CN111860501A - Image recognition method of high-speed rail height adjustment rod falling out fault based on shape matching - Google Patents

Abstract

An image recognition method for a high-speed rail height adjustment rod falling out fault based on shape matching belongs to the technical field of high-speed rail detection. The present invention is to solve the problem that the existing detection method has a long detection time and cannot take into account the problem of the detection time and the accuracy. The invention firstly extracts the image including the detection part area in the image, and extracts the corresponding template image and creates the image feature pyramid, respectively carries out edge extraction and most discrete sampling of the feature points; extracts the feature point gradient information from the feature points; The coarse positioning result obtained by establishing the feature pyramid of the image, the similarity between the image edge points of the template image and the image edge points of the search image is calculated, and the original template contour is converted into a corrected contour line; then the least squares method is used for further correction, and finally based on the template Images and search results, through logic analysis to make fault determinations on finely positioned component images. Mainly used for image recognition of high-speed rail height adjustment rods falling out.

Description

Image recognition method of high-speed rail height adjustment rod falling out fault based on shape matching

technical field

The invention relates to an image recognition method for a high-speed rail height adjustment rod falling out fault. It belongs to the technical field of high-speed rail detection.

Background technique

High-speed trains have the characteristics of fast speed and stable operation, and have become one of the main means of transportation for people at present. The operation safety of high-speed rail is the top priority, so it needs special attention. The disengagement of the height adjustment rod is a kind of fault that endangers the driving safety, which will lead to serious consequences such as the rollover of the train during the traveling process.

In traditional fault detection methods, fault detection is usually performed by manually checking images. Due to the fact that the inspectors are prone to fatigue and omissions during their work, resulting in missed inspections and wrong inspections, affecting driving safety.

In view of the above problems, with the rapid development of image processing technology and deep learning technology, the automatic image recognition detection method can be introduced to detect the height adjustment rod coming out fault. The automatic image recognition method has the advantages of high detection efficiency and good stability. . When using neural network processing technology for detection and detection, if you want to ensure a high recognition accuracy, you must build a network model and train the model in a targeted manner, which not only requires a lot of manpower and material resources, and may not be able to achieve better results. Moreover, the detection time is relatively long when the neural network is used for detection, which needs to be further improved. At the same time, if a better effect is to be achieved, the depth of the model is generally required to be larger, which will further extend the detection time, and at the same time, the hardware requirements are relatively high.

SUMMARY OF THE INVENTION

The present invention is to solve the problem that the existing detection method has a long detection time and cannot take into account the problem of the detection time and the accuracy.

An image recognition method for high-speed rail height adjustment rod disengagement faults based on shape matching, including the following steps:

s1. Collect an image and extract an image containing the detection component area; and extract a template image corresponding to the detection component area;

s2. Create an image feature pyramid for the template image;

s3. After the number of pyramid layers is determined, for each layer of image feature pyramids, edge extraction is performed respectively and feature points are most discretely sampled; feature point gradient information is extracted for feature points;

s4. On each layer of image feature pyramid, rotate the extracted most discrete edge points;

s5. Build a feature pyramid for the search image;

Take the template image of the highest level of the pyramid and perform a full-angle sliding window traversal at the highest level of the search image pyramid; take the one with the highest similarity as the matching result; map this result to the next level of the pyramid until the first level of the image feature pyramid, that is Image original resolution layer to complete coarse positioning;

s6. For the coarse positioning result obtained after the pyramid search, perform affine transformation on the result of the rough positioning of the template image contour point to the corresponding point on the search image as P, and compare the main gradient of point P and the edge point of the search image. The main gradient is used to calculate the similarity, and the point on the search image with the highest similarity is used as the edge point A, and then P is replaced by A, and the original template outline is converted into a corrected outline;

s7. Use the least squares method for further correction, and the correction takes the shortest distance from the feature point to the feature line as the optimization goal to iterate; the feature point is the edge point P of the template image obtained after the edge point of the template image is deformed and corrected; the feature line is The tangent corresponding to the adjacent edge point on the search object closest to the point P on the search image;

s8. Based on the template image and the search result, perform fault determination on the finely positioned component image through logical analysis.

Preferably, the similarity calculation described in step s6 is calculated based on the following similarity measure function;

表示d′_i的转置；e_i′为搜索图片上被模板覆盖的子区域中i的对应点i′的梯度；n表示点i的总数；(t′_i，u′_i)及(v，w)分别为d′_i与e_i′的x、y方向上的梯度分量。Among them, d' _i is the gradient of point i in the template image,

represents the transposition of d'_i; e _i' is the gradient of the corresponding point i' of i in the sub-region covered by the template on the search image; n represents the total number of points i; (t' _i , u' _i ) and (v , w) are the gradient components in the x and y directions of d′ _i and e _i′ respectively.

Preferably, the process of creating an image feature pyramid for the template image in the step s2 includes the following steps:

a. Perform feature extraction on the input image to obtain a feature map;

b. Set the initial number of pyramid layers L=0;

c. For the feature map, create an image feature pyramid layer by layer starting from the L layer;

d. Perform feature sampling on the current layer pyramid image according to the initial sampling rate ratio;

e. If the number of edge points N is greater than the first threshold and less than the second threshold, jump out of the process and use the current pyramid level;

f. If the number of edge points N is greater than the second threshold, update L to L+1, and jump to step c;

g. If the number of edge points N is greater than 0 and less than the first threshold, then L=0, at the same time divide the sampling rate by 1.5, and jump to step c;

h. Loop steps e to g until step e is satisfied.

Preferably, the first threshold is 40, and the second threshold is 60.

Preferably, the process of performing fault determination on the finely positioned component images through logical analysis described in step s8 includes the following steps:

1) Obtain subgraphs according to prior information such as wheelbase;

2) Match the height adjustment rod and the adjustment rod mounting base block respectively;

3) Calculate the coordinates of the lower edge of the adjustment rod (x1, y1) and the coordinates of the lower side of the mounting block (x2, y2), and the installation height;

4) Calculate the ratio ratio=(y2-y1)/height;

5) If ratio>0.3, it is judged as failure.

Preferably, step s2 uses the maximum value pyramid for the image feature pyramid described in creating the image feature pyramid for the template image.

Preferably, the specific process of extracting feature point gradient information for feature points described in step s3 includes the following steps:

For the coordinates of feature points, gradient quantization is performed on 9 points in 3 neighborhoods including the current feature point;

Gradient quantization refers to quantizing the gradient direction into 8 representative directions, namely 11.25°, 33.75°, 56.25°, 78.75°, 101.25°, 123.75°, 146.25° and 168.75°;

For the current feature point, adjust the gradient directions of 9 points in 3 neighborhoods around its coordinates to 8 quantization directions. If the gradient direction of the current point falls within the range of 0° to 22.5°, use 11.25° as a representative; If the gradient direction of the current point falls within the range of 22.5° to 45°, use 33.75° as the representative; if the gradient direction of the current point falls within the range of 45° to 67.5°, use 56.25° as the representative; if the gradient direction of the current point falls In the range of 67.5° to 90°, use 78.75° as a representative; if the gradient direction of the current point falls within the range of 67.5° to 90°, take 22.5° as a unit, and for the range of 90° to 180°, use 8 The specific values in each representative direction are regularized;

For the current feature point, gradient transfer is performed:

Gradient transfer is to merge the gradient directions of the 9 neighbor points to the center point, and take the average direction after the convergence as the final gradient direction.

Preferably, the process of rotating the extracted most discrete edge points described in step s4 includes the following steps:

The rotation step size is set to 0.5° for the bottom layer, and it is multiplied by 2 as the number of layers increases. If it exceeds 8°, it remains unchanged at 8°.

Preferably, the characteristic line in s7 is the tangent line corresponding to the adjacent edge point on the search object closest to the point P on the search image. The process of determining the adjacent edge point on the search object closest to the point P includes the following steps:

Draw a rectangle with the transformed template image edge point P as the center, traverse all the search image edge points within the rectangle, calculate the distance between each edge point and point P, and take the point with the smallest distance as the closest edge point A;

or,

Taking the transformed template image edge point P as the starting point, the intersection of the straight line where the gradient of this point is located and the edge of the search image is taken as the nearest edge point B.

Preferably, before creating the image feature pyramid, the template image and the search image need to be preprocessed, and the preprocessing process includes dilation, erosion, and filtering morphological processing.

Beneficial effects:

1. Introduce automatic identification technology into high-speed rail fault detection to realize automatic fault identification and alarm, and only need to confirm the alarm results manually, which effectively saves labor costs and improves operation quality and operation efficiency.

2. The shape-based template matching algorithm is applied to the automatic identification of the failure of the height adjustment rod of the high-speed rail, which has higher accuracy and stability than the traditional machine vision detection method. At the same time, the invention has the advantages of fast speed and strong stability, and can effectively improve the detection accuracy while ensuring the detection speed.

3. The present invention improves the original algorithm and adds a local deformation correction and matching method, which can effectively solve the problem of unrecognizable image deformation caused by changes in vehicle speed or different camera angles of view.

4. The present invention improves the original algorithm, adds the function of automatically determining the number of feature pyramid layers, and further improves the degree of recognition automation.

Description of drawings

Fig. 1 is a schematic diagram of the search process;

Figure 2 is a schematic diagram of gradient quantization;

Figure 3 is a schematic diagram of gradient transfer;

Fig. 4 is the mapping schematic diagram of template outline;

Figure 5 is a schematic diagram of a deformation correction process;

6 is a schematic diagram of an image feature pyramid;

Fig. 7 is a schematic diagram of a comparison diagram of the effect of the mean value pyramid and the maximum value pyramid;

8 is a schematic diagram of the closest point acquisition;

Figure 9 is a schematic diagram of the identification process.

Detailed ways

Embodiment 1: This embodiment is described with reference to FIG. 9 ,

The shape matching-based image recognition method for a high-speed rail height adjustment rod disengagement fault described in this embodiment includes the following steps:

1. Line array image acquisition:

Build high-definition equipment around the train tracks, and obtain high-definition train images after the train passes through the equipment. The line scanning method can form a two-dimensional image with a wide field of view and high precision.

2. Interception of candidate regions to be detected:

According to the train wheelbase information and vehicle type information, the two-dimensional picture is intercepted, and the local area image including the detection component (height adjustment rod) is intercepted from the train image as the detection candidate area image, which can effectively reduce the time required for fault identification and improve the accuracy of identification. Rate.

3. Detection of detection components (height adjustment rod):

In general, template matching is a search process in which the template image of the detection component (height adjustment rod) is compared with the sub-region of the search image, and the sub-region in the search image that is most similar to the template is used as the final matching result. The search process is shown in the figure. 1 shown. The whole process is divided into two stages: offline stage and online stage.

Among them, the creation of the template of the object to be recognized in the offline stage requires manual design of image features, and the template image is required to be clean and tidy, with obvious features and clear outlines. In some embodiments, the difference between the background grayscale and the foreground grayscale can be 50-100; online The stage is the automatic identification stage, and human intervention cannot be carried out, and it is required to be real-time.

Shape-based template matching focuses on the shape information of objects, uses the gradient of edge points as basic data for similarity comparison, and defines a similarity metric function as an evaluation standard for similarity, which can solve uneven illumination changes, obvious contrast changes, and strong noise. , there are problems such as foreign objects and occlusion, and it has high robustness; the most discrete sampling of edge points and the step-by-step matching strategy based on image feature pyramid from coarse to fine have high real-time performance; The pose precision correction method has high matching accuracy. The specific identification process and improvement are as follows:

1. Definition of similarity measure function:

Taking the similarity metric function as the similarity evaluation criterion, the shape-based template matching similarity metric function is as follows:

表示d‘_i的转置；e_i′为搜索图片上被模板覆盖的子区域中i的对应点i′的梯度；n表示点i的总数；(t′_i，u′_i)及(v，w)分别为d′_i与e_i′的x、y方向上的梯度分量。Among them, d' _i is the gradient of point i in the template image,

represents the transposition of d'_i; e _i' is the gradient of the corresponding point i' of i in the sub-region covered by the template on the search image; n represents the total number of points i; (t' _i , u' _i ) and (v , w) are the gradient components in the x and y directions of d′ _i and e _i′ respectively.

2. Obtaining the template image:

The obtained high-definition two-dimensional images are screened, and an ideal image with no foreign matter, no occlusion, and clear outline is selected as a template.

3. Basic morphological processing:

The template image is subjected to morphological processing such as expansion, corrosion, and filtering to remove the influence of holes and noise; the image to be detected of the detection component is used as the search image, and the same processing is also performed.

4. Create a feature pyramid:

Create an image feature pyramid on the template image. The image feature pyramid is a tower-like structure obtained by continuously down-sampling the image and stacking it layer by layer from large to small, as shown in Figure 6. The image feature pyramid can reduce the complexity of the algorithm on an exponential scale and greatly improve the computational efficiency. The image feature pyramid has features that the higher the number of layers, the lower the resolution and the blurrier the shape information; the lower the number of layers, the higher the resolution and the more obvious the shape information. Based on this feature, it is required that the highest level of the image feature pyramid still has obvious contour features.

The traditional method determines the number of pyramid layers by manual input. This method has the following disadvantages:

On the one hand, for objects of different sizes and shapes, a single tower building layer is no longer applicable. For example, for large objects, a 5-layer pyramid is suitable, but for small objects, only 3 layers may be needed. If a 5-layer pyramid is built for small objects, the basic outline information of objects on the top layer of the pyramid may have disappeared; on the other hand, due to the inapplicability of a single tower layer, manual adjustment is required, which undoubtedly interrupts the production line. Streamlined operations reduce the degree of automation.

Based on the above problems, the present invention proposes a method for automatically determining the number of pyramid levels based on the number of top-level feature points. The basic process is as follows:

a. Perform feature extraction on the input image to obtain a feature map;

b. Set the initial number of pyramid layers L=0;

c. For the feature map, create an image feature pyramid layer by layer starting from the L layer;

d. Perform feature sampling on the current layer pyramid image according to the initial sampling rate ratio;

e. If the number of edge points N is greater than the first threshold (preferably 40) and smaller than the second threshold (preferably 60), jump out of the process and use the current pyramid level;

f. If the number of edge points N is greater than the second threshold (preferably 60), then L+=1 (update L to L+1), and jump to step c;

g. If the number of edge points N is greater than 0 and less than the first threshold (preferably 40), then L=0, at the same time divide the sampling rate by 1.5, and jump to step c;

h. Cycle step c to step g until step e is satisfied.

The present invention automatically obtains the number of image feature pyramid layers, and has the following advantages compared to the prior art:

First, the traditional method is to manually set the number of pyramid layers, and there are many problems in setting the same number of layers for images of different sizes. For example, using the scaling parameters of large-sized images for small-sized images will cause the small-sized images to lose texture information, and using the scaling parameters of small-sized images for large-sized images cannot achieve an ideal scaling ratio and cannot maximize acceleration.

The method for automatically determining the number of image feature pyramid layers proposed by the present invention is to decide whether to use the current layer as the highest layer of the image feature pyramid according to the number of feature points in each layer of the feature pyramid. Moreover, the parameter of the feature point screening ratio is further proposed. Although the complexity is higher, the obtained pyramid layers and the number of feature points on them are more reasonable. Moreover, in the process of edge extraction and the most discrete sampling of feature points, the number of feature points of the current layer is screened, and the most discrete N points that can represent the shape information of the object (the most discrete refers to the largest degree of discreteness) are screened. Reduce the calculation amount of the similarity measure, on the other hand, remove the influence of redundant feature points (to get the matching position of an object does not require all feature points to participate in the calculation), use a small number of but more refined and more accurate feature points Information to determine the number of pyramid levels, the number of levels obtained will be more accurate.

Second, the method for automatically determining the number of pyramid layers proposed by the present invention is self-comparative, that is, only the feature information of the image itself is referred to, and no comparison with other images is performed, so that the pictures can be independent, and each picture retains its own original integrity The obtained image texture information, the obtained layer number information is very suitable for its own image characteristics.

Third, the present invention changes the traditional mean value pyramid (the average value of adjacent four pixels is taken as the corresponding point of the next layer of pyramid) into the maximum value pyramid (the most value of adjacent four pixels is taken as the corresponding point of the next layer of pyramid), so do to preserve the shape information to the greatest extent possible. The comparison of the effect of the mean pyramid and the maximum value pyramid is shown in Figure 7. It can be clearly seen from the figure that the maximum value pyramid proposed by the present invention still has clear object outlines on the third and fourth layers, while the traditional mean pyramid is in Layers 3 and 4 can barely see the outline of objects.

5. After the number of pyramid layers is determined, for each layer of the image feature pyramid, edge extraction is performed and the feature points are most discretely sampled, and feature point gradient information is extracted from the feature points; the specific process includes the following steps:

5.1. Use an edge extraction algorithm, such as canny, to extract the edge of the picture.

5.2. Perform the most discrete sampling on the edge points, and extract N points that can represent the shape information of the object. The extracted feature points are required to be scattered throughout the outline of the object as much as possible. The most discrete feature point screening can remove redundant edge points, reduce the amount of matching calculation, and improve the matching speed.

5.3. Extract feature point gradient information for subsequent matching:

Traditional methods directly use gradient information on edge points as matching data. Due to the change of train speed and the difference of shooting angle, horizontal stretching and vertical deformation will inevitably occur, resulting in confusion of local edge gradients. Finding the correspondence between the edge points of the template and the edge points of the search image results in a low score of the final similarity measure, which eventually leads to missed detection. Traditional shape matching can no longer be applied. Based on this problem, the present invention proposes a local deformation matching method based on local gradient quantization and gradient transfer to solve the problem.

For the coordinates of the feature points, gradient quantization is performed on 9 points in the 3 neighborhoods including the current feature point (the nine-square grid in Figure 3 represents 9 points in the 3 neighborhoods);

Gradient quantization refers to quantizing the gradient direction into 8 representative directions, namely 11.25°, 33.75°, 56.25°, 78.75°, 101.25°, 123.75°, 146.25° and 168.75°, which are the positions pointed by the solid line in Figure 2;

For the current feature point, adjust the gradient directions of 9 points in 3 neighborhoods around its coordinates to 8 quantization directions. If the gradient direction of the current point falls within the range of 0° to 22.5°, use 11.25° as a representative; If the gradient direction of the current point falls within the range of 22.5° to 45°, use 33.75° as the representative; if the gradient direction of the current point falls within the range of 45° to 67.5°, use 56.25° as the representative; if the gradient direction of the current point falls In the range of 67.5° to 90°, use 78.75° as a representative; if the gradient direction of the current point falls within the range of 67.5° to 90°, take 22.5° as a unit, and for the range of 90° to 180°, use 8 The specific values in each representative direction are normalized; the specific quantification method is shown in Figure 2.

For the current feature point, gradient transfer is performed:

Gradient transfer is to merge the gradient directions of 9 neighboring points (the nine-square grid in Figure 3) to the center point, and take the average direction after the convergence as the final gradient direction. The specific transfer method is shown in Figure 3.

Through gradient quantization and gradient transfer, the gradient in the local range can be represented by a main gradient, which eliminates the influence of the gradient direction confusion caused by local deformation (reduces the similarity measure), so that the subsequent matching process has a good matching effect.

In the process of gradient quantization and transfer, in addition to recording the main gradient, each original real gradient is also recorded, which is recorded as the component gradient; the component gradient is used in the subsequent steps to determine the tangent of the edge point on the search object;

After the rough positioning operation is completed, the template contour can be mapped to the search image according to the affine transformation operation. At this time, the template contour can be deformed and corrected to the actual search object contour according to the quantization gradient, and finally the deformation matching can be realized. The specific process is shown in Figure 4 and shown in Figure 5.

6. Rotation feature modeling:

On each layer of image feature pyramid, the most discrete edge points extracted are rotated to solve the matching problem of rotating objects. The rotation step size is set to 0.5° for the bottom layer, and it is multiplied by 2 as the number of layers increases. If it exceeds 8°, it remains unchanged at 8°. If it exceeds 8°, some instances will be lost due to excessive rotation, affecting the final matching stability.

7. Coarse-to-fine step-by-step search strategy based on image feature pyramid:

This stage belongs to the online automatic identification stage.

First, build a feature pyramid for the search image.

Take the template image of the highest level of the pyramid and perform a full-angle sliding window traversal at the highest level of the search image pyramid. The highest similarity is used as the matching result (x, y, θ). Map this result to the next level of the pyramid, that is, multiply the coordinates by 2, ie (2x, 2y, θ). Considering the instability of high-level information, the search area of the next layer is located in the range around the mapping coordinates (2x, 2y, θ), that is, the search area of the next layer (2x, 2y, θ) is expanded by a circle As the search area range, then continue searching within the area range, and take the maximum value of the results in the area range as the matching result of the current layer and continue to map to the next layer; this process is repeated until the first layer of the image feature pyramid, that is, the image The original resolution layer, complete the coarse positioning.

The process of locating the search area of the next layer in the range around the mapping coordinates (2x, 2y, θ) can be illustrated by the following embodiments:

Perform a full traversal search at the top layer of the pyramid (the layer with low resolution). If the size is 512x512, you need to slide the template 512x512 positions to get the result (x, y, θ), and then map the result to the next layer to get ( 2x, 2y, θ), called the anchor point, at this time the size of the current layer picture is (2048, 2048), the search range at this time can be set in the 3x3 range centered on the anchor point, considering the inaccuracy of coarse positioning If deta is 1, the template needs to traverse only 5x5, which is a very large reduction compared to 2048x2048.

As can be seen from Figure 4, the traditional matching method, namely deformation-free matching, the final result is a point coordinate (r, c) and a rotation angle θ. The template contour can be mapped to the search image through affine transformation. target location. After the mapping is completed, the correspondence between feature points and feature edges has been basically confirmed.

8. Perform deformation correction:

The deformation correction process is shown in Figure 5. It can be seen from Figure 5 that the template image contour has been mapped to the search image through affine transformation, but due to the existence of local deformation, the gradient direction of the feature points in the deformation area is confused and cannot be corresponded, as shown in Figure 5 5 in Component Gradient 1 - Component Gradient 6. The quantized gradient can be normalized to the main gradient by gradient transfer, see main gradient 1 and main gradient 2 in Figure 5. Using the main gradient 1 and the main gradient 2 for matching, the corresponding point P and the corresponding point A can be determined:

Perform affine transformation of the contour points of the template image through the rough positioning of the pyramid to the corresponding points on the search image, denoted as P, and calculate the similarity between the main gradient of point P and the main gradient of the edge points of the search image, and search for the highest similarity. The point on the picture is used as the edge point A; then P is replaced with A, so as to convert the original template outline (original edge 1 in Figure 5) into a corrected outline (corrected edge 1 in Figure 5).

Correcting other edge points (moving the deformation area window, correcting other deformed contours), finally achieves the template contour correction coincides with the actual contour, and completes the deformation correction matching.

Generally, the correction can be completed after 6-7 iterations, so that the corrected template outline basically coincides with the actual object outline.

9. Pose fine correction based on least squares:

After deformation correction, there is still a slight deviation between this result and the real result, which needs to be corrected. The present invention adopts the least square method for correction, which can obtain higher precision.

The correction takes the shortest distance from the feature point to the feature line as the optimization goal to iterate; the feature point is the position point obtained after the edge point of the template image is deformed and corrected, that is, the edge point P of the template image; the feature line is the distance on the search image. The tangent corresponding to the adjacent edge point A or B on the nearest search object of point P;

There are two ways to obtain the nearest edge point A or B of point P, see Figure 8:

Method 1: Draw a rectangle with the transformed template image edge point P as the center, traverse all search image edge points within the rectangle, calculate the distance between each edge point and point P, and take the point with the smallest distance as the closest edge point A .

Method 2: After the deformation is corrected, the correspondence between the transformed template edge and the search image edge has been confirmed. Therefore, taking the transformed template image edge point P as the starting point, the intersection of the straight line where the gradient of this point is located and the edge of the search image is taken as the nearest edge point B.

10. Based on the template image and search results, the fault determination is performed on the finely positioned component images through logical analysis;

The first is to identify the fault image of the height adjustment rod of the high-speed rail. In fact, the above process can be used for fault identification of the image of other parts of the high-speed rail or train, and only some details can be adjusted, for example: The preferred values in step e, step f, step g, etc. are all for the limitation of the high-speed rail height adjustment rod. At this time, the number of pyramid layers in the identification process of the high-speed rail height adjustment rod has a very good effect on the recognition effect. For the fault identification of the images of other components, adjustment can be made according to the solution of the present invention.

The specific embodiment is only an explanation and description of the technical solution of the present invention, and cannot be used to limit the protection scope of the right. Any changes made according to the claims and description of the present invention are only partial changes, which should still fall within the protection scope of the present invention.

Specific implementation two:

In the shape matching-based method for identifying a fault image of a height adjustment rod of a high-speed rail falling out, the process of performing fault determination on a finely positioned component image through logical analysis described in step 10 includes the following steps:

1) According to the prior information such as the wheelbase, the sub-graph (including the small graph of the component) is obtained;

2) Match the height adjustment rod and the adjustment rod mounting base block respectively;

3) Calculate the coordinates of the lower edge of the adjustment rod (x1, y1) and the coordinates of the lower side of the mounting block (x2, y2), and the installation height;

4) Calculate the ratio ratio=(y2-y1)/height;

5) If ratio>0.3, it is judged as failure.

Other steps and parameters are the same as in the first embodiment.

It should be noted that the specific embodiments are only explanations and descriptions of the technical solutions of the present invention, and cannot be used to limit the protection scope of the rights. Any changes made according to the claims and description of the present invention are only partial changes, which should still fall within the protection scope of the present invention.

Claims (10)

1. The high-speed rail height adjustment rod disengagement fault image recognition method based on shape matching is characterized in that, comprises the following steps: s1. Collect an image and extract an image containing the detection component area; and extract a template image corresponding to the detection component area; s2. Create an image feature pyramid for the template image; s3. After the number of pyramid layers is determined, for each layer of image feature pyramids, edge extraction is performed respectively and feature points are most discretely sampled; feature point gradient information is extracted for feature points; s4. On each layer of image feature pyramid, rotate the extracted most discrete edge points; s5. Build a feature pyramid for the search image; Take the template image of the highest level of the pyramid and perform a full-angle sliding window traversal at the highest level of the search image pyramid; take the one with the highest similarity as the matching result; map this result to the next level of the pyramid until the first level of the image feature pyramid, that is Image original resolution layer to complete coarse positioning; s6. For the coarse positioning result obtained after the pyramid search, perform affine transformation on the result of the rough positioning of the template image contour point to the corresponding point on the search image as P, and compare the main gradient of point P and the edge point of the search image. The main gradient is used to calculate the similarity, and the point on the search image with the highest similarity is used as the edge point A, and then P is replaced by A, and the original template outline is converted into a corrected outline; s7. Use the least squares method for further correction, and the correction takes the shortest distance from the feature point to the feature line as the optimization goal to iterate; the feature point is the edge point P of the template image obtained after the edge point of the template image is deformed and corrected; the feature line is The tangent corresponding to the adjacent edge point on the search object closest to the point P on the search image; s8. Based on the template image and the search result, perform fault determination on the finely positioned component image through logical analysis. 2. The shape matching-based high-speed rail height adjustment rod disengagement fault image recognition method according to claim 1, wherein the similarity calculation in step s6 is calculated based on the following similarity measure function;

Among them, d' _i is the gradient of point i in the template image,

represents the transposition of d'_i; e _i' is the gradient of the corresponding point i' of i in the sub-region covered by the template on the search image; n represents the total number of points i; (t' _i , u' _i ) and (v , w) are the gradient components in the x and y directions of d′ _i and e _i′ respectively. 3. The high-speed rail height adjustment rod disengagement fault image recognition method based on shape matching according to claim 1, is characterized in that, the process that described step s2 creates image feature pyramid to template image comprises the following steps: a. Perform feature extraction on the input image to obtain a feature map; b. Set the initial number of pyramid layers L=0; c. For the feature map, create an image feature pyramid layer by layer starting from the L layer; d. Perform feature sampling on the current layer pyramid image according to the initial sampling rate ratio; e. If the number of edge points N is greater than the first threshold and less than the second threshold, jump out of the process and use the current pyramid level; f. If the number of edge points N is greater than the second threshold, update L to L+1, and jump to step c; g. If the number of edge points N is greater than 0 and less than the first threshold, then L=0, at the same time divide the sampling rate by 1.5, and jump to step c; h. Loop steps e to g until step e is satisfied. 4 . The shape matching-based image recognition method for a high-speed rail height adjustment rod disengagement fault according to claim 3 , wherein the first threshold is 40, and the second threshold is 60. 5 . 5. The shape matching-based high-speed rail height adjustment rod disengagement fault image recognition method according to claim 3, wherein the process of performing fault determination on the finely positioned component image by logical analysis described in step s8 comprises the following steps: 1) Obtain subgraphs according to prior information such as wheelbase; 2) Match the height adjustment rod and the adjustment rod mounting base block respectively; 3) Calculate the coordinates of the lower edge of the adjustment rod (x1, y1) and the coordinates of the lower side of the mounting block (x2, y2), and the height of the installation; 4) Calculate the ratio ratio=(y2-y1)/height; 5) If ratio>0.3, it is judged as failure. 6. The shape matching-based high-speed rail height adjustment rod prolapse fault image recognition method according to claim 1, 2, 3, 4 or 5, wherein step s2 creates the image feature pyramid described in the image feature pyramid for the template image Use the maximum pyramid. 7. The shape matching-based high-speed rail height adjustment rod disengagement fault image recognition method according to claim 6, wherein the specific process of extracting feature point gradient information to feature points described in step s3 comprises the following steps: For the coordinates of feature points, gradient quantization is performed on 9 points in 3 neighborhoods including the current feature point; Gradient quantization refers to quantizing the gradient direction into 8 representative directions, namely 11.25°, 33.75°, 56.25°, 78.75°, 101.25°, 123.75°, 146.25° and 168.75°; For the current feature point, adjust the gradient directions of 9 points in 3 neighborhoods around its coordinates to 8 quantization directions. If the gradient direction of the current point falls within the range of 0° to 22.5°, use 11.25° as a representative; If the gradient direction of the current point falls within the range of 22.5° to 45°, use 33.75° as the representative; if the gradient direction of the current point falls within the range of 45° to 67.5°, use 56.25° as the representative; if the gradient direction of the current point falls In the range of 67.5° to 90°, use 78.75° as a representative; if the gradient direction of the current point falls within the range of 67.5° to 90°, take 22.5° as a unit, and for the range of 90° to 180°, use 8 The specific values in each representative direction are regularized; For the current feature point, gradient transfer is performed: Gradient transfer is to merge the gradient directions of the 9 neighbor points to the center point, and take the average direction after the convergence as the final gradient direction. 8. The shape matching-based method for identifying a fault image of the height adjustment rod of a high-speed rail falling out, wherein the process of rotating the extracted most discrete edge points in step s4 comprises the following steps: The rotation step size is set to 0.5° for the bottom layer, and it is multiplied by 2 as the number of layers increases. If it exceeds 8°, it remains unchanged at 8°. 9. The high-speed rail height adjustment rod detachment fault image recognition method based on shape matching according to claim 7, is characterized in that, the characteristic line in s7 is corresponding to the adjacent edge point on the search object closest to the point P on the search image. The process of determining the adjacent edge points on the search object closest to the point P of the tangent line includes the following steps: Draw a rectangle with the transformed template image edge point P as the center, traverse all the search image edge points within the rectangle, calculate the distance between each edge point and point P, and take the point with the smallest distance as the closest edge point A; or, Taking the transformed template image edge point P as the starting point, the intersection of the straight line where the gradient of this point is located and the edge of the search image is taken as the nearest edge point B. 10 . The shape matching-based image recognition method for high-speed rail height adjustment rod disengagement faults according to claim 9 , wherein the template image and the search image need to undergo preprocessing before creating an image feature pyramid, and the preprocessing The process includes dilation, erosion, filtering and morphological processing.

Images