3 includes: further subdividing the fault types of the feature map.

5. The intelligent recognition method for leakage coaxial cable fault images in subway tunnels according to claim 1, characterized in that, The specific steps for training the softmax classifier in Step S3 include: Use the cross-entropy loss function to calculate the difference between the predicted probability distribution and the true label. Calculate the gradient through the backpropagation algorithm and update the network parameters using the gradient descent algorithm. Use the validation set to evaluate the performance of the softmax classifier, including accuracy and recall rate metrics. Adjust the network structure and learning rate parameters according to the evaluation results to improve the classification performance.

6. The intelligent recognition method for leakage coaxial cable fault images in subway tunnels according to claim 1, characterized in that, Step S4 includes: processing the boundary by interpolation or smoothing filtering methods.

7. The intelligent recognition method for leakage coaxial cable fault images in subway tunnels according to claim 1, characterized in that, Step S4 includes: adopting a block processing and parallel computing method to improve the calculation speed.

8. An intelligent recognition system for leakage coaxial cable fault images in subway tunnels, characterized in that, Including: An image preprocessing module for preprocessing the collected image to be detected. A feature extraction module for constructing a convolutional neural network model, collecting preprocessed image samples of subway leaky coaxial cables in various types and states, inputting them into the model for training, enabling it to learn feature extraction to obtain a mature convolutional neural network model, and inputting the preprocessed image to be detected into the model to output a feature map. The classification and recognition module is used to design and train a mature softmax classifier including a fully connected layer and a softmax layer, and is used to input the feature map into the softmax classifier to classify and recognize the feature map. The classification and recognition first divides the pixels in the feature map into faulty pixels and non-faulty pixels, marks the pixel value of the faulty pixels as 1, and marks the pixel value of the non-faulty pixels as 0 to obtain a binary image; The fault location module is used to traverse the pixel values in the binary image, check the adjacent pixels of each faulty pixel, identify the pixels with the pixel value marked as 1 as part of the fault area, identify the pixels with the pixel value marked as 0 as the boundary of the fault area, and then use different colors or lines to mark the boundary of the fault area in the original image.

Images