Unmanned aerial vehicle image-based high-speed railway sound barrier unsupervised anomaly detection method

Through the analysis of HSV color space connectivity area of ​​the drone image and the unsupervised abnormality detection network of the teacher-student-autocoder architecture, the problem of low efficiency and inability to detect sound barriers in high-speed railways is solved, and high-precision and fast multi-scale anomaly detection, suitable for complex railway environments.

CN119068360APending Publication Date: 2024-12-03CHINA STATE RAILWAY GRP CO LTD +1

Patent Information

Application Number
CN202411065491.8
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Priority Date
2024-07-05
Filing Date
2024-08-05
Publication Date
2024-12-03

AI Technical Summary

Technical Problem

The detection efficiency of high-speed railway acoustic barriers in the prior art is low and cannot be fully detected. In particular, supervised learning methods can only detect specific defect categories that have been trained, and due to the limited defect sample size, it cannot effectively detect in complex environments. multi-scale unknown exception.

Method used

Using the HSV color space connectivity area analysis based on drone images, combined with the GhostConv structure and the jump-connected autoencoder to construct an unsupervised anomaly detection network of the teacher-student-autocoder architecture, perform abnormal detection of the acoustic barrier area, and generate local and Global exception graph.

Benefits of technology

It realizes the ability to detect multi-scale unknown anomalies with high accuracy and rapid detection in complex railway environments, without manual data annotation and abnormal sample training, improves the accuracy and efficiency of sound barrier detection, and can obtain high quality without affecting railway operation. image.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN119068360A_ABST
    Figure CN119068360A_ABST
Patent Text Reader

Abstract

The invention discloses a high-speed railway sound barrier unsupervised anomaly detection method based on an unmanned aerial vehicle image, and belongs to the field of sound barrier anomaly detection, and the method comprises the steps: carrying out the communication region analysis based on the color distribution of the unmanned aerial vehicle image in an HSV color space, and extracting a sound barrier region; a GhostConv structure is integrated into a student network, a jump connection auto-encoder is utilized to guide student network training, and an unsupervised anomaly detection network based on a teacher-student-auto-encoder architecture is constructed; and according to the sound barrier area, carrying out sound barrier anomaly detection by using an unsupervised anomaly detection network based on a teacher-student-auto-encoder architecture to obtain a local anomaly graph and a global anomaly graph, and completing high-speed railway sound barrier unsupervised anomaly detection based on the unmanned aerial vehicle image. The problems that in the prior art, the detection efficiency is low, and comprehensive detection cannot be achieved are solved.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] The present invention belongs to the field of sound barrier anomaly detection, and in particular relates to an unsupervised anomaly detection method for high-speed railway sound barriers based on drone images. Background Art

[0002] In recent years, the rapid expansion of high-speed railways has brought tremendous pressure on the inspection and maintenance of railway infrastructure. As a key facility along the railway, the sound barrier is gradually damaged by the complex natural environment and the strong air pressure generated by the operation of high-speed trains. These abnormal conditions may cause the sound barrier to collapse in the track area, which may lead to serious accidents. In the past few decades, the abnormal detection of sound barriers has mainly relied on manual work, through manual on-site patrols or browsing videos taken by inspection vehicles to identify sound barrier diseases. These methods are mainly carried out at night and can only detect sound barrier defects facing the track side, resulting in low detection efficiency.

[0003] The automatic detection method based on drones has the advantages of high efficiency, low cost, and no influence from terrain. The supervised learning-based method can only detect specific defect categories that have been trained, and the limited number of defect samples also prevents the supervised learning method from becoming a more popular field application. Summary of the invention

[0004] In view of the above-mentioned deficiencies in the prior art, the present invention provides an unsupervised anomaly detection method for high-speed railway sound barriers based on drone images, which solves the problems of low detection efficiency and inability to conduct comprehensive detection in the prior art.

[0005] In order to achieve the above-mentioned invention object, the technical solution adopted by the present invention is: an unsupervised anomaly detection method for high-speed railway sound barriers based on drone images, comprising the following steps:

[0006] S1. Based on the color distribution of drone images in the HSV color space, connected region analysis is performed to extract the sound barrier area;

[0007] S2, integrate the GhostConv structure into the student network, and use the jump connection autoencoder to guide the student network training, and build an unsupervised anomaly detection network based on the teacher-student-autoencoder architecture;

[0008] S3. According to the sound barrier area, the unsupervised anomaly detection network based on the teacher-student-autoencoder architecture is used to perform sound barrier anomaly detection, obtain local anomaly maps and global anomaly maps, and complete the unsupervised anomaly detection of high-speed railway sound barriers based on drone images.

[0009] Furthermore, the unsupervised anomaly detection network in step S2 includes a teacher network, a student network and a skip connection autoencoder network.

[0010] Furthermore, the teacher network includes a normalization layer, a first convolutional layer, a first adaptive average pooling layer, a second convolutional layer, a second adaptive average pooling layer, a third convolutional layer and a fourth convolutional layer connected in sequence.

[0011] Furthermore, the student network includes a normalization layer, a fifth convolution layer, a third adaptive average pooling layer, a sixth convolution layer, a seventh adaptive average pooling layer, a seventh convolution layer, a first Ghost convolution layer, an eighth convolution layer and a ninth convolution layer, which are connected in sequence.

[0012] Further, the jump connection autoencoder network includes a normalization layer, a tenth convolutional layer, an eleventh convolutional layer, a twelfth convolutional layer, a thirteenth convolutional layer, a fourteenth convolutional layer, a fifteenth convolutional layer, a sixteenth convolutional layer, a first transposed convolutional layer, a seventeenth convolutional layer, a second transposed convolutional layer, an eighteenth convolutional layer, a third transposed convolutional layer, a nineteenth convolutional layer, a fourth transposed convolutional layer, a twentieth convolutional layer, a first bilinear interpolation operation, a twenty-first convolutional layer, a second bilinear interpolation operation, a twenty-second convolutional layer, a twenty-third convolutional layer, a third bilinear interpolation operation, a twenty-fourth convolutional layer and a twenty-fifth convolutional layer, which are connected in sequence; the tenth convolutional layer, the eleventh convolutional layer, the twelfth convolutional layer, the thirteenth convolutional layer, the fourteenth convolutional layer and the fifteenth convolutional layer are respectively connected to the seventeenth convolutional layer, the eighteenth convolutional layer, the nineteenth convolutional layer, the twentieth convolutional layer, the twenty-first convolutional layer and the twenty-second convolutional layer in a one-to-one correspondence.

[0013] Furthermore, the loss function of the unsupervised anomaly detection network in step S2 is:

[0014] L=L ST +L AE +L STAE

[0015]

[0016] Where L is the loss function of the unsupervised anomaly detection network; L ST is the teacher-student network loss function; L AE is the teacher-autoencoder network loss function; L STAE is the student-autoencoder network loss function; T() c,w,h and S() c,w,h are the teacher network and the student network respectively; I is the training image; is the quantile; C is the number of channels of the image; W is the width of the image; H is the height of the image; is the square of the Frobenius norm; S(P) cis the output of the student network channel c after the input image P; c is the channel in the selected feature matrix; P is a randomly selected image extracted from the pre-training dataset ImageNet; T(I) c is the output of the teacher network after inputting the training image; A(I) c After inputting the training image, the output of the jump-connected autoencoder network is S'(I) c The output of the additional output channel of the student network after inputting the training image.

[0017] Furthermore, the step S3 is specifically as follows:

[0018] The unsupervised anomaly detection network is used to perform sound barrier detection in the sound barrier area, and the square output feature map of the teacher network, the square output feature map of the student network and the square output feature map of the autoencoder are obtained. Based on the difference between the square output feature map of the teacher network and the square output feature map of the student network, a local anomaly map is obtained. Based on the difference between the square output feature map of the student network and the square output feature map of the autoencoder, a global anomaly map is obtained, completing the unsupervised anomaly detection of high-speed railway sound barriers based on drone images.

[0019] The beneficial effects of the present invention are as follows: by using drones to photograph the sound barrier, images of the inside and outside of the sound barrier with better image quality can be obtained without affecting the operation of the line and without considering the skylight period; the sound barrier positioning based on the HSV color space does not require cumbersome data labeling and model training, and can accurately extract the sound barrier area. The unsupervised anomaly detection network based on the teacher-student-autoencoder architecture does not require abnormal samples for training and manual data labeling, and can detect unknown anomalies at multiple scales, thereby improving the accuracy of sound barrier anomaly detection and achieving high detection rate and fast processing speed in complex railway environments. BRIEF DESCRIPTION OF THE DRAWINGS

[0020] Figure 1 The present invention is a flow chart of the method.

[0021] Figure 2 This is a diagram of the sound barrier data processing process in an embodiment of the present invention.

[0022] Figure 3 It is the average HSV histogram of the sound barrier image in the embodiment of the present invention.

[0023] Figure 4 : is a network structure diagram of UADNet in an embodiment of the present invention.

[0024] Figure 5 Schematic diagram of the GhostConv module in an embodiment of the present invention.

[0025] Figure 6This is a visualization diagram of the detection effect of the model in an embodiment of the present invention on the MVTec AD, VisA and sound barrier datasets.

[0026] in, Figure 2 (a) is the original image of the sound barrier. Figure 2 (b) is the false detection effect diagram. Figure 2 (c) is a schematic diagram of region segmentation. Figure 2 (d) is the effect diagram of connected area analysis. DETAILED DESCRIPTION

[0027] The specific implementation modes of the present invention are described below so that those skilled in the art can understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific implementation modes. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention as defined and determined by the attached claims, these changes are obvious, and all inventions and creations utilizing the concept of the present invention are protected.

[0028] like Figure 1 As shown, in one embodiment of the present invention, a method for unsupervised anomaly detection of high-speed railway sound barriers based on drone images includes the following steps:

[0029] S1. Based on the color distribution of drone images in the HSV color space, connected region analysis is performed to extract the sound barrier area;

[0030] S2, integrate the GhostConv structure into the student network, and use the jump connection autoencoder to guide the student network training, and build an unsupervised anomaly detection network based on the teacher-student-autoencoder architecture;

[0031] S3. According to the sound barrier area, the unsupervised anomaly detection network based on the teacher-student-autoencoder architecture is used to perform sound barrier anomaly detection, obtain local anomaly maps and global anomaly maps, and complete the unsupervised anomaly detection of high-speed railway sound barriers based on drone images.

[0032] In this embodiment, Figure 2 As shown in (a), there are a lot of irrelevant areas at the top and bottom of the sound barrier image collected by the drone. Figure 2 As shown in (b), irrelevant background interference can easily lead to a high false detection rate. Figure 2 (d) shows the effect of connected region analysis. Therefore, the extraction of sound barrier region is the first step of anomaly detection.

[0033] In image color extraction, the HSV color space is simpler and more intuitive than the RGB color space. The average HSV histogram of the sound barrier dataset is as follows: Figure 3 As shown, each row corresponds to Figure 2(c) Pixel distribution of Hue Histogram, Saturation Histogram, and Value Histogram from left to right in a different region. It is obvious that the sound barrier structure mainly presents blue tones, and there is a significant difference in the histogram distribution compared with the other two regions. Figure 3 Pixel Count is the number of pixels.

[0034] The sound barrier area localization algorithm mainly includes three steps: (1) Analyze the connected regions of the image and identify Figure 2 (d) The blue area shown; (2) Calculate the bottom layer position of the blue area and determine the lower boundary of the sound barrier area; (3) Scan upwards line by line to confirm the upper boundary of the sound barrier area.

[0035] The unsupervised anomaly detection network in step S2 includes a teacher network, a student network and a skip connection autoencoder network.

[0036] The teacher network includes a normalization layer, a first convolutional layer, a first adaptive average pooling layer, a second convolutional layer, a second adaptive average pooling layer, a third convolutional layer and a fourth convolutional layer which are connected in sequence.

[0037] The student network includes a normalization layer, a fifth convolution layer, a third adaptive average pooling layer, a sixth convolution layer, a seventh adaptive average pooling layer, a seventh convolution layer, a first Ghost convolution layer, an eighth convolution layer and a ninth convolution layer, which are connected in sequence.

[0038] The jump connection autoencoder network includes a normalization layer, a tenth convolutional layer, an eleventh convolutional layer, a twelfth convolutional layer, a thirteenth convolutional layer, a fourteenth convolutional layer, a fifteenth convolutional layer, a sixteenth convolutional layer, a first transposed convolutional layer, a seventeenth convolutional layer, a second transposed convolutional layer, an eighteenth convolutional layer, a third transposed convolutional layer, a nineteenth convolutional layer, a fourth transposed convolutional layer, a twentieth convolutional layer, a first bilinear interpolation operation, a twenty-first convolutional layer, a second bilinear interpolation operation, a twenty-second convolutional layer, a twenty-third convolutional layer, a third bilinear interpolation operation, a twenty-fourth convolutional layer and a twenty-fifth convolutional layer, which are connected in sequence; the tenth convolutional layer, the eleventh convolutional layer, the twelfth convolutional layer, the thirteenth convolutional layer, the fourteenth convolutional layer and the fifteenth convolutional layer are respectively connected to the seventeenth convolutional layer, the eighteenth convolutional layer, the nineteenth convolutional layer, the twentieth convolutional layer, the twenty-first convolutional layer and the twenty-second convolutional layer in one-to-one correspondence.

[0039] In this embodiment, Figure 4The specific network architecture of the unsupervised anomaly detection network UADNet based on the teacher-student-autoencoder architecture is given, which includes three main branches: teacher network, student network and skip connection autoencoder network. Tables 1 and 2 provide the detailed configuration of the teacher-student network. The GhostConv structure is integrated into the student network to further reduce the student's ability to imitate the anomalies detected by the teacher. Figure 5 As shown in Figure 1, GhostConv consists of two convolutions, 3x3 and 1x1, which eliminates a lot of redundancy in the intermediate feature maps. Figure 4 In the figure, Patch Description Network is the teacher network; Student Network: GPDN-S is the student network integrated with the GhostConv structure; Skip Connection Autoencoder is the skip connection autoencoder network; Anomaly score is the anomaly score; Local Map is the local anomaly map; Global Map is the global anomaly map; Norm_batch is the data normalization layer; Conv is the convolution layer; avgpool is the adaptive average pooling layer; GhostConv is the Ghost convolution layer; ConvTrans is the transposed convolution layer; interpolate is the bilinear interpolation operation; diff is the pixel-by-pixel square difference between the two sets of output feature maps. Figure 5 Input is input; Conv2d is 2D convolution; ReLU is activation function; Concat is feature fusion layer. Table 1 is the teacher network parameter table; Table 2 is the student network parameter table.

[0040] During the training of the student network, non-anomaly training samples are used to mimic the output of the teacher network. During model inference, since the student network generalizes poorly outside the cluster where the non-anomaly samples are distributed and produces incorrect predictions, an anomaly score for each input pixel can be calculated based on the regression error of the student network relative to the teacher network.

[0041] Table 1

[0042]

[0043] Table 2

[0044]

[0045] A skip-connected autoencoder is built in UADNet to complement the teacher-student network from a global perspective. Figure 4In the network structure shown, an autoencoder is trained to predict the output of the teacher network. Since the autoencoder has difficulty reconstructing fine details in the teacher's feature space, the student network can be enhanced by doubling the output channels of the shared parameters. This adjustment allows the student network to predict the output of both the teacher and the autoencoder.

[0046] The encoder of the skip-connected autoencoder uses convolution operations to shrink the intermediate feature maps. In the decoder, upsampling is performed using transposed convolution and bilinear interpolation. In addition, the feature maps of the same scale of the encoder and decoder are concatenated to enhance feature fusion and reduce the blurry reconstruction of the autoencoder. Unlike the student network whose receptive field is limited to small patches of the image, the skip-connected autoencoder learns from global image features. Therefore, the student network learns the global reconstruction error of the skip-connected autoencoder on normal images. The difference between the output of the autoencoder and the output of the student network is used to calculate the global anomaly map. Table 3 shows the network parameters of the skip-connected autoencoder.

[0047] Table 3

[0048]

[0049]

[0050] The loss function of the unsupervised anomaly detection network in step S2 is:

[0051] L=L ST +L AE +L STAE

[0052]

[0053] Where L is the loss function of the unsupervised anomaly detection network; L ST is the teacher-student network loss function; L AE is the teacher-autoencoder network loss function; L STAE is the student-autoencoder network loss function; T() c,w,h and S() c,w,h are the teacher network and the student network respectively; I is the training image; is the quantile; C is the number of channels of the image; W is the width of the image; H is the height of the image; is the square of the Frobenius norm; S(P) c is the output of the student network channel c after the input image P; c is the channel in the selected feature matrix; P is a randomly selected image extracted from the pre-training dataset ImageNet; T(I) c is the output of the teacher network after inputting the training image; A(I) cAfter inputting the training image, the output of the jump-connected autoencoder network is S'(I) c The output of the additional output channels of the student network after inputting the training image.

[0054] The loss function of the UADNet network and the abnormal graph normalization operation:

[0055] In UADNet, local and global anomaly maps are calculated based on the squared output feature maps of the teacher network and the student network, as well as the squared output feature maps of the student network and the autoencoder. Therefore, the loss function is mainly composed of L ST , L AE and L STAE It consists of three parts. As shown in the formula:

[0056]

[0057] Where, L ST is the teacher-student network loss function; L AE is the teacher-autoencoder network loss function; L STAE is the student-autoencoder network loss function; I is the training image; T() c,w,h and S() c,w,h They are teacher network and student network respectively; is the quantile, set to 0.999, and only the average value exceeding the quantile is calculated; C is the number of channels of the image; W is the width of the image; H is the height of the image; P is a randomly selected image extracted from the pre-training dataset ImageNet, which is used to prevent the student network from generalizing its imitation of the teacher network to images outside the distribution; is the square of the Frobenius norm; c is the channel in the selected feature matrix; A() c For the skip connection autoencoder network, S'() c Additional output channels for the student network. The total loss function is the sum of these three components.

[0058] The step S3 is specifically as follows:

[0059] The unsupervised anomaly detection network is used to perform sound barrier detection in the sound barrier area, and the square output feature map of the teacher network, the square output feature map of the student network and the square output feature map of the autoencoder are obtained. Based on the difference between the square output feature map of the teacher network and the square output feature map of the student network, a local anomaly map is obtained. Based on the difference between the square output feature map of the student network and the square output feature map of the autoencoder, a global anomaly map is obtained, completing the unsupervised anomaly detection of high-speed railway sound barriers based on drone images.

[0060] Two types of anomaly maps are generated in the UADNet network: the local anomaly map is calculated based on the difference between the feature maps of the teacher network and the student network, and the global anomaly map is determined by the difference between the feature map of the jump connection autoencoder and the student network. When the ratio is set to (0.5, 0.5), UADNet obtains the best result, indicating the final parameter selection.

[0061] As shown in Table 4, due to the presence of obvious abnormal features, except for DRAEM based on the reconstruction network, the models perform well on the MVTec AD dataset. UADNet has the highest pixel_AUPRO (The Area Under Per-Region Overlap curve) and pixel_AUROC (The Area Under Receiver Operator Characteristic) indexes, which are 95.1% and 98.4%, respectively. On the VisA dataset, where anomaly detection is more difficult, the UADNet proposed in this embodiment can obtain a score of more than 90% on pixel_AUPRO, which shows that UADNet has obvious advantages in detecting anomalies of small targets. Considering the more important sound barrier dataset used in railway engineering, UADNet shows excellent performance in both image anomaly detection and positioning. Table 4 is a comparison table of the anomaly detection performance of UADNet on public and sound barrier datasets.

[0062] Table 4

[0063]

[0064]

[0065] The visual detection effects of each model on the data set are as follows: Figure 6 As shown in the figure, the MVTec AD, VisA, and Noise barrier datasets are arranged from top to bottom. For large anomaly targets in the MVTec AD dataset, the models all show effective anomaly recognition, except for some areas missed by the CFLOW algorithm. In addition, the DRAEM algorithm shows a higher false positive rate, which is consistent with the quantitative detection results shown in Table 4. The noise barrier dataset poses challenges to anomaly detection, such as small targets, multi-scale targets, and diverse anomaly shapes. Even the STFPM and ReverseDistillation algorithms that perform well on public datasets still have missed detections and false detections. Compared with other models, UADNet provides more accurate anomaly localization, minimizing the probability of missed and false detections of anomalies in actual railway applications.

Claims

1. An unsupervised anomaly detection method for high-speed railway sound barriers based on drone images, characterized in that: The following steps are involved: S1. Based on the color distribution of drone images in the HSV color space, connected region analysis is performed to extract the sound barrier area; S2, integrate the GhostConv structure into the student network, and use the jump connection autoencoder to guide the student network training, and build an unsupervised anomaly detection network based on the teacher-student-autoencoder architecture; S3. According to the sound barrier area, the unsupervised anomaly detection network based on the teacher-student-autoencoder architecture is used to perform sound barrier anomaly detection, obtain local anomaly maps and global anomaly maps, and complete the unsupervised anomaly detection of high-speed railway sound barriers based on drone images.

2. The unsupervised anomaly detection method for high-speed railway sound barriers based on drone images according to claim 1 is characterized in that: The unsupervised anomaly detection network in step S2 includes a teacher network, a student network and a skip connection autoencoder network.

3. The unsupervised anomaly detection method for high-speed railway sound barriers based on drone images according to claim 2 is characterized in that: The teacher network includes a normalization layer, a first convolutional layer, a first adaptive average pooling layer, a second convolutional layer, a second adaptive average pooling layer, a third convolutional layer and a fourth convolutional layer which are connected in sequence.

4. The unsupervised anomaly detection method for high-speed railway sound barriers based on drone images according to claim 2 is characterized in that: The student network includes a normalization layer, a fifth convolution layer, a third adaptive average pooling layer, a sixth convolution layer, a seventh adaptive average pooling layer, a seventh convolution layer, a first Ghost convolution layer, an eighth convolution layer and a ninth convolution layer, which are connected in sequence.

5. The unsupervised anomaly detection method for high-speed railway sound barriers based on drone images according to claim 2 is characterized in that: The jump connection autoencoder network includes a normalization layer, a tenth convolutional layer, an eleventh convolutional layer, a twelfth convolutional layer, a thirteenth convolutional layer, a fourteenth convolutional layer, a fifteenth convolutional layer, a sixteenth convolutional layer, a first transposed convolutional layer, a seventeenth convolutional layer, a second transposed convolutional layer, an eighteenth convolutional layer, a third transposed convolutional layer, a nineteenth convolutional layer, a fourth transposed convolutional layer, a twentieth convolutional layer, a first bilinear interpolation operation, a twenty-first convolutional layer, a second bilinear interpolation operation, a twenty-second convolutional layer, a twenty-third convolutional layer, a third bilinear interpolation operation, a twenty-fourth convolutional layer and a twenty-fifth convolutional layer, which are connected in sequence; the tenth convolutional layer, the eleventh convolutional layer, the twelfth convolutional layer, the thirteenth convolutional layer, the fourteenth convolutional layer and the fifteenth convolutional layer are respectively connected to the seventeenth convolutional layer, the eighteenth convolutional layer, the nineteenth convolutional layer, the twentieth convolutional layer, the twenty-first convolutional layer and the twenty-second convolutional layer in one-to-one correspondence.

6. The unsupervised anomaly detection method for high-speed railway sound barriers based on drone images according to claim 1 is characterized in that: The loss function of the unsupervised anomaly detection network in step S2 is: L=L ST +L AE +L STAE Where L is the loss function of the unsupervised anomaly detection network; L ST is the teacher-student network loss function; L AE is the teacher-autoencoder network loss function; L STAE is the student-autoencoder network loss function; T() c,w,h and S() c,w,h are the teacher network and the student network respectively; I is the training image; is the quantile; C is the number of channels of the image; W is the width of the image; H is the height of the image; is the square of the Frobenius norm; S(P) c is the output of the student network channel c after the input image P; c is the channel in the selected feature matrix; P is a randomly selected image extracted from the pre-training dataset ImageNet; T(I) c is the output of the teacher network after inputting the training image; A(I) c After inputting the training image, the output of the jump-connected autoencoder network is S'(I) c The output of the additional output channels of the student network after inputting the training image.

7. The unsupervised anomaly detection method for high-speed railway sound barriers based on drone images according to claim 1 is characterized in that: The step S3 is specifically as follows: The unsupervised anomaly detection network is used to perform sound barrier detection in the sound barrier area, and the square output feature map of the teacher network, the square output feature map of the student network and the square output feature map of the autoencoder are obtained. Based on the difference between the square output feature map of the teacher network and the square output feature map of the student network, a local anomaly map is obtained. Based on the difference between the square output feature map of the student network and the square output feature map of the autoencoder, a global anomaly map is obtained, completing the unsupervised anomaly detection of high-speed railway sound barriers based on drone images.

Citation Information

Patent Citations

  • Image processing method and device and electronic equipment

    CN114463245A

Cited By

  • Multi-source unsupervised track anomaly detection method and system

    CN120411893A

  • Sound barrier defect detection method based on deep learning

    CN121190493A

  • A deep learning-based sound barrier defect detection method

    CN121190493B