CN114119577B - Detection method of leaky cable buckle in high-speed railway tunnel - Google Patents

Abstract

The invention discloses a high-speed railway tunnel cable leakage buckle detection method, and relates to the technical field of image detection methods. The detection method comprises the following steps: improving a CS-LBP algorithm and calculating gradient direction characteristics of a high-speed railway tunnel cable leakage buckle image sampling circle domain one by one to obtain a complete primary gradient direction characteristic diagram; performing two continuous downsampling on the feature map and extracting continuous gradient features of the two sampled images respectively; and (3) connecting the two groups of gradient features in series to serve as descriptors, and completing the classification work of the fault buckle images by using the SVM. The method can improve the feature extraction efficiency, reduce the latitude of the description sub-features and more accurately reflect the outline features of the leaky cable buckle.

Description

Detection method of leaky cable buckle in high-speed railway tunnel

technical field

The invention relates to the technical field of image processing methods, in particular to a method for detecting leaky cable buckles in high-speed railway tunnels.

Background technique

At present, railway transportation is still the main mode of transportation to solve the problem of population transfer and material transportation in our country, especially as the carrying capacity continues to increase, it is particularly important to ensure the normal operation of the railway communication system. The communication network in the mountainous tunnel section is mainly realized by covering the leaky cable, which is usually hung on the tunnel wall and fixed with a special buckle. The air pressure and energy waves generated when the high-speed train passes through the tunnel will affect the buckle, and the wet geological environment will also accelerate the loosening or even detachment of the buckle. The main detection method for leaky cable buckles is manual fixed-point regular inspection one by one. This detection method is not only severely restricted by environmental factors and human factors, but also has great potential safety hazards. With the increasing mileage of railway tunnels in recent years, it is difficult for this traditional inspection method to meet the actual needs, and it has become an inevitable trend to realize the automatic detection of leaky cable snaps. The leaky cable buckle detection technology combined with computer vision mainly has the following two links: first, a high-speed camera needs to be placed at the window of the train, and the entire section of the picture taken during the tunnel operation is collected along with the train; then the leaky cable is processed by the computer Snap data, extract features frame by frame, and complete snap inspection. At present, the data acquisition part has been realized, but the detection part is still in the stage of manual playback video troubleshooting. The local binary pattern-based deep mining algorithm and multi-feature fusion algorithm that appear in the prior art also have the problem that the descriptor is not well targeted and the feature dimension is too high.

Contents of the invention

The technical problem to be solved by the present invention is how to provide a high-speed railway tunnel leaky cable buckle detection method that can improve the feature extraction efficiency, reduce the feature latitude of the descriptor, and reflect the profile features of the leaky cable buckle more accurately.

In order to solve the above-mentioned technical problems, the technical solution adopted by the present invention is: a method for detecting the buckling of leaky cables in high-speed railway tunnels, which is characterized in that it comprises the following steps:

Improve the CS-LBP algorithm and calculate the gradient direction characteristics of the image sampling circle of the leaky cable buckle of the high-speed railway tunnel one by one, and obtain a complete preliminary gradient direction feature map;

Perform two continuous downsampling on this feature map and extract the continuous gradient features of the two sampled images respectively;

The two groups of gradient features are connected in series as descriptors, and SVM is used to complete the classification of fault snap images.

A further technical solution is to improve the CS-LBP algorithm as follows:

The number of samples in the image sampling circle is P=8, the sampling radius is R=1, and the threshold T=10; the gray value of all sampling points in the circular neighborhood is first determined by bilinear interpolation method; and then each group of centrosymmetric The gray level difference of the sampling point; finally, compare the relationship between the gray level difference and the threshold T to determine the direction of the gradient.

A further technical solution is: if the absolute value of the grayscale difference is greater than the threshold T, then the counting bit corresponding to the sampling point with a large grayscale value is marked as 1, and this counting bit is the gradient direction of this group of sampling pixel pairs; Mark the count bit corresponding to the sampling point with a small gray value as 0; if the absolute value of the gray difference value is not greater than the threshold T, the count bits corresponding to the two sampling points are marked as 0, indicating that this group of sampling pixel pairs has no Gradient, repeat the above operation on the entire Gaussian filtered image to obtain a complete preliminary gradient direction feature map, and the value of each pixel uniquely corresponds to a one-dimensional array composed of eight binary numbers 0 or 1.

A further technical solution is that the calculation method of key gradient features is as follows:

The key gradient direction is used to represent the overall gradient trend of the pixel block, sorted according to the gradient azimuth map, and the numbers 1 to 8 distributed in the counterclockwise direction correspond to 8 gradient directions respectively, and the binary pixel values of all pixels in the pixel block are divided into the same After the count bits are accumulated one by one, the count and the largest count bit are the key gradient direction of the pixel block.

A further technical solution is: a Cell unit is a pixel block composed of four pixels a, b, c, and d, and the method for obtaining the key gradient direction feature of this Cell is as follows:

First, the improved CS-LBP algorithm is used to sequentially calculate the one-dimensional array pixel values corresponding to these four pixels; then the four one-dimensional pixel values are accumulated one by one according to the same counting position, and the gradient direction with the most counting direction is the key Gradient direction; finally get the gradient direction of the Cell and the output encoding value of the key gradient direction, the calculation formula is defined as follows:

X _i ＝[x _i1 ,x _i2 ,…x _ij …,x _im ]

Where X _i represents a one-dimensional array, Y represents the accumulation of all one-dimensional arrays X _i in the Cell structure, x _ij ∈ {0, 1}, i ∈ {1, n}, j ∈ {1, m}, i and j Both are integers; the function max_id(Y) indicates the position corresponding to the maximum value in the output one-dimensional array Y, that is, the column coordinate j of the maximum value at this time; repeat the above steps for the entire preliminary gradient direction feature map, and take the stride as 2, so the gradient feature map after the first downsampling is obtained;

The same iteration is performed on the feature map after the first downsampling to obtain the second downsampling feature map.

The further technical solution is: the length and width of the Cell structure are set to 2×2, the number of channels is 8, and the sliding step size is 2, which can facilitate the setting of relevant parameters during continuous gradient feature extraction, and at the same time The length and width of the sampling feature map are proportional to the size of the smallest partition sub-image obtained by the traditional LBP and its variant algorithms.

A further technical solution is to define a 3×3 window as a block when extracting continuous gradient features on the sampling feature map, the number of channels is 8, and the sliding step is 1.

The further technical solution is: in the block, perform circular sorting in the counterclockwise direction of the gradient azimuth map, count the maximum continuous sequence length of non-central pixels of all reference channels and adjacent channels, first determine the channels with central pixels one by one and use this channel As a benchmark, if there is no central pixel in all channels, the Block does not output continuous gradient features.

A further technical solution is: the calculation steps and constraints of the continuous gradient feature include the following three points:

In the 8-channel Block, first determine the channel with the central pixel, and if there is no central pixel, the feature will not be output;

Based on the channel with the central pixel, judge whether the maximum continuous sequence length in its neighborhood is not less than 3, and if not, then do not output the feature;

The maximum continuous sequence length of the channel adjacent to the reference channel is counted separately. The encoding form of the continuous gradient feature output is the reference channel number and the maximum continuous sequence length of the three corresponding channels. The total number of encodings of this strategy is 768.

The beneficial effect of adopting the above technical solution is that the method described in this application has a significant advantage compared with the prior art in that the scale transformation of the buckle contour features is realized through continuous down-sampling feature maps, which greatly improves the The efficiency of feature extraction also plays a role in reducing the latitude of descriptor features. At the same time, the continuous gradient feature can also more accurately reflect the contour features of the leaky cable buckle, especially the gradient information at the inflection point and the sideline, and has a better detection effect in distinguishing faulty buckle images in the high-speed rail tunnel scene.

Description of drawings

The present invention will be further described in detail below in conjunction with the accompanying drawings and specific embodiments.

Fig. 1 is LBP and CS-LBP sampling mode diagram in the embodiment of the present invention;

Fig. 2 is the calculation preliminary gradient direction figure in the embodiment of the present invention;

Figure 3a is a gradient orientation diagram in an embodiment of the present invention;

Fig. 3b is a Cell structure diagram in the embodiment of the present invention;

Fig. 4 is the calculation key gradient figure in the embodiment of the present invention;

Fig. 5 is a downsampling process diagram in the embodiment of the present invention;

Fig. 6 is the characteristic map of calculating continuous gradient in the embodiment of the present invention;

Fig. 7 is the experimental flowchart of the method described in the embodiment of the present invention;

Fig. 8 is the ROC curve figure in the embodiment of the present invention;

Fig. 9 is a flow chart of the method described in the embodiment of the present invention;

Detailed ways

The technical solutions in the embodiments of the present invention are clearly and completely described below in conjunction with the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of them. Based on the embodiments of the present invention, all other embodiments obtained by persons of ordinary skill in the art without making creative efforts belong to the protection scope of the present invention.

In the following description, a lot of specific details are set forth in order to fully understand the present invention, but the present invention can also be implemented in other ways different from those described here, and those skilled in the art can do it without departing from the meaning of the present invention. By analogy, the present invention is therefore not limited to the specific examples disclosed below.

As shown in Figure 9, in general, the embodiment of the present invention discloses a method for detecting leaky cable snapping of a high-speed railway tunnel with multi-scale continuous gradient features, and the method includes the following steps:

First, improve the CS-LBP algorithm and calculate the gradient direction features of the sampling circle one by one to obtain a complete preliminary gradient direction feature map; then perform two consecutive downsampling on this feature map and extract the continuous gradients of the two sampled images respectively features; finally, the two sets of gradient features are concatenated as descriptors, and SVM is used to complete the classification of fault snap images.

The method can realize the scale transformation of the buckle contour feature and effectively reduce the feature dimension of the descriptor, and improve the classification accuracy of the image by extracting continuous gradient features. The experiment compares the detection of various variant algorithms such as CS-LBP and HOG. The results show that the recall rate and precision of the algorithm proposed by the method reach 0.923 and 0.857, respectively. Compared with these comparison algorithms, there are obvious differences. Advantage.

The above method will be described in detail below in conjunction with specific steps

Improved CS-LBP Algorithm

Traditional LBP and CS-LBP:

The traditional LBP descriptor was originally proposed by Ojala et al. It is still widely used in face recognition because it can keenly capture the grayscale difference between fine textures, and also has good rotation invariance and grayscale invariance. and texture classification domains. The core idea of the algorithm is to compare the gray value between the center pixel and P sampling pixels in the neighborhood in a circular neighborhood with R as the radius, and then use the Boolean function to calculate a string of binary numbers containing only 0 and 1 Sequence, and finally converted into decimal LBP coded values in a unified order, the corresponding calculation formula is:

n _i represents the gray value of the ith sampling point on the circular neighborhood, its size can be determined by bilinear interpolation, n _c represents the gray value of the central pixel. In order to enhance the expression ability of the LBP descriptor in the spatial direction, Heikkila et al. proposed the CS-LBP descriptor, which replaced the LBP method of encoding the difference information between the neighborhood sampling points and the central pixel, and only for four groups of radially distributed The centrosymmetric sampling points are encoded, and the corresponding calculation formula becomes:

It is necessary to compare the feature extraction effects of multiple groups of experiments in order to have a better evaluation of the threshold T in the above formula (4). Usually it can be determined according to the global grayscale of the input image or the grayscale of the neighborhood sampled pixels. The method chooses the CS-LBP operator as the improved basic model mainly for the following three reasons:

1) Under the same sampling number and sampling radius, as shown in Figure 1, the feature vector encoded by CS-LBP is more compact, and the feature dimension of the descriptor is much lower than that of LBP.

2) CS-LBP inherits the good texture characteristics of LBP, and at the same time expands four groups of radially distributed gradient difference information.

3) CS-LBP can have stronger stability in the gentle area of the processing image. Many experiments on image matching have also proved that the CS-LBP operator is more robust than using the gradient operator or the amplitude operator alone. sex.

Improve the directionality of CS-LBP:

The traditional CS-LBP only shows that there is a gradient difference in a certain diameter direction or some diameter directions in the sampling circle, but it does not clearly indicate which side the gradient direction is on. Therefore, on this basis, the method retains the CS-LBP coding method for comparing the gray value of the centrally symmetrical sampling points, but at the same time must clearly identify the side with the higher gray value of each group of centrally symmetrical sampling points to obtain More specific gradient direction information, that is, the number of binary codes has been extended to 8 bits. As the preliminary gradient direction feature map output in this link, the pixel value is no longer converted to decimal, but the binary form of the encoding method is directly output.

Figure 2 shows the process of calculating the preliminary gradient direction using the improved CS-LBP algorithm. The number of samples in the circle is P=8, the sampling radius R=1, and the threshold T=10. Use the bilinear interpolation method to first determine the gray value of all sampling points in the circular neighborhood; then calculate the gray value difference of each group of centrally symmetrical sampling points in turn; finally compare the relationship between the difference and the threshold T to determine the gradient direction. If the absolute value is greater than the threshold T, the count bit corresponding to the sampling point with a large gray value is marked as 1, and the count bit is the gradient direction of this group of sampling pixel pairs. At the same time, the count bit corresponding to the sampling point with a small gray value is marked as 0. If the absolute value is not greater than the threshold T, the count bits corresponding to the two sampling points are marked as 0, indicating that this group of sampling pixel pairs has no gradient. After calculation, it can be concluded that there are only three gradient directions of 45°, 90°, and 315° in Figure 2, so only the count bits corresponding to these three directions are marked as 1, and the other direction bits are all marked as 0. Therefore, starting from 0°, it is coded counterclockwise, and 10000110 is the binary code value sought. Repeat the above operations on the entire Gaussian filtered image to obtain a complete preliminary gradient direction feature map. The value of each pixel uniquely corresponds to a one-dimensional array composed of eight binary numbers 0 or 1.

Gradient direction feature map:

In the obtained preliminary gradient direction feature map, each one-dimensional array pixel value has 81 (3 ⁴ ) binary arrangements. If the continuous gradient feature is directly extracted from this feature map, the obtained descriptor feature dimension will be too large. Therefore, the described method proposes an effective dimension reduction mode, which can reduce the total number of binary pixel value arrangements on the one hand, and realize the scale transformation of the feature map on the other hand.

The key gradient direction is used to represent the overall gradient trend of the pixel block. The purpose is to convert the original 81 values of a single pixel into 8 effective values of the pixel block, which is convenient for encoding continuous gradient features. Sorting according to the gradient azimuth map in Figure 3a, the numbers 1 to 8 distributed in the counterclockwise direction correspond to 8 gradient directions respectively, and after accumulating the binary pixel values of all pixels in the pixel block according to the same count bit one by one, the count sum is the largest The counting bit is the key gradient direction of the pixel block. The Cell unit in Figure 3b is a pixel block composed of four pixels a, b, c, and d. Taking the key gradient direction feature of this Cell as an example, Figure 4 shows the calculation process of the key gradient. First, the improved CS-LBP algorithm is used to sequentially calculate the one-dimensional array pixel values corresponding to these four pixels; and then the four one-dimensional pixel values are accumulated one by one according to the same counting position, and it can be found that the count in the 90° direction is the largest, and the Four times; finally, the gradient direction of the Cell is 90°, that is, the output code value of the key gradient direction is 2, and the calculation formula is defined as follows:

X _i ＝[x _i1 ,x _i2 ,…x _ij …,x _im ] (5)

Where X _i represents a one-dimensional array, Y represents the accumulation of all one-dimensional arrays X _i in the Cell structure, x _ij ∈ {0, 1}, i ∈ {1, n}, j ∈ {1, m}, i and j are all integers. The function max_id(Y) indicates the position corresponding to the maximum value in the output one-dimensional array Y, that is, the column coordinate j of the maximum value at this time. Repeat the above steps for the entire preliminary gradient direction feature map, take the stride as 2, and then obtain the gradient feature map after the first downsampling.

The same iteration is performed on the feature map after the first downsampling to obtain the second downsampling feature map. The length and width of the Cell structure are set to 2×2, the number of channels is 8, and the sliding step size is 2, which not only facilitates the setting of relevant parameters during continuous gradient feature extraction, but also the length and The width is in multiples of the size of the smallest partition sub-image obtained by the traditional LBP and its variant algorithms, which is convenient for comparison. In addition, it is also different from the maximum pooling in deep learning, here is for value suppression. Figure 5 is a schematic diagram of the output results of the downsampling process. Except for the input image is a single-channel grayscale image, the feature maps of other links are 8-channel, and each channel represents a gradient direction in Figure 3a. After scale transformation, the typical gradient features at the target contour will be preserved, but some inflection points, edges, and noise that do not change significantly will be removed during the downsampling process.

Continuous gradient features:

Some deep mining operators of LBP and this series of fusion feature operators basically only analyze the texture characteristics of the gray value difference of local sampling points, which has obvious limitations. Therefore, the method extracts continuous gradient features on the feature map after twice downsampling, and improves the quality of the descriptor through the gradient change relationship of the target contour at the inflection point and the edge. In order to distinguish it from the Cell unit in the downsampling process, a 3×3 window is defined as a Block when extracting continuous gradient features on the sampling feature map, the number of channels is 8, and the sliding step is 1. The reason why the processing unit chooses 3×3 is because the 8 field sampling points are adjacent to the central sampling point, which can better reflect the key feature of gradient continuity, and is more in line with coding habits.

Perform circular sorting in the counterclockwise direction of the gradient orientation map in the block, and count the maximum continuous sequence length of non-center pixels of all reference channels and adjacent channels. First, determine the channel with the central pixel one by one and use this channel as a reference. If there is no central pixel in all channels, the block will not output continuous gradient features. In Figure 6, only channels 2 and 4 have central pixels, so feature extraction is only performed on these two channels. The maximum continuous sequence length of the 8 neighboring sampling pixels in channel 2 is 4, the maximum continuous sequence length of adjacent channel 1 is 1, and the maximum continuous sequence length of channel 3 is 2, so the continuous gradient feature of channel 2 is obtained, and the encoding The result is 2412, which means that based on channel 2, the maximum continuous sequence lengths of channels 2, 1, and 3 are 4, 1, and 2, respectively. Since the maximum continuous sequence length of channel 4 itself is less than 3, this channel also does not output features. The calculation steps and constraints of continuous gradient features are summarized into the following three points:

1) In the 8-channel Block, first determine the channel with the central pixel, if there is no central pixel, no feature is output.

2) Based on the channel with the central pixel, judge whether the maximum continuous sequence length in its neighborhood is not less than 3, and if not, then do not output the feature.

3) Count the maximum continuous sequence lengths of the channels adjacent to the reference channel respectively. The encoding form of the continuous gradient feature output is the reference channel number and the maximum continuous sequence length of the three corresponding channels. The total number of encodings for this strategy is 768.

experiment procedure:

The CPU used in the experimental running environment is Intel Xeon E5-2697 v2, the running memory is 98GB, and the hard disk storage space is 8TB. Leaky cable buckle image data is taken with a two-line CMOS camera throughout the whole process, including 195 faulty buckle images and 900 normal buckle images. 70% of the input image data set is used as a training set, and the remaining 30% is used as a test set. Figure 7 briefly describes the complete process from obtaining continuous gradient features from raw data to using SVM to complete the fault snap detection task. After preprocessing, the size of the input image becomes 256×256, the basic unit for extracting preliminary gradient direction features and continuous gradient features is 3×3, and the basic unit for twice downsampling is 2×2.

Calculate the adaptive threshold T:

In order to obtain a suitable threshold T, a total of 50 images with 5-6 global gray mean values distributed in different intervals were selected from the data set, and the corresponding relationship between gray and threshold was analyzed through experiments. Use a 9×9 Gaussian filter to preprocess the original input image, and then compare the imaging effects of their preliminary gradient direction feature maps under different threshold conditions. Under the premise of meeting the following two requirements, the threshold T should be as large as possible big.

1) It must be ensured that the preliminary gradient direction feature map can completely and clearly show the outline of the leaky cable buckle.

2) To a certain extent, it can filter the background noise of the tunnel wall.

Table 1 Analysis of imaging effects of different gray levels-thresholds

In order to facilitate the comparison of the imaging effects of different gray levels and thresholds, the global mean gray level is divided into 8 intervals, corresponding to 8 threshold levels. Table 1 lists the imaging effect of the preliminary gradient direction feature map of some snapping images under different thresholds. The specific calculation formula is:

When the global average gray value is between 80 and 220, use the rounding function to calculate the corresponding threshold T; if the average gray value is less than 80, it is the lowest threshold level T=4; if the average gray value is greater than 220, it is the highest threshold level T=11 . According to the formula (8), the improved CS-LBP algorithm is used to calculate the preliminary gradient direction features one by one for the input image, and the corresponding preliminary gradient direction feature map is obtained.

Compute continuous gradient features:

Some typical features such as the inflection point or edge of the leaky cable buckle are not obvious in the feature map of the third downsampling, or a large number of continuous gradient features are no longer expressed in the feature map of the third and subsequent sampling, so Downsampling was performed only twice during the experiment. The continuous gradient features are extracted from the feature maps after twice downsampling respectively, and the two sets of continuous gradient features are concatenated to obtain the final fusion feature descriptor with a feature dimension of 1536, and there is no need to divide the image into blocks.

Experimental results and analysis:

The experiment uses the recall rate Recall, precision precision, feature dimension Bin and AUC value of the descriptor as the evaluation indicators of the detection results. Among them, Recall and Precision respectively refer to the proportion of buckles correctly judged as faults to all faulty buckles, and the percentage of buckles correctly judged as faults to all buckles judged to be faulty, respectively. The specific calculation formula is:

In order to prove the superiority of the method algorithm in extracting the profile features of leaky cable clips, it is also compared with various variant algorithms such as CS-LBP and HOG, and the experimental results in Table 2 are obtained. The parameter settings of these comparison algorithms The situation is as follows:

1) LBP sampling radius R is 1, sampling number P is 8, and bilinear interpolation method is used for sampling. In the experiment, the image needs to be divided into 4×4 without overlapping, and the feature dimension of the descriptor is 4096.

(完备的二值模式)用双线性插值法进行采样，采样半径R＝1，采样个数P＝8。每一个采样圆域内的局部符号算子CLBP_S和局部幅值算子CLBP_M均使用旋转不变均匀模式编码。实验中该方法对图像进行4×4无重叠划分，每个分区子图像的特征维度为2(P+2)²＝200，输出的描述子特征维度为3200，在后面的表格和图中将以CLBP代指该方法。2)

(Complete binary mode) Sampling is performed by bilinear interpolation method, sampling radius R=1, sampling number P=8. The local sign operator CLBP_S and the local magnitude operator CLBP_M in each sampling circle are encoded using the rotation-invariant uniform mode. In the experiment, this method divides the image into 4×4 without overlapping. The feature dimension of each sub-image is 2(P+2) ² =200, and the output descriptor feature dimension is 3200. In the following tables and figures, This method is denoted by CLBP.

3) HOG features use the nearest neighbor interpolation method to first reduce the size of the input image to 128×128, and then perform feature extraction. The Cell is 8×8, the Block is 16×16, and the Stride is 16. The bins of the gradient component in each Block unit are 9, so the feature dimension of the output descriptor is 2304.

(4) Both TPLBP and ARCS-LBP use bilinear interpolation method for sampling, the sampling radius R=1, the number of samples P=8, the image needs to be divided into 4×4 without overlapping in the experiment, and the feature dimensions of the descriptors are both for 4096.

5) HOG_LBP and HOG_CLBP refer to concatenating the previously obtained HOG descriptor with the LBP descriptor and CLBP descriptor respectively to obtain the fusion features of the two, and the feature dimensions of the corresponding descriptors are directly accumulated.

Through the comparison of the detection results in Table 2 and the ROC curve in Figure 8, it can be seen that the multi-scale continuous gradient feature algorithm MSG-LBP proposed by the method has a significant advantage in the detection results of leaky cable buckles. Among them, Recall and Precision reached 0.923 and 0.857 respectively, and the AUC value of MSG-LBP is also the highest in the ROC curve. Even though the fusion feature of HOG_LBP can make the Recall reach 0.889, the Precision is only 0.615, which shows that a large number of normal snaps are misjudged as faulty snaps. The MSG-LBP descriptor does not divide the image into blocks, but the feature map of the first layer of downsampling is consistent with the size of the smallest partition sub-image of the image divided by 2×2, and the feature map of the second layer of downsampling is 4 The sub-images of the smallest partition divided by ×4 have the same size, which greatly reduces the computational load of feature extraction. The two downsampled feature maps in the experiment not only contain the key gradient information of the snap contour feature, but also realize the scale transformation of the snap contour feature. The dimensionality of the multi-scale continuous gradient feature descriptor is lower than that of most of the deep mining descriptors and fusion feature descriptors of this series, and the detection ability is obviously stronger.

Table 2 Comparison of test results

In summary, the significant advantage of the described method compared with this series of algorithms is that the scale transformation of the buckle contour features is realized through continuous downsampling feature maps, which greatly improves the feature extraction efficiency and also reduces the number of descriptors. The role of feature latitude. At the same time, the continuous gradient feature can also more accurately reflect the contour features of the leaky cable buckle, especially the gradient information at the inflection point and the edge.

Claims (7)

1. A high-speed rail tunnel leaky cable buckle detection method is characterized in that comprising the steps: Improve the CS-LBP algorithm and calculate the gradient direction characteristics of the image sampling circle of the leaky cable buckle of the high-speed railway tunnel one by one, and obtain a complete preliminary gradient direction feature map; The steps of the improved CS-LBP algorithm are as follows: the number of samples in the image sampling circle is P=8, the sampling radius R=1, and the threshold T=10; the gray values of all sampling points in the circular neighborhood are first determined by bilinear interpolation method. Then calculate the gray level difference of each group of centrally symmetrical sampling points in turn; finally compare the relationship between the gray level difference and the threshold T to determine the direction of the gradient; If the absolute value of the grayscale difference is greater than the threshold T, the counting bit corresponding to the sampling point with a large grayscale value is marked as 1, and this counting bit is the gradient direction of this group of sampling pixel pairs; at the same time, the grayscale value is small The count bit corresponding to the sampling point is marked as 0; if the absolute value of the gray difference value is not greater than the threshold T, the count bit corresponding to the two sampling points is marked as 0, indicating that this group of sampling pixel pairs has no gradient, and the entire Gaussian The filtered image repeats the above operations to obtain a complete preliminary gradient direction feature map, and the value of each pixel uniquely corresponds to a one-dimensional array composed of eight binary numbers 0 or 1; Perform two continuous downsampling on this feature map and extract the continuous gradient features of the two sampled images respectively; The two groups of gradient features are connected in series as descriptors, and SVM is used to complete the classification of fault snap images. 2. The high-speed rail tunnel leaky cable buckle detection method as claimed in claim 1, is characterized in that: the calculation method of key gradient feature is as follows: The key gradient direction is used to represent the overall gradient trend of the pixel block, sorted according to the gradient azimuth map, and the numbers 1~8 distributed in the counterclockwise direction correspond to 8 gradient directions respectively, and the binary pixel values of all pixels in the pixel block are divided into the same After the count bits are accumulated one by one, the count and the largest count bit are the key gradient direction of the pixel block. 3. the high-speed rail tunnel leaky cable buckle detection method as claimed in claim 2, is characterized in that, Cell unit is the pixel block that is made up of a, b, c, d four pixels, obtains the key gradient direction characteristic of this Cell Methods as below: First, the improved CS-LBP algorithm is used to sequentially calculate the one-dimensional array pixel values corresponding to these four pixels; then the four one-dimensional pixel values are accumulated one by one according to the same counting position, and the gradient direction with the most counting direction is the key Gradient direction; finally get the gradient direction of the Cell and the output encoding value of the key gradient direction, the calculation formula is defined as follows:

, Where X _i represents a one-dimensional array, Y represents the accumulation of all one-dimensional arrays X _i in the Cell structure, x _ij ∈ {0, 1}, i ∈ {1, n}, j ∈ {1, m}, i and j are all integers; the function

Indicates the position corresponding to the maximum value in the output one-dimensional array Y , that is, the column coordinate j of the maximum value at this time; repeat the above steps for the entire preliminary gradient direction feature map, and take the stride as 2, so the first downsampling is obtained After the gradient feature map; The same iteration is performed on the feature map after the first downsampling to obtain the second downsampling feature map. 4. The method for detecting leaky cable snaps in high-speed railway tunnels as claimed in claim 3, characterized in that: the length and width of the Cell structure are set to 2×2, the number of channels is 8, and the sliding step stride is 2, which can facilitate The setting of relevant parameters during continuous gradient feature extraction, and the length and width of each layer of downsampled feature maps are in multiples of the size of the smallest sub-image obtained by traditional LBP and its variant algorithms. 5. The high-speed railway tunnel leaky cable buckle detection method as claimed in claim 1, characterized in that: when extracting continuous gradient features on the sampling feature map, define a 3 × 3 window as a Block, the number of channels is 8, and the sliding step is is 1. 6. The high-speed rail tunnel leaky cable buckle detection method as claimed in claim 5, characterized in that: in the Block, do circular sorting in the counterclockwise direction of the gradient azimuth map, and count the maximum values of all reference channels and adjacent channel non-central pixels For the length of the continuous sequence, first determine the channel with the central pixel one by one and use this channel as a reference. If there is no central pixel in all channels, the block will not output continuous gradient features. 7. The high-speed rail tunnel leaky cable buckle detection method as claimed in claim 5, characterized in that: the calculation steps and limiting conditions of the continuous gradient feature include the following three points: In the 8-channel Block, first determine the channel with the central pixel, and if there is no central pixel, the feature will not be output; Based on the channel with the central pixel, judge whether the maximum continuous sequence length in its neighborhood is not less than 3, and if not, then do not output the feature; The maximum continuous sequence length of the channel adjacent to the reference channel is counted separately. The encoding form of the continuous gradient feature output is the reference channel number and the maximum continuous sequence length of the three corresponding channels. The total number of encodings of this strategy is 768.

Images