Rail corrugation identification method and system, and storage medium

Abstract

The embodiment of the application provides a rail corrugation identification method, system and storage medium, and belongs to the technical field of railway track detection. The rail corrugation identification method comprises the following steps: inputting preprocessed rail corrugation true value image data into an image segmentation identification model to extract global semantic features and local texture features; generating edge features through a Sobel operator after performing channel averaging on the output of the intermediate feature extraction stage, and fusing the global semantic features, the local texture features and the edge features to obtain a segmentation mask graph; inputting ROI image segments cut based on the segmentation mask graph and features extracted by the intermediate feature extraction stage for target detection into a target detection network, performing regression calculation based on a bounding box overlap optimization loss, and outputting a rail corrugation detection result. Through multi-scale feature extraction and channel space attention fusion, the scheme effectively improves the positioning accuracy and segmentation robustness of the rail corrugation area.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of railway track detection, in particular to a rail corrugation identification method, a rail corrugation identification system and a storage medium. BACKGROUND

[0002] As a key component of railway track, rail is subjected to high-frequency load and complex contact stress from repeated action of wheels during train operation. Due to rolling, sliding and flange lateral action of wheel tread and rail surface, periodic uneven plastic deformation and wear will occur on the length direction of rail surface, which will form a wave-shaped geometric irregularity on the rail head tread, and this phenomenon is called rail corrugation. Rail corrugation not only causes vibration and noise during train operation, but also accelerates fatigue damage of track structure components, and thus affects train safety and ride comfort, so it is listed as a key monitoring object in railway maintenance operation.

[0003] Currently, the detection methods of rail corrugation mainly include manual inspection and mechanical measurement based on a corrugation instrument. Manual inspection usually relies on maintenance personnel to use a corrugation ruler to conduct spot inspection on rails on site, which not only has low detection efficiency and high labor intensity, but also is easily affected by human factors, and has the risk of missed detection and misjudgment. Although the detection device based on the corrugation instrument can reduce labor input and improve measurement accuracy, it usually adopts contact measurement or one-string N-point string measurement method, which needs special hardware equipment, and the measurement data needs to be processed offline to obtain the corrugation distribution after being collected, so it cannot realize real-time and rapid determination and positioning. In addition, such device has poor adaptability to external light changes, rail surface pollution and environmental vibration, and has insufficient stability under harsh operating conditions such as high-speed railway.

[0004] The existing rail corrugation identification method based on image processing improves the detection efficiency to some extent, but the traditional image processing relies on manual design of features and fixed threshold, which is difficult to adapt to rail corrugation identification tasks under different lines, different illuminations and different wear patterns, and has limited ability to capture long-distance continuous features and local subtle textures at the same time, resulting in insufficient recognition accuracy and robustness. In summary, the existing rail corrugation detection technology still has obvious deficiencies in real-time performance and adaptability, and needs to be improved. SUMMARY

[0005] The purpose of the embodiments of the present application is to provide a rail corrugation identification method, system and storage medium to at least solve the problem of insufficient real-time performance and adaptability of the existing rail corrugation detection.

[0006] In order to achieve the above object, the present application provides a rail corrugation identification method, which comprises: generating rail corrugation ground truth image data based on collected rail original image data and performing preprocessing; inputting the preprocessed image data into an image segmentation identification model with a multi-scale hollow convolution feature extraction structure and a CBAM attention mechanism introduced after the output of multiple feature layers in the intermediate feature extraction stage, so as to extract global semantic features and local texture features; generating edge features after channel averaging the output of the intermediate feature extraction stage and passing through a Sobel operator, and fusing the global semantic features, local texture features and edge features to obtain a segmentation mask image; inputting ROI image segments cropped based on the segmentation mask image and features extracted by the intermediate feature extraction stage for target detection into a target detection network, performing regression calculation based on bounding box overlap optimization loss, and outputting rail corrugation detection results.

[0007] Optionally, the rail corrugation ground truth image data generated based on the collected rail original image data comprises: calling an artificial labeling tool in the rail original image data to perform pixel-by-pixel contour sketching on the corrugation area, and generating a labeling file containing corrugation area position information after labeling is completed; performing pixel-level registration on the labeling file and the corresponding rail original image data to accurately distinguish the corrugation area and the background area in the two-dimensional pixel matrix of the original image; marking the pixels belonging to the corrugation area as a first pixel category and marking the pixels belonging to the background area as a second pixel category in the registration result, so as to construct the rail corrugation ground truth image data reflecting the shape and position relationship of the corrugation area.

[0008] Optionally, the preprocessing rule for the rail corrugation ground truth image data is that in the rail corrugation ground truth image data, the corrugation area and the background area corresponding to the labeling file are identified, all pixel values in the corrugation area are replaced with a preset first value, and all pixel values in the background area are replaced with a preset second value; after the value replacement is completed, linear normalization operation is performed on the pixel matrix of the whole image, all pixel values are mapped to the interval (0, 1) in proportion, and preprocessed image data is obtained.

[0009] Optionally, the image segmentation recognition model comprises: a backbone network provided with a multi-scale hollow convolution feature extraction structure, configured to extract global semantic features from an input image; a CBAM channel attention module and a spatial attention module arranged at the output of a plurality of feature layers of an intermediate feature extraction stage of the backbone network, configured to enhance the wave abrasion region related feature response; a wave abrasion feature branch arranged in parallel with the backbone network, configured to extract local texture features; an edge feature generation branch connected to the intermediate feature extraction stage of the backbone network, configured to perform Sobel operator operation on the channel-averaged intermediate features to obtain edge features; and a feature fusion unit connected to the wave abrasion feature branch and the edge feature generation branch, configured to fuse the global semantic features, the local texture features and the edge features in the channel dimension and output a segmentation mask map.

[0010] Optionally, the extraction rule of the global semantic features is that, in the backbone network of the image segmentation recognition model, multi-scale convolution operation is performed on the input image by using the multi-scale hollow convolution feature extraction structure, and features of each scale are converged to extract and output the global semantic features. The extraction rule of the local texture features is that, in the wave abrasion feature branch arranged in parallel with the backbone network of the image segmentation recognition model, convolution, normalization and nonlinear activation operations are sequentially performed to extract and output the local texture features from the feature representation of the rail original image.

[0011] Optionally, the edge features generated after the channel-averaged output of the intermediate feature extraction stage by the Sobel operator comprise: performing average operation on the multi-channel feature map output by the intermediate feature extraction stage in the channel dimension to obtain a single-channel feature map; respectively calling a horizontal gradient operator and a vertical gradient operator with fixed parameters to perform convolution operation on the obtained single-channel feature map to extract gradient response maps along the horizontal direction and the vertical direction; and performing amplitude calculation on the gradient response maps to fuse the gradient information of the horizontal and vertical directions to obtain edge features reflecting the position and trend of the rail wave abrasion boundary.

[0012] Optionally, the global semantic features, the local texture features and the edge features are fused to obtain a segmentation mask map, comprising: sequentially concatenating the global semantic features, the local texture features and the edge features in the channel dimension to form a fusion feature tensor; inputting the fusion feature tensor into a convolution dimension reduction unit to compress the channel number to a preset dimension by one-dimensional convolution operation and perform batch normalization processing; applying a nonlinear activation function to the batch normalization result to obtain a fused feature map; and inputting the fused feature map into an output convolution layer of the image segmentation recognition model to generate a segmentation mask map corresponding to the rail region in the spatial position and distinguishing the wave abrasion region and the background region in the pixel category.

[0013] Optionally, the ROI image segment cropped based on the segmentation mask image is input into a target detection network together with the features extracted by the intermediate feature extraction stage for target detection, regression calculation based on a bounding box overlap optimization loss is performed, and a rail corrugation detection result is output, including: determining the boundary coordinates of the corrugation region in the segmentation mask image through connected domain analysis, and performing outer expansion on the boundary coordinates according to a preset proportion to form a cropping box; extracting an ROI image segment containing the corrugation region in the original rail image by using the cropping box, and aligning the ROI image segment and the preprocessed and adapted target detection features output by the intermediate feature extraction stage in the feature dimension; inputting the aligned ROI image segment and the target detection features into a target detection network based on a deep convolutional neural network to generate a candidate bounding box set; performing bounding box regression optimization based on DIoU Loss on the candidate bounding box set to adjust the position and shape of each candidate box, and outputting a rail corrugation detection result containing the corrugation position and range after the optimization is completed.

[0014] The second aspect of the present application provides a rail corrugation identification system, which comprises: an acquisition unit for generating rail corrugation ground truth image data and performing preprocessing based on acquired rail original image data; an extraction unit for inputting the preprocessed image data into an image segmentation identification model having a multi-scale hollow convolution feature extraction structure and introducing a CBAM attention mechanism after outputting a plurality of feature layers in an intermediate feature extraction stage, so as to extract global semantic features and local texture features; a fusion unit for generating edge features by a Sobel operator after channel averaging the output of the intermediate feature extraction stage, and fusing the global semantic features, local texture features and edge features to obtain a segmentation mask image; and an output unit for inputting an ROI image segment cropped based on the segmentation mask image and features extracted by the intermediate feature extraction stage for target detection into a target detection network together, performing regression calculation based on a bounding box overlap optimization loss, and outputting a rail corrugation detection result.

[0015] In another aspect, the present application provides a computer-readable storage medium having instructions stored thereon, which, when executed on a computer, cause the computer to perform the above-mentioned rail corrugation identification method.

[0016] By the technical scheme, the rail original image data is constructed and preprocessed to obtain a corrugation true value image data, an image segmentation and recognition model with a multi-scale hollow convolution feature extraction structure and a CBAM attention mechanism is introduced, edge features generated by a Sobel operator are combined, global semantic features, local texture features and edge features are fused to generate a segmentation mask image, and based on the mask image, a ROI is cropped and combined with intermediate features in the segmentation stage to perform target detection and bounding box optimization, so that accurate positioning and recognition of the rail corrugation area are realized. The technical effect is that the extraction of long-distance continuous features and fine texture features can be considered under complex light and diversified wear forms, the accuracy and stability of the detection result are improved, feature reuse is used to reduce calculation redundancy, and the real-time performance and adaptability of the overall detection are improved.

[0017] Other features and advantages of the present application will be described in detail in the following detailed description. BRIEF DESCRIPTION OF DRAWINGS

[0018] The accompanying drawings are included to provide a further understanding of the embodiments of the application, and constitute a part of the specification, and are used together with the following detailed description to explain the embodiments of the application, but do not constitute a limitation on the embodiments of the application. In the drawings:

[0019] Figure 1 is a step flow chart of a rail corrugation recognition method provided by an embodiment of the present application;

[0020] Figure 2 is a structure schematic diagram of an image segmentation and recognition model provided by an embodiment of the present application;

[0021] Figure 3 is a convolution module structure schematic diagram provided by an embodiment of the present application;

[0022] Figure 4 is an ASPP module structure schematic diagram provided by an embodiment of the present application;

[0023] Figure 5 is a CBAM attention mechanism working principle diagram provided by an embodiment of the present application;

[0024] Figure 6 is a system structure diagram of a rail corrugation recognition system provided by an embodiment of the present application. DETAILED DESCRIPTION

[0025] The specific embodiments of the present application will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are only used to illustrate and explain the present application, and are not used to limit the present application.

[0026] Figure 1is a method flowchart of a rail corrugation identification method provided by an embodiment of the present application. As shown in Figure 1 The present application provides a rail corrugation identification method, which comprises the following steps:

[0027] Step S10: generating rail corrugation ground truth image data based on the collected rail original image data and performing preprocessing.

[0028] Specifically, the rail corrugation ground truth image data is generated based on the collected rail original image data, which comprises the following steps: calling an artificial labeling tool to perform pixel-by-pixel contour sketching on the corrugation area in the rail original image data, and generating a labeling file containing the position information of the corrugation area after labeling is completed; performing pixel-level registration on the labeling file and the corresponding rail original image data to accurately distinguish the corrugation area and the background area in the two-dimensional pixel matrix of the original image; marking the pixels belonging to the corrugation area as a first pixel category and marking the pixels belonging to the background area as a second pixel category in the registration result to construct the rail corrugation ground truth image data reflecting the shape and position relationship of the corrugation area.

[0029] In the embodiment of the present application, when the rail corrugation ground truth image data is generated based on the collected rail original image data, first, an artificial labeling tool with pixel-by-pixel labeling function is called in the collected rail original image to perform fine-grained contour sketching on the position of the corrugation defect in the image. During the labeling process, the real boundary of the corrugation defect needs to be described point by point, and all wave crests, wave troughs and edge transition areas need to be covered to avoid missing small protrusions, depressions or broken places, so as to ensure the integrity of the subsequent feature extraction stage. After the labeling is completed, a labeling file containing the pixel coordinate information and closed contour description of the corrugation area is generated, which maintains a one-to-one correspondence with the original image in file name or index to ensure the traceability of data matching.

[0030] After obtaining the labeling file, the labeling file and the corresponding rail original image are registered at the pixel level. During the registration process, the contour of the corrugation area is accurately mapped to the actual pixel position of the original image by comparing the pixel coordinates recorded in the labeling file with the two-dimensional pixel matrix index position of the original image. This step ensures the complete consistency of the labeling information and the original image in the spatial position, that is, each corrugation boundary point can be accurately corresponded to the pixel unit of the original image, so as to avoid label deviation or shape distortion caused by registration error.

[0031] In the registration result, all pixels located inside the profile of the corrugation region are uniformly assigned a preset first pixel category code, and all background pixels located outside the profile are uniformly assigned a preset second pixel category code. In this way, the corrugation region and the background region form a clear classification label relationship in data representation. The two-dimensional pixel matrix obtained through the above steps is the steel rail corrugation true value image data, which presents a feature pattern in which the corrugation region and the background region are significantly separated in visualization, and can be directly used as a supervised signal input for an image segmentation and recognition model, thereby providing high-quality training and verification data for subsequent feature extraction, segmentation and detection steps.

[0032] Through the pixel-by-pixel labeling, accurate registration and category coding process, the stability and consistency of the corrugation region label can be maintained under different lighting, wear conditions and surface contamination conditions, thereby effectively improving the accuracy and robustness of the subsequent model in the corrugation recognition task.

[0033] Preferably, the preprocessing rule of the steel rail corrugation true value image data is as follows: in the steel rail corrugation true value image data, the corrugation region and the background region corresponding to the annotation file are identified, all pixel values in the corrugation region are replaced with a preset first value, and all pixel values in the background region are replaced with a preset second value; after the value replacement is completed, a linear normalization operation is performed on the pixel matrix of the entire image, all pixel values are mapped to the interval (0, 1) in proportion, and the preprocessed image data is obtained.

[0034] In the embodiment of the present application, after the steel rail corrugation true value image data is generated based on the collected steel rail original image data, the true value image data is preferably preprocessed so that the subsequent model can stably and accurately identify the corrugation region. First, the pixel positions corresponding to the corrugation region and the pixel positions corresponding to the background region are identified one by one according to the coordinate information recorded in the annotation file. The annotation file is usually generated after the profile of the corrugation region is manually drawn pixel by pixel in the early stage, and contains the boundary information and closed region coordinates of each corrugation region, so as to accurately match the true value image data.

[0035] After the identification of the corrugation region and the background region is completed, the values in the pixel matrix are replaced. Specifically, all pixel values inside the corrugation region are uniformly replaced with a preset first value, and all pixel values in the background region are uniformly replaced with a preset second value. The values of the first value and the second value should ensure that there is sufficient difference between them, so that the model can easily distinguish between the two types of regions during training, reducing confusion in the feature extraction stage. The value replacement adopts an overall replacement manner instead of a pixel-by-pixel calculation generation, so as to ensure the processing efficiency and avoid the ambiguity of the category boundary caused by the slight deviation introduced by the floating point operation.

[0036] After the numerical replacement is completed, a linear normalization operation is performed on the pixel matrix of the entire image. The normalization process is to map the pixel values to the (0, 1) interval in proportion, that is, to scale by the difference between the minimum value and the maximum value, to ensure that all pixel values are in the same dimension range. In this way, not only can the numerical stability of the model input be improved, but also the problem of training divergence caused by the difference in brightness and contrast under different image acquisition conditions can be avoided. After the normalization processing is completed, the preprocessed image data not only maintains the class distinction between the rail wear area and the background area, but also meets the input specification of the model in terms of numerical range.

[0037] In another possible implementation, in the steel rail detection operation of part of the railway line, due to insufficient light in night operation, more attachments on the surface of the steel rail, and channel compression of the equipment, the true value image data generated after the original rail image collected is labeled may have blurred boundaries and uneven pixel grayscale. Although the conventional pixel value replacement and linear normalization can achieve class distinction, they are prone to grayscale jitter in the boundary transition area, resulting in unstable boundary prediction of the model during the training process.

[0038] To solve the above problems, a weighted smoothing process can be performed on the boundary between the rail wear area and the background area before performing numerical replacement. Specifically, first, the Euclidean distance between the boundary pixels of the rail wear area and the background area pixels is calculated, and a weight decay coefficient is assigned to the pixels within the preset threshold range, so that their pixel values are linearly interpolated between the first value and the second value according to the distance ratio. For example, a point 1 pixel away from the boundary can be set to an interpolation weight of 0.8, a point 2 pixels away can be set to an interpolation weight of 0.6, and so on until the decay to the pure value of the background or wear. In this way, in the true value image, the boundary area will form a grayscale gradient zone, rather than a sudden binary boundary.

[0039] This gradient processing will be mapped to the (0, 1) interval in the normalization stage, forming a label distribution closer to the actual wear edge. Although this method introduces continuous values in the labeled data, it can significantly reduce the gradient oscillation of the model in the boundary area, improving the stability of boundary positioning, especially when there is noise interference or surface contamination, which can effectively avoid the over-segmentation or under-segmentation of the rail wear area.

[0040] Step S20: The preprocessed image data is input into an image segmentation and recognition model having a multi-scale hollow convolution feature extraction structure and introducing a CBAM attention mechanism after outputting multiple feature layers in the intermediate feature extraction stage, to extract global semantic features and local texture features.

[0041] Specifically, the image segmentation identification model comprises: a backbone network provided with a multi-scale hollow convolution feature extraction structure, used for extracting global semantic features from an input image; a CBAM (Convolutional Block Attention Module) channel attention module and a spatial attention module arranged after a plurality of feature layers output in an intermediate feature extraction stage of the backbone network, used for enhancing the response of the rail corrugation area related features; a rail corrugation feature branch arranged in parallel with the backbone network, used for extracting local texture features; an edge feature generation branch connected with the intermediate feature extraction stage of the backbone network, used for performing Sobel operator operation on the channel-averaged intermediate features to obtain edge features; and a feature fusion unit connected with the rail corrugation feature branch and the edge feature generation branch, used for fusing the global semantic features, the local texture features and the edge features in the channel dimension and outputting a segmentation mask map.

[0042] In the embodiment of the present application, the backbone network of the image segmentation identification model is provided with a multi-scale hollow convolution feature extraction structure. By arranging different hollow rates in different convolution branches, the receptive field covers the long-distance features along the rail direction and the vertical detail features while maintaining the spatial resolution. In the rail corrugation identification scene, the rail surface defects often appear as periodic and long-wave textures, so the multi-scale hollow convolution can capture both the overall shape of the macro wave and the microscopic edge transition, effectively improving the feature expression ability.

[0043] After the plurality of feature layers output in the intermediate feature extraction stage of the backbone network, a CBAM channel attention module and a spatial attention module are arranged in sequence. The channel attention module can calculate the channel weight based on the global average pooling and the maximum pooling result, thereby highlighting the channel features related to the rail corrugation texture; the spatial attention module generates a spatial weight distribution through convolution, further strengthens the position response of the rail corrugation area and suppresses the background interference. This joint attention mechanism of channel and space can still focus on the real rail corrugation area when there are interference factors such as uneven illumination, rust or oil stains on the rail surface.

[0044] The rail corrugation feature branch arranged in parallel with the backbone network is used to extract local texture features. This branch maintains a high spatial resolution and, through multi-layer convolution, normalization and activation operations, it strengthens the elongated texture pattern of the rail corrugation, so as to provide fine structure information when the features are fused. At the same time, the edge feature generation branch connected with the intermediate feature extraction stage of the backbone network performs channel averaging on the intermediate features to eliminate channel redundancy, and then uses a Sobel operator with fixed parameters to extract gradient information in the horizontal and vertical directions, generating an edge feature map that can accurately reflect the position and direction of the rail corrugation boundary.

[0045] The outputs of the wave feature branch and the edge feature generation branch are concatenated in the channel dimension with the global semantic features of the backbone network in a feature fusion unit, and are subjected to convolution dimension reduction, normalization and activation to output a segmentation mask map. The specificity of this structure in the steel rail wave identification scene is that it can not only consider multi-scale features of long-distance waveforms and local textures, but also accurately depict the boundary form of the wave by supplementing the edge features, thereby achieving high-precision defect segmentation in complex environments.

[0046] In a possible implementation manner, as shown in Figure 2 , a ResNet backbone network-based image feature extraction and segmentation structure is shown. The input is a steel rail image with a size of 2048xHxWx3, where H and W represent the height and width of the input steel rail image, respectively, and their values are determined by the imaging size of the actual acquisition device, used to represent the size of the input image in the spatial dimension. First, multi-layer features are extracted by the ResNet backbone network, including Layer (layer) 1 (256xH / 4xW / 4), Layer2 (512xH / 8xW / 8), Layer3 (1024xH / 16xW / 16), and Layer4 (2048xH / 16xW / 16). Among them, the outputs of Layer2 and Layer3 are connected to the CBAM attention module to enhance the channel and spatial features related to the wave area; the output of Layer3 is also used to generate edge features through the Sobel operator. The output of Layer4 is connected to the ASPP module (Atrous Spatial Pyramid Pooling, ASPP) to capture multi-scale context information, and is combined with the wave feature branch (Conv-BN-ReLU structure, channel number 256). After all the features are fused in the channel dimension, they are compressed to 1024 channels through 1x1 convolution + BN (Batch Normalization, BN) + ReLU, and then a 1x1 convolution classifier is used to generate a segmentation result. Finally, the result is bilinearly upsampled to restore the original image size, and a segmentation mask map containing the wave area and the background area is output. This structure realizes the multi-source fusion of global semantic features, local texture features and edge features, and improves the accuracy and robustness of steel rail wave identification. The structure of the convolution module is shown in Figure 3 .

[0047] Among them, as shown in Figure 4The input features of the ASPP module are firstly subjected to a two-dimensional convolution to generate basic features, and then are subjected to three 3x3 convolution operations with different hole rates (6, 12 and 18) respectively to capture context information of different scales. Meanwhile, the input features are subjected to global average pooling to generate global context information. The four feature outputs are spliced in the channel dimension together with the original convolution features to realize multi-scale information fusion. Finally, feature fusion and compression are performed through a 1x1 convolution to obtain feature representation with local details and global semantics, thereby improving the accuracy and robustness of segmentation or detection.

[0048] The working principle of the CBAM attention module is as shown in Figure 5 The input features are fed into the channel attention module, channel weights are generated through global average pooling and global maximum pooling, the original features are weighted in the channel dimension, and channel information related to the task is highlighted. Then, the channel weighted features are fed into the spatial attention module, and the attention map is calculated in the spatial dimension (usually combined with the average pooling and maximum pooling results) to highlight the key information area and suppress the irrelevant background. Both weightings adopt element-wise multiplication operation, and finally the optimized feature map is output. The module can retain important semantic information while enhancing the response to the target area, reducing background interference and improving feature representation capability.

[0049] Preferably, the extraction rule of the global semantic feature is that, in the backbone network of the image segmentation and recognition model, a multi-scale convolution operation is performed on the input image by using a multi-scale dilated convolution feature extraction structure, and features of different scales are converged to extract and output the global semantic feature.

[0050] In the embodiment of the present application, in the backbone network of the image segmentation and recognition model, a convolution branch group with a multi-scale dilated convolution feature extraction structure is constructed, and the convolution kernels of different branches are set with different hole rates while keeping the same size, so as to expand the receptive field range without increasing the parameter amount. Specifically, the branch with a larger hole rate can capture the long-distance waveform periodic feature of the rail extending along the rail direction, and the branch with a smaller hole rate can retain the detail information and texture change of the local area of the rail surface, thereby forming a feature representation that takes into account both macroscopic morphology and local details.

[0051] After the multi-scale convolution operation is completed, the feature maps output by the branches are spliced or pixel-by-pixel weighted and converged in the channel dimension to fuse the feature information from different scales. The convergence process can adopt a channel-by-channel weighting method to dynamically adjust the proportion of features of different scales in the fusion result according to the response strength of the features, avoid the weak features being submerged or the strong features being over-dominated, and maintain the balance of the global features.

[0052] The global semantic features extracted in the above manner can accurately reflect the spatial distribution trend and overall morphology of the corrugation defects in the entire length of the rail, and can maintain the stability and robustness of the features even in the case of uneven illumination, rail surface pollution or variable wear morphology, thereby providing a solid semantic foundation for subsequent feature fusion and mask generation.

[0053] Preferably, the extraction rule of the local texture features is that, in a corrugation feature branch of the backbone network of the image segmentation and recognition model, convolution, normalization and nonlinear activation operations are sequentially performed to extract and output the local texture features from the feature representation of the original rail image.

[0054] In the embodiment of the present application, in the corrugation feature branch of the backbone network of the image segmentation and recognition model, a convolution processing link with high spatial resolution is constructed. The input of this branch is a feature map synchronized with the output of the intermediate feature extraction stage in the backbone network, the first layer of convolution is used to extract the basic texture pattern of the local region, and the batch normalization operation is used to standardize the feature distribution to reduce the statistical difference between different batches of input; then a nonlinear activation function (such as ReLU or LeakyReLU) is introduced to enhance the nonlinear ability of feature expression, thereby strengthening the slender texture and periodic waveform details in the corrugation region.

[0055] In this branch, multiple layers of convolution-normalization-activation units can be further stacked to gradually expand the receptive field while maintaining the integrity of the local structural features. The size, step and padding mode of each layer of convolution kernel can be designed according to the average width and pitch of the corrugation stripes in the rail image to ensure that the directionality and continuity of the rail surface texture can be accurately captured.

[0056] The local texture features extracted by this rule can complement the deficiencies of the global semantic features in the detail level, making the subsequent fusion of edge features and global semantic features more fine, so that high-precision segmentation and positioning can still be achieved in the case of ambiguous corrugation boundary and complex wear morphology.

[0057] Step S30: generating edge features by performing channel averaging on the output of the intermediate feature extraction stage and then using a Sobel operator, and fusing the global semantic features, local texture features and edge features to obtain a segmentation mask map.

[0058] Specifically, the edge feature is generated by performing channel averaging on the output of the intermediate feature extraction stage and then passing the output through a Sobel operator, including: performing averaging operation on the multi-channel feature map output by the intermediate feature extraction stage in the channel dimension to obtain a single-channel feature map; respectively calling a fixed-parameter horizontal gradient operator and a fixed-parameter vertical gradient operator to perform convolution operation on the obtained single-channel feature map, to extract gradient response maps along the horizontal direction and the vertical direction; performing amplitude calculation on the gradient response maps to fuse the gradient information of the horizontal direction and the vertical direction, to obtain an edge feature reflecting the boundary position and trend of the rail corrugation.

[0059] Further, the global semantic feature, the local texture feature and the edge feature are fused to obtain a segmentation mask map, including: sequentially concatenating the global semantic feature, the local texture feature and the edge feature in the channel dimension to form a fusion feature tensor; inputting the fusion feature tensor into a convolution dimension reduction unit, compressing the channel number to a preset dimension through one-dimensional convolution operation, and performing batch normalization processing; applying a nonlinear activation function to the batch normalization result to obtain a fused feature map; inputting the fused feature map into an output convolution layer of the image segmentation and recognition model to generate a segmentation mask map corresponding to the rail region in the spatial position and distinguishing the corrugation region and the background region in the pixel category.

[0060] In the embodiment of the present application, when the edge feature is generated by performing channel averaging on the output of the intermediate feature extraction stage and then passing the output through a Sobel operator, first, a multi-channel feature map output by the image segmentation and recognition model in the intermediate feature extraction stage is obtained. The feature map is usually derived from the intermediate layer output after a certain number of layers of the main network and the parallel branch, and contains high-dimensional feature information processed by multi-scale atrous convolution and attention mechanism. These features may contain dozens or even hundreds of feature channels in the channel dimension, and each channel responds to different convolution kernel filtering characteristics and receptive fields, so they contain semantic information of the corrugation region, texture details and background noise information.

[0061] In order to extract feature information with boundary indication from these features, channel averaging operation is needed. Specifically, the multi-channel feature map is averaged pixel by pixel in the channel dimension, that is, for each pixel at a spatial position, the average value of all channels at the position is calculated to generate a single-channel feature map. The purpose of this step is to compress the response intensity information of the multi-channel into a single gray distribution, so as to eliminate the directional bias caused by the response difference of different channels, while retaining the overall structure contour information, making the subsequent gradient calculation more stable.

[0062] After obtaining the single-channel feature map, the Sobel horizontal gradient operator and the Sobel vertical gradient operator with fixed parameters are called to perform convolution operations respectively. The horizontal gradient operator is used to detect the gray level change of the image in the horizontal direction, that is, to capture the transverse edge features formed when the rail corrugation extends longitudinally; the vertical gradient operator detects the gray level change of the image in the vertical direction, which is used to capture the change features of the corrugation edge in the cross-sectional direction of the rail head. The convolution operation adopts a standard two-dimensional convolution form, and the convolution kernel size is usually 3x3, which is fixed and does not participate in training, so as to ensure the stability and consistency of edge detection.

[0063] After obtaining the gradient response maps in the horizontal and vertical directions respectively, amplitude calculation needs to be performed to fuse the two. The amplitude calculation can be realized by squaring and square root, that is, the square sum of the horizontal gradient value and the vertical gradient value of each pixel point is calculated, and then the square root is taken to obtain the gradient amplitude of the pixel point. The gradient amplitude map obtained in this way can reflect the change intensity of the edge in two directions at the same time, forming a comprehensive description of the boundary position and trend of the rail corrugation. After this processing, the output edge feature map visually presents a high-light area distributed along the corrugation contour, which forms a clear contrast with the background area, and can preserve the boundary details. Even in the case of uneven lighting or surface stains, it also has strong robustness.

[0064] Further, when the global semantic features, local texture features and the above generated edge features are fused to obtain a segmentation mask map, the three types of features are sequentially spliced in the channel dimension first to form a fusion feature tensor. The global semantic features mainly include the overall structure of the rail and the large-scale distribution information of the corrugation defects, the local texture features include the elongated texture and local pattern of the corrugation, and the edge features provide accurate boundary contour information. The splicing method of the three is usually to directly stack the channels in the channel dimension, that is, to merge the channel numbers of each type of feature into a higher-dimensional feature set, so that the subsequent convolution operation can simultaneously perceive the spatial distribution relationship of different types of features.

[0065] After the formation of the fusion feature tensor, it is input into the convolution dimension reduction unit, and the channel number is compressed to a preset dimension through one-dimensional convolution (1x1 convolution) operation. Convolution dimension reduction not only reduces the subsequent calculation amount, but also learns the optimal combination method between different feature channels through the weights of the convolution kernel, retains useful information while suppressing redundant features. After dimension reduction, batch normalization processing is performed to adjust the distribution difference of the features between different batches of input, so that the training process is more stable and the convergence speed is faster.

[0066] A nonlinear activation function such as ReLU or LeakyReLU is applied on the batch-normalized results to introduce nonlinear mapping capability and enhance the model's expression ability for complex feature combinations. Through nonlinear activation, the model can better separate the wave-worn area and the background area, even if they are close in grayscale or texture distribution, and maintain a good classification interval.

[0067] The fused feature map after dimension reduction, normalization and activation is input into the output convolution layer of the image segmentation recognition model, and the number of convolution kernels of the output convolution layer is consistent with the number of classifications (usually 2, corresponding to the wave-worn area and the background area). After convolution operation, a two-dimensional feature map consistent with the input image in spatial position is output, and a segmentation mask graph is generated through a normalization function such as Softmax or Sigmoid. The segmentation mask graph can clearly distinguish the wave-worn area and the background area in the pixel category, providing accurate area reference for subsequent ROI cropping and target detection.

[0068] Through the above edge feature generation and multi-feature fusion rules, in the steel rail wave-worn recognition scene, global distribution information, local texture details and accurate boundary information can be effectively combined, so that in the case of complex wave-worn morphology, variable lighting conditions and strong background interference, a segmentation mask graph with accurate boundary and complete contour can still be output.

[0069] Step S40: input the ROI image segment cropped based on the segmentation mask graph and the features for target detection extracted in the intermediate feature extraction stage into the target detection network, perform regression calculation based on the bounding box overlap optimization loss, and output the steel rail wave-worn detection result.

[0070] Specifically, the boundary coordinates of the wave-worn area are determined in the segmentation mask graph through connected component analysis, and the boundary coordinates are expanded outward according to a preset proportion to form a cropping frame; the ROI (Region of Interest) image segment containing the wave-worn area is extracted from the original rail image using the cropping frame, and the ROI image segment and the preprocessed and adapted target detection features output by the intermediate feature extraction stage are aligned in the feature dimension; the aligned ROI image segment and target detection features are input into the target detection network based on the deep convolutional neural network, and a candidate bounding box set is generated; the candidate bounding box set is subjected to bounding box regression optimization based on DioULoss (distance intersection over union loss) to adjust the position and shape of each candidate box, and after optimization, the steel rail wave-worn detection result containing the wave-worn position and range is output.

[0071] In the embodiment of the present application, when the ROI image segment cropped based on the segmentation mask is input into the target detection network together with the features extracted by the intermediate feature extraction stage for target detection, first, a connected component analysis is performed in the segmentation mask to identify all independent corrugation regions. The connected component analysis aggregates pixels with the same value and belonging to the corrugation region into the same connected component through pixel-by-pixel scanning, and assigns a unique identification number to each connected component. For each connected component, the boundary coordinates of the minimum bounding rectangle thereof are calculated, which consist of the minimum row index, the maximum row index, the minimum column index, and the maximum column index, and can accurately envelope all pixels of the connected component.

[0072] To avoid the cropped ROI image segment being too close to the boundary of the corrugation region and resulting in insufficient context information, after determining the minimum bounding rectangle, the boundary coordinates are expanded outward by a preset proportion. The proportion of the expansion can be set according to the typical size of the corrugation feature and the input size requirement of the subsequent detection network, for example, expanding the rectangle width and height by 5% to 15% in each of the up, down, left, and right directions. The expanded rectangle is the final cropping frame, which can contain the corrugation region while additionally retaining a certain amount of rail surface background information to improve the context understanding ability of the detection stage.

[0073] Using the cropping frame, a corresponding ROI image segment is extracted from the original rail image. The extraction process needs to maintain the original resolution and proportion of the image to avoid scaling or interpolation introducing deformation affecting feature quality. At the same time, in the feature map output by the intermediate feature extraction stage, the feature region corresponding in space to the ROI image segment is extracted according to the spatial position corresponding to the cropping frame. In order to ensure the consistency of the ROI image segment and the target detection feature in dimension, a preprocessing adaptation operation is performed on the extracted feature region, such as bilinear interpolation size adjustment, channel alignment, and normalization processing, so that it can be one-to-one corresponding with the ROI image segment in feature dimension.

[0074] After alignment, the ROI image segment and the adapted target detection feature are jointly input into the target detection network constructed based on a deep convolutional neural network. The front end of the network usually includes multiple convolutional layers and down-sampling layers for further refining the feature information within the ROI and generating a set of multi-scale candidate bounding boxes. These candidate boxes may have redundancies in position, size, and aspect ratio, so they need to be screened and optimized after being generated.

[0075] For the candidate bounding box set, DIoULoss-based bounding box regression optimization is performed. DIoULoss is based on the calculation of the IoU (intersection over union) of the predicted box and the real box, and additionally introduces a Euclidean distance penalty term between the center points of the predicted box and the real box, so as to consider the matching degree of the overlapping area and the center position in the optimization process. The optimization process iteratively adjusts the network parameters through back propagation, so that the predicted box gradually approaches the boundary of the real rail corrugation region in position and shape.

[0076] After optimization, the detection results obtained include the position coordinates of each rail corrugation region in the original image, the size of the circumscribed rectangle, and the corresponding confidence score. The final output of the rail corrugation detection result can accurately label the position and range of the rail corrugation region in space, and when the adjacent defects are close or the morphology is complex, it can still maintain the detection effect of clear boundary separation and accurate position.

[0077] In another possible implementation, in actual railway field detection, due to the strong directionality of the rail corrugation morphology extending along the track, and the slight tilt or shaking of the camera under high-speed acquisition conditions, the connected domain generated by the segmentation mask image may have a slight rotation deviation in the spatial coordinates. If the axis-aligned rectangular crop frame is directly used for ROI extraction, it will often contain too much irrelevant background or crop the end of the rail corrugation region, thereby affecting the positioning accuracy of the subsequent detection network.

[0078] Therefore, the minimum circumscribed rotating rectangular cropping rule is introduced in the connected domain analysis stage. Specifically, after determining the pixel set of each rail corrugation region, the principal direction angle of the pixel set is calculated using the principal component analysis (PCA)-based minimum circumscribed rectangle algorithm, and the minimum circumscribed rectangle is generated with the angle as the rotation reference. Then the rectangular boundary is extended by a preset proportion in the rotated coordinate system, and the cropped frame obtained not only conforms to the direction of the rail corrugation region, but also maximizes the reduction of the inclusion amount of the background interference region.

[0079] When the rotated rectangular coordinates are mapped back to the original image space, the ROI image segment is extracted in a rotating manner through affine transformation, and the feature region with the same rotation angle and position is extracted in the intermediate feature map, maintaining the consistency of the two in geometric shape. In this way, the ROI image segment and the corresponding features input into the target detection network are perfectly aligned in the direction of the rail corrugation, significantly reducing the modeling pressure of the detection network in the learning process.

[0080] This method is particularly suitable for long wave distance, non-vertical boundary, and slight camera tilt in the rail corrugation detection scene, and can significantly improve the positioning accuracy and boundary fitting effect, and reduce the shape deviation of the detection frame.

[0081] Embodiment:

[0082] Step one, the image segmentation model preprocessing process is as follows.

[0083] 1) Sample railway image data and construct railway ground truth image data; the railway image ground truth image data is the rail wear profile image data drawn according to the railway image data.

[0084] 2) Binary processing of railway ground truth image data: in the rail wear profile image data, set the pixel value of the background to 0 and the pixel value of the rail wear to 1, to represent the rail wear profile image data in numerical form.

[0085] 3) Normalize the railway ground truth according to the following formula:

[0086]

[0087] Where i, j represent the row number and column number of the railway image respectively, C(i,j) represents the pixel value corresponding to the original railway image data, I(i,j) represents the pixel value of the normalized railway image, I(i,j) belongs to (0,1).

[0088] Step two, the identification process of rail wear is as follows.

[0089] Feature extraction stage, input: original rail image (3200x2400x3) after standardization processing (μ=[0.485, 0.456, 0.406], σ=[0.229, 0.224, 0.225])

[0090] 1) Backbone network: extract multi-scale features through ResNet50 Layer1-3; insert CBAM attention module after Layer2 (256 dimensions) and Layer3 (512 dimensions) to calculate channel weight (GAP-MLP) and spatial weight (Conv7x7) respectively; Layer4 outputs 2048-dimensional features into improved ASPP module:

[0091] Asymmetric dilation rate (4,8) 3x3 hollow convolution is adopted; 1x1 convolution and global average pooling branch are retained; wear feature enhancement.

[0092] 2) Special branch: from the ASPP output, draw a wear feature branch (2 layers of 3x3 Conv+BN+ReLU) - output 256-dimensional local texture features.

[0093] 3) Edge Assistance: Channel averaging is performed on Layer 3 features - Sobel edge detection (fixed X / Y direction gradient kernel) - generating an edge map; Feature Fusion and Segmentation: ASPP main features (2048-dimensional), ripple features (256-dimensional), and edge features (upsampled to the same size) are concatenated along the channel dimension, compressed to 1024 dimensions through 1×1 convolution – BatchNorm-ReLU activation. Finally, a 3-channel segmentation map is output through 3×3 convolution.

[0094] Step 3: Target Detection Stage.

[0095] 1) ROI generation: Perform connected component analysis on the segmentation results - extract the minimum bounding rectangle of the rail region; expand the ROI boundary by a factor of 1.2 - crop the corresponding region of the original image.

[0096] 2) YOLOv5+ detection: Shares ResNet50 Layer 3 features (avoids redundant calculations); uses DIoULoss to optimize bounding box regression.

[0097] Figure 6 This is a system structure diagram of a rail corrugation identification system provided in one embodiment of the present invention. Figure 6 As shown, this invention provides a rail corrugation recognition system. The system includes: an acquisition unit for generating ground truth image data of rail corrugation based on acquired raw rail image data and performing preprocessing; an extraction unit for feeding the preprocessed image data into an image segmentation and recognition model with a multi-scale dilated convolutional feature extraction structure and introducing a CBAM attention mechanism after the output of multiple feature layers in the intermediate feature extraction stage, to extract global semantic features and local texture features; a fusion unit for averaging the output of the intermediate feature extraction stage through channels and generating edge features using the Sobel operator, and fusing the global semantic features, local texture features, and edge features to obtain a segmentation mask; and an output unit for inputting ROI image fragments cropped based on the segmentation mask and the features extracted for target detection in the intermediate feature extraction stage into a target detection network, performing regression calculation based on bounding box overlap optimization loss, and outputting rail corrugation detection results.

[0098] The present invention also provides a computer-readable storage medium storing instructions that, when executed on a computer, cause the computer to perform the above-described rail corrugation identification method.

[0099] Those skilled in the art can understand that all or part of the steps of the method for implementing the above-mentioned embodiments can be completed by programs instructing relevant hardware, the programs are stored in a storage medium, and the programs include a plurality of instructions for enabling a single-chip microcomputer, a chip or a processor to execute all or part of the steps of the method described in each embodiment of the present application. The aforementioned storage medium includes a U disk, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and various storage media capable of storing program codes.

[0100] The optional embodiments of the present application are described in detail above in combination with the drawings, but the embodiments of the present application are not limited to the specific details in the above-described embodiments. Within the technical concept scope of the embodiments of the present application, various simple modifications can be made to the technical solutions of the embodiments of the present application, and these simple modifications all belong to the protection scope of the embodiments of the present application. In addition, it should be noted that each specific technical feature described in the above-described specific embodiments can be combined in any appropriate manner without contradiction. In order to avoid unnecessary repetition, the embodiments of the present application will not be described again for various possible combinations.

[0101] In addition, various different embodiments of the present application can also be combined in any manner, as long as it does not deviate from the idea of the embodiments of the present application, and it should also be considered as disclosed by the embodiments of the present application.

Claims

1. A method for identifying rail corrugation, characterized in that, The rail corrugation identification method includes: Based on the acquired raw rail image data, true image data of rail corrugation is generated and preprocessing is performed; The preprocessed image data is fed into an image segmentation and recognition model with a multi-scale dilated convolutional feature extraction structure, and a CBAM attention mechanism is introduced after the output of multiple feature layers in the intermediate feature extraction stage, in order to extract global semantic features and local texture features; among which, The image segmentation and recognition model includes: a backbone network with a multi-scale dilated convolutional feature extraction structure for extracting global semantic features from the input image; a CBAM channel attention module and a spatial attention module set after the output of multiple feature layers in the intermediate feature extraction stage of the backbone network for enhancing the feature response of the wavy region; a wavy feature branch set in parallel with the backbone network for extracting local texture features; an edge feature generation branch connected to the intermediate feature extraction stage of the backbone network for performing Sobel operator operations on the channel-averaged intermediate features to obtain edge features; and a feature fusion unit connected to the wavy feature branch and the edge feature generation branch for fusing global semantic features, local texture features, and edge features in the channel dimension and outputting a segmentation mask map. The extraction rules for global semantic features are as follows: In the backbone network of the image segmentation and recognition model, a multi-scale dilated convolution feature extraction structure is used to perform multi-scale convolution operations on the input image and aggregate features at each scale to extract and output global semantic features; The extraction rules for local texture features are as follows: In the ripple feature branch set in parallel in the backbone network of the image segmentation and recognition model, convolution, normalization and nonlinear activation operations are performed in sequence to extract and output local texture features from the feature representation of the original rail image; After channel averaging of the output of the intermediate feature extraction stage, edge features are generated by the Sobel operator, and the global semantic features, local texture features and edge features are fused to obtain a segmentation mask image. The ROI image fragments cropped based on the segmentation mask image and the features extracted for target detection in the intermediate feature extraction stage are input into the target detection network. Regression calculation based on bounding box overlap optimization loss is performed, and the rail corrugation detection result is output.

2. The rail corrugation identification method according to claim 1, characterized in that, Based on the acquired raw rail image data, true image data of rail corrugation is generated, including: In the original image data of the rail, a manual annotation tool is used to draw the outline of the corrugated area pixel by pixel, and an annotation file containing the location information of the corrugated area is generated after the annotation is completed. The annotation file is registered pixel-level with the corresponding original rail image data to accurately distinguish the corrugated area from the background area in the two-dimensional pixel matrix of the original image. In the registration results, pixels belonging to the corrugated area are labeled as the first pixel category, and pixels belonging to the background area are labeled as the second pixel category, in order to construct true image data of rail corrugation that reflects the morphology and positional relationship of the corrugated area.

3. The rail corrugation identification method according to claim 1, characterized in that, The preprocessing rules for true image data of rail corrugation are as follows: In the true image data of rail corrugation, the corrugated area and background area corresponding to the annotation file are identified, and all pixel values ​​in the corrugated area are replaced with a preset first value, and all pixel values ​​in the background area are replaced with a preset second value. After the numerical replacement is completed, a linear normalization operation is performed on the pixel matrix of the entire image to map all pixel values ​​to the (0,1) interval proportionally, thus obtaining the preprocessed image data.

4. The rail corrugation identification method according to claim 1, characterized in that, The output of the intermediate feature extraction stage is averaged by channels and then processed by the Sobel operator to generate edge features, including: The multi-channel feature map output from the intermediate feature extraction stage is averaged along the channel dimension to obtain a single-channel feature map. Convolution operations are performed on the obtained single-channel feature maps by calling horizontal and vertical gradient operators with fixed parameters respectively, in order to extract gradient response maps along the horizontal and vertical directions. The gradient response map is subjected to amplitude calculation to fuse gradient information in the horizontal and vertical directions, thereby obtaining edge features that reflect the position and orientation of the rail corrugation boundary.

5. The rail corrugation identification method according to claim 1, characterized in that, The global semantic features, local texture features, and edge features are fused to obtain a segmentation mask image, including: Global semantic features, local texture features, and edge features are sequentially concatenated along the channel dimension to form a fused feature tensor. The fused feature tensor is input into the convolutional dimensionality reduction unit, and the number of channels is compressed to a preset dimension through one-dimensional convolution operation, and batch normalization is performed. A nonlinear activation function is applied to the batch normalization result to obtain the fused feature map; The fused feature map is input into the output convolutional layer of the image segmentation and recognition model to generate a segmentation mask map that corresponds to the rail region in spatial location and distinguishes the corrugated region from the background region in pixel category.

6. The rail corrugation identification method according to claim 1, characterized in that, The ROI image fragments cropped based on the segmentation mask image and the features extracted for target detection in the intermediate feature extraction stage are input into the target detection network. Regression calculation based on bounding box overlap optimization loss is performed, and the rail corrugation detection results are output, including: In the segmentation mask image, the boundary coordinates of the erosion region are determined by connected component analysis, and the boundary coordinates are expanded outward according to a preset ratio to form a clipping frame; The cropping box is used to extract the ROI image fragment containing the corrugated area from the original rail image, and the ROI image fragment is aligned with the preprocessed and adapted target detection features output from the intermediate feature extraction stage in the feature dimension. The aligned ROI image fragments and object detection features are input together into a deep convolutional neural network-based object detection network to generate a set of candidate bounding boxes. Perform DIoU Loss-based bounding box regression optimization on the candidate bounding box set to adjust the position and shape of each candidate box, and output the rail corrugation detection results containing the location and range of corrugation after optimization.

7. A rail corrugation identification system, characterized in that, The rail corrugation identification system is used to execute the rail corrugation identification method according to any one of claims 1-6, and the rail corrugation identification system includes: The acquisition unit is used to generate true image data of rail corrugation based on the acquired raw rail image data and to perform preprocessing. The extraction unit is used to feed the preprocessed image data into an image segmentation and recognition model with a multi-scale dilated convolutional feature extraction structure and a CBAM attention mechanism introduced after the output of multiple feature layers in the intermediate feature extraction stage, so as to extract global semantic features and local texture features. The fusion unit is used to perform channel averaging on the output of the intermediate feature extraction stage and then generate edge features using the Sobel operator, and fuse the global semantic features, local texture features and edge features to obtain a segmentation mask image; The output unit is used to input the ROI image fragment cropped based on the segmentation mask image and the features for target detection extracted in the intermediate feature extraction stage into the target detection network, perform regression calculation based on the bounding box overlap optimization loss, and output the rail corrugation detection result.

8. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores instructions that, when executed on a computer, cause the computer to perform the rail corrugation identification method according to any one of claims 1-6.

Images