Steel rail corrugation recognition method and system and storage medium

Abstract

The embodiment of the invention provides a steel rail corrugation recognition method and system and a storage medium, and belongs to the technical field of railway track detection. The steel rail corrugation identification method comprises the following steps: sending preprocessed steel rail corrugation true value image data into an image segmentation identification model to extract global semantic features and local texture features; performing channel averaging on the output of the intermediate feature extraction stage, generating edge features through a Sobel operator, and fusing the global semantic features, the local texture features and the edge features to obtain a segmented mask graph; and inputting the ROI image segments cut based on the segmentation mask image and the features extracted in the intermediate feature extraction stage and used for target detection into a target detection network together, executing regression calculation based on bounding box overlapping degree optimization loss, and outputting a rail corrugation detection result. According to the scheme of the invention, through multi-scale feature extraction and channel space attention fusion, the positioning precision and segmentation robustness of the rail corrugation region are effectively improved.

Description

Technical Field

[0001] This invention relates to the field of railway track inspection technology, specifically to a method for identifying rail corrugation, a system for identifying rail corrugation, and a storage medium. Background Technology

[0002] As a crucial component of railway tracks, rails bear the high-frequency loads and complex contact stresses repeatedly applied by train wheels during operation. Due to the rolling and sliding between the wheel tread and the rail surface, as well as the lateral action of the wheel flange, the rail surface undergoes periodic, uneven plastic deformation and wear along its length, resulting in a wavy geometric irregularity on the rail head tread surface—a phenomenon known as rail corrugation. Rail corrugation not only causes vibration and noise during train operation but also accelerates fatigue damage to track structural components, affecting driving safety and passenger comfort. Therefore, it is a key monitoring target in railway maintenance operations.

[0003] Currently, the main methods for detecting rail corrugation include manual inspection and mechanical measurement based on corrugation measuring instruments. Manual inspection typically relies on maintenance personnel using corrugation gauges to randomly inspect the rails on-site. This method is not only inefficient and labor-intensive, but the measurement results are also easily affected by human factors, posing a risk of missed detections and misjudgments. While corrugation measuring instruments can reduce manpower and improve measurement accuracy, they usually employ contact measurement or a one-chord N-point chord measurement method, requiring specialized hardware. Furthermore, the measurement data needs complex offline processing after acquisition to obtain the corrugation distribution, making real-time and rapid judgment and location impossible. In addition, these devices have poor adaptability to changes in external lighting, rail surface contamination, and environmental vibration, and lack stability under harsh operating conditions such as high-speed railways.

[0004] Existing image processing-based corrugation detection methods have improved detection efficiency to some extent. However, traditional image processing relies heavily on manually designed features and fixed thresholds, making it difficult to adapt to corrugation detection tasks under different track conditions, lighting conditions, and wear patterns. Furthermore, its ability to simultaneously capture long-distance continuous features and localized fine textures is limited, resulting in insufficient recognition accuracy and robustness. In summary, existing rail corrugation detection technologies still have significant shortcomings in terms of real-time performance and adaptability, and urgently need improvement. Summary of the Invention

[0005] The purpose of this invention is to provide a method, system, and storage medium for identifying rail corrugation, so as to at least solve the problems of insufficient real-time performance and adaptability of existing rail corrugation detection methods.

[0006] To achieve the above objectives, the first aspect of the present invention provides a method for identifying rail corrugation, the method comprising: generating ground truth image data of rail corrugation based on acquired original rail image data and performing preprocessing; feeding the preprocessed image data into an image segmentation and recognition model having 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; averaging the output of the intermediate feature extraction stage by 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; inputting the ROI image fragment 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.

[0007] Optionally, generating true image data of rail corrugation based on the acquired original rail image data includes: using a manual annotation tool to draw the corrugated area pixel by pixel in the original rail image data, and generating an annotation file containing the location information of the corrugated area after annotation; performing pixel-level registration between the annotation file and 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 result, marking pixels belonging to the corrugated area as the first pixel category and pixels belonging to the background area as the second pixel category, so as to construct true image data of rail corrugation reflecting the morphology and positional relationship of the corrugated area.

[0008] Optionally, the preprocessing rules for the true image data of rail corrugation are as follows: In the true image data of rail corrugation, the corrugated area and the background area corresponding to the annotation file are identified, 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 value replacement is completed, a linear normalization operation is performed on the pixel matrix of the entire image, and all pixel values ​​are mapped to the (0,1) interval proportionally to obtain the preprocessed image data.

[0009] Optionally, 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 relevant 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 operation 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.

[0010] Optionally, the global semantic feature extraction rule is 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 local texture feature extraction rule is as follows: In the convection feature branch set in parallel in the backbone network of the image segmentation and recognition model, convolution, normalization, and nonlinear activation operations are performed sequentially to extract and output local texture features from the feature representation of the original rail image.

[0011] Optionally, the output of the intermediate feature extraction stage is averaged across channels and then processed by the Sobel operator to generate edge features. This includes: averaging the multi-channel feature map output from the intermediate feature extraction stage along the channel dimension to obtain a single-channel feature map; performing convolution operations on the obtained single-channel feature map using horizontal and vertical gradient operators with fixed parameters to extract gradient response maps along the horizontal and vertical directions; and performing amplitude calculations on the gradient response maps to fuse the gradient information in the horizontal and vertical directions to obtain edge features reflecting the position and orientation of the rail corrugation boundary.

[0012] Optionally, the global semantic features, local texture features, and edge features are fused to obtain a segmentation mask image, including: sequentially concatenating the global semantic features, local texture features, and edge features along the channel dimension to form a fused feature tensor; inputting the fused feature tensor into a convolutional dimensionality reduction unit, compressing the number of channels to a preset dimension through one-dimensional convolution operations, and performing batch normalization processing; applying a nonlinear activation function to the batch normalization result to obtain the fused feature map; and inputting the fused feature map into the output convolutional layer of the image segmentation and recognition model to generate a segmentation mask image that corresponds to the rail region in spatial location and distinguishes the corrugated region from the background region in pixel category.

[0013] Optionally, 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 to perform regression calculation based on bounding box overlap optimization loss, and output the rail corrugation detection result. This includes: determining the boundary coordinates of the corrugation region in the segmentation mask image through connected component analysis, and expanding the boundary coordinates outward according to a preset ratio to form a cropping box; extracting ROI image fragments containing the corrugation region from the original rail image using the cropping box, and aligning the ROI image fragments with the preprocessed and adapted target detection features output from the intermediate feature extraction stage in the feature dimension; inputting the aligned ROI image fragments and target detection features into the 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 the rail corrugation detection result containing the location and range of the corrugation after optimization.

[0014] A second aspect of the present invention provides a rail corrugation recognition system, comprising: an acquisition unit for generating ground truth image data of rail corrugation based on acquired original rail image data and performing preprocessing; an extraction unit for feeding the preprocessed image data into an image segmentation and recognition model having 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 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.

[0015] On the other hand, the present invention 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.

[0016] Through the above technical solution, this invention constructs and preprocesses ground truth image data of corrugated rail from the original image data, introduces an image segmentation and recognition model with a multi-scale dilated convolution feature extraction structure and CBAM attention mechanism, and combines edge features generated by the Sobel operator to fuse global semantic features, local texture features, and edge features to generate a segmentation mask map. Then, based on the mask map, ROIs are cropped and target detection and bounding box optimization are performed in conjunction with intermediate features from the segmentation stage, achieving accurate localization and recognition of rail corrugated areas. The advantage of this technology lies in its ability to extract both long-distance continuous features and subtle texture features even under complex lighting and diverse wear patterns, improving the accuracy and stability of the detection results. Furthermore, feature reuse reduces computational redundancy, enhancing the real-time performance and adaptability of the overall detection.

[0017] Other features and advantages of the embodiments of the present invention will be described in detail in the following detailed description section. Attached Figure Description

[0018] The accompanying drawings are provided to further illustrate embodiments of the present invention and form part of the specification. They are used together with the following detailed description to explain the embodiments of the present invention, but do not constitute a limitation thereof. In the drawings: Figure 1 This is a flowchart of the steps of a rail corrugation identification method provided in one embodiment of the present invention; Figure 2 This is a schematic diagram of the image segmentation and recognition model structure provided in one embodiment of the present invention; Figure 3 This is a schematic diagram of a convolution module structure provided in one embodiment of the present invention; Figure 4 This is a schematic diagram of the ASPP module structure provided in one embodiment of the present invention; Figure 5 This is a schematic diagram illustrating the working principle of the CBAM attention mechanism provided in one embodiment of the present invention; Figure 6 This is a system structure diagram of a rail corrugation identification system provided in one embodiment of the present invention. Detailed Implementation

[0019] The specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0020] Figure 1 This is a flowchart of a rail corrugation identification method provided in one embodiment of the present invention. Figure 1 As shown, an embodiment of the present invention provides a method for identifying rail corrugation, the method comprising: Step S10: Generate true image data of rail corrugation based on the acquired original rail image data and perform preprocessing.

[0021] Specifically, generating true image data of rail corrugation based on the collected original rail image data includes: using a manual annotation tool to delineate the corrugated area pixel by pixel in the original rail image data, and generating an annotation file containing the location information of the corrugated area after annotation; performing pixel-level registration between the annotation file and 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 result, pixels belonging to the corrugated area are marked as the first pixel category, and pixels belonging to the background area are marked as the second pixel category, so as to construct true image data of rail corrugation reflecting the morphology and positional relationship of the corrugated area.

[0022] In this embodiment of the invention, when generating true image data of rail corrugation based on the acquired original rail image data, a manual annotation tool with pixel-by-pixel annotation function is first invoked in the acquired original rail image to finely delineate the contours of locations with corrugation defects in the image. During the annotation process, it is necessary to draw point by point along the true boundary of the corrugation defect, covering all peaks, valleys, and edge transition areas, avoiding omission of minor protrusions, depressions, or breaks to ensure the integrity of the subsequent feature extraction stage. After the annotation is completed, an annotation file containing pixel coordinate information of the corrugated area and a description of the closed contour is generated. This file maintains a one-to-one correspondence with the original image in terms of file name or index, ensuring the traceability of data matching.

[0023] After obtaining the annotation file, it is pixel-level registered with the corresponding original rail image. During the registration process, the contour of the corrugated area is accurately mapped to the actual pixel position in the original image by comparing the pixel coordinates recorded in the annotation file with the two-dimensional pixel matrix index position in the original image. This step ensures complete spatial consistency between the annotation information and the original image, meaning that each corrugated boundary point can accurately correspond to a pixel unit in the original image, thereby avoiding label offset or morphological distortion caused by registration errors.

[0024] In the registration results, all pixels located inside the corrugated area contour are uniformly assigned a preset first pixel category code, and all background pixels located outside the contour are uniformly assigned a preset second pixel category code. This method establishes a clear classification label relationship between the corrugated area and the background area in the data representation. The two-dimensional pixel matrix obtained through the above steps is the ground truth image data of the rail corrugated area. Visually, it presents a feature pattern showing a significant separation between the corrugated area and the background area, and can be directly used as a supervisory signal input for the image segmentation and recognition model, providing a high-quality training and validation data foundation for subsequent feature extraction, segmentation, and detection steps.

[0025] Through this process of pixel-by-pixel annotation, precise registration, and category encoding, the stability and consistency of the erosion area labels can be maintained under different lighting conditions, wear patterns, and surface contamination conditions, thereby effectively improving the accuracy and robustness of subsequent models in erosion identification tasks.

[0026] Preferably, the preprocessing rule for the true image data of rail corrugation is as follows: In the true image data of rail corrugation, the corrugated area and the background area corresponding to the annotation file are identified, 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 value replacement is completed, a linear normalization operation is performed on the pixel matrix of the entire image, and all pixel values ​​are mapped to the (0,1) interval proportionally to obtain the preprocessed image data.

[0027] In this embodiment of the invention, after generating ground truth image data of rail corrugation based on the acquired raw rail image data, preferably, preprocessing is performed on the ground truth image data so that the subsequent model can stably and accurately identify the corrugated areas. First, according to the coordinate information recorded in the annotation file, the pixel positions corresponding to the corrugated areas and the pixel positions corresponding to the background areas are identified one by one. The annotation file is usually generated after manually drawing the contours of the corrugated areas pixel by pixel in the early stage, and contains the boundary information and closed area coordinates of each corrugated area, thus enabling accurate matching with the ground truth image data.

[0028] After identifying the wavy region and the background region, a replacement operation is performed on the pixel matrix. Specifically, all pixel values ​​within the wavy 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 first and second values ​​should have sufficient numerical difference to ensure that the model can easily distinguish between the two types of regions during training, reducing confusion in the feature extraction stage. The numerical replacement is performed globally, rather than by calculating and generating values ​​pixel by pixel, to ensure processing efficiency and avoid blurring of class boundaries due to minor deviations introduced by floating-point operations.

[0029] After numerical replacement, linear normalization is performed on the pixel matrix of the entire image. Normalization maps pixel values ​​to the (0,1) interval proportionally, scaling by the difference between the minimum and maximum values ​​to ensure all pixel values ​​are within the same dimension. This not only improves the numerical stability of the model input but also avoids training non-convergence issues caused by differences in brightness and contrast under different image acquisition conditions. After normalization, the preprocessed image data maintains the distinction between the wavy region and the background region while meeting the model's input specifications in terms of numerical range.

[0030] In another possible implementation, during rail inspection operations on some railway lines, factors such as insufficient lighting during nighttime operations, numerous deposits on the rail surface, and compressed equipment acquisition channels can lead to blurred boundaries and uneven pixel grayscale in the ground truth image data generated after annotation of the original rail images. While conventional pixel value replacement and linear normalization can achieve class differentiation, they are prone to grayscale fluctuations in boundary transition areas, causing unstable boundary predictions during model training.

[0031] To address the aforementioned issue, a weighted smoothing process can be applied to the boundary between the wavy region and the background region before performing numerical replacement. Specifically, the Euclidean distance between the boundary pixels of the wavy region and the background pixels is first calculated. Pixels within a preset threshold range are assigned a weighted attenuation coefficient, causing their pixel values ​​to be linearly interpolated between a first value and a second value proportional to the distance. For example, a point 1 pixel away from the boundary can be assigned an interpolation weight of 0.8, a point 2 pixels away can be assigned an interpolation weight of 0.6, and so on, until the value is attenuated to the pure value of the background or wavy region. In this way, in the ground truth image, the boundary region will form a gray-scale gradient band, rather than an abrupt binary boundary.

[0032] This gradient processing is mapped to the (0,1) interval during the normalization stage, forming a label distribution that is closer to the actual wear edge. Although this method introduces continuous values ​​into the labeled data, it can significantly reduce gradient oscillations in the boundary region of the model and improve the stability of boundary localization. Especially in the presence of noise interference or surface contamination, it can effectively avoid the over- or under-segmentation of the wear region.

[0033] Step S20: 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 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.

[0034] Specifically, 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 (Convolutional Block Attention Module) 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 related to 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.

[0035] In this embodiment of the invention, the backbone network of the image segmentation and recognition model is equipped with a multi-scale dilated convolutional feature extraction structure. This structure, by setting different dilation rates in different convolutional branches, allows the receptive field to cover both long-distance features along the rail direction and detailed features in the vertical direction of rail corrugation while maintaining spatial resolution. In the corrugation recognition scenario, rail surface defects often manifest as periodic and long-extended waveform textures. Therefore, multi-scale dilated convolution can simultaneously capture the overall shape of the macroscopic waveform and the microscopic edge transitions, effectively improving feature representation capabilities.

[0036] After the output of multiple feature layers in the intermediate feature extraction stage of the backbone network, a CBAM channel attention module and a spatial attention module are sequentially set. The channel attention module calculates the weights of each channel based on the results of global average pooling and max pooling, thereby highlighting the channel features related to the corrugation texture; the spatial attention module generates a spatial weight distribution through convolution, further enhancing the positional response of the corrugated area and suppressing background interference. This joint channel and spatial attention mechanism can maintain feature focus on the real corrugated area even when there is uneven lighting on the rail surface or interference factors such as rust or oil stains.

[0037] A corrugated feature branch, set up in parallel with the backbone network, is dedicated to extracting local texture features. This branch maintains high spatial resolution and enhances the elongated texture pattern of rail corrugations through multi-layer convolution, normalization, and activation operations, providing fine structural information during feature fusion. Simultaneously, an edge feature generation branch, connected to the intermediate feature extraction stage of the backbone network, eliminates channel redundancy by averaging intermediate features 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 accurately reflects the position and orientation of the corrugation boundaries.

[0038] The outputs of the corrugation feature branch and the edge feature generation branch are concatenated with the global semantic features of the backbone network in the feature fusion unit along the channel dimension. After convolutional dimensionality reduction, normalization, and activation, a segmentation mask image is output. The uniqueness of this structure in the rail corrugation recognition scenario lies in its ability to not only take into account the multi-scale features of long-distance waveforms and local textures, but also to accurately characterize the boundary morphology of corrugation through the supplementation of edge features, thereby achieving high-precision defect segmentation in complex environments.

[0039] In one possible implementation, such as Figure 2 This paper demonstrates an image feature extraction and segmentation structure based on a ResNet backbone network. The input is a rail image with dimensions of 2048×H×W×3, where H and W represent the height and width of the input rail image, respectively. Their values ​​are determined by the imaging size of the actual acquisition device and are used to characterize the spatial size of the input image. First, the ResNet backbone network extracts features from multiple layers, including Layer 1 (256×H / 4×W / 4), Layer 2 (512×H / 8×W / 8), Layer 3 (1024×H / 16×W / 16), and Layer 4 (2048×H / 16×W / 16). The outputs of Layer 2 and Layer 3 are respectively connected to the CBAM attention module to enhance the channel and spatial features related to the corrugated region; the output of Layer 3 also generates edge features through the Sobel operator. The output of Layer 4 is fed into the ASPP (Dilated Convolutional Pyramid) module to capture multi-scale contextual information and combined with the corrugation feature branch (Conv-BN-ReLU structure, 256 channels). After all features are fused along the channel dimension, they are compressed to 1024 channels by 1×1 convolution + BN (Batch Normalization) + ReLU, then a 1×1 convolution classifier is used to generate the segmentation result. Finally, bilinear upsampling is used to restore the original image size, outputting a segmentation mask map containing the corrugated region and the background region. This structure achieves multi-source fusion of global semantic features, local texture features, and edge features, improving the accuracy and robustness of rail corrugation recognition. The structure of the convolution module is as follows: Figure 3 As shown.

[0040] Among them, such as Figure 4The ASPP module's input features are first processed through a two-dimensional convolution to generate basic features. Then, they are subjected to three sets of 3×3 dilated convolutions with different dilation rates (6, 12, and 18) to capture contextual information at different scales. Simultaneously, the input features are also processed by global average pooling to generate global contextual information. These four feature outputs are then concatenated with the original convolutional features along the channel dimension to achieve multi-scale information fusion. Finally, a 1×1 convolution is used for feature fusion and compression, resulting in a feature representation that combines local details with global semantics, thereby improving the accuracy and robustness of segmentation or detection.

[0041] The working principle of the CBAM attention module is as follows: Figure 5 As shown, the input features are fed into the channel attention module, where global average pooling and global max pooling are used to generate channel weights, weighting the original features along the channel dimension to highlight task-relevant channel information. Next, the channel-weighted features are fed into the spatial attention module, where attention maps are calculated in the spatial dimension (typically combining the results of average pooling and max pooling) to highlight key information regions and suppress irrelevant background. Both weighting operations use element-wise multiplication, ultimately outputting an optimized feature map. This module can enhance the response to the target region, reduce background interference, and improve feature representation capabilities while preserving important semantic information.

[0042] Preferably, the global semantic feature extraction rule is 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 the features at each scale to extract and output global semantic features.

[0043] In this embodiment of the invention, the backbone network of the image segmentation and recognition model is constructed with convolutional branch groups possessing a multi-scale dilated convolutional feature extraction structure. Different branches have convolutional kernels of the same size but with different dilation rates, in order to expand the receptive field without increasing the number of parameters. Specifically, branches with larger dilation rates can capture the long-distance waveform periodic features extending along the rail direction, while branches with smaller dilation rates retain detailed information and texture variations in local areas of the rail surface, thus forming a feature representation that considers both macroscopic morphology and local details.

[0044] After the multi-scale convolution operation is completed, the feature maps output by each branch are concatenated or weighted pixel-wise along the channel dimension to fuse feature information from different scales. The convergence process can use a channel-wise weighting method to dynamically adjust the proportion of features at each scale in the fusion result according to the response intensity of the features, avoiding weak features being overwhelmed or strong features being overly dominant, and maintaining the balance of global features.

[0045] The global semantic features extracted in the above manner can accurately reflect the spatial distribution trend and overall morphology of corrugation defects along the entire length of the rail. Even under conditions of uneven lighting, rail surface contamination, or varied wear patterns, the features can still maintain stability and robustness, providing a solid semantic foundation for subsequent feature fusion and mask generation.

[0046] Preferably, the extraction rule for local texture features is as follows: in the corrugated 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.

[0047] In this embodiment of the invention, a convolutional processing link that maintains high spatial resolution is constructed in the wavy feature branch set in parallel in the backbone network of the image segmentation and recognition model. The input of this branch is a feature map synchronized with the output of the intermediate feature extraction stage of the backbone network. The basic texture pattern of the local region is extracted through the first layer of convolution, and the feature distribution is standardized by batch normalization to reduce the statistical differences between different batches of input. Subsequently, a nonlinear activation function (such as ReLU or LeakyReLU) is introduced to enhance the nonlinearity of feature expression, thereby strengthening the elongated texture and periodic waveform details in the wavy region.

[0048] In this branch, multiple layers of convolution-normalization-activation units can be stacked to progressively expand the receptive field while maintaining the integrity of local structural features. The size, stride, and filling method of each convolutional kernel can be designed based on the average width and spacing of the corrugated stripes in the rail image to ensure accurate capture of the directionality and continuity of the rail surface texture.

[0049] The local texture features extracted by this rule can supplement the shortcomings of the global semantic features at the detail level, making the subsequent fusion with edge features and global semantic features more refined. Thus, even when the ripple boundary is relatively blurred and the wear pattern is complex, high-precision segmentation and positioning can still be achieved.

[0050] Step S30: 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.

[0051] Specifically, the output of the intermediate feature extraction stage is averaged across channels and then processed by the Sobel operator to generate edge features. This includes: averaging the multi-channel feature map output from the intermediate feature extraction stage along the channel dimension to obtain a single-channel feature map; performing convolution operations on the obtained single-channel feature map using horizontal and vertical gradient operators with fixed parameters to extract gradient response maps along the horizontal and vertical directions; and performing amplitude calculations on the gradient response maps to fuse the gradient information in the horizontal and vertical directions to obtain edge features reflecting the position and orientation of the rail corrugation boundary.

[0052] Furthermore, the global semantic features, local texture features, and edge features are fused to obtain a segmentation mask image, including: sequentially concatenating global semantic features, local texture features, and edge features along the channel dimension to form a fused feature tensor; inputting the fused feature tensor into a convolutional dimensionality reduction unit, compressing the number of channels to a preset dimension through one-dimensional convolution operations, and performing batch normalization processing; applying a nonlinear activation function to the batch normalization result to obtain the fused feature map; and inputting the fused feature map into the output convolutional layer of the image segmentation and recognition model to generate a segmentation mask image that corresponds to the rail region in spatial location and distinguishes the corrugated region from the background region in pixel category.

[0053] In this embodiment of the invention, when generating edge features using the Sobel operator after channel averaging of the output of the intermediate feature extraction stage, the multi-channel feature map output by the image segmentation and recognition model in the intermediate feature extraction stage is first obtained. This feature map typically originates from the intermediate layer output of the backbone network and parallel branches after a certain number of layers, and contains high-dimensional feature information processed by multi-scale dilated convolution and attention mechanisms. These features may contain dozens or even hundreds of feature channels in the channel dimension, each channel responding to different convolution kernel filtering characteristics and receptive fields. Therefore, they contain both semantic information of the rippled region and texture details and background noise information.

[0054] To extract boundary-indicating feature information from these features, channel averaging is first required. Specifically, the multi-channel feature map is averaged pixel-by-pixel along the channel dimension; that is, for each pixel at a spatial location, the average value of all channels at that location is calculated to generate a single-channel feature map. This step compresses the multi-channel response intensity information into a single grayscale distribution, thereby eliminating the directional bias caused by differences in channel responses, while preserving the overall structural contour information, making subsequent gradient calculations more stable.

[0055] After obtaining the single-channel feature map, convolution operations are performed using the Sobel horizontal and vertical gradient operators with fixed parameters. The horizontal gradient operator detects grayscale changes in the horizontal direction of the image, capturing the lateral edge features formed when the rail corrugation extends longitudinally; the vertical gradient operator detects grayscale changes in the vertical direction of the image, capturing the variation features of the corrugation edge in the rail head cross-section direction. The convolution operation adopts a standard two-dimensional convolution form, with a kernel size typically of 3×3. The parameters are fixed and not used in training to ensure the stability and consistency of edge detection.

[0056] After obtaining the gradient response maps in the horizontal and vertical directions respectively, amplitude calculation is required to fuse them. Amplitude calculation can be achieved by summing the squares and taking the square root of the horizontal and vertical gradient values ​​for each pixel, and then taking the square root to obtain the gradient amplitude of that pixel. The resulting gradient amplitude map can simultaneously reflect the intensity of edge changes in both directions, forming a comprehensive description of the position and orientation of the rail corrugation boundary. After this processing, the output edge feature map visually appears as a bright area distributed along the corrugation contour, forming a clear contrast with the background area, while preserving boundary details and exhibiting strong robustness even under uneven lighting or surface stains.

[0057] Furthermore, when fusing the global semantic features, local texture features, and the generated edge features to obtain the segmentation mask, these three types of features are first concatenated sequentially along the channel dimension to form a fused feature tensor. The global semantic features mainly contain information on the overall structure of the rail and the large-scale distribution of corrugation defects; the local texture features contain the fine textures and local patterns of the corrugations; and the edge features provide precise boundary contour information. The concatenation method typically involves directly superimposing channels along the channel dimension, merging the number of channels for each type of feature into a higher-dimensional feature set, so that subsequent convolutional operations can simultaneously perceive the spatial distribution relationships of different types of features.

[0058] After the fused feature tensor is formed, it is input into a convolutional dimensionality reduction unit, where the number of channels is compressed to a preset dimension through one-dimensional convolution (1×1 convolution). Convolutional dimensionality reduction not only reduces subsequent computation but also learns the optimal combination of different feature channels through the weights of the convolution kernel, preserving useful information while suppressing redundant features. After dimensionality reduction, batch normalization is performed to adjust the distribution differences of features across different batches of input, making the training process more stable and accelerating convergence.

[0059] Applying a non-linear activation function, such as ReLU or LeakyReLU, to the batch-normalized results introduces non-linear mapping capabilities, enhancing the model's ability to express complex feature combinations. Through non-linear activation, the model can better separate rippled regions from background regions, maintaining a good classification margin even when they are similar in grayscale or texture distribution.

[0060] The fused feature map, after dimensionality reduction, normalization, and activation processing, is input into the output convolutional layer of the image segmentation and recognition model. The number of convolutional kernels in this output convolutional layer is consistent with the number of categories (usually 2, corresponding to the wavy region and the background region). After the convolution operation, a two-dimensional feature map with the same spatial location as the input image is output, and a segmentation mask map is generated by normalization functions such as Softmax or Sigmoid. This segmentation mask map can clearly distinguish the wavy region and the background region in terms of pixel categories, providing an accurate region reference for subsequent ROI cropping and object detection.

[0061] By using the aforementioned rules for edge feature generation and multi-feature fusion, in the rail corrugation recognition scenario, global distribution information, local texture details, and precise boundary information can be effectively combined. Thus, even under conditions of complex corrugation morphology, variable lighting conditions, and strong background interference, it can still output a segmentation mask image with accurate boundaries and complete contours.

[0062] Step S40: Input the ROI image fragment cropped based on the segmentation mask image and the features extracted for target detection in the intermediate feature extraction stage into the target detection network, perform regression calculation based on bounding box overlap optimization loss, and output the rail corrugation detection result.

[0063] Specifically, in the segmentation mask image, the boundary coordinates of the corrugated region are determined through connected component analysis, and the boundary coordinates are expanded outward according to a preset ratio to form a clipping box. Using the clipping box, a Region of Interest (ROI) image fragment containing the corrugated region is extracted 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 fragment and the target detection features are input together into a target detection network based on a deep convolutional neural network to generate a candidate bounding box set. A bounding box regression optimization based on DioULoss (Distance Intersection over Union) loss is performed on the candidate bounding box set to adjust the position and shape of each candidate box, and after optimization, a rail corrugation detection result containing the location and extent of the corrugation is output.

[0064] In this embodiment of the invention, when the ROI image fragment cropped based on the segmentation mask map and the features extracted for target detection in the intermediate feature extraction stage are input into the target detection network, connected component analysis is first performed on the segmentation mask map to identify all independent wavy regions. Connected component analysis uses a pixel-by-pixel scanning method to aggregate pixels with the same adjacent pixel value and belonging to the wavy region category into the same connected component, and assigns a unique identifier to each connected component. For each connected component, the boundary coordinates of its minimum bounding rectangle are calculated. These coordinates consist of the minimum row index, the maximum row index, the minimum column index, and the maximum column index, accurately encompassing all pixels of the connected component.

[0065] To avoid the cropped ROI image fragments being too close to the boundary of the erosion region, resulting in insufficient contextual information, the boundary coordinates are expanded outward according to a preset ratio after determining the minimum bounding rectangle. The expansion ratio can be set according to the typical size of the erosion feature and the input size requirements of the subsequent detection network, for example, expanding the width and height of the rectangle by 5% to 15% in each direction. The expanded rectangle is the final cropping box. This cropping box can retain a certain amount of track surface background information while ensuring that the erosion region is included, thereby improving the contextual understanding ability in the detection stage.

[0066] Using the cropping box, corresponding Region of Interest (ROI) image segments are extracted from the original rail image. The extraction process must maintain the original resolution and aspect ratio of the image to avoid distortion introduced by scaling or interpolation that could affect feature quality. Simultaneously, feature regions corresponding to the ROI image segments are extracted from the feature map output from the intermediate feature extraction stage, based on the spatial location of the cropping box. To ensure dimensional consistency between the ROI image segments and the target detection features, preprocessing adaptation operations are performed on the extracted feature regions, such as bilinear interpolation to adjust size, channel alignment, and normalization, ensuring a one-to-one correspondence between them and the ROI image segments in terms of feature dimensions.

[0067] After alignment, the ROI image fragments and the adapted object detection features are input into an object detection network built on a deep convolutional neural network. The front end of this network typically contains multiple convolutional layers and downsampling layers to further refine the feature information within the ROI and generate a multi-scale set of candidate bounding boxes. These candidate boxes may have redundancy in position, size, and aspect ratio, therefore they need to be filtered and optimized after generation.

[0068] For the candidate bounding box set, bounding box regression optimization based on DIoULoss is performed. DIoULoss, in addition to calculating the IoU (Intersection over Union) between the predicted and ground truth bounding boxes, introduces an extra Euclidean distance penalty term between the center points of the predicted and ground truth bounding boxes, thus considering both the overlap area and the matching degree of the center positions during the optimization process. The optimization process iteratively adjusts the network parameters through backpropagation, causing the predicted bounding boxes to gradually approximate the boundary of the actual corrugated region in both position and shape.

[0069] After optimization, the resulting detection results include the position coordinates of each corrugated area in the original image, the size of the bounding rectangle, and the corresponding confidence score. The final output rail corrugation detection results can accurately mark the location and extent of the corrugated areas in space, and maintain clear boundary separation and precise location detection even when adjacent defects are close or have complex shapes.

[0070] In another possible implementation, during actual railway field inspections, the morphology of rail corrugation exhibits a strong directionality extending along the rail axis. Furthermore, under high-speed acquisition conditions, the camera may experience slight tilt or jitter, leading to a slight rotational deviation in the connected components generated by the segmentation mask image in spatial coordinates. If this deviation is directly extracted using axis-aligned rectangular clipping boxes for ROI extraction, it often results in the inclusion of excessive irrelevant background or the clipping of the corrugated region's ends, thus affecting the positioning accuracy of subsequent detection networks.

[0071] To address this, a minimum bounding rectangle clipping rule is introduced during the connected component analysis phase. Specifically, after determining the pixel set for each wavy region, a minimum bounding rectangle algorithm based on Principal Component Analysis (PCA) is used to calculate the principal direction angle of the pixel set, and this angle is used as a rotation reference to generate the minimum bounding rectangle. Subsequently, the rectangle boundary is expanded outward by a preset ratio in the rotating coordinate system. The resulting clipping box not only aligns with the direction of the wavy region but also minimizes the inclusion of background interference regions.

[0072] When the coordinates of the rotated rectangle are returned to the original image space, they are precisely mapped through an affine transformation, and the ROI image fragments are extracted in a rotated manner. At the same time, feature regions with the same rotation angle and position are extracted in the intermediate feature map, maintaining the geometric consistency between the two. In this way, the ROI image fragments and corresponding features input to the object detection network can be perfectly aligned in the ripple direction, significantly reducing the modeling pressure on unnecessary backgrounds during the learning process of the detection network.

[0073] This method is particularly suitable for long wavelengths, non-vertical boundaries, and slightly tilted cameras in rail corrugation detection scenarios. It can significantly improve positioning accuracy and boundary fitting effect, and reduce shape deviation of the detection frame.

[0074] Example: Step 1: The image segmentation model preprocessing process is as follows.

[0075] 1) Sample railway image data and construct true railway image data; the true railway image data is rail corrugation profile image data drawn based on railway image data.

[0076] 2) Binarize the true railway image data: In the rail corrugation profile image data, set the background pixel value to 0 and the rail corrugation pixel value to 1, and represent the rail corrugation profile image data in numerical form.

[0077] 3) Normalize the true value of the railway according to the following formula:

[0078] Where i and j represent the row and column numbers 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, and I(i,j) belongs to (0,1).

[0079] Step 2: The process of identifying rail corrugation is as follows.

[0080] In the feature extraction stage, the input is the original rail image (3200×2400×3) after standardization (μ=[0.485,0.456,0.406],σ=[0.229,0.224,0.225]). 1) Backbone Network: Multi-scale features are extracted using ResNet50 Layers 1-3; CBAM attention modules are inserted after Layer 2 (256-dimensional) and Layer 3 (512-dimensional) to calculate channel weights (GAP-MLP) and spatial weights (Conv7×7), respectively; Layer 4 outputs 2048-dimensional features which are then fed into the improved ASPP module. A 3×3 dilated convolution with asymmetric dilationrate (4,8) is used; 1×1 convolution and global average pooling branches are preserved; and ripple feature enhancement is achieved.

[0081] 2) Dedicated branch: Draw a wave-moistening feature branch (2 layers 3×3Conv+BN+ReLU) from the ASPP output - output 256-dimensional local texture features.

[0082] 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.

[0083] Step 3: Target Detection Stage.

[0084] 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.

[0085] 2) YOLOv5+ detection: Shares ResNet50 Layer 3 features (avoids redundant calculations); uses DIoULoss to optimize bounding box regression.

[0086] 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.

[0087] 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.

[0088] Those skilled in the art will understand that all or part of the steps in the methods of the above embodiments can be implemented by a program instructing related hardware. This program is stored in a storage medium and includes several instructions to cause a microcontroller, chip, or processor to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes various media capable of storing program code, such as a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk.

[0089] The optional embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details described above. Within the scope of the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the protection scope of the embodiments of the present invention. It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any suitable manner without contradiction. To avoid unnecessary repetition, the embodiments of the present invention will not further describe the various possible combinations.

[0090] Furthermore, various different embodiments of the present invention can be combined in any way, as long as they do not violate the spirit of the embodiments of the present invention, they should also be regarded as the content disclosed by the embodiments of the present invention.

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 after the output of multiple feature layers in the intermediate feature extraction stage, in order to extract global semantic features and local texture features. 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 image segmentation and recognition model includes: A backbone network with a multi-scale dilated convolutional feature extraction structure is configured to extract global semantic features from the input image. The CBAM channel attention module and spatial attention module are set after the output of multiple feature layers in the intermediate feature extraction stage of the backbone network to enhance the relevant feature response of the erosion region. A wave-moistening feature branch, set in parallel with the backbone network, is used to extract local texture features; The edge feature generation branch, which is connected to the intermediate feature extraction stage of the backbone network, is used to perform Sobel operator operation on the channel-averaged intermediate features to obtain edge features; The feature fusion unit, connected to the wave-moistening feature branch and the edge feature generation branch, is used to fuse global semantic features, local texture features and edge features in the channel dimension and output a segmentation mask map.

5. The rail corrugation identification method according to claim 4, characterized in that, The rules for extracting 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 rules for extracting local texture features are as follows: In the corrugated 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.

6. 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.

7. 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.

8. 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.

9. A rail corrugation identification system, characterized in that, 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.

10. 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 1-8.

Images