Catenary laser imaging identification method combining multilayer perceptron and image processing

Abstract

The invention relates to a contact network laser imaging identification method combining a multilayer perceptron and image processing. The method comprises the following steps: step 1, collecting samples; 2, carrying out threshold segmentation; step 3, splitting a binarization region; 4, closing a view blind area; step 5, merging contact network areas; step 6, dividing labels; 7, sample self-increasing; 8, model training; and 9, identifying the contact network. According to the railway catenary laser imaging recognition method combining the multilayer perceptron and the digital image processing technology, the catenary can be efficiently recognized in laser imaging, good stability is achieved under different working conditions, and the requirement for hardware cost is low; the method is high in practicability and has high popularization significance.

Description

Technical Field

[0001] The present invention relates to the technical field of train tracks, and in particular, to a catenary laser imaging recognition method combining a multi-layer perceptron and image processing. Background Art

[0002] The various geometric parameters of the catenary are important indicators for studying and analyzing the pantograph-catenary system. Currently, the mainstream on-vehicle non-contact geometric parameter measurement devices mainly use visual measurement technology. Its main light source is laser, which projects laser of a specific wavelength band onto the catenary, and the image is received and formed by a camera. The geometric parameter information of the catenary is calculated based on the change of the structured light of the catenary in the image. Therefore, accurately identifying and extracting the position of the catenary in the image is the key factor affecting the measurement accuracy of geometric parameters. However, in actual detection conditions, it is inevitable to have optical noise points or scan structured light images of non-catenary types. In the area outside the tunnel, it will also be affected by direct sunlight interference, resulting in irregular light spots, which affect the extraction of the catenary area by the program. Especially at characteristic positions such as riveted sections or turnouts on the line, it is necessary to simultaneously identify and extract double-branch catenaries, and the influence of optical noise and the structured light of other objects is particularly obvious. Summary of the Invention

[0003] Based on this, it is necessary to provide a catenary laser imaging recognition method combining a multi-layer perceptron and image processing for the deficiencies in the prior art.

[0004] A catenary laser imaging recognition method combining a multi-layer perceptron and image processing includes the following steps:

[0005] Step 1: Sample collection, collecting laser image samples containing the catenary through a camera;

[0006] Step 2: Threshold segmentation, finding the optimal threshold by the maximum inter-class variance method, and performing binary threshold segmentation on the sample image using the binary segmentation formula to separate the laser imaging part from the sample image. The binary region obtained after segmentation contains the laser structure information of the catenary, optical noise points, and other non-catenary objects;

[0007] Step 3: Splitting the binary region, using the labeling algorithm to split the binary region, and dividing the laser imaging part in the image into multiple sub-regions including catenary types and non-catenary types;

[0008] Step 4: Closing the field of view blind area, performing dilation operations on each sub-region using the dilation formula. After dilation processing, the catenary busbar sub-regions that were originally separated into multiple parts will intersect and have connectivity, which can close the field of view blind area generated when the pull-out value of the catenary is large;

[0009] Step 5: Merge the catenary regions. Mark multiple sub-regions of the same catenary busbar divided by the visual blind area as equivalent regions to achieve the merging of catenary regions;

[0010] Step 6: Divide the labels. Divide all sub-regions in the image into two types of labels: catenary type and non-catenary type;

[0011] Step 7: Sample increment. Perform small-range matrix scaling and rotation processing on the pixel areas of the catenary type regions. The gray values of the scaled regions are obtained by the nearest neighbor interpolation method. Mark the catenary region samples after scaling and rotation processing as the catenary type to increase the number of samples and enhance the robustness of the samples;

[0012] Step 8: Model training. After cropping out the pixel information corresponding to all sub-regions in the image, divide them into catenary type and non-catenary type according to the labels, process all the sample pictures, and provide the classified samples to the multi-layer perceptron model for training;

[0013] Step 9: Identify the catenary. According to the same image processing method as in Steps 2 to 5, split the laser structure in the image to be recognized into multiple sub-regions, and classify the pixel information of all cropped sub-region parts through the trained multi-layer perceptron model, extract all regions classified as catenary, and filter other regions classified as non-catenary structures.

[0014] Furthermore, in the said Step 2, the maximum inter-class variance method is to divide the gray difference degree between two types of pixels according to the threshold t where when is the largest, the threshold t is the optimal threshold t * ; the binary segmentation formula is specifically where R is the binary region after threshold segmentation.

[0015] Furthermore, in the said Step 3, the marking algorithm is specifically to scan the pixels starting from the upper left corner of the image. When encountering an unmarked pixel , assign a label l and start expanding. Check the 8-connected pixels of p i . If there are connected pixels , then assign the same label l and continue to check the 8-connected pixels of p j . Repeat multiple times until there are no new unmarked and connected pixels belonging to the region R; if encountering a pixel and it already has a label l1, and one of its connected pixels belongs to R and has been labeled as l, then the label l is equivalent to the label l1; finally, process the equivalent labels, count the types of different labels, and multiple sub-regions {S1, S2,..., S n} can be obtained, where n is the number of sub-regions. The sub-regions have the following properties: (There is no overlapping part in the sub - regions),

[0016] Furthermore, in the step 4, the dilation formula is specifically where A is the defined dilation element and S is the sub - region to be dilated.

[0017] Furthermore, in the step 7, the algorithm formulas for matrix scaling and rotation processing are specifically where X ′ is the matrix after scaling and rotation, X is the matrix composed of the pixel point coordinates in the catenary sub - region of the image, k x is the scaling ratio along the x - axis, k y is the scaling ratio along the y - axis, and θ is the rotation angle; the algorithm formula for the nearest - neighbor interpolation method is specifically where G ′ is the gray value of the scaled region, and G is the gray value of the original region.

[0018] In summary, the beneficial effects of the catenary laser imaging recognition method combining a multi - layer perceptron and image processing of the present invention are as follows: By designing a railway catenary laser imaging recognition method that combines the technologies of a multi - layer perceptron and digital image processing, it can efficiently recognize the catenary in laser imaging, has good stability under different working conditions, and has low requirements for hardware costs; the present invention has strong practicability and great significance for popularization. Specific Embodiments

[0019] To further understand the features, technical means, specific purposes, and functions achieved by the present invention, and to analyze the advantages and spirit of the present invention, a further understanding can be obtained through the detailed description of the present invention by specific embodiments.

[0020] The present invention provides a catenary laser imaging recognition method combining a multi - layer perceptron and image processing, which includes the following steps:

[0021] Step 1: Sample collection, collecting laser image samples containing catenaries through a camera;

[0022] Step 2: Threshold segmentation, finding the optimal threshold through the maximum inter - class variance method, that is, dividing the gray - level difference degree between two types of pixels according to the threshold t where when is the largest, the threshold t is the optimal threshold t * ; and using the binary segmentation formula Perform threshold segmentation for binarizing the sample image, and segment the laser imaging part from the sample image, where R is the binarized region after threshold segmentation. The obtained binarized region contains the laser structure information of the catenary, optical noise, and other non-catenary objects.

[0023] Step 3: Split the binarized region. Use the labeling algorithm to split the binarized region and segment the laser imaging part in the image into multiple sub-regions including catenary types and non-catenary types. The labeling algorithm specifically scans the pixels starting from the upper left corner of the image. When encountering an unlabeled pixel , assign a label l and start expanding. Check the 8-connected pixels of p i . If there are connected pixels , then assign the same label l and continue to check the 8-connected pixels of p j . Repeat this process multiple times until there are no new unlabeled and connected pixels belonging to region R. If encountering a pixel with an existing label l1, and one of its connected pixels belongs to R and has been labeled l, then label l is equivalent to label l1. Finally, process the equivalent labels and count the types of different labels to obtain multiple sub-regions {S1, S2, …, S n} where n is the number of sub-regions. The sub-regions have the following properties: (The sub-regions have no overlapping parts),

[0024] Step 4: Close the field of view blind area. Use the dilation formula to perform dilation operations on each sub-region. After dilation processing, the catenary busbar sub-regions that were originally separated into multiple parts will have intersections and have connectivity properties, which can close the field of view blind area generated when the catenary pull-out value is large. The dilation formula is specifically where A is the defined dilation element and S is the sub-region to be dilated;

[0025] Step 5: Merge the catenary regions. Mark the multiple sub-regions of the same catenary busbar divided by the field of view blind area as equivalent regions to achieve the merging of the catenary regions;

[0026] Step 6: Divide the labels. Divide all the sub-regions in the image into two types of labels: catenary type and non-catenary type;

[0027] Step 7: Sample increment. Perform small-range matrix scaling and rotation processing on the regional pixels of the catenary type. The algorithm formulas for matrix scaling and rotation are specifically where X ′ is the matrix after scaling and rotation, X is the matrix composed of the pixel point coordinates in the catenary sub-region of the image, and k xis the scaling ratio in the x-axis direction, and k y is the scaling ratio in the y-axis direction, and θ is the rotation angle; the gray value of the scaled region is obtained by the nearest neighbor interpolation method. The algorithm formula of the nearest neighbor interpolation method is specifically where G ′ is the gray value of the scaled region, and G is the gray value of the original region; the catenary region samples after scaling and rotation processing are also labeled as catenary class to increase the number of samples and enhance the robustness of the samples;

[0028] Step 8: Model training. After cropping the pixel information corresponding to all sub-regions in the image, divide them into catenary class and non-catenary class according to the labels, process all sample pictures, and provide the classified samples to the multi-layer perceptron model for training;

[0029] Step 9: Identify the catenary. In the same image processing manner as in Steps 2 to 5, split the laser structure in the image to be identified into multiple sub-regions, and classify the pixel information of all cropped sub-region parts through the trained multi-layer perceptron model, extract all regions classified as catenary, and filter other regions classified as non-catenary structures.

[0030] In summary, the beneficial effects of the catenary laser imaging recognition method combining a multi-layer perceptron and image processing according to the present invention are as follows: By designing a railway catenary laser imaging recognition method that combines the technologies of a multi-layer perceptron and digital image processing, the catenary can be efficiently recognized in laser imaging, has good stability under different working conditions, and has low requirements for hardware costs; the present invention has strong practicability and has strong popularization significance.

[0031] The above embodiments only represent one implementation manner of the invention, and the description is relatively specific and detailed, but it should not be construed as a limitation on the scope of the invention patent. It should be noted that for those of ordinary skill in the art, without departing from the inventive concept, several deformations and improvements can still be made, and these all belong to the protection scope of the invention. Therefore, the protection scope of the invention patent should be subject to the appended claims.

Claims

1. A method for identifying catenary laser imaging by combining a multi-layer perceptron and image processing, characterized in that, It includes the following steps: Step 1: Sample collection, collecting laser image samples containing catenaries through a camera; Step 2: Threshold segmentation, finding the optimal threshold by the maximum inter-class variance method and performing binary threshold segmentation on the sample image using the binary segmentation formula to segment the laser imaging part from the sample image. The binary region obtained after segmentation contains the laser structure information of the catenary, optical noise points, and other non-catenary objects; Step 3: Splitting the binary region, using the labeling algorithm to split the binary region and segment the laser imaging part in the image into multiple sub-regions containing catenary types and non-catenary types; Step 4: Closing the visual field blind area, performing dilation operations on each sub-region using the dilation formula. After dilation processing, the catenary busbar sub-regions that were originally separated into multiple parts will intersect and have connectivity, which can close the visual field blind area generated when the catenary pull-out value is large; Step 5: Merging the catenary regions, marking multiple sub-regions of the same catenary busbar divided by the visual field blind area as equivalent regions to achieve the merging of the catenary regions; Step 6: Labeling, dividing all sub-regions in the image into two types of labels: catenary class and non-catenary class; Step 7: Sample increment, performing small-range matrix scaling and rotation processing on the regional pixels of the catenary class. The gray value of the scaled region is obtained by the nearest neighbor interpolation method, and the catenary region samples after scaling and rotation processing are also marked as the catenary class to increase the sample quantity and enhance the robustness of the samples; Step 8: Model training, after cropping out the pixel information corresponding to all sub-regions in the image, dividing them into the catenary class and the non-catenary class according to the labels, processing all sample pictures, and providing the classified samples to the multi-layer perceptron model for training; Step 9: Identifying the catenary, splitting the laser structure in the image to be identified into multiple sub-regions in the same image processing manner as in Steps 2 to 5, and classifying the pixel information of all cropped sub-region parts through the trained multi-layer perceptron model, extracting all regions classified as catenaries, and filtering other regions classified as non-catenary structures.

2. The catenary laser imaging recognition method combining a multi-layer perceptron and image processing according to claim 1, wherein: In the said step 2, the Otsu method is to divide the gray-scale difference degree between two types of pixels according to the threshold t where when is the largest, the threshold t is the optimal threshold t * ; the binary segmentation formula is specifically where R is the binary region after threshold segmentation.

3. The method for identifying catenary laser imaging by combining a multi-layer perceptron with image processing according to claim 2, wherein: In the step 3, the marking algorithm specifically scans the pixels starting from the upper left corner of the image. When encountering an unmarked pixel it assigns a label l and starts to expand, checking the 8-connected pixels of p i if there are connected pixels it assigns the same label l and continues to check the 8-connected pixels of p j repeating multiple times until there are no new unmarked connected pixels belonging to the region R; if encountering a pixel with an existing label l1, and one of its connected pixels belongs to R and has been labeled l, then the label l is equivalent to the label l1; finally, process the equivalent labels, count the types of different labels, and multiple sub-regions {S1, S2, …, S n} can be obtained, where n is the number of sub-regions. The sub-regions have the following properties:

4. The catenary laser imaging recognition method combining a multi-layer perceptron and image processing according to claim 1, characterized in that: In the step 4, the dilation formula is specifically where A is the defined dilation element and S is the sub-region to be dilated.

5. The catenary laser imaging recognition method combining a multi-layer perceptron and image processing according to claim 1, characterized in that: In the step 7, the algorithm formulas for matrix scaling and rotation processing are specifically as follows where X ′ is the matrix after scaling and rotation, X is the matrix composed of the pixel point coordinates in the catenary sub-region in the image, k x is the scaling ratio along the x-axis, k y is the scaling ratio along the y-axis, and θ is the rotation angle; the algorithm formula for the nearest neighbor interpolation method is specifically as follows where G ′ is the gray value of the scaled region, and G is the gray value of the original region.

Images