A method and system for locating and identifying wheel tread defects in rail vehicles

By using a dual-candidate segmentation model and grid window technology, the spatial propagation of window classification results is limited, solving the problems of detection efficiency and consistency in the inspection of wheel treads of rail vehicles, and achieving accurate defect localization and type identification.

CN123118607APending Publication Date: 2026-11-17LANZHOU JIAOTONG UNIV
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202611516870.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-09-28
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

In the detection of defects in the wheel tread of rail vehicles, the efficiency and consistency of the detection are affected by personnel experience, workload and on-site environment, and the window classification results may change the already determined defect location boundary.

Method used

A dual-candidate segmentation model is used to fuse candidate masks to generate candidate windows. A supplementary window is generated by the intersection of the grid window and the set of visible tread surfaces outside the candidate mask. The spatial propagation range of window classification is limited. The defect type is determined by combining deep features and local image features, forming a set of background, defect and unknown, and fixing the type discrimination within the defect localization range.

Benefits of technology

It effectively reduces the propagation of window results outside the candidate area, provides accurate defect location and type identification, and improves the automation and consistency of detection.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123118607A_ABST
    Figure CN123118607A_ABST
Patent Text Reader

Abstract

This invention discloses a method and system for locating and identifying defects in the tread surface of rail vehicles. The method involves acquiring the region of interest (ROI) on the tread surface and fusing the foreground responses of two candidate segmentation models trained on different training sets to form a candidate mask. Candidate windows are generated from the candidate connected regions, and supplementary windows are generated outside the candidate mask in the visible tread surface. The candidate windows provide classification scores only to pixels within their corresponding connected regions, while the supplementary windows provide classification scores only to visible pixels outside the candidate mask. Based on valid votes, a set of defects, background, and unknowns is determined, with locations outside the tread surface considered invalid regions. These sets are then fixed, and the window and pixel conditional type scores are fused only within the defect set to determine wear or peeling, outputting the defect location, type, and two conditional type scores. This method limits the spatial propagation of window classification results, ensuring that type discrimination does not alter the predetermined defect location boundaries, and can assist manual identification and inspection of defects in the tread surface of rail vehicles.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to the field of visual inspection and digital image processing technology for rail vehicles, and in particular to a method and system for locating and identifying the type of defects in the wheel tread of rail vehicles. Background Technology

[0002] In some rail vehicle maintenance scenarios, the detection of defects in rail vehicle wheelset treads still requires manual observation or verification. The efficiency and consistency of judgments are easily affected by factors such as personnel experience, workload, and the on-site environment. Visual inspection technology can automatically process images of rail vehicle wheelset treads, providing auxiliary information for identifying defect locations and types. Rail vehicle wheelset treads may exhibit appearance anomalies such as scratches, abrasions, peeling, and wear. The image representation and processing requirements for different defects vary; this implementation method uses wear and peeling as the target defect types. Visual inspection of rail vehicle wheelset treads typically includes tread area extraction, defect location determination, and defect type discrimination. The tread area is used to define the inspection object, and the defect area within the tread is used to indicate the location of the anomaly; these two belong to different processing levels.

[0003] Chinese patent application CN116385545A discloses a method for locating and extracting train wheelset tread images, which obtains a binary image of the tread through a segmentation network and fits the boundary to output the tread area for subsequent detection; Chinese patent CN108520274B discloses a method for detecting defects on highly reflective surfaces, which obtains local patches through image processing and then uses a neural network to determine whether the patches are defects.

[0004] When relying solely on candidate segmentation results, low-response defects may not be included in subsequent discrimination. When using overlapping windows to supplement discrimination, if the window result is assigned to the entire rectangular window, the classification result may propagate to the background or other candidate regions that have no corresponding relationship with the window context. Pixels without windows providing effective information also have different meanings from pixels that have been judged as background by the window.

[0005] Furthermore, window context and pixel local details can provide type information at different scales. If the type classification result is re-involved in background and defect judgment, it may change the already formed defect range, causing type error to have a reverse impact on localization. Therefore, it is necessary to establish a clear relationship between candidate coverage, window score space scope, unknown states, and type discrimination within the localization range. Summary of the Invention

[0006] This invention provides a method and system for locating and identifying defects in the tread surface of rail vehicles, thereby limiting the spatial propagation range of window classification scores and ensuring that defect type identification does not change the already determined defect location boundary.

[0007] The method first obtains the region of interest on the tread surface, and then fuses the foreground responses of two candidate segmentation models to form a candidate mask. Candidate windows are generated based on the candidate connected regions, and a mesh window is generated within the region of interest on the tread surface. A supplementary window is retained based on the intersection of the mesh window and the set of visible tread surfaces outside the candidate mask.

[0008] For a candidate window, its effective support range is limited to the intersection of the window and the bound candidate connected regions; for a supplementary window, its effective support range is limited to the intersection of the window and the set of visible pixels outside the candidate mask. The window can read the surrounding image as classification context, but only provides classification scores to pixels within its effective support range.

[0009] For each pixel, the number of valid votes is recorded, and the background, wear, and stripping scores are summarized. Pixels with zero votes within the visible tread area are retained as unknown, while pixels outside the tread area are considered invalid regions. Pixels with votes are determined as either background or defects based on the summarized scores, thus forming a mutually distinguishable set of backgrounds. B Defect set D and unknown sets U .

[0010] Fixing the above three sets, the wear and stripping scores in the window summary scores are converted into window condition type scores according to equation (5); pixel condition type scores are obtained from deep features and local image features at the same location; only when... D The internal fusion of two condition types of scores and selection of wear or peeling, in B and U Type rewriting is not performed internally.

[0011] Candidate connected region support constraints can reduce the propagation of window results outside the candidate region; supplementary windows provide discrimination opportunities for visible locations not covered by the candidate mask; unknown states are used to retain locations without valid window evidence; fixed location sets ensure that type discrimination does not expand or shrink the defect range when using pixel details.

[0012] This invention also provides a corresponding system, including a tread image acquisition module 101, a candidate and supplementary coverage module 102, a window positioning module 103, a pixel type module 104, a positioning constraint fusion module 105, and a result output module 106. These modules are software functional modules divided according to processing functions, implemented by a processor executing programs in the memory, and transmitting images, supporting sets, fractional images, and masks under the same tread region of interest coordinates. This invention also provides an electronic device, including a memory and a processor. The memory stores a computer program, and when the processor executes the computer program, it implements the above-described method for locating and identifying the type of defects in the tread of a rail vehicle wheelset. This invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described method for locating and identifying the type of defects in the tread of a rail vehicle wheelset. Attached Figure Description

[0013] Figure 1 This is a module relationship diagram of a rail vehicle wheelset tread defect location and type identification system.

[0014] Figure 2 The overall flowchart for candidate coverage, supplementary coverage, defect localization, and type discrimination within the localization range.

[0015] Figure 3 This diagram illustrates the effective support range of the candidate window and the supplementary window, as well as the pixel-by-pixel voting process.

[0016] Figure 4 A schematic diagram illustrating the joint discrimination of window and pixel type for fixing the defect location range.

[0017] Figure 5 This is a diagram showing the relationship between the fusion of dual-candidate models and the fusion of two types of conditional information.

[0018] Figure 6 This is a diagram showing the correspondence between the actual effective range of the tread surface, candidate masks, and supplementary support sets.

[0019] Figure 7 This diagram shows the actual candidate window, supplementary window, effective support scope, number of votes, and window stage positioning results.

[0020] Figure 8 Example diagram of type fusion and consistency of location range within a fixed defect location range.

[0021] Figure 9 A schematic diagram of the network structure for defect localization and window-pixel joint type recognition in a dual-candidate model. Detailed Implementation

[0022] An embodiment of the present invention is described below with reference to the accompanying drawings. This method processes visible light images of the wheel tread of a rail vehicle, defining wear as the continuous abrasion or texture change on the tread surface due to friction, and delamination as the local peeling or layering of the tread material; the specific determination of the two categories is made by corresponding labeled samples.

[0023] like Figure 9 As shown, the network structure consists of a dual-candidate defect localization part and a window-pixel joint type recognition part within a fixed defect localization range. The dual-candidate defect localization part includes candidate segmentation network A and candidate segmentation network B, which have the same structure but independent training parameters, and output foreground responses respectively. and The larger value of the two responses is taken as the fused response. P Subsequently, candidate masks are formed according to the threshold processing, connectivity growth, and area screening in step S2. C The joint type identification part includes window branching and pixel branching. Window branching extracts contextual features from multi-scale windows, while pixel branching extracts multi-layer deep features and local statistical features from the same tread region of interest. Window voting forms and fixes the defect set. D After that, only D The condition type score of the two branches is merged to determine wear or stripping. The MAX value in the figure represents the pixel-wise representation of the condition type. and Take the larger value at the same pixel position. Figure 9 The model modules and data transfer relationships are represented. Candidate mask post-processing, effective window support range, and pixel-by-pixel voting relationships are executed according to steps S2 and S3, respectively.

[0024] like Figure 1 and Figure 2 As shown, step S1 uses U²-Net to segment the wheelset image, blacking out pixels outside the tread and cropping the region of interest (ROI) by the bounding box of the tread mask, maintaining the original aspect ratio. Subsequent steps read the ROI, and the tread segmentation mask is not used as a defect mask.

[0025] Let the width and height of the ROI be respectively W and H The pixel position is x , y The foreground responses of the two candidate models are and Candidate masks are C It can be seen that the set of tread pixels is V Supplementary support set is F ; Indicates the first j One window, This represents the valid support set for the window. Indicates window to category c The score, n Indicates the number of valid votes. Indicates the total pixel score; c The values ​​of 0, 1, and 2 correspond to background, wear, and peeling, respectively. B , D , U These represent the sets of background, defects, and unknown pixels within the visible tread area, respectively. E Indicates that the ROI does not belong to V Invalid region.

[0026] In step S2, both the first and second candidate segmentation models employ a binary U-shaped segmentation network to merge wear and peeling into candidate foreground images. In one training embodiment, the network input is a 640×640 pixel three-channel image. The number of channels in the four-level encoding stage is 32, 64, 128, and 256, respectively; the number of channels in the bottleneck layer is 512; and the number of channels in the four-level decoding stage is 256, 128, 64, and 32, respectively. Features are passed to the decoding end through same-scale skip connections, and finally, a 1×1 convolution outputs two channels: background and candidate foreground. The first candidate segmentation model uses a basic candidate training set, while the second candidate segmentation model adds confirmed background samples of reflection, rust, and difficult tread edges to the training configuration. The corresponding training sets contain 158 and 159 training images, respectively, and 6 validation images are used. Training employs the sum of cross-entropy loss and Dice loss, using the Adam optimizer with an initial learning rate of 5×10⁻⁶. -5 The batch size is 1, and the maximum number of training epochs is 40. Difficult background samples are selected from high-response regions generated by the model in non-target areas, and added to the second training set after verification as non-target defects according to annotation rules. The aforementioned 158 and 159 training images and 6 validation images are used only for training and validation of the two candidate segmentation models. The two candidate segmentation models have their parameters saved separately, using the same image transformation and output coordinates, and the maximum value is taken for the foreground response at the same pixel location.

[0027] According to the fusion response P Using high-threshold pixels as seeds, eight-neighbor connectivity growth is performed on pixels corresponding to the growth threshold, and candidate masks are generated through area screening. C In one embodiment, the seed threshold is 0.45 and the growth threshold is 0.25; the candidate mask is restored to the ROI size using nearest neighbor interpolation. Full-image prediction can be fused with overlapping slice prediction in the same coordinate system. When the reflection is strong and the initial candidate area is small, the brightness-adjusted image can be used to supplement the prediction.

[0028] right CPerform connectivity analysis and save region identifiers. Larger candidate connected regions are covered by overlapping windows, while other regions are formed into candidate windows using bounding boxes; each candidate window is bound to one candidate connected region. Its effective support range equal and The intersection of the regions. Even if the same window contains other connected regions, it does not provide the score of that window to those other connected regions.

[0029] The set of visible tread pixels is formed by pixels in the ROI whose maximum color channel value is greater than the near-black threshold. V ROI that does not belong V The pixels were determined to be invalid regions. E and from V Excluded C The prospects are supported by a supplementary set of data:

[0030] like Figure 6 As shown, V , E , C and F Determined within the same ROI coordinate system. In one embodiment, the near-black threshold is set to 3. E It does not participate in window voting, background judgment, or type discrimination. Category masks can be represented using the same stored value. E And zero vote positions, but when reading the results, through V Distinguishing: Belongs to V And the pixel with zero votes is U It does not belong to V The number of pixels is E .

[0031] Generate overlapping mesh windows within the ROI at a preset step size, preserving their overlap with... F A window whose intersection pixel count and intersection area ratio satisfy the lower limit is used as a supplementary window, and its effective support range is set to... and F The intersection of the ROI. In one embodiment, the window side length is 0.24 times the shorter side of the ROI and is limited to 96 to 768 pixels, and the window overlap ratio is 0.5; the number of intersection pixels is not less than 32 and the proportion of the window area is not less than 0.01. F Fix it before the window is generated so that the same pixel can receive votes from multiple supplementary windows.

[0032] In step S3, image patches are cropped around each window at adjacent, local, and wider scales, proportionally scaled, and filled to the encoder input size. In one embodiment, the three cropping scales have expansion ratios of 0, 0.20, and 0.60 relative to the maximum side of the window, respectively, and the cropped result is filled to 224×224 pixels. ResNet34 with frozen parameters is used to extract 512-dimensional features for each scale. The features from the three scales are concatenated, standardized, and then input into a three-class logistic regression classifier. The regularization parameters of this classifier... C The maximum number of iterations is set to 2500, and the output consists of three non-negative classification scores: background, wear, and stripping. The sum of these three scores is 1. These scores are used for comparisons within the same classifier output and for cross-window arithmetic summation, and the context-based pruning does not expand the output. .

[0033] The window classifier is trained using explicitly labeled background, worn, and stripped windows. In one embodiment, windows with a support ratio of less than 10% within the unexpanded window for positive class labeling are not used as corresponding positive class supervision; the remaining training windows are weighted according to class and source image, and divided into five groups based on source images for training and validation, ensuring that windows from the same ROI remain within the same data group. Manual annotation is used only for model training and validation; the windows and effective support range of the test image are automatically determined by the test image and candidate masks. The five training and validation groups based on source images are for the window classifier, and their data statistics are independent of the candidate segmentation model.

[0034] like Figure 3 As shown, each window only... The score is provided internally. The number of valid votes for the same pixel is:

[0035] in For the characteristic function, the pixel belongs to If the value is 1, then 0. Only the scores of windows containing that pixel are summed and averaged based on the number of valid votes.

[0036] exist V inside, when n When it is greater than 0, in , and Select the category corresponding to the highest score, composed of worn or peeled pixels. D Background pixels B When two or more summary scores are tied for the highest value, the category that appears first is selected according to a fixed category index order, namely background, wear, and peeling. n When the value is 0, the pixel is assigned to... UNo comparison of the three maximum values ​​is performed; E The above judgment is not applied. This distinguishes between invalid positions, positions without valid window evidence, and positions with existing window evidence that are judged as background.

[0037] Step S4 receives a fixed B , D , U And the three categories of total scores. For D For each inner pixel, the sum of the wear and tear window scores is converted into a window condition type score according to equation (5). :

[0038] In the formula ε To preset a lower limit for positive numbers, the denominator is taken as... and sum and ε The larger value in c Only values ​​of 1 and 2 are used. This value serves as a condition type score for detected locations between wear and peeling. It is not required that the sum of the two scores be exactly 1 when numerical protection is active, nor is it used to determine whether there are defects in the background or unknown locations. In one embodiment, ε is set to 10. -12 .

[0039] The pixel branch uses the same ROI as input and extracts features from four residual layers using a frozen pre-trained ResNet34. Layers with more than 64 channels are seeded and compressed to 64 channels using random linear projection independent of the image and label, and then aligned to the grid of the first residual layer using bilinear interpolation. The resulting 256-dimensional deep features are concatenated with the grayscale mean and standard deviation of RGB, Lab, grayscale, 3×3, 9×9, and 21×21 neighborhoods, as well as the horizontal and vertical gradients, to form a 271-dimensional pixel description.

[0040] Based on the positions of wear and peeling labeled pixels in the feature grid, bilinear interpolation is used to obtain pixel descriptions, which are then standardized to train a binary logistic regression classifier. In one embodiment, a maximum of 512 supervised pixels are extracted for each image type, and regularization parameters are trained using three pre-determined random seeds. C A classifier with a value of 0.1 and a maximum number of iterations of 1500 is used. The arithmetic mean of the two-class outputs of the three classifiers is taken by category to obtain the pixel conditional type score. Training and validation are performed using source image groups, without splitting pixels from the same image into different groups. The three random seeds are used to repeatedly extract supervised pixels from the same pixel type training image pool and train three classifiers, without representing the addition of independent images; each training iteration is still performed using source image groups.

[0041] like Figure 4 As shown, fixed before type fusionB , D , U .by α The coefficient representing the pixel condition type score, minus 1. α The coefficient representing the window condition type score is used to obtain the merged condition type score:

[0042] Two coefficients are non-negative and their sum is 1. In one embodiment... α Take 0.75. Only when... D Internal selection and The larger one is taken as the final defect type; B Maintain background within. U The internal structure remains unknown. The fusion score is not re-compared with the background score in the three categories; therefore, the type branch cannot be expanded, narrowed, or moved. D The arithmetic weighting described above is not a Bayesian inference, and the resulting score is not interpreted as a calibrated probability of actual defect occurrence.

[0043] Step S5 outputs a category mask of the same size as the ROI and a conditional type score map with two channels. The category mask encodes background, wear, stripping, and unknown, respectively. The two channels of the score map correspond to wear and stripping, respectively, and are one-to-one with the pixel row and column coordinates of the ROI. When reading the score map, it is combined with... D To limit its meaning.

[0044] when D If empty, output the results of no detected defects within the processed range. U The status remains unknown. If image reading, model loading, or window construction fails, the generation of the current valid result is aborted and the exception is logged. The failed result is not replaced with a full background mask.

[0045] Figure 3 Given pixels p It belongs to two valid support sets at the same time, therefore it receives two sets of window scores and averages them; pixels q It belongs to only one valid support set, and therefore only one set of window fractions is used; pixels r If a pixel does not belong to any valid support set, its vote count is zero and remains unknown. This example illustrates that voting occurs based on the correspondence between pixels and valid support sets, not by directly filling the entire rectangular window with the window category.

[0046] Candidate Mask C With the final defect set D They can be different: candidate windows can be categorized and judged as background. C Visible locations outside of these areas can also be identified as defects via supplementary windows; the type stage is fixed by the window classification process. DInstead of candidate masks C .

[0047] Figure 6 The plot is drawn using the actual ROI, candidate connected components saved by the program, and supplementary support sets, for illustration. V , E , C and F Spatial correspondence, not E This is interpreted as an unknown location within the tread surface. Each region in the diagram is identified by both a sub-figure number and a set letter; color is used only for visual aids.

[0048] Figure 7 The window record is drawn using the same actual historical ROI. The rectangular range of the candidate window is used to read the image, and the yellow effective support range is the intersection of the window and the bound candidate connected region. The supplementary window is generated in the visible tread outside the candidate mask, and the blue effective support range is the intersection of the window and the visible region F outside the candidate. The vote count map is obtained by pixel-by-pixel summing of the effective support ranges of all windows in both branches. Figure 7 (f) is the background, defect and unknown set formed after the candidate branch is judged by the window, which is used to show the correspondence between the candidate support range and the positioning result; whether the supplementary branch is written into the defect set is determined by the supplementary window classification score and the preset judgment conditions. Figure 7 The black outline in (f) is the manually marked wear boundary, used only for result comparison and not for model inference.

[0049] Figure 8 The left column shows Example 1, and the right column shows Example 2. The binary range of defects formed during the windowing stage is fixed before type fusion. The window condition type score and pixel condition type score are only used for wear and peeling selection within the defect set D. In the figure, the number of pixels differing in the binary range of defects before and after fusion is zero, indicating that type discrimination does not change the defect set D. Example 1 has a black manually annotated wear boundary superimposed for result comparison; this annotation does not participate in model inference. Example 2 does not have a manually annotated ground truth value superimposed. These examples are used to illustrate type fusion and its data relationships within a fixed positioning range, and are not used to compare the positioning accuracy or classification performance of the two examples. Each sub-figure uses a title, figure number, and grayscale difference to identify the category, and color is used to assist in observing whether defects exist at unknown locations.

Claims

1. A method for locating and identifying the type of defects in the tread surface of rail vehicles wheelsets, characterized in that, The image processing device performs the following steps: S1, acquires an image of the region of interest on the wheel tread of the rail vehicle, and determines the set of visible tread pixels in the image. V S2, candidate foreground responses are obtained using the first and second candidate segmentation models respectively. Candidate foreground responses at the same pixel position are fused and then subjected to thresholding and connected component analysis to obtain a candidate mask. C Candidate connected regions; candidate windows are generated based on the candidate connected regions, and a grid window is generated within the region of interest image of the tread surface, which is then compared with the supplementary support set. F A grid window whose intersection pixel count and intersection area ratio meet preset conditions is determined as a supplementary window, and the supplementary support set F The set of visible tread pixels V The middle is not a candidate mask C The set of pixels in the foreground; The effective support range of a candidate window is limited to the intersection of the window and the candidate connected regions it is bound to, and the effective support range of a supplementary window is limited to the window and the supplementary support set. F The intersection of; S3, using a window classifier to obtain the background, wear, and stripping scores of each window, accumulating the scores only for pixels within the effective support range of the corresponding window, and recording the number of votes each pixel receives for the effective window score; in the set of visible tread pixels. V Within this set, pixels with zero votes are defined as the unknown set. U The background set is determined by selecting the pixels with a vote count greater than zero and the largest total background score. B The pixel with the highest number of votes (greater than zero) and the largest total wear or stripping score is identified as the defect set. D ; The region of interest on the tread is defined as the set of pixels that do not belong to the visible tread surface. V The pixels were determined to be invalid regions. E The invalid region E It does not participate in window voting, background judgment, and type discrimination; S4, the defect set D The window condition type score is obtained by dividing the sum of the wear and peeling window scores by the larger of their sum and a preset positive lower limit; the pixel condition type score is obtained based on the image features corresponding to the pixel positions in the region of interest image of the tread surface; a fixed background set is used. B Defect set D and unknown sets U Only in the defect set D The inner fusion window condition type score and pixel condition type score are used to determine wear or peeling based on the fusion result; S5, the defect location, defect type, and two condition type scores corresponding to the image coordinates of the region of interest of the tread surface are output, and the background set is maintained during the type discrimination process. B and unknown sets U The state remains unchanged.

2. The method according to claim 1, characterized in that, The first candidate segmentation model is trained from the basic candidate training set, and the second candidate segmentation model is trained from the second training set containing confirmed difficult background samples. The difficult background samples include at least one high-response region among the reflective region, rust region, or tread edge region that is confirmed as a non-target defect by the labeling rules of the detection task. Both candidate segmentation models merge wear and peeling into candidate foregrounds and take the maximum value of the foreground response at the same pixel position.

3. The method according to claim 1, characterized in that, Using pixels whose candidate foreground response is not lower than the seed threshold as seeds, eight-neighbor connectivity growth is performed on the set of pixels whose candidate foreground response is not lower than the growth threshold, and candidate masks are obtained by area filtering. C The seed threshold is higher than the growth threshold; The candidate window stores the identifier of the candidate connected region it is bound to. Larger candidate connected regions are covered by overlapping windows, while other candidate connected regions are formed into candidate windows using bounding boxes.

4. The method according to claim 1, characterized in that, Within the region of interest on the tread surface, an overlapping mesh window is generated according to a preset step size, preserving and supplementing the support set. F A window whose intersection pixel count and intersection area ratio satisfy the lower limit is used as a supplementary window; supplementary support sets F The decision is made before the window is generated, and the pixel is not reduced due to the already generated supplementary windows, so that the same pixel can receive valid votes from multiple supplementary windows.

5. The method according to claim 1, characterized in that, Each window is cropped to obtain an image patch according to at least two outward scaling scales. The encoder, which is input to the freeze parameters, extracts the window features, and the background, wear, and stripping scores are output by a three-class classifier. The outward scaling range used to read the window context does not expand the effective support range of the window.

6. The method according to claim 1, characterized in that, The image features corresponding to the pixel location include deep features at multiple levels of spatial alignment, color and grayscale features, grayscale mean and standard deviation at at least two neighborhood scales, and directional gradients. The image features are input into at least two pre-trained binary type classifiers, which are trained from wear-marked pixels and stripped-marked pixels. The scores output by the type classifiers at the same location are arithmetically averaged by category to obtain the pixel conditional type score.

7. The method according to claim 1, characterized in that, The window condition type score is obtained by dividing the combined window scores of wear and peeling by the larger of their sum and a preset positive lower limit; the window condition type score and the pixel condition type score are arithmetically weighted using non-negative coefficients that sum to 1, and the weighted scores are used only for the defect set. D The type selection within the score does not re-compare the maximum value with the background score.

8. A system for locating and identifying the type of defects in the tread surface of rail vehicles, characterized in that, include: The tread image acquisition module is used to acquire the tread region of interest image, the set of visible tread pixels, and the invalid region; The candidate and supplementary coverage module is used to form candidate masks, candidate windows, supplementary windows, and the effective support range of each window; the window positioning module is used to summarize background, wear and stripping scores and vote counts only within the effective support range, and form background set, defect set, and unknown set. The pixel type module is used to output the wear and peeling condition type scores corresponding to the pixel position; the positioning constraint fusion module is used to fix the three sets and fuse the window condition type scores and pixel condition type scores only within the defect set; the result output module is used to output the corresponding defect positions, defect types and two types of condition type scores.

9. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes a computer program capable of locating and identifying defects in the tread surface of a rail vehicle wheel set, it implements the method of any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program for locating and identifying defects in the wheel tread of a rail vehicle is executed by the processor, it implements the method of any one of claims 1 to 7.

Citation Information

Patent Citations

  • A Highly Reflective Surface Defect Detection Method Based on Image Processing and Neural Network Classification

    CN108520274B

  • Train wheel set tread image positioning and extracting method

    CN116385545A