A CNN-based mobile train image fault detection method

Abstract

The present invention discloses a mobile terminal train image fault detection method based on CNN, and the steps of the method include: S1: preprocessing train images to obtain a data set to be detected; S2: inputting the pre-trained weights and the data set in step S1 to a PC end for training a mobile terminal train image fault detection model LMTD, wherein the LMTD model consists of a lightweight feature extraction network LFENet, a residual aggregation network RPAN and a coupling module; S3: deploying the LMTD model trained in step S2 to a mobile terminal device, inputting the train image to be detected, detecting train faults, and displaying the detection results of the LMTD model detecting train faults; LFENet in the LMTD model of the present invention is used to improve the model detection accuracy and reduce the computing cost; RPAN in the LMTD model obtains multi-level feature maps to detect targets of different sizes, improves the expression ability of the model, and improves the detection accuracy; in summary, the present invention improves the train image detection accuracy and speed, and realizes mobile terminal practice train image fault detection.

Description

Technical Field

[0001] The present invention relates to the technical field of train image fault detection, and in particular to a CNN-based mobile terminal train image fault detection method. Background Art

[0002] Traditionally, train fault detection is performed manually by train inspectors, who inspect and repair key components while the train is parked at the platform. This method is inefficient and susceptible to extreme weather conditions. Advances in computing technology and artificial intelligence algorithms have led to the emergence of various detection models for fault detection of train components.

[0003] Existing train component fault detection methods, such as data-driven train image fault detection methods, have improved train image fault detection performance to a certain extent. However, traditional signal processing methods are not very practical and rely too heavily on technicians' extensive fault detection experience and railway system knowledge. Another existing train component fault detection method, such as a deep learning-based train image fault detection method, has been proposed by scholars as a fault diagnosis method based on cyclic spectrum correlation two-dimensional graphs and convolutional neural networks. First, the two-dimensional graph of the vibration signal is analyzed through cyclic spectrum analysis to provide bearing discrimination patterns for specific fault types. Then, domain-specific diagnostic knowledge is leveraged to reduce the difficulty of feature learning in the deep diagnostic model. Finally, a CNN model is constructed to learn high-level feature representations for fault classification.

[0004] The detection efficiency of the above train fault detection method has been greatly improved, but the number of train images that need to be detected every day is too large, the detection speed cannot meet the demand, and in the absence of high-performance servers outdoors, it is difficult to complete the real-time train fault detection task.

[0005] Therefore, it is currently necessary to design an efficient and effective train image fault detection method model to improve the accuracy and speed of train image detection and realize mobile terminal practice train image fault detection.

[0006] The invention patent with publication number CN105501248A discloses a railway line inspection system, which includes train information acquisition, data and fault processing, flight control and data acquisition subsystems. The train information acquisition subsystem determines the train information of the railway line and sends the train information to the data and fault processing subsystem. The data acquisition subsystem is set on the drone to collect field data of the railway line and send the field data to the data and fault processing subsystem for processing. The field data includes radar scanning data and field image data of the railway line. The detection method of this invention improves the train image fault detection performance to a certain extent. However, due to the large number of train images that need to be detected every day, the train fault detection speed of this invention is difficult to meet the demand, and this invention is not convenient for use outdoors in the absence of high-performance servers. Summary of the Invention

[0007] In response to the above-mentioned technical problems, the present invention provides a CNN-based mobile terminal train image fault detection method to quickly and accurately complete train image fault detection.

[0008] The present invention adopts the following specific technical solutions:

[0009] A CNN-based mobile terminal train image fault detection method, characterized in that the method comprises the following steps:

[0010] S1: Preprocess the train images, scale the train images to the appropriate size, and use data augmentation methods to expand the dataset to be tested.

[0011] S2: Input the pre-trained weights and the dataset from step S1 to the PC to train the mobile train image fault detection model LMTD and output the model weights. The LMTD model consists of a lightweight feature extraction network LFENet, a residual aggregation network RPAN, and a coupling module.

[0012] S3: The training process of the LMTD model is as follows:

[0013] S3.1: Send the preprocessed image to the lightweight feature extraction network LFENet in the LMTD model to obtain the feature map of the train image;

[0014] S3.2: Use the residual aggregation network RPAN to fuse the adjacent feature maps obtained in the above step S2.1 and obtain multi-level feature maps for multi-scale feature prediction and detection of targets of different scales.

[0015] S3.3: The train image features extracted in step S2.2 are input into the coupling detector in the coupling module for multi-scale train fault detection, and the SimOTA dynamic label assignment strategy is used to optimize the training process of the LMTD model to obtain classification and regression results.

[0016] S4: Deploy the LMTD model trained in step S3 to the mobile device, input the train image to be detected, detect train faults, and display the detection results of the LMTD model for train faults.

[0017] Furthermore, the lightweight feature extraction network LFENet of step S3.1 is composed of a lightweight block LB stacked by a Ghost module and a CBAM module.

[0018] Furthermore, the Ghost module uses the residual structure in ResNet, which is composed of point convolution and depth convolution. The Ghost module uses a small number of parameters to generate a large amount of feature information.

[0019] Furthermore, the CBAM is a lightweight attention mechanism embedded in CNN. CBAM obtains key information in the image by weighting the channel dimension and spatial dimension of the feature map, so that LFENet focuses on the key information in the image.

[0020] Furthermore, the LFENet mainly includes a convolutional layer and three LB modules, outputting feature maps of three scales.

[0021] Furthermore, the process of the residual aggregation network RPAN in step S2 for feature map processing is as follows: adjusting the number of channels of the feature map by convolution operation, adjusting the size of the feature map by upsampling or downsampling, and performing feature fusion on the resized feature map using a residual structure to obtain feature maps of different sizes.

[0022] Furthermore, the convolution operation ensures that the number of channels of all feature maps is consistent with the minimum number of channels in the input feature map.

[0023] Furthermore, the coupled detector in step S2.3 couples the classification branch and the regression branch.

[0024] Furthermore, during the training process of the optimized LMTD model in step S2.2, SimOTA determines the candidate region of the positive sample through the center point of the prior box, calculates the intersection-and-union ratio of each predicted box generated by the candidate region of the positive sample and the true box, sorts the calculated intersection-and-union ratios from large to small, sums the first n intersection-and-union ratios to obtain the parameters, and calculates the cost matrix of the candidate region.

[0025] Furthermore, the cost matrix of the candidate region is calculated as follows:

[0026]

[0027] Where i, j represent the rows and columns of the matrix, λ is the balance coefficient, is the classification loss between the predicted box and the true box, It is the regression loss between the predicted box and the true box, and the cost matrix represents the relationship between the current true box and the predicted box.

[0028] The beneficial effects of the present invention are:

[0029] The present invention designs a novel lightweight feature extraction network LFENet to improve the model detection accuracy, reduce the computational cost, and enhance the model's robustness to illumination changes and noise; secondly, the present invention proposes a residual aggregation network RPAN to obtain multi-level feature maps to detect targets of different sizes and improve the expressiveness of the model; finally, the coupled detector classifies and locates train faults, introduces the SimOTA dynamic label allocation strategy, optimizes the training process of the LMTD model, reduces the negative impact on the model detection effect caused by improper allocation of positive and negative sample labels, and completes the task of train image fault detection; in summary, the present invention provides an efficient and effective train image fault detection method, improves the train image detection accuracy and speed, and realizes mobile terminal practice train image fault detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0030] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments or the description of the prior art. Obviously, the drawings described below are merely embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on the provided drawings without paying any creative work.

[0031] Figure 1 This is the overall flow chart of the CNN-based mobile terminal train image fault detection method;

[0032] Figure 2 This is the overall framework diagram of the LMTD model;

[0033] Figure 3 This is the structure diagram of lightweight block LB;

[0034] Figure 4 This is the structure diagram of the lightweight attention mechanism CBAM;

[0035] Figure 5 This is the structure diagram of the depth-separable convolution DSConv;

[0036] Figure 6 This is the structure diagram of the residual aggregation network RPAN. DETAILED DESCRIPTION

[0037] The present invention will be further explained and illustrated below with reference to the embodiments, but the specific embodiments do not limit the present invention in any form.

[0038] Example 1

[0039] like Figures 1 to 6 As shown, the embodiment of the present invention discloses a mobile terminal train image fault detection method based on CNN, comprising the following steps:

[0040] A CNN-based mobile terminal train image fault detection method, characterized in that the method comprises the following steps:

[0041] S1: Preprocess the train images, scale the train images to the appropriate size, and use data augmentation methods to expand the dataset to be tested.

[0042] S2: Input the pre-trained weights and the data set in step S1 to the PC to train the train image fault detection model LMTD on the mobile terminal and output the model weights, such as Figure 1 As shown in Figure 2, its LMTD model consists of a lightweight feature extraction network LFENet, a residual aggregation network RPAN, and a coupling module;

[0043] S3: The training process of the LMTD model is as follows:

[0044] S3.1: Send the preprocessed image to the lightweight feature extraction network LFENet in the LMTD model to obtain the feature map of the train image;

[0045] like Figure 3 As shown in Figure 2, the lightweight feature extraction network LFENet is composed of a lightweight block LB stacked by the Ghost module and the CBAM module. The Ghost module uses the residual structure in ResNet and consists of point convolution and depth convolution. The Ghost module uses a small number of parameters to generate a large amount of feature information. Figure 4 As shown in the figure, CBAM is a lightweight attention mechanism embedded in CNN. CBAM obtains the key information in the image by weighting the channel dimension and spatial dimension of the feature map, so that LFENet focuses on the key information in the image and ignores irrelevant background information, thereby improving the model detection accuracy.

[0046] LFENet uses deep separable convolution DSConv instead of ordinary convolution, which greatly reduces the computational complexity of the network. The structure of DSConv is as follows: Figure 5As shown in the figure, DSConv consists of depth convolution dwConv and point convolution pwConv. dwConv is the convolution kernel that performs convolution calculation on the channel of the feature map, and pwConv is the weighted summation of the convolution kernel in the channel dimension of the feature map.

[0047] In addition, LFENet adjusts the size of feature maps using a width factor γ and a resolution factor η to further reduce model parameters. LFENet primarily consists of a convolutional layer and three LB modules, ultimately outputting feature maps of three scales: {(13×13×384), (26×26×192), and (52×52×96)}. The scale of a feature map is described by (H×W×C), where H, W, and C represent the height, width, and number of channels of the feature map, respectively. Detailed LFENet network parameters are shown in Table 1.

[0048]

[0049] Table 1

[0050] Where Conv represents the convolution layer, MaxPooling represents the maximum pooling layer, LB represents the lightweight block, RPAN represents the residual aggregation network, Softmax represents the softmax classification function, DScov represents the depthwise separable convolution, Upsample represents upsampling, s1 represents the convolution step size of 1, and s2 represents the convolution step size of 2.

[0051] S3.2: Use the residual aggregation network RPAN to fuse the adjacent feature maps obtained in the above step S2.1 and obtain multi-level feature maps for multi-scale feature prediction and detection of targets of different scales.

[0052] The present invention uses RPAN to fuse adjacent feature maps to obtain multi-scale feature maps. The structure of RPAN is as follows: Figure 6As shown in the figure, Conv represents the convolution layer, US represents upsampling, DSConv represents depth-wise separable convolution, and Res represents the residual structure. The process of RPAN for feature map processing is as follows: the number of channels of the feature map is adjusted by a 1×1 convolution operation, and then the size of the feature map is adjusted by upsampling or downsampling. Finally, the resized feature map is fused using the residual structure to obtain feature maps of four sizes, namely {(7×7×96), (13×13×96), (26×26×96), (52×52×96)}. RPAN uses a 1×1 convolution operation to make the number of channels of all feature maps consistent with the minimum number of channels in the input feature map, which greatly reduces the computational cost. RPAN uses the residual structure to fuse adjacent feature maps, enhance the feature fusion effect, and prevent the model from overfitting. In addition, RPAN adds an additional feature map and obtains a larger receptive field by adjusting the size of the convolution kernel to detect more targets. RPAN further improves the detection accuracy of LMTD with a small increase in parameters.

[0053] S3.3: The train image features extracted in step S2.2 are input into the coupling detector in the coupling module for multi-scale train fault detection, and the SimOTA dynamic label assignment strategy is used to optimize the training process of the LMTD model to obtain classification and regression results.

[0054] In the coupling module stage, the coupled detector couples the classification branch and the regression branch. The coupled detector can achieve higher detection accuracy without reducing the number of channels. In addition, in the coupling module stage, the LMTD model uses a depthwise separable convolution with a convolution kernel of 5×5 to expand the model perception. The number of depthwise separable convolutions can be set to 2, 4, or 8. The overall model results are shown in Figure 2. Figure 1 shown.

[0055] During the training of the optimized LMTD model in step S2.2, SimOTA determines the candidate region of the positive sample by the center point of the prior box, calculates the intersection-over-union (IoU) of each predicted box generated by the candidate region of the positive sample with the true box, sorts the calculated IoUs from large to small, sums the first n IoUs to obtain the parameters, and calculates the cost matrix of the candidate region.

[0056] Furthermore, the cost matrix of the candidate region is calculated as follows:

[0057]

[0058] Where i, j represent the rows and columns of the matrix, λ is the balance coefficient, is the classification loss between the predicted box and the true box, The cost matrix represents the regression loss between the predicted and true boxes. The cost matrix represents the relationship between the current true box and the predicted box. The smaller the cost value, the better the predicted box. Finally, the values in the cost matrix are sorted from smallest to largest. The predicted box with the lowest cost is used as the positive sample for the current true box, and the remaining predicted boxes are used as negative samples. SimOTA can automatically assign positive samples to the true boxes, accelerating the model training process.

[0059] S4: Deploy the LMTD model trained in step S3 to the mobile device, input the train image to be detected, detect train faults, and display the detection results of the LMTD model for train faults.

[0060] Example 2

[0061] like Figures 1 to 6 As shown, the embodiment of the present invention discloses a mobile terminal train image fault detection method based on CNN, comprising the following steps:

[0062] A CNN-based mobile terminal train image fault detection method, characterized in that the method comprises the following steps:

[0063] S1: Preprocess the train images, scale the train images to the appropriate size, and use data augmentation methods to expand the dataset to be tested.

[0064] S2: Input the pre-trained weights and the data set in step S1 to the PC to train the train image fault detection model LMTD on the mobile terminal and output the model weights, such as Figure 1 As shown in Figure 2, its LMTD model consists of a lightweight feature extraction network LFENet, a residual aggregation network RPAN, and a coupling module;

[0065] S3: The training process of the LMTD model is as follows:

[0066] S3.1: Send the preprocessed image to the lightweight feature extraction network LFENet in the LMTD model to obtain the feature map of the train image;

[0067] like Figure 3 As shown in Figure 2, the lightweight feature extraction network LFENet is composed of a lightweight block LB stacked by the Ghost module and the CBAM module. The Ghost module uses the residual structure in ResNet and consists of point convolution and depth convolution. The Ghost module uses a small number of parameters to generate a large amount of feature information. Figure 4As shown in the figure, CBAM is a lightweight attention mechanism embedded in CNN. CBAM obtains the key information in the image by weighting the channel dimension and spatial dimension of the feature map, so that LFENet focuses on the key information in the image and ignores irrelevant background information, thereby improving the model detection accuracy.

[0068] LFENet uses deep separable convolution DSConv instead of ordinary convolution, which greatly reduces the computational complexity of the network. The structure of DSConv is as follows: Figure 5 As shown in the figure, DSConv consists of depth convolution dwConv and point convolution pwConv. dwConv is the convolution kernel that performs convolution calculation on the channel of the feature map, and pwConv is the weighted summation of the convolution kernel in the channel dimension of the feature map.

[0069] In addition, LFENet further reduces model parameters by adjusting the feature map size using a width factor γ and a resolution factor η. LFENet primarily consists of a convolutional layer and three LB modules, ultimately outputting feature maps of three scales: {(13×13×384), (26×26×192), and (52×52×96)}. Detailed LFENet network parameters are shown in Table 1.

[0070]

[0071] Table 1

[0072] Where Conv represents the convolution layer, MaxPooling represents the maximum pooling layer, LB represents the lightweight block, RPAN represents the residual aggregation network, Softmax represents the softmax classification function, DScov represents the depthwise separable convolution, Upsample represents upsampling, s1 represents the convolution step size of 1, and s2 represents the convolution step size of 2.

[0073] S3.2: Use the residual aggregation network RPAN to fuse the adjacent feature maps obtained in the above step S2.1 and obtain multi-level feature maps for multi-scale feature prediction and detection of targets of different scales.

[0074] The present invention uses RPAN to fuse adjacent feature maps to obtain multi-scale feature maps. The structure of RPAN is as follows: Figure 6As shown in the figure, Conv represents the convolution layer, US represents upsampling, DSConv represents depth-wise separable convolution, and Res represents the residual structure. The process of RPAN for feature map processing is as follows: the number of channels of the feature map is adjusted by a 1×1 convolution operation, and then the size of the feature map is adjusted by upsampling or downsampling. Finally, the resized feature map is fused using the residual structure to obtain feature maps of four sizes, namely {(7×7×96), (13×13×96), (26×26×96), (52×52×96)}. RPAN uses a 1×1 convolution operation to make the number of channels of all feature maps consistent with the minimum number of channels in the input feature map, which greatly reduces the computational cost. RPAN uses the residual structure to fuse adjacent feature maps, enhance the feature fusion effect, and prevent the model from overfitting. In addition, RPAN adds an additional feature map and obtains a larger receptive field by adjusting the size of the convolution kernel to detect more targets. RPAN further improves the detection accuracy of LMTD with a small increase in parameters.

[0075] S3.3: The train image features extracted in step S2.2 are input into the coupling detector in the coupling module for multi-scale train fault detection, and the SimOTA dynamic label assignment strategy is used to optimize the training process of the LMTD model to obtain classification and regression results.

[0076] In the coupling module stage, the coupled detector couples the classification branch and the regression branch. The coupled detector can achieve higher detection accuracy without reducing the number of channels. In addition, in the coupling module stage, the LMTD model uses a depthwise separable convolution with a convolution kernel of 5×5 to expand the model perception. The number of depthwise separable convolutions can be set to 2, 4, or 8. The overall model results are shown in Figure 2. Figure 1 shown.

[0077] During the training of the optimized LMTD model in step S2.2, SimOTA determines the candidate region of the positive sample by the center point of the prior box, calculates the intersection-over-union (IoU) of each predicted box generated by the candidate region of the positive sample with the true box, sorts the calculated IoUs from large to small, sums the first n IoUs to obtain the parameters, and calculates the cost matrix of the candidate region.

[0078] Furthermore, the cost matrix of the candidate region is calculated as follows:

[0079]

[0080] Where i, j represent the rows and columns of the matrix, λ is the balance coefficient, is the classification loss between the predicted box and the true box, The cost matrix represents the regression loss between the predicted and true boxes. The cost matrix represents the relationship between the current true box and the predicted box. The smaller the cost value, the better the predicted box. Finally, the values in the cost matrix are sorted from smallest to largest. The predicted box with the lowest cost is used as the positive sample for the current true box, and the remaining predicted boxes are used as negative samples. SimOTA can automatically assign positive samples to the true boxes, accelerating the model training process.

[0081] Furthermore, LMTD uses the Varifocal loss function as the classification loss function, and the GIoU and DistributionFocal loss functions as the regression loss functions. The calculation formula of the loss function is as follows:

[0082] loss=loss vfl +αloss giou +βloss dfl (2)

[0083] S4: Deploy the LMTD model trained in step S3 to the mobile device, input the train image to be detected, detect train faults, and display the detection results of the LMTD model for train faults.

[0084] Example 3

[0085] In order to verify the feasibility of the LMTD model in the CNN-based mobile terminal train image fault detection method described in Example 1, the experimental results of the LMTD model of the present invention and other existing advanced models such as YOLOv3-Tiny, YOLOv4-Tiny, YOLOX and YOLOv5 on the same train image dataset are as follows:

[0086] As shown in Table 2, Table 2 compares the detection results of different detection models on the same train image dataset.

[0087]

[0088] Table 2

[0089] According to Table 2, the mean average precision (mAP) of the LMTD model in the detection results is 85.35%, which is 8.83%, 5.49%, 7.89%, and 5.31% higher than that of YOLOv3-Tiny, YOLOv4-Tiny, YOLOX, and YOLOv5, respectively. The detection delay using the LMTD model is only 16.86ms, which is lower than that of other target detection models. The above experimental results further demonstrate the feasibility of the method of the present invention.

[0090] The LMTD model of the present invention introduces the CBAM module and the Ghost module, which can enhance the expression ability of the model and improve the detection accuracy of the model; RPAN can enrich feature information, predict multi-scale feature maps, and detect targets of different sizes; in addition, the SimOTA strategy can quickly and accurately assign the most appropriate real box to each target, thereby improving the detection accuracy of LMTD. LMTD has a simple and efficient feature extraction network LFENet, which can reduce computing costs and speed up model reasoning. RPAN keeps the number of channels of all feature maps consistent with the minimum number of channels in the feature maps, further reducing the number of parameters of the model. In summary, the present invention provides a model of an efficient and effective train image fault detection method, which improves the accuracy and speed of train image detection and realizes mobile terminal practice train image fault detection.

[0091] Obviously, the above embodiments of the present invention are merely examples for the purpose of clearly illustrating the present invention, and are not intended to limit the embodiments of the present invention. Those skilled in the art will appreciate that other variations or modifications can be made based on the above description. It is not necessary and impossible to enumerate all embodiments here. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the claims of the present invention.

Claims

1. A CNN-based mobile terminal train image fault detection method, characterized in that: The method comprises the following steps: S1: Preprocess the train images, scale the train images to an appropriate size, and use data augmentation methods to expand the dataset to be tested; S2: Input the pre-trained weights and the dataset from step S1 to the PC to train the mobile train image fault detection model LMTD and output the model weights. The LMTD model consists of a lightweight feature extraction network LFENet, a residual aggregation network RPAN, and a coupling module. S3: The training process of the LMTD model is as follows: S3.1: Send the preprocessed image to the lightweight feature extraction network LFENet in the LMTD model to obtain the feature map of the train image; S3.2: Use the residual aggregation network (RPAN) to fuse the adjacent feature maps obtained in step S3.1 above and obtain multi-level feature maps for multi-scale feature prediction and detection of objects at different scales. S3.3: The train image features extracted in step S3.2 are input into the coupled detector in the coupling module for multi-scale train fault detection. The SimOTA dynamic label assignment strategy is used to optimize the training process of the LMTD model to obtain classification and regression results. S4: Deploy the LMTD model trained in step S3 to the mobile device, input the train image to be detected, detect train faults, and display the detection results of the LMTD model for train faults.

2. The CNN-based mobile terminal train image fault detection method according to claim 1 is characterized in that: The lightweight feature extraction network LFENet of step S3.1 is composed of a stack of lightweight blocks LB consisting of a Ghost module and a CBAM module.

3. The CNN-based mobile terminal train image fault detection method according to claim 2 is characterized in that: The Ghost module uses the residual structure in ResNet and consists of point convolution and depth convolution. The Ghost module uses a small number of parameters to generate a large amount of feature information.

4. The CNN-based mobile terminal train image fault detection method according to claim 2 is characterized in that: The CBAM is a lightweight attention mechanism embedded in CNN. CBAM obtains key information in the image by weighting the channel dimension and spatial dimension of the feature map, allowing LFENet to focus on the key information in the image.

5. The CNN-based mobile terminal train image fault detection method according to claim 2, characterized in that: The LFENet consists of a convolutional layer and three LB modules, outputting feature maps of three scales.

6. The CNN-based mobile terminal train image fault detection method according to claim 1, characterized in that: The process of the residual aggregation network RPAN in step S2 for feature map processing is as follows: adjusting the number of channels of the feature map by convolution operation, adjusting the size of the feature map by upsampling or downsampling, and fusing the resized feature map using the residual structure to obtain feature maps of different sizes.

7. The CNN-based mobile terminal train image fault detection method according to claim 6 is characterized in that: The convolution operation ensures that the number of channels of all feature maps is consistent with the minimum number of channels in the input feature map.

8. The CNN-based mobile terminal train image fault detection method according to claim 1, characterized in that: The coupled detector in step S3.3 couples the classification branch and the regression branch.

9. The CNN-based mobile terminal train image fault detection method according to claim 1, characterized in that: During the training process of the optimized LMTD model in step S3.3, SimOTA determines the candidate area of the positive sample by the center point of the prior frame, calculates the intersection-and-union ratio of each predicted box generated by the candidate area of the positive sample and the true box, sorts the calculated intersection-and-union ratios from large to small, and sums the first n intersection-and-union ratios to obtain the parameter , and calculate the cost matrix of the candidate region.

10. A CNN-based mobile terminal train image fault detection method according to claim 9, characterized in that: The calculation formula of the cost matrix of the candidate region is as follows: in represents the rows and columns of the matrix, is the balance coefficient, is the classification loss between the predicted box and the true box, It is the regression loss between the predicted box and the true box, and the cost matrix represents the relationship between the current true box and the predicted box.

Images