Method for on-line detection and classification of valve sealing surface defects based on machine vision

Abstract

This invention belongs to the field of image processing technology, specifically relating to an online detection and classification method for valve sealing surface defects based on machine vision. The method includes: acquiring the boundary of the sealing contact zone, phase fringe reflection images, and multi-incident angle laser speckle images to construct functional coordinates of the sealing contact zone; generating a reflection phase residual map and phase curvature intensity based on the phase fringe reflection images, and constraining speckle correlation processing to obtain a speckle angle decorrelation map; generating a phase speckle cross-verification risk tensor based on the reflection phase residual map and speckle angle decorrelation map, determining candidate defect regions, and constructing a directed leakage risk map; performing multi-threshold through-scanning on the directed leakage risk map to obtain through-path information, and outputting the defect classification result and leakage risk level. This invention enables the comprehensive utilization of phase-side geometric perturbations and speckle-side microscopic scattering anomalies in a unified coordinate system, enhancing the consistency between defect classification and leakage risk assessment.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and more specifically to a method for online detection and classification of defects on valve sealing surfaces based on machine vision. Background Technology

[0002] The valve sealing surface is a critical surface for achieving a proper sealing fit. Scratches, pitting, porosity, inclusions, roughness, or residual contaminants on its surface can affect the continuity of the sealing contact zone. Relevant detection techniques typically use industrial cameras to capture images of the valve sealing surface, and then identify suspected defect areas through grayscale thresholding, edge detection, texture feature extraction, or defect classification models. While these methods can detect some surface anomalies with significant grayscale differences, they primarily rely on single reflection intensity, single texture response, or single classification results for judgment.

[0003] In practical testing, valve sealing surfaces typically exhibit characteristics such as high metallic reflectivity, narrow annular contact bands, continuous machining textures, and significant differences in the scale of local defects. Detection methods based solely on grayscale or texture are easily affected by illumination angle, surface reflectivity, clamping posture, and machining texture direction, making it difficult to reliably distinguish between genuine defects and normal machining textures. While some techniques can identify defect areas, these areas usually only indicate the location of local anomalies and do not further determine whether the anomaly forms a continuous leakage path from the high-pressure side to the low-pressure side along the width of the sealing contact band. Other techniques process morphological and scattering information separately, lacking a mechanism for correlation at the same sealing contact band location, making it difficult to cross-verify geometric disturbances and microscopic scattering anomalies. Consequently, the detection results may remain at the level of local defect identification, failing to reliably output classification results and risk levels directly related to the risk of sealing leakage. Summary of the Invention

[0004] This invention provides a machine vision-based online detection and classification method for valve sealing surface defects, which at least solves the problem of how to perform phase speckle cross-verification identification of sealing contact zone defects and output leakage risk level under conditions of high reflectivity and processing texture interference on valve sealing surfaces.

[0005] This invention provides a machine vision-based online detection and classification method for valve sealing surface defects, the method comprising: Acquire images of the sealing contact strip boundary, phase fringe reflection, and multi-incident angle laser speckle images, and construct functional coordinates of the sealing contact strip based on the sealing contact strip boundary; The reflection phase residual map and phase curvature intensity are generated from the phase fringe reflection image. Based on the phase curvature intensity, speckle correlation processing is constrained to obtain the speckle angle decorrelation map. A phase speckle cross-verification risk tensor is generated based on the reflection phase residual map and speckle angle decorrelation map. Candidate defect regions are determined based on the phase speckle cross-verification risk tensor, and a directed leakage risk map is constructed based on the candidate defect regions. Multi-threshold through-scanning is performed on the directed leakage risk map to obtain through-path information. Based on the phase speckle cross-validation risk tensor and through-path information, the defect classification results and leakage risk level are output.

[0006] In one possible implementation, the sealing contact zone functional coordinates are constructed based on the sealing contact zone boundary, including: determining the contact zone width direction, the contact zone width direction range, and the circumferential direction reference according to the sealing contact zone boundary; mapping the image position of the valve sealing surface to the coordinate region defined by the contact zone width direction and the circumferential direction reference to obtain the sealing contact zone functional coordinates.

[0007] In one possible implementation, generating a reflection phase residual map and phase curvature intensity based on a phase stripe reflection image includes: generating a reflection phase map based on phase stripe reflection images under multiple phase states; performing circumferential reference residual calculation on the reflection phase map based on the functional coordinates of the sealing contact strip to obtain the reflection phase residual map; and determining the phase curvature intensity based on local changes in the reflection phase residual map.

[0008] In one possible implementation, speckle correlation processing based on phase curvature intensity constraints includes: determining the window size, window orientation, and window shape of the speckle correlation window according to the phase curvature intensity; calculating the local correlation between multi-incident angle laser speckle images under different incident angles within the speckle correlation window; and generating a speckle angle decorrelation map based on the local correlation.

[0009] In one possible implementation, a phase speckle cross-verification risk tensor is generated based on the reflection phase residual map and the speckle angle decorrelation map, including: establishing a circumferential self-reference baseline for the same valve sealing surface in the functional coordinates of the sealing contact zone; and generating the phase speckle cross-verification risk tensor based on the phase deviation of the reflection phase residual map relative to the circumferential self-reference baseline and the speckle angle decorrelation map relative to the circumferential self-reference baseline.

[0010] In one possible implementation, generating the phase speckle cross-verification risk tensor includes: determining the geometric perturbation response based on the phase deviation; determining the microscopic scattering response based on the speckle deviation; and associating the geometric perturbation response, the microscopic scattering response, and the cross-verification term determined based on the geometric perturbation response and the microscopic scattering response as tensor elements at the same location to generate the phase speckle cross-verification risk tensor.

[0011] In one possible implementation, candidate defect regions are determined based on the phase speckle cross-verification risk tensor, including: determining local cross-verification risk in the functional coordinates of the sealing contact zone according to the phase speckle cross-verification risk tensor; merging adjacent regions with local cross-verification risk not less than the candidate defect threshold to obtain candidate defect regions, and using the candidate defect regions as input for constructing a directed leakage risk map.

[0012] In one possible implementation, a directed leakage risk map is constructed, including: discretizing candidate defect regions into risk elements; determining risk nodes based on the positions of the risk elements in the functional coordinates of the sealing contact zone; and determining the high-pressure side boundary zone and the low-pressure side boundary zone in the functional coordinates of the sealing contact zone; determining directed edges and their connection strengths based on the distance between adjacent risk nodes, the local mutual verification risks associated with adjacent risk nodes, and the consistency between the connection direction of the risk nodes and the leakage direction from the high-pressure side boundary zone to the low-pressure side boundary zone; and generating a directed leakage risk map based on the risk nodes, directed edges, and their connection strengths.

[0013] In one possible implementation, risk nodes with local mutual verification risk not less than the corresponding risk threshold are retained according to multiple risk thresholds, and directed edges with connection strength not less than the corresponding risk threshold are retained, resulting in multiple risk subgraphs; in each risk subgraph, directed connected paths from the high-voltage side boundary area to the low-voltage side boundary area are searched; the searched directed connected paths and the threshold ranges in which the directed connected paths are located are determined as the through path information.

[0014] In one possible implementation, the defect classification result and leakage risk level are output based on the phase speckle cross-verification risk tensor and the penetration path information, including: extracting the defect response fingerprint of the candidate defect region based on the phase speckle cross-verification risk tensor, wherein the defect response fingerprint includes phase perturbation features and speckle angle decorrelation features; determining the defect classification result based on the defect response fingerprint; determining the leakage risk level based on the defect classification result and the penetration path information, and outputting the defect classification result and leakage risk level.

[0015] Compared with the prior art, the advantages and beneficial effects of the present invention are as follows: By constructing functional coordinates of the sealing contact zone based on the boundary of the sealing contact zone, the registration of different imaging data of the valve sealing surface at the same contact zone position is realized, enabling phase response, speckle response and candidate defect area to be uniformly processed according to the contact zone width direction and circumferential direction.

[0016] By generating a reflection phase residual map and phase curvature intensity based on the phase fringe reflection image, local geometric perturbations can be extracted from the reflection phase change of the metal sealing surface, and the phase curvature intensity can constrain speckle correlation processing, reducing the averaging effect of the fixed window on slender defects, narrow depressions or directional defects.

[0017] By generating a phase speckle cross-validation risk tensor based on the reflection phase residual map and speckle angle decorrelation map, the geometric perturbation response and the microscopic scattering response form a cross-validation relationship at the same coordinate position, which can reduce the judgment bias caused by a single gray level, a single texture or a single speckle response.

[0018] By constructing a directed leakage risk map based on candidate defect regions and performing multi-threshold through-scanning on the directed leakage risk map, local defect identification can be transformed into a directed connectivity judgment from the high-voltage side to the low-voltage side, so that the defect classification results and leakage risk level have a clear path basis. Attached Figure Description

[0019] Figure 1 This is a schematic flowchart of the method of the present invention; Figure 2 This is a coordinate diagram of the sealing contact strip function in an embodiment of the present invention; Figure 3 This is a phase speckle cross-verification risk tensor diagram in an embodiment of the present invention; Figure 4 This is a directed leakage risk map and threshold scan result map in an embodiment of the present invention; Figure 5 This is a defect response fingerprint and result write-back diagram in an embodiment of the present invention. Detailed Implementation

[0020] To enable those skilled in the art to better understand the technical solutions of this invention, the technical solutions of this invention will be clearly and completely described below with reference to the accompanying drawings of one or more embodiments. Obviously, the described embodiments are only a part of the embodiments of this invention, and not all of them. Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of this invention.

[0021] In this embodiment, machine vision is used to convert the spatial imaging response of the valve sealing surface into computable detection data. The detection station acquires positioning images, phase fringe reflection images, and multi-incident-angle laser speckle images for the same valve sealing surface. The positioning image is used to determine the boundary of the sealing contact zone, the phase fringe reflection image is used to characterize the local reflection phase changes of the sealing surface, and the multi-incident-angle laser speckle image is used to characterize the microscopic scattering changes under different illumination angles. After the above image data is registered with the functional coordinates of the same sealing contact zone, it forms the data basis for subsequent reflection phase residual calculation, speckle angle decorrelation analysis, phase speckle cross-verification risk expression, leakage path continuity judgment, and defect classification output.

[0022] like Figure 1As shown, a machine vision-based online detection and classification method for valve sealing surface defects includes the following processing steps.

[0023] In this embodiment, after the valve sealing surface enters the testing station, the testing equipment acquires images of the sealing contact zone boundary, phase stripe reflection, and multi-incident angle laser speckle.

[0024] The sealing contact zone boundary is used to define the annular detection area within the valve sealing surface that actually participates in the sealing fit. Phase fringe reflection images are used to subsequently obtain changes in the surface reflection phase, and multi-incident angle laser speckle images are used to subsequently obtain microscopic scattering changes.

[0025] The image processing unit uses the boundary of the sealing contact zone as a geometric constraint to transform the image position of the valve sealing surface into a unified functional coordinate system of the sealing contact zone. This functional coordinate system serves as the basis for subsequent positional registration of phase information, speckle information, and candidate defect regions, enabling data obtained from different imaging methods to be compared and transferred at the same detection location.

[0026] In one embodiment, after receiving a positioning image of the valve sealing surface, the image processing unit identifies the boundary of the sealing contact zone in the positioning image. The positioning image can be acquired under coaxial illumination or low-angle supplementary lighting conditions. The acquisition method only needs to be able to clearly distinguish the sealing contact zone from the non-sealing area, and is not required to be limited to a certain fixed light source arrangement.

[0027] The boundary of the sealing contact zone can be determined by the changes in edge grayscale, the continuity of the annular edge, and the consistency of boundary fitting. The image processing unit sets multiple boundary search positions along the circumference of the valve sealing surface, extracting inner and outer edge points at each position. The inner and outer edge points are then fitted with a circular or near-circular curve to form the inner and outer boundaries of the sealing contact zone. For discrete edge points caused by localized reflections, oil stains, burr shadows, or fixture obstructions, the image processing unit can remove them based on the deviation between the edge point and the fitted boundary. The deviation judgment threshold can be determined by the boundary fitting deviation distribution of qualified samples, used to constrain whether abnormal edge points participate in boundary fitting.

[0028] After obtaining the inner and outer boundaries, the image processing unit determines the contact strip width direction in the direction from the inner boundary to the outer boundary. The range of the contact strip width direction is defined by the effective detection distance between the inner and outer boundaries, and is used to represent the relative distribution of image positions across the width of the sealing contact strip. This range is not a typical image cropping range, but rather the basis for subsequent judgments on whether defects are close to the high-pressure side, the low-pressure side, and whether they might cross the sealing contact strip.

[0029] The circumferential reference is used to determine the angular starting point for the circumferential position of the valve sealing surface. The circumferential reference can be determined by the positioning marks of the workpiece fixture, the zero position recorded at the rotary station, the boundary fitting center of the valve sealing surface, and the fixed reference direction. If the positioning marks are clear, the image processing unit uses the direction of the positioning marks as the circumferential reference; if the positioning marks are obscured, the image processing unit can use the zero position recorded at the rotary station as the circumferential reference and correct the angular starting point using the boundary fitting center. Once the circumferential reference is determined, the circumferential position of the same valve sealing surface can remain consistent in different acquired images.

[0030] During coordinate transformation, the image processing unit maps the image position of the valve sealing surface to a coordinate region defined by the width direction of the contact strip and the circumferential direction reference. Within this coordinate region, one direction represents the width position of the detection position between the inner and outer boundaries, and the other direction represents the circumferential position of the detection position relative to the circumferential direction reference. After mapping, each valid detection position in the positioning image forms a position index in the functional coordinates of the sealing contact strip.

[0031] Before being processed, both the phase fringe reflection image and the multi-incident angle laser speckle image are resampled or registered using the same location index. For locations falling outside the sealing contact zone boundary, with insufficient boundary fitting confidence, or missing resampling data, the image processing unit marks these locations as invalid detection locations. Invalid detection locations are not included in subsequent reflection phase residual map generation, speckle angle decorrelation map generation, or candidate defect region determination.

[0032] Boundary fit confidence can be determined based on edge continuity length, boundary fit deviation, and boundary smoothness at adjacent circumferential positions. Edge continuity length is used to determine whether the boundary points are sufficient to support a stable fit; boundary fit deviation is used to determine whether local edge points deviate from the normal boundary; and boundary smoothness at adjacent circumferential positions is used to determine whether there are abrupt changes in the boundary. These determinations collectively ensure that the functional coordinates of the sealing contact zone only cover the actual and effective sealing detection area.

[0033] After completing the functional coordinates of the sealing contact strip, subsequent stages can read the reflection response corresponding to the phase fringe reflection image and the speckle response corresponding to the multi-incident angle laser speckle image at the same coordinate position. Thus, the phase-side data and speckle-side data are spatially comparable and transferred to the reflection phase residual calculation and speckle correlation calculation stages.

[0034] In this embodiment, the image processing unit receives a phase fringe reflection image and a multi-incident angle laser speckle image under the same sealing contact zone functional coordinates.

[0035] Phase fringe reflection images are formed by the reflection of fringe patterns in different phase states from the valve sealing surface. The image processing unit performs phase calculations on multiple phase fringe reflection images at the same coordinate position to form a reflection phase map. After generating the reflection phase map, the image processing unit establishes a circumferential reference along the circumferential direction and calculates the residual between the reflection phase map and the circumferential reference to form a reflection phase residual map.

[0036] The reflection phase residual map is used to characterize local deviations in the continuity of reflection on the sealing contact strip surface. The image processing unit further determines the phase curvature intensity based on the local changes in the reflection phase residual map and uses the phase curvature intensity to constrain the speckle correlation window. Local correlation calculations are performed on the multi-incident-angle laser speckle images within the speckle correlation window to generate a speckle angle decorrelation map, which is then passed to the subsequent phase speckle cross-verification calculation stage.

[0037] In one embodiment, the phase stripe reflection image can be obtained by sequentially displaying sinusoidal stripe patterns of multiple phase states by a stripe display unit, and by acquiring the image reflected from the valve sealing surface by a camera. The image processing unit registers the positions of the images under different phase states in the functional coordinates of the sealing contact zone, so that the grayscale value at the same coordinate position comes from the same actual detection position on the valve sealing surface.

[0038] When using four phase states, the reflection phase diagram can be determined by the following formula: Φ(a,b)=atan2(I_270 (a,b)-I_90 (a,b),I_0 (a,b)-I_180 (a,b)) Where Φ(a,b) is the reflection phase value at position (a,b) in the functional coordinate system of the sealing contact strip, a is the position in the width direction of the contact strip, b is the position in the circumferential direction, I_0 (a,b), I_90 (a,b), I_180 (a,b), and I_270 (a,b) are the image grayscale values ​​at the same position under the four phase states, respectively, and atan2 is the arctangent operation used to determine the phase quadrant. This calculation relationship is used to convert multiple stripe reflection images into a reflection phase map, and the output result is used to calculate the circumferential reference residual.

[0039] After the reflection phase map is formed, the image processing unit statistically analyzes the phase values ​​at positions along the same contact strip width direction along the circumferential direction to obtain the circumferential reference. The circumferential reference can be determined using the circumferential median, the circumferential mean after removing outliers, or the circumferential reference value obtained from calibration of qualified samples. The principle for setting the circumferential reference is to maintain the normal reflection trend at the same width position while reducing the impact of local scratches, pores, inclusions, and residual contaminants on the reference.

[0040] The image processing unit calculates the difference between the reflection phase map and the circumferential reference to obtain the reflection phase residual map. The values ​​in the reflection phase residual map reflect the deviation of the local position relative to the circumferential reference at the same width position. For slow changes caused by circumferential tilt, installation angle deviation, or camera viewing angle, the circumferential reference residual calculation can reduce its interference with the judgment of local defects.

[0041] The phase curvature intensity is determined by local variations in the reflection phase residual map. The image processing unit sets up a local neighborhood around each valid detection location and compares the residual variations in the contact strip width and circumferential directions, respectively. If continuous phase changes, abrupt phase changes, or directional ridge-valley structures appear within the local neighborhood, the image processing unit identifies the corresponding location as having a higher phase curvature intensity. The phase curvature intensity is used to control the subsequent speckle correlation window and is not directly output as the final defect classification result.

[0042] The speckle correlation window is generated by phase curvature intensity constraints. In regions with low phase curvature intensity, the window size can be larger to reduce the impact of speckle noise on correlation calculations. In regions with high phase curvature intensity, the window size can be reduced, the window orientation can be set along the extension direction of local phase changes, and the window shape can be elongated or elliptical to reduce the averaging of fine scratches or narrow pits by the window. The window size, orientation, and shape can be calibrated by the phase curvature distribution of qualified and defective samples, or jointly set by the resolution of the inspection station, the width of the sealing contact zone, and the minimum detectable defect size.

[0043] After the multi-incident-angle laser speckle image is entered into the speckle correlation window, the image processing unit calculates the correlation between local speckle patterns at different incident angles. For the speckle correlation window at position (a, b), the local correlation can be determined according to the following formula: R(a,b)=(∑_(q∈Ω(a,b))▒(S_1 (q)-S ‾_1 )(S_2 (q)-S ‾_2 ) ) / (√(∑_(q∈Ω(a,b))▒(S_1 (q)-S ‾_1 )^2 ) √(∑_(q∈Ω(a,b))▒(S_2 (q)-S ‾_2 )^2 )+ε) Where R(a,b) represents the local correlation at position (a,b), Ω(a,b) is the speckle correlation window obtained by phase curvature intensity constraint, q is the pixel position within the speckle correlation window, S_1(q) is the speckle gray value at pixel position q under the first incident angle, S_2(q) is the speckle gray value at pixel position q under the second incident angle, S‾_1 is the average gray value within the speckle correlation window under the first incident angle, S‾_2 is the average gray value within the speckle correlation window under the second incident angle, and ε is a stability constant to prevent the denominator from being zero. This calculation relationship is used to obtain the degree of local correlation of speckle patterns under different incident angles.

[0044] The image processing unit generates a speckle angle decorrelation map based on local correlation. Lower local correlation indicates a more significant change in the speckle pattern at that location as the incident angle changes, resulting in a stronger speckle angle decorrelation response at that location. If three or more incident angles are used, the image processing unit can calculate the local correlation between the reference incident angle and other incident angles separately, and generate the speckle angle decorrelation map by taking the minimum local correlation, average local correlation, or weighted local correlation. The weights can be determined based on the magnitude of the incident angle change, the stability of the light source, and the effective response distribution in the calibration sample, used to limit the influence of low-quality incident angle images on the final speckle angle decorrelation map.

[0045] During speckle correlation processing, the image processing unit marks windows with overexposure saturation, low speckle gray levels, or those falling into invalid detection positions. These marked windows do not participate in the effective response calculation of the speckle angle decorrelation map and are treated as invalid positions in the subsequent phase speckle cross-verification calculation stage. After the speckle angle decorrelation map is generated, the reflection phase residual map, phase curvature intensity, and speckle angle decorrelation map are all retained in the sealing contact zone functional coordinates. In subsequent stages, the phase-side response and speckle-side response are read from the same coordinate positions to generate the phase speckle cross-verification risk tensor.

[0046] In this embodiment, the image processing unit receives the reflection phase residual map and speckle angle decorrelation map formed in the previous stage, and performs position alignment in the same sealing contact zone functional coordinate system.

[0047] The reflection phase residual diagram reflects the deviation of the local geometric reflection continuity of the valve sealing surface, while the speckle angle decorrelation diagram reflects the microscopic scattering changes at the same location under different incident angles.

[0048] The image processing unit establishes a circumferential self-reference baseline of the same valve sealing surface in the functional coordinates of the sealing contact zone, calculates the phase deviation and speckle deviation respectively, and forms a phase-speckle mutual verification risk tensor from the two.

[0049] The phase speckle cross-verification risk tensor is used to record the geometric perturbation response, microscopic scattering response, and the cross-verification term between them at the same detection location. The image processing unit determines the local cross-verification risk based on this tensor, merges adjacent regions that meet the candidate defect threshold into candidate defect regions, and converts these candidate defect regions into risk nodes and directed edges, generating a directed leakage risk map. This directed leakage risk map is then passed to the subsequent multi-threshold through-scan stage.

[0050] In one embodiment, the image processing unit establishes circumferential self-reference baselines for both phase-side and speckle-side data in the functional coordinates of the sealing contact zone. The circumferential self-reference baseline represents a normal response reference formed circumferentially at the same contact zone width location on the same valve sealing surface.

[0051] When establishing the circumferential self-reference baseline, the image processing unit uses the position along the width direction of the contact strip as an index to read the residual values ​​in the reflection phase residual map along the circumferential direction, and removes local points that significantly deviate from the circumferential distribution. The removal criteria can be determined by the circumferential residual distribution in qualified samples, the median deviation of residual values ​​in the same workpiece, and boundary validity markings. After removing outliers, the image processing unit calculates the circumferential reference value at each width position to form the circumferential self-reference baseline on the phase side.

[0052] The circumferential self-reference baseline on the speckle side is established using the same coordinate index. The image processing unit reads the response value from the speckle angle decorrelation map along the circumferential direction, removes exposure saturation, insufficient speckle grayscale, invalid detection positions, and isolated local anomalies, and then calculates the speckle reference value at each width position. In this way, both the phase side and the speckle side use the circumferential stable response of the same valve sealing surface itself as a reference, avoiding the direct impact of material differences, overall reflectivity differences, and clamping angle differences between different workpieces on the judgment of local defects.

[0053] The phase deviation and speckle deviation can be determined by the following formula: D_p (a,b)=|P_r (a,b)-B_p (a)| D_s (a,b)=|S_d (a,b)-B_s (a)| Wherein, D_p(a,b) is the phase deviation at position (a,b), D_s(a,b) is the speckle deviation at position (a,b), a is the position in the contact band width direction, b is the position in the circumferential direction, P_r(a,b) is the residual value of the reflection phase residual map at position (a,b), S_d(a,b) is the response value of the speckle angle decorrelation map at position (a,b), B_p(a) is the phase-side circumferential self-reference baseline at position a in the contact band width direction, and B_s(a) is the speckle-side circumferential self-reference baseline at position a in the contact band width direction. The phase deviation is included in the geometric perturbation response calculation, and the speckle deviation is included in the microscopic scattering response calculation.

[0054] In one embodiment, the image processing unit determines the geometric perturbation response based on the phase deviation. The geometric perturbation response can be obtained by normalizing the phase deviation relative to the calibration upper limit of the phase-side circumferential self-reference baseline, and corrected by incorporating the continuity of ridges, valleys, or abrupt boundaries within the local neighborhood, ensuring the geometric perturbation response value falls within the range of 0 to 1. The geometric perturbation response represents the local undulations, breaks, depressions, or scratches on the valve sealing surface at the reflection phase level. A larger phase deviation, especially when it presents continuous ridges, valleys, or abrupt boundaries within the neighborhood, results in a higher geometric perturbation response value. When the phase deviation is large but only occurs near isolated invalid locations, the image processing unit can reduce the geometric perturbation response at that location to avoid misjudgments caused by boundary mismatches or strong local reflections.

[0055] The image processing unit determines the microscopic scattering response based on the speckle deviation. The microscopic scattering response can be obtained by normalizing the speckle deviation or speckle angle decorrelation response relative to the upper limit of the speckle side calibration, and then corrected by combining exposure saturation markers, under-grayscale markers, and incident angle image quality markers to ensure the microscopic scattering response value falls within the range of 0 to 1. The microscopic scattering response is used to represent the change in speckle stability of the valve sealing surface under multi-incident angle illumination. If the speckle pattern rapidly loses correlation at a certain location when the incident angle changes, and this location is not marked as exposure saturation or under-grayscale, the image processing unit determines that location to have a high microscopic scattering response. When the microscopic scattering response is high but the phase deviation is low, the image processing unit still retains the speckle side record at that location for identifying inclusions, residual contamination, or minor roughness defects.

[0056] The geometric perturbation response and the microscopic scattering response are correlated at the same functional coordinate location of the sealed contact zone, and the image processing unit further determines the mutual verification term between them. The mutual verification term is used to represent the joint risk when both phase-side anomalies and speckle-side anomalies occur simultaneously at the same location. The mutual verification term does not replace the geometric perturbation response and the microscopic scattering response, but is written as a third type of record at the same location into the phase speckle mutual verification risk tensor.

[0057] The risk of partial mutual verification can be determined using the following formula: Q(a,b)=αG(a,b)+βM(a,b)+γG(a,b)M(a,b) Where Q(a,b) represents the local cross-verification risk at location (a,b), G(a,b) represents the geometric perturbation response at location (a,b), M(a,b) represents the microscopic scattering response at location (a,b), G(a,b)M(a,b) represents the cross-verification term between the geometric perturbation response and the microscopic scattering response, α represents the weight of the geometric perturbation response, β represents the weight of the microscopic scattering response, and γ represents the weight of the cross-verification term. These weights can be calibrated based on the response distribution in qualified and defective samples, or they can be set based on the width of the sealing contact strip, the reflectivity of the material, and the stability of the testing station. The local cross-verification risk is used for screening candidate defect areas.

[0058] In one embodiment, the image processing unit traverses the local cross-verification risk within the functional coordinates of the sealing contact zone and marks locations where the local cross-verification risk is not less than a candidate defect threshold as candidate locations. The candidate defect threshold can be set based on the local cross-verification risk distribution of qualified samples, ensuring that normal processing textures and random speckle fluctuations do not enter the candidate defect area. For different valve models, the candidate defect threshold can also be adjusted in conjunction with the sealing contact zone width and the minimum detectable defect size.

[0059] After candidate locations are formed, the image processing unit merges regions according to spatial adjacency relationships. These relationships can be determined using four-neighborhood, eight-neighborhood, or by setting adjacency distances along both the contact zone width and circumferential directions. For adjacent candidate locations, if the local mutual verification risk is continuously distributed and there is no obvious break in the phase-side or speckle-side response direction, the image processing unit merges these locations into the same candidate defect region. For regions containing only a few isolated candidate locations and lacking continuous phase-side or speckle-side responses, the image processing unit can either retain them as low-confidence candidate regions or reduce their participation weight when subsequently constructing the directed leakage risk map.

[0060] After candidate defect regions are generated, the image processing unit records the width range, circumferential range, local cross-validation risk statistics, and main response type of each candidate defect region in the functional coordinates of the sealing contact zone. The main response type can be determined by the relative magnitudes of geometric perturbation response, microscopic scattering response, and cross-validation terms, and is used for subsequent defect classification result generation. Candidate defect regions are not directly used as the final defect output, but rather as input for constructing a directed leakage risk map, enabling subsequent processing to determine whether local defects have a leakage path trend extending from the high-pressure side boundary region to the low-pressure side boundary region.

[0061] In one embodiment, the image processing unit discretizes the candidate defect region into risk units. A risk unit can consist of several adjacent coordinate positions within the candidate defect region, or it can consist of local segments after the candidate defect region has been skeletonized or meshed. The scale of the risk unit is determined based on the width of the sealing contact strip, the image resolution, and the minimum detectable defect size, to ensure that fine scratches, pitting accumulation areas, and inclusion pits can all be converted into graph structural elements that can participate in path search.

[0062] The position of each risk element in the functional coordinate system of the sealing contact zone is determined as a risk node. Risk nodes carry information such as location, local mutual verification risk, geometric disturbance response, microscopic scattering response, and candidate defect region attribution. The image processing unit determines the high-pressure side boundary region and the low-pressure side boundary region within the same functional coordinate system of the sealing contact zone. The high-pressure side boundary region can be determined based on the valve sealing mating direction, the installation direction of the inspection fixture, or a pre-input sealing pressure difference direction. The low-pressure side boundary region is located on the opposite side of the sealing contact zone from the high-pressure side boundary region. If different valve models have different assembly directions, the image processing unit can determine the positions of the high-pressure side boundary region and the low-pressure side boundary region using workpiece model information or inspection station configuration parameters.

[0063] Whether a directed edge is established between adjacent risk nodes is determined by distance, local mutual verification risk, and leakage direction consistency. Distance is used to limit the connection range between discontinuous defect segments; the distance threshold can be set based on the minimum detectable defect length, the width of the sealing contact zone, and the image resolution. Local mutual verification risk is used to indicate the strength of the defect response of adjacent risk nodes. Leakage direction consistency is used to determine whether the connection direction of risk nodes is close to the direction from the high-pressure side boundary area to the low-pressure side boundary area.

[0064] The strength of a directed edge connection can be determined by the following formula: L_ij=(Q(j) C_ij) / (1+d_ij⁄d_0 ) Where L_ij is the directed edge connection strength from risk node i to risk node j, Q(j) is the local mutual verification risk associated with risk node j, d_ij is the distance between risk node i and risk node j, d_0 is the reference distance calibrated by the minimum detectable defect length, image resolution, or sealing contact strip width, and C_ij is the consistency between the risk node connection direction and the leakage direction. C_ij can be determined based on the angle between the two directions; the smaller the angle, the higher the consistency. This connection strength is used to determine whether the directed edge should be retained and the path search priority in subsequent multi-threshold through-scanning.

[0065] The image processing unit generates a directed leakage risk map based on risk nodes, directed edges, and the connection strength of directed edges. Nodes in the directed leakage risk map represent local risk units within a candidate defect region that may participate in a leakage path, and directed edges represent the possible relationships between risk units that form a continuous defect chain along the leakage direction. For risk node combinations with excessively large distances, connection directions significantly deviating from the leakage direction, or local mutual verification risks below the graph structure preservation condition, the image processing unit does not establish directed edges. After construction, the directed leakage risk map is passed to the multi-threshold through-scan stage to determine whether a candidate defect region forms a risk path from the high-pressure side boundary region to the low-pressure side boundary region.

[0066] In this embodiment, the image processing unit receives the directed leakage risk map and the phase speckle cross-verification risk tensor formed in the previous stage.

[0067] The directed leakage risk graph records the risk nodes and directed edges obtained from the transformation of candidate defect regions. Risk nodes represent local risk units that may participate in the leakage path, and directed edges represent the possible relationship of adjacent risk nodes forming a continuous connection along the high-pressure side boundary region to the low-pressure side boundary region.

[0068] The image processing unit performs a step-by-step filtering of the directed leakage risk map according to multiple risk thresholds, and searches for directed connected paths in the filtering results at each level. The search results form connected path information. The image processing unit then extracts the defect response fingerprints of candidate defect regions from the phase speckle cross-verification risk tensor, and outputs the defect classification results and leakage risk level by combining them with the connected path information.

[0069] In one embodiment, the image processing unit generates a risk threshold sequence before performing a multi-threshold through scan. The risk threshold sequence can be determined based on the local cross-verification risk distribution in qualified and defective samples, or it can be set based on the defect recall range allowed by the inspection station, the width of the sealing contact strip, and the minimum detectable defect size. The risk thresholds are used to control which risk nodes and directed edges can enter the current scan level.

[0070] The image processing unit processes the directed leakage risk graph sequentially according to the risk threshold sequence. At any given risk threshold, risk nodes with a local mutual verification risk not less than that threshold are retained; directed edges connected to retained risk nodes with a connection strength not less than that risk threshold are also retained. The retained risk nodes and directed edges together form a risk subgraph. The risk subgraph represents a local graph structure that still has the potential for a continuous defect chain at the current risk threshold.

[0071] After forming the risk subgraph, the image processing unit starts searching for directed connected paths from the risk nodes corresponding to the high-voltage side boundary area. The search direction is limited to propagation along directed edges and cannot cross directed edges in reverse. If a path can reach the low-voltage side boundary area from the high-voltage side boundary area, the image processing unit records the path as a directed connected path. If there are multiple directed connected paths in the same risk subgraph, the image processing unit can record the path with fewer nodes, higher connection strength, or higher cumulative local mutual verification risk value as the primary connected path, while retaining other paths as alternative connected paths.

[0072] The path information records at least the risk nodes traversed by the directed connected path, the candidate defect regions where the risk nodes are located, the path's entry point into the high-pressure side boundary zone, the path's arrival point in the low-pressure side boundary zone, and the threshold range within which the directed connected path exists. The threshold range represents the extent to which the same or similar directed connected path persists under different risk thresholds. If a directed connected path has already appeared at a higher risk threshold and persists under multiple adjacent risk thresholds, the image processing unit records this path as a stable connected path. If a directed connected path only appears at a lower risk threshold, the image processing unit records this path as a weakly connected path.

[0073] If no directed connected path is found from the high-pressure side boundary to the low-pressure side boundary at any given risk threshold, the image processing unit continues processing the next risk threshold. If no directed connected path is found at any of the risk thresholds, the path information is recorded as "not connected," and the local risk record of the candidate defect area is retained. The "not connected" state does not delete the candidate defect area; instead, it is used to distinguish between local defects and through-type defects in subsequent leakage risk level determination.

[0074] In one embodiment, the image processing unit extracts a defect response fingerprint of the candidate defect region based on the phase speckle cross-verification risk tensor. The defect response fingerprint describes the joint performance of the candidate defect region on the phase side and the speckle side. The phase perturbation features can be determined by the geometric perturbation response, phase deviation distribution, phase abrupt change direction, and region extension morphology within the candidate defect region. The speckle angle decorrelation features can be determined by the microscopic scattering response, speckle deviation distribution, and local correlation changes at different incident angles within the candidate defect region.

[0075] When extracting defect response fingerprints, the image processing unit not only statistically analyzes the response strength within candidate defect regions but also records the spatial morphology of the response in the functional coordinates of the sealing contact zone. For candidate defect regions extending along the width of the contact zone, with continuous phase perturbation and synchronously enhanced speckle angle decorrelation response, the defect response fingerprint exhibits linear phase perturbation and continuous speckle instability. This type of defect response fingerprint can be used to identify scratch-type defects. For candidate defect regions with strong speckle angle decorrelation response in localized areas and phase perturbation exhibiting point-like or concave variations, the defect response fingerprint can be used to identify porosity-type or inclusion-type defects. For candidate defect regions with weak phase perturbation but a patchy speckle angle decorrelation response, the image processing unit can classify them into the category of surface roughness anomalies or residual contamination awaiting verification.

[0076] Defect classification rules can be determined jointly by sample calibration data and manual verification results. Sample calibration data provides typical distributions of different defect types in phase perturbation features and speckle angle decorrelation features, while manual verification results correct the boundaries between easily confused categories. Defect classification rules do not need to be limited to a single fixed classification model; they can be implemented using rule tables, classifiers, or a combination of both. Regardless of the method used, the input is the defect response fingerprint, and the output is the defect classification result.

[0077] The leakage risk level is determined jointly by the defect classification results and the penetration path information. When determining the leakage risk level, the image processing unit considers whether the candidate defect region participates in a stable penetration path, whether it crosses the high-pressure side boundary zone to the low-pressure side boundary zone, and the threshold range within which the penetration path lies. For scratch-type defects, porosity-type defects, or inclusion-type defects located on a stable penetration path, a higher leakage risk level is assigned. If the candidate defect region does not form a penetration path but has a high local mutual verification risk, an intermediate leakage risk level can be assigned. If the candidate defect region only exhibits a weak penetration path or is in an unpenetrated state, a lower leakage risk level can be assigned, or the process should proceed to review.

[0078] In the output stage, the image processing unit writes the defect classification results, leakage risk level, candidate defect area location, and through-path information into the inspection record. The inspection record can be transmitted to the quality assessment unit, review station, or production line control. The quality assessment unit determines whether the valve sealing surface should enter the re-inspection, rework, or release process based on the leakage risk level. For cases where the speckle image quality is insufficient, the risk sub-image lacks effective nodes, or the high-pressure side boundary area and low-pressure side boundary area cannot be determined, the image processing unit marks the relevant inspection results as invalid and initiates a re-image or manual review process for the workpiece.

[0079] In this embodiment, the object of inspection is the metal valve sealing surface with sample number V260410-017. After the valve sealing surface enters the online inspection station, the input data is obtained by the positioning camera, the phase fringe reflection acquisition unit, and the multi-incident angle laser speckle acquisition unit, respectively.

[0080] A positioning camera acquires positioning images of the sealing contact strip, and an image processing unit identifies the inner and outer boundaries in these images. After boundary fitting, the width range of the sealing contact strip is 0–2.40 mm, and the circumferential reference is determined to be 0° by the fixture zero position. The image processing unit maps the image position of the valve sealing surface to a coordinate region defined by the width and circumferential references of the contact strip, forming the functional coordinates of the sealing contact strip.

[0081] like Figure 2 As shown, the horizontal coordinate represents the circumferential position, and the vertical coordinate represents the position along the width of the contact zone. The high-pressure side boundary area is located at 0–0.20 mm in the width direction, and the low-pressure side boundary area is located at 2.20–2.40 mm in the width direction. C1 is the candidate area for a penetrating scratch extending along the width direction, C2 is the candidate area for pitting pores, and C3 is the contamination residue verification area. The right-hand field area records the sample number V260410-017, the contact zone width (0–2.40 mm), the phase state (0°, 90°, 180°, 270°), the speckle incident angle (20°, 35°, 50°), the candidate threshold Q≥0.45, and the risk threshold (0.45–0.65). The lower flowchart area represents the functional coordinates of the sealing contact zone formed after boundary fitting of the positioning image, and the registration of phase and speckle data is completed within these functional coordinates.

[0082] The phase fringe reflection acquisition unit acquires phase fringe reflection images at four phase states: 0°, 90°, 180°, and 270°, for the same valve sealing surface. The multi-incident-angle laser speckle acquisition unit acquires laser speckle images at three incident angles: 20°, 35°, and 50°. All of the above images are resampled and registered according to the functional coordinates of the sealing contact zone, so that the phase grayscale sequence and speckle grayscale sequence at the same coordinate position correspond to the same actual detection position.

[0083] The image processing unit generates a reflection phase map based on the grayscale values ​​of the four phase states and establishes a circumferential reference along the circumferential direction. Taking the circumferential data at a position of 1.20 mm in the width direction as an example, the median deviation of the circumferential residual of the qualified sample is 0.08 rad. This sample exhibits continuous phase deviation within the angle range of 136°–150°, with a maximum phase deviation of 0.42 rad. The image processing unit converts the phase deviation into a geometric perturbation response: the geometric perturbation response for region C1 is 0.78, for region C2 it is 0.50, and for region C3 it is 0.18.

[0084] On the speckle side, the image processing unit determines the speckle correlation window based on the phase curvature intensity. Region C1 exhibits local phase changes extending along its width; the speckle correlation window is a long strip window arranged along the defect extension direction, with a size of 9×21 pixels. Region C2 shows localized point-like phase changes; the speckle correlation window is a 13×13 pixel window. Region C3 shows weaker phase changes but more pronounced speckle fluctuations; the window size remains 15×15 pixels to reduce the impact of random speckle noise on correlation calculations.

[0085] The image processing unit calculates the local correlation at different incident angles within the speckle correlation window. The average local correlation for region C1 between incident angles of 20° and 50° is 0.28, corresponding to a microscopic scattering response of 0.72; the average local correlation for region C2 is 0.18, corresponding to a microscopic scattering response of 0.82; and the average local correlation for region C3 is 0.32, corresponding to a microscopic scattering response of 0.68. Because region C3 has a low geometric perturbation response, the image processing unit temporarily stores it as a verification object instead of directly outputting it as a high-risk defect.

[0086] The image processing unit establishes a phase speckle cross-verification risk tensor within the same functional coordinate system of the sealed contact zone. Tensor elements record the geometric perturbation response G, the microscopic scattering response M, the cross-verification term G×M, and the local cross-verification risk Q. The local cross-verification risk is determined by weighting the geometric perturbation response, microscopic scattering response, and cross-verification term at 0.35, 0.35, and 0.30, respectively. This weighting is calibrated using the response distributions of 20 qualified samples and 12 defective samples. The geometric perturbation response and microscopic scattering response are used to retain unilateral anomaly information, while the cross-verification term is used to enhance the risk contribution of bilateral anomalies at the same location.

[0087] Based on the above calculation relationships, the geometric perturbation response of region C1 is 0.78, the microscopic scattering response is 0.72, the mutual verification term is 0.56, and the local mutual verification risk is 0.69; the geometric perturbation response of region C2 is 0.50, the microscopic scattering response is 0.82, the mutual verification term is 0.41, and the local mutual verification risk is 0.58; the geometric perturbation response of region C3 is 0.18, the microscopic scattering response is 0.68, the mutual verification term is 0.12, and the local mutual verification risk is 0.34. The candidate defect threshold is set to 0.45, which is determined based on the 95th percentile of the local mutual verification risk of qualified samples (0.39) plus a safety margin of 0.06. C1 and C2 meet the candidate defect threshold and enter the candidate defect region merging process; C3 is below the candidate defect threshold and is not used as a candidate defect region for constructing the directed leakage risk map; however, the speckle-side response of C3 reaches 0.68, so the image processing unit retains the response fingerprint of C3 as a review object and puts C3 into the cleaning and re-image review process.

[0088] like Figure 3As shown, the matrix on the left represents the local mutual verification risk distribution of the phase speckle mutual verification risk tensor in the functional coordinates of the sealing contact zone. The horizontal axis of the matrix represents the circumferential position, and the vertical axis represents the width position. The values ​​in the matrix grid points represent the local mutual verification risk of the corresponding local area. In the figure, red grid points represent high-risk areas, orange grid points represent candidate areas, yellow grid points represent verification areas, and light green grid points represent normal fluctuation areas. The local mutual verification risk matrix on the right lists the key values ​​of C1, C2, C3, and candidate thresholds. The flowchart in the lower right corner shows the relationship between the phase deviation amount generating geometric perturbation G, the speckle deviation amount generating microscopic scattering M, and the mutual verification term G×M generating local mutual verification Q. Figure 3 C1 forms a continuous high-risk distribution within a range of approximately 120°–180° in the circumferential direction, C2 forms a local candidate distribution near approximately 240° in the circumferential direction, and C3 exhibits a core area distribution near approximately 300° in the circumferential direction.

[0089] After the candidate defect region is generated, the image processing unit discretizes C1 into 5 risk units, corresponding to risk nodes N1 to N5. N1 is located at 0.18mm in width and 132° in circumference, with a local mutual verification risk of 0.68; N2 is located at 0.70mm in width and 136° in circumference, with a local mutual verification risk of 0.71; N3 is located at 1.15mm in width and 141° in circumference, with a local mutual verification risk of 0.75; N4 is located at 1.72mm in width and 146° in circumference, with a local mutual verification risk of 0.69; and N5 is located at 2.20mm in width and 151° in circumference, with a local mutual verification risk of 0.63.

[0090] The image processing unit establishes directed edges based on the distance between adjacent risk nodes, the local mutual verification risk associated with adjacent risk nodes, and the consistency between the node connection direction and the direction from the high-pressure side boundary to the low-pressure side boundary. Within region C1, the distance between nodes N1 to N5 is less than 0.65 mm, and the connection direction is consistent with the leakage direction, thus forming a directed edge chain of N1→N2→N3→N4→N5. Region C2 forms two risk nodes, P1 and P2, but these nodes are only distributed within a width of 0.72–0.94 mm and are not close to the low-pressure side boundary, therefore a through path cannot be formed.

[0091] The image processing unit performs a multi-threshold interconnection scan of the directed leakage risk graph using five risk thresholds: 0.45, 0.50, 0.55, 0.60, and 0.65. The risk thresholds are set progressively upwards from the candidate defect threshold of 0.45 to determine whether defect chains remain connected under different risk intensities. During the scan, risk nodes with local mutual verification risks not less than the current risk threshold, as well as directed edges with connection strengths not less than the current risk threshold, are retained, forming corresponding risk subgraphs.

[0092] like Figure 4 As shown, the left side displays the risk nodes and directed edges in region C1. The arrows indicate the direction of the leakage path, with the purple arrows indicating the leakage direction from the high-pressure side boundary to the low-pressure side boundary. The table on the right shows the results of the multi-threshold through-scan. When the risk threshold is 0.45, 0.50, 0.55, and 0.60, a through-path extending from N1 to N5 exists in all risk subgraphs. When the risk threshold is 0.65, the local mutual verification risk of N5 (0.63) does not meet the threshold requirement, and the path cannot reach the low-pressure side boundary. Figure 4 The “N1→N5” in the diagram represents the path summary. The main path recorded in the through-path information box is N1→N2→N3→N4→N5, with a threshold range of 0.45–0.60 and a width range of 0.18–2.20 mm. Region C2 does not have a directed connected path from the high-pressure side boundary to the low-pressure side boundary and is recorded as not connected.

[0093] After generating the through-path information, the image processing unit extracts the defect response fingerprints of candidate defect regions and verification objects from the phase speckle cross-verification risk tensor. The defect response fingerprint of C1 has a phase perturbation of 0.78 with a linear and continuous distribution, a speckle angle decorrelation of 0.72 with continuous enhancement along the width direction, and a stable through-path. The classification result is determined to be a through scratch, and the leakage risk level is determined to be high. The defect response fingerprint of C2 has a phase perturbation of 0.50 with point-like variations, a speckle angle decorrelation of 0.82 with a concentrated distribution, and no through-path is formed. The classification result is determined to be pitting pores, and the leakage risk level is determined to be medium. The defect response fingerprint of C3 has a phase perturbation of 0.18, a speckle angle decorrelation of 0.68, and a patchy distribution. The classification result is determined to be residual contamination, and the leakage risk level is determined to be low, triggering cleaning and re-image capture. C4, as a normal texture region, has a phase perturbation of 0.12 and a speckle decorrelation of 0.19. The classification result is normal texture, and the processing destination is record archiving.

[0094] like Figure 5 As shown, the left-hand flow represents the data transfer relationship between the risk tensor, defect response fingerprint, through-path information, and detection record write-back. The upper right table records the phase perturbation, speckle decorrelation, classification results, risk level, and processing destination for C1 to C4. In the middle result write-back field, the sample number is V260410-017, the candidate region is C1, the status code is RISK_HIGH, the main path is N1-N5, the threshold range is 0.45-0.60, and the interface is QC_WRITEBACK. The lower left review trigger area indicates that medium-level defects enter manual review, contamination residue enters cleaning and re-photographing, and invalid images enter re-acquisition. The lower right quality inspection output destinations include re-inspection, manual review, cleaning and re-photographing, and release archiving.

[0095] In the above implementation process, the reflection phase residual map provides evidence of surface geometric perturbation, and the multi-incident-angle laser speckle image provides evidence of microscopic scattering anomalies. The phase speckle mutual verification risk tensor establishes a correspondence between the two in the same functional coordinate system of the sealing contact zone. The directed leakage risk map further transforms the local defect response into a path judgment object from the high-pressure side boundary zone to the low-pressure side boundary zone, so that the final output not only reflects the existence of local defects, but also reflects whether the defects have the spatial connectivity conditions to form a sealing leakage channel.

[0096] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0097] The various embodiments in this invention are described in a progressive manner. Similar or identical parts between embodiments can be referred to mutually. Each embodiment focuses on describing the differences from other embodiments. In particular, the system embodiments are basically similar to the method embodiments, so the description is relatively simple; relevant parts can be referred to the descriptions in the method embodiments.

Claims

1. A method for online detection and classification of valve sealing surface defects based on machine vision, characterized in that, The method includes: Acquire images of the sealing contact strip boundary, phase fringe reflection, and multi-incident angle laser speckle images, and construct functional coordinates of the sealing contact strip based on the sealing contact strip boundary; Based on the phase fringe reflection image, a reflection phase residual map and a phase curvature intensity are generated. Based on the phase curvature intensity, speckle correlation processing is constrained to obtain a speckle angle decorrelation map. A phase speckle cross-verification risk tensor is generated based on the reflection phase residual map and the speckle angle decorrelation map. Candidate defect regions are determined based on the phase speckle cross-verification risk tensor, and a directed leakage risk map is constructed based on the candidate defect regions. The directed leakage risk map is subjected to multi-threshold through-scan to obtain through-path information. Based on the phase speckle cross-validation risk tensor and the through-path information, the defect classification result and leakage risk level are output.

2. The method according to claim 1, characterized in that, The construction of the sealing contact zone functional coordinates based on the sealing contact zone boundary includes: The contact strip width direction, contact strip width direction range, and circumferential direction reference are determined based on the boundary of the sealing contact strip. The image position of the valve sealing surface is mapped to a coordinate region defined by the width direction of the contact strip and the circumferential direction reference to obtain the functional coordinates of the sealing contact strip.

3. The method according to claim 1, characterized in that, The step of generating a reflection phase residual map and phase curvature intensity based on the phase fringe reflection image includes: A reflection phase map is generated based on the phase stripe reflection images under multiple phase states; Based on the functional coordinates of the sealing contact strip, the circumferential reference residual of the reflection phase diagram is calculated to obtain the reflection phase residual diagram; The phase curvature intensity is determined based on the local changes in the reflected phase residual map.

4. The method according to claim 3, characterized in that, The speckle correlation processing based on the phase curvature intensity constraint includes: The window size, window orientation, and window shape of the speckle correlation window are determined based on the phase curvature intensity. Within the speckle correlation window, the local correlation between the multi-incident-angle laser speckle images at different incident angles is calculated, and the speckle angle decorrelation map is generated based on the local correlation.

5. The method according to claim 4, characterized in that, The step of generating a phase speckle cross-verification risk tensor based on the reflection phase residual map and the speckle angle decorrelation map includes: Establish a circumferential self-reference baseline for the same valve sealing surface in the functional coordinates of the sealing contact zone; The phase speckle cross-verification risk tensor is generated based on the phase deviation of the reflection phase residual map relative to the circumferential self-reference baseline and the speckle angle decorrelation map relative to the circumferential self-reference baseline.

6. The method according to claim 5, characterized in that, The generation of the phase speckle cross-verification risk tensor includes: The geometric disturbance response is determined based on the phase deviation. The microscopic scattering response is determined based on the speckle deviation. The phase speckle mutual verification risk tensor is generated by associating the geometric perturbation response, the microscopic scattering response, and the mutual verification terms determined based on the geometric perturbation response and the microscopic scattering response as tensor elements at the same location.

7. The method according to claim 1, characterized in that, The determination of candidate defect regions based on the phase speckle cross-verification risk tensor includes: In the functional coordinates of the sealing contact strip, the local mutual verification risk is determined based on the phase speckle mutual verification risk tensor; The adjacent regions with local mutual verification risks not less than the candidate defect threshold are merged to obtain the candidate defect region, and the candidate defect region is used as the input for constructing the directed leakage risk map.

8. The method according to claim 7, characterized in that, The construction of the directed leakage risk graph includes: The candidate defect region is discretized into risk units, and risk nodes are determined based on the position of the risk units in the functional coordinates of the sealing contact strip. The high-pressure side boundary region and the low-pressure side boundary region are also determined in the functional coordinates of the sealing contact strip. The directed edge and its connection strength are determined based on the distance between adjacent risk nodes, the local mutual verification risk associated with the adjacent risk nodes, and the consistency between the connection direction of the risk nodes and the leakage direction from the high-pressure side boundary area to the low-pressure side boundary area. The directed leakage risk graph is generated based on the risk nodes, the directed edges, and the connection strength of the directed edges.

9. The method according to claim 8, characterized in that, The step of performing a multi-threshold through-scan on the directed leakage risk map to obtain through-path information includes: Based on multiple risk thresholds, risk nodes with local mutual verification risk not less than the corresponding risk threshold are retained, and directed edges with connection strength not less than the corresponding risk threshold are also retained, resulting in multiple risk subgraphs. Search for directed connectivity paths from the high-pressure side boundary region to the low-pressure side boundary region in each risk subgraph; The directed connected path obtained from the search and the threshold range in which the directed connected path is located are determined as the through path information.

10. The method according to claim 1, characterized in that, The step of outputting defect classification results and leakage risk levels based on the phase speckle cross-verification risk tensor and the penetration path information includes: Based on the phase speckle cross-verification risk tensor, the defect response fingerprint of the candidate defect region is extracted, and the defect response fingerprint includes phase perturbation features and speckle angle decorrelation features. The defect classification result is determined based on the defect response fingerprint; The leakage risk level is determined based on the defect classification result and the penetration path information, and the defect classification result and the leakage risk level are output.

Images