Vision-based precision part topography image analysis method

Abstract

This invention relates to the field of machine vision and precision component image processing technology, specifically a vision-based method for analyzing the morphology of precision components. The method includes: acquiring a two-dimensional image sequence of the target component using a vision acquisition device according to initial vision acquisition parameters; extracting pixel grayscale distribution information and spatial neighborhood topological relationships; performing two-dimensional manifold spatial mapping; initializing a virtual energy field and extracting the topological semantic decoupling degree representing edge differences based on its dynamic evolution; dynamically configuring the morphology analysis process; calculating and obtaining the global morphology skeleton and local structural abrupt change features; and adjusting the vision acquisition parameters in real time accordingly, forming a closed loop of acquisition, analysis, and control. This invention achieves the separation of real geometric morphology edges from optical pseudo-edges.

Description

Technical Field

[0001] This invention relates to the field of machine vision and precision component image processing technology, specifically a vision-based method for analyzing the shape images of precision components. Background Technology

[0002] With the rapid development of precision manufacturing and online inspection technologies, more and more precision parts need to be visually identified in real time during the processing. This is especially true for structural features such as blades, hole edges, chamfers, and minute defects, where image analysis results directly affect the judgment of processing quality and subsequent process adjustments. Therefore, how to stably extract the true geometric shape edges of precision parts under conditions of complex surface reflection, partial occlusion, and continuous motion acquisition is a technical problem that urgently needs to be solved in the field of visual inspection.

[0003] Traditional visual morphology analysis of precision components currently relies mainly on the following methods: edge extraction based on grayscale thresholds, contour detection based on fixed operators, and defect recognition based on static images;

[0004] However, grayscale thresholding, fixed operator contour detection, and static image defect recognition all have certain drawbacks. For example, grayscale thresholding is easily affected by high-gloss reflections and false edge interference from metal surfaces, making it difficult to distinguish between real geometric boundaries and optical interference. Fixed operator contour detection is not adaptable to complex curved surfaces, local fractures, and continuous reflective areas, and is prone to problems such as discontinuous skeletons or false detections. Static image defect recognition methods lack consideration for the evolution of image structure and the feedback adjustment of acquisition parameters, making it difficult to balance the stability of morphological analysis and real-time control capabilities in online processing and inspection scenarios. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides a vision-based method for analyzing the morphology of precision components from images. Specifically, the technical solution of this invention includes:

[0006] The target component's two-dimensional image sequence is acquired using a visual acquisition device according to the initial visual acquisition parameters, and pixel grayscale distribution information and spatial neighborhood topology are extracted.

[0007] Two-dimensional manifold space mapping is performed based on pixel grayscale distribution information and spatial neighborhood topology to obtain the initial two-dimensional image manifold space;

[0008] The virtual energy field distribution matrix is ​​initialized in the manifold space of the initial two-dimensional image based on the pixel grayscale distribution information. The topological semantic decoupling degree representing the edge difference is extracted according to the dynamic evolution state of the initial virtual energy field in the space.

[0009] Based on the topological semantic decoupling degree, the target energy field evolution convergence rate, the target manifold gradient continuity index, and the target structure isomorphism preservation rate are dynamically configured to generate an adaptive morphology analysis process.

[0010] The initial two-dimensional image manifold space is solved according to the adaptation morphology analysis process to obtain the global morphology skeleton and local structural mutation features;

[0011] Control instructions are generated based on the global topography skeleton and local structural mutation features, and the visual acquisition parameters of the visual acquisition device are controlled in real time through a programmable logic controller.

[0012] Among them, the virtual energy field is a distribution matrix constructed by combining the initial energy distribution values ​​of the boundary nodes with the pixel grayscale distribution information; the target energy field evolution convergence rate is used to control the number of iteration steps of the initial virtual energy field in the initial two-dimensional image manifold space; the target manifold gradient continuity index is used to constrain the adaptive repair of local incomplete features; and the target structure isomorphism preservation rate is used to constrain the isomorphic mapping verification of local high-frequency abrupt change regions.

[0013] Preferably, a two-dimensional manifold space mapping is performed based on pixel grayscale distribution information and spatial neighborhood topological relationships to obtain an initial two-dimensional image manifold space, including:

[0014] Using pixel grayscale distribution information and spatial neighborhood topology as input variables, a two-dimensional image manifold space mapping function is constructed.

[0015] The initial two-dimensional image manifold space is generated by performing matrix transformation on the two-dimensional image sequence of the target component through the two-dimensional image manifold space mapping function.

[0016] Preferably, the initial two-dimensional image manifold space is initialized with a virtual energy field distribution matrix based on pixel grayscale distribution information, and the topological semantic decoupling degree representing edge differences is extracted according to the dynamic evolution state of the initial virtual energy field in space, including:

[0017] The mapping nodes corresponding to the preset image edge regions obtained based on the initial high gradient detection are extracted from the manifold space of the initial two-dimensional image as boundary nodes. The boundary nodes are assigned preset initial energy distribution values, and a virtual energy field distribution matrix is ​​constructed by combining pixel grayscale distribution information as the initial virtual energy field.

[0018] Calculate the manifold curvature of each mapping node in the initial two-dimensional image manifold space; guide the initial virtual energy field to perform numerical diffusion iteration based on the manifold curvature of the initial two-dimensional image manifold space to obtain the steady-state distribution matrix of the virtual energy field;

[0019] The steady-state distribution matrix of the virtual energy field is processed to separate the real geometric shape edge from the optical pseudo edge, and the energy distribution difference between the real geometric shape edge and the optical pseudo edge is calculated as the topological semantic decoupling degree.

[0020] Preferably, the steady-state distribution matrix of the virtual energy field is processed to separate the real geometric shape edges from the optical pseudo-edges, and the energy distribution difference between the real geometric shape edges and the optical pseudo-edges is calculated as the topological semantic decoupling degree, including:

[0021] Extract energy convergence paths and energy diversion nodes from the steady-state distribution matrix of the virtual energy field;

[0022] The edge continuity score of the energy convergence path is calculated based on the ratio of the length of the energy convergence path to the number of energy diversion nodes.

[0023] If the edge continuity score is greater than the preset continuity threshold, the edge corresponding to the energy convergence path will be judged as the true geometric shape edge.

[0024] If the edge continuity score is less than or equal to the preset continuity threshold, the edge corresponding to the energy convergence path will be identified as an optical pseudo-edge.

[0025] The ratio of the energy density of the real geometric edge to the energy density of the optical pseudo-edge is used as the topological semantic decoupling degree.

[0026] Preferably, the target energy field evolution convergence rate, target manifold gradient continuity index, and target structural isomorphism preservation rate are dynamically configured based on the topological semantic decoupling degree to generate an adaptive morphology analysis process, including:

[0027] Obtain historical component image analysis logs. Based on the historical component image analysis logs, configure a decoupling parameter mapping table. The decoupling parameter mapping table contains at least one mapping relationship, and each mapping relationship includes the sample topological semantic decoupling threshold range and the corresponding sample energy field evolution convergence rate, sample manifold gradient continuity index, and sample structural isomorphism preservation rate.

[0028] Using the decoupling parameter mapping table, the corresponding target energy field evolution convergence rate, target manifold gradient continuity index, and target structure isomorphism preservation rate are obtained based on topological semantic decoupling degree matching.

[0029] The convergence rate of the target energy field evolution, the continuity index of the target manifold gradient, and the isomorphism preservation rate of the target structure are combined as the adaptation morphology analysis process.

[0030] Preferably, the initial two-dimensional image manifold space is solved according to the adaptive morphology analysis process to obtain the global morphology skeleton and local structural abrupt change features, including:

[0031] According to the convergence rate of the target energy field evolution in the adaptation morphology analysis process, the number of iteration steps of the initial virtual energy field in the initial two-dimensional image manifold space is controlled to obtain the converged energy field distribution map.

[0032] Extract the incomplete features from the convergence energy field distribution map, and perform adaptive repair processing on the incomplete features according to the target manifold gradient continuity index in the adaptation morphology analysis process to extract the global morphology skeleton.

[0033] High-frequency abrupt change regions are extracted from the convergent energy field distribution map, and isomorphic mapping verification is performed on the high-frequency abrupt change regions according to the target structure isomorphic retention rate in the adaptation morphology analysis process to extract local structural abrupt change features.

[0034] Preferably, incomplete features are extracted from the convergent energy field distribution map, and adaptive repair processing is performed on the incomplete features according to the target manifold gradient continuity index in the adaptation morphology analysis process to extract the global morphology skeleton, including:

[0035] Calculate the manifold gradient values ​​for each local region in the convergent energy field distribution map;

[0036] Regions where the gradient value of the local manifold is less than the gradient continuity index of the target manifold are identified as gradient break regions, and these gradient break regions are used as incomplete features.

[0037] By utilizing the energy flow direction of the surrounding area of ​​the incomplete feature, path interpolation is performed on the incomplete feature to generate continuous morphological feature lines.

[0038] A global topological skeleton is constructed by topologically connecting continuous feature lines.

[0039] Preferably, high-frequency abrupt change regions are extracted from the convergent energy field distribution map, and isomorphic mapping verification is performed on these regions according to the target structure isomorphic preservation rate in the adaptation morphology analysis process. Local structural abrupt change features are then extracted, including:

[0040] In the convergent energy field distribution map, regions with energy distribution variance greater than a preset variance threshold are extracted as high-frequency mutation regions.

[0041] Extract a two-dimensional topological graph of high-frequency mutation regions, where the two-dimensional topological graph includes nodes and connected components;

[0042] Obtain a two-dimensional projection topology of the target component's preset three-dimensional design model from the current visual perspective of the visual acquisition device;

[0043] The topological similarity of nodes and connected components between a two-dimensional topological relation graph and a two-dimensional projected topological graph is used as the graph theory isomorphism matching degree.

[0044] If the graph theory isomorphism matching degree is greater than or equal to the target structure isomorphism retention rate, then the high-frequency mutation region is identified as a local structural mutation feature.

[0045] If the graph theory isomorphism matching degree is less than the target structure isomorphism retention rate, then the high-frequency abrupt change region is identified as optical interference noise and is removed.

[0046] Preferably, control instructions are generated based on the global topographic skeleton and local structural abrupt changes, and the visual acquisition parameters of the visual acquisition device are controlled in real time by a programmable logic controller, including:

[0047] The integrity score of the global topography skeleton is calculated based on the number of connected branches of the global topography skeleton, and the signal-to-noise ratio of the local structural mutation features is calculated based on the ratio of the energy density of the local structural mutation features to the preset background energy density extracted from the background region.

[0048] Among them, the visual acquisition parameters include the lighting conditions;

[0049] If the integrity score is less than the preset integrity threshold or the signal-to-noise ratio is less than the preset signal-to-noise ratio threshold, a light source adjustment command is generated and sent to the light source controller included in the vision acquisition device. The light source controller then adjusts the illumination status of the target component in real time.

[0050] If the integrity score is greater than or equal to the preset integrity threshold and the signal-to-noise ratio is greater than or equal to the preset signal-to-noise ratio threshold, then keep the current visual acquisition parameters unchanged.

[0051] Compared with the prior art, the present invention has the following beneficial effects:

[0052] 1. By adopting the above technical solution, the present invention simultaneously extracts pixel grayscale distribution information and spatial neighborhood topology from the two-dimensional image sequence of the target component within a preset acquisition period, and performs two-dimensional manifold space mapping and virtual energy field dynamic evolution analysis accordingly, effectively solving the problem that the traditional grayscale thresholding method is difficult to distinguish between real geometric boundaries and optical interference under the conditions of high light reflection and pseudo-edge interference on metal surfaces.

[0053] 2. This invention transforms the original planar pixel information into an initial two-dimensional image manifold space that better highlights local continuity and global structural relationships. Combined with boundary node energy initialization, steady-state energy diffusion, separation of energy convergence paths and branching nodes, and topological semantic decoupling calculation, it ensures that the edges of real geometric shapes can be effectively distinguished from optical pseudo-edges in terms of continuous propagation stability, thereby improving the edge recognition accuracy under complex curved surfaces and continuous reflective areas.

[0054] 3. Based on the topological semantic decoupling degree, this invention further matches the convergence rate of the target energy field evolution, the continuity index of the target manifold gradient, and the isomorphism preservation rate of the target structure. When these are applied to global morphological skeleton extraction, adaptive repair of incomplete features, and isomorphic mapping verification of local high-frequency abrupt change regions, the system can take into account both the restoration of main boundary connectivity and reliable screening of local anomalies. This effectively improves the defects of fixed operator contour detection, which is prone to skeleton discontinuity, false detection, and insufficient adaptability to complex structures.

[0055] 4. This invention generates light source adjustment instructions based on the global morphological skeleton integrity score and the signal-to-noise ratio of local structural mutation features, and performs real-time control of illumination state, incident angle and polarization direction, constructing a closed-loop link of acquisition, analysis and control. This effectively makes up for the lack of structural evolution relationship and the deficiency of acquisition parameter feedback adjustment in static image defect recognition. Ultimately, it can significantly improve the stability of real morphological recognition of precision parts in online processing and inspection scenarios, the reliability of defect extraction and the adaptive control capability of visual acquisition parameters. Attached Figure Description

[0056] Figure 1 This is a schematic flowchart of a vision-based precision component topography image analysis method provided in an embodiment of this application. Detailed Implementation

[0057] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments.

[0058] Vision-based image analysis methods for precision component topography include:

[0059] The target component's two-dimensional image sequence is acquired using a visual acquisition device according to the initial visual acquisition parameters, and pixel grayscale distribution information and spatial neighborhood topology are extracted.

[0060] Two-dimensional manifold space mapping is performed based on pixel grayscale distribution information and spatial neighborhood topology to obtain the initial two-dimensional image manifold space;

[0061] The virtual energy field distribution matrix is ​​initialized in the manifold space of the initial two-dimensional image based on the pixel grayscale distribution information. The topological semantic decoupling degree representing the edge difference is extracted according to the dynamic evolution state of the initial virtual energy field in the space.

[0062] Based on the topological semantic decoupling degree, the target energy field evolution convergence rate, the target manifold gradient continuity index, and the target structure isomorphism preservation rate are dynamically configured to generate an adaptive morphology analysis process.

[0063] The initial two-dimensional image manifold space is solved according to the adaptation morphology analysis process to obtain the global morphology skeleton and local structural mutation features;

[0064] Control instructions are generated based on the global topography skeleton and local structural mutation features, and the visual acquisition parameters of the visual acquisition device are controlled in real time through a programmable logic controller.

[0065] Among them, the virtual energy field is a distribution matrix constructed by combining the initial energy distribution values ​​of the boundary nodes with the pixel grayscale distribution information; the target energy field evolution convergence rate is used to control the number of iteration steps of the initial virtual energy field in the initial two-dimensional image manifold space; the target manifold gradient continuity index is used to constrain the adaptive repair of local incomplete features; and the target structure isomorphism preservation rate is used to constrain the isomorphic mapping verification of local high-frequency abrupt change regions.

[0066] This embodiment provides a vision-based mechanism for analyzing the shape of precision components from images, such as... Figure 1 As shown; specifically, the method is deployed at a precision turbine blade online machining and inspection station, which includes an industrial camera, an adjustable ring light source, a polarizer assembly, an edge computing unit, and a light source controller for receiving control commands;

[0067] The target component is the leading edge section of a turbine blade that has just been precision milled. Its surface has both real geometric shape edges and optical pseudo edges caused by metal reflection. Therefore, shape recognition needs to be completed in a closed-loop process of continuous acquisition, continuous analysis and continuous dimming.

[0068] Specifically, the visual acquisition device acquires a sequence of two-dimensional images of the target component within a preset acquisition period according to initial visual acquisition parameters; the initial parameters may include exposure time, gain, illumination intensity, incident angle, and polarization direction; for ease of explanation, it is assumed that four frames of images are continuously acquired within the preset acquisition period. to Take one image per frame The pixel region serves as an example of local extrapolation, where... The grayscale matrix can be represented as:

[0069] In this matrix, the high-grayscale column on the right corresponds to the suspected edge region; the matrix elements represent the grayscale values ​​of the corresponding pixels, which are dimensionless pure numbers; and it is set that... This indicates the first element in the local image matrix. line, number Column of pixel nodes;

[0070] At this point, the system not only records the grayscale value of each pixel, but also extracts the spatial neighborhood topology, such as marking the connectivity between the center pixel and its surrounding pixels using an 8-neighborhood approach; if the center pixel With the right pixel Top right pixel bottom right pixel If all relationships form a high gradient transition, then these relationships are written into the adjacency table to form the basic input for subsequent manifold mappings;

[0071] The system performs two-dimensional manifold space mapping based on pixel grayscale distribution information and spatial neighborhood topological relationships to obtain an initial two-dimensional image manifold space; this can be understood as converting the original planar pixel arrangement into a matrix representation that better highlights local continuity and global structural relationships.

[0072] For ease of explanation, if the original The region is mapped to a matrix ,but Instead of directly storing the original grayscale values, the system stores the structural values ​​after coupling with the neighborhood; for example, the center position is updated from 30 to 0.42, the suspected edge point on the right is updated to 0.91, and the background point in the lower left is updated to 0.18, thus forming:

[0073] In this matrix, when the value is greater than the preset manifold structure threshold, it indicates that the node is located at the structural boundary or energy accumulation region in the manifold space.

[0074] The system initializes the initial two-dimensional image manifold space with a virtual energy field distribution matrix based on pixel grayscale distribution information to obtain an initial virtual energy field. Specifically, nodes with high grayscale values ​​and located in preset edge regions can be assigned an initial energy greater than or equal to a preset high energy threshold, while background regions can be assigned an initial energy less than or equal to a preset low energy threshold. For example, an initial energy matrix can be generated as follows:

[0075] Subsequently, the system dynamically evolves according to the topological connectivity in the manifold space, causing energy to propagate along continuous boundaries and to split or attenuate at optical pseudo-edges;

[0076] Assume that the steady-state energy matrix is ​​obtained after three rounds of iteration:

[0077] If the high-energy column on the right remains continuous along the longitudinal direction, and the energy gradient propagating from the center to the left is lower than the preset attenuation threshold, then the right region can be determined as the true geometric shape boundary.

[0078] The system further extracts the topological semantic decoupling degree, which characterizes edge differences, from the energy flow path, the splitting situation, and the density comparison. For example, the average energy density of the real geometric shape edge is 0.86, and the average energy density of the optical pseudo edge is 0.43. The decoupling degree can be taken as the ratio of the two to 2.0 or the normalized 0.67, which is used to indicate that the real geometric shape edge and optical interference in the current image are distinguishable.

[0079] Based on this, the system dynamically configures and adapts the morphology parsing process according to the topological semantic decoupling degree; if the decoupling degree is greater than or equal to the first preset decoupling degree threshold, it means that the gradient value of the real structure boundary in the current image is greater than the preset boundary gradient threshold, and the target energy field convergence rate greater than the first preset parameter threshold, the target manifold gradient continuity requirement greater than the second preset parameter threshold, and the target structure isomorphism preservation rate greater than the third preset parameter threshold can be adopted.

[0080] If the decoupling degree is less than the first preset decoupling degree threshold, it means that the energy distribution variance corresponding to the optical interference is greater than the preset interference variance threshold. Then, the preset convergence speed less than the first preset parameter threshold, the local repair threshold less than the second preset parameter threshold, and the topology consistency review parameter greater than the preset review intensity threshold are called.

[0081] For ease of explanation, if the decoupling degree is 0.67, the system can match the following parameter combination: the convergence rate of the target energy field evolution is 0.75, the continuity exponent of the target manifold gradient is 0.58, and the isomorphism preservation rate of the target structure is 0.81. This combination constitutes the adaptation morphology analysis process. The system solves the initial two-dimensional image manifold space according to this process to obtain the global morphology skeleton and local structural mutation features.

[0082] The global topography skeleton is used to characterize the overall topography of the blade's main boundary, cooling hole contour connection lines, etc.; local structural abrupt changes are used to characterize local anomalies such as chipping, scratches, burrs, etc.; it is assumed that after calculation, the system extracts a main skeleton line that runs through the leading edge. and two local high-frequency mutation regions , ;in Consistent with design expectations, this can be classified as a normal chamfer feature. If the topology is inconsistent with the design, it may indicate a manufacturing defect.

[0083] The system adjusts the visual acquisition parameters in real time based on the global morphological skeleton and the abrupt changes in local structures. If the main skeleton is broken or the contrast of the local abrupt change area is less than the preset contrast threshold, the system automatically increases lateral supplementary lighting or adjusts the polarization direction. If the morphology is intact and the signal-to-noise ratio of local features is high, the current acquisition parameters are kept unchanged, thus forming a closed loop of acquisition-analysis-control.

[0084] As an anomaly handling mechanism, if an image frame within a preset acquisition period is overexposed, causing the area of ​​high grayscale regions to exceed a preset upper limit, such as exceeding 60% of the entire image, then this frame will not directly enter manifold mapping, but will be marked as an abnormal frame and trigger re-exposure sampling; if it is still abnormal after two consecutive re-samplings, the previous stable parameter combination will be maintained, and the abnormal frame will be written to the analysis log for subsequent parameter mapping updates; if the acquired image has occlusion, resulting in a significant loss of topological connectivity, only low-confidence skeleton results will be output, and aggressive dimming will not be directly performed to avoid misjudgment during the robotic arm movement phase;

[0085] In online inspection of the turbine blade leading edge, the initial exposure time was set to 4 milliseconds, the ring light source brightness was 70%, and the camera continuously acquired 4 frames of images within a 0.5-second acquisition cycle. Due to the reflection from the blade's curved surface, the decoupling degree obtained in the first round of analysis was only 0.41. Based on this, the system selected a lower convergence rate and a higher isomorphism preservation constraint to first suppress the influence of specular false edges. After completing local repair, the connectivity of the leading edge main skeleton was improved, and the local scratched areas were cleared. The ratio of energy density to background energy density reaches 3.2, which meets the conditions for anomaly feature extraction;

[0086] The system sends an adjustment command to the light source controller to change the incident angle from 35 degrees to 28 degrees and rotate the polarization direction by 15 degrees, and obtain a clearer image in the next acquisition cycle, thereby stabilizing the output of the global skeleton and local defect boundaries of the blade leading edge.

[0087] The purpose of this step is to transform traditional static image analysis into a dynamic analytical process based on the evolution of manifold space and virtual energy field, thereby achieving the separation of real geometric shape edges from optical pseudo-edges, and further supporting subsequent adaptive parameter control and online shape reconstruction.

[0088] Furthermore, based on pixel grayscale distribution information and spatial neighborhood topology, a two-dimensional manifold space mapping is performed to obtain an initial two-dimensional image manifold space, including: using pixel grayscale distribution information and spatial neighborhood topology as input variables to construct a two-dimensional image manifold space mapping function; and performing matrix transformation on the two-dimensional image sequence of the target component through the two-dimensional image manifold space mapping function to generate the initial two-dimensional image manifold space.

[0089] This embodiment provides a construction step for a two-dimensional image manifold spatial mapping. Specifically, in the aforementioned online turbine blade detection scenario, simply using the original grayscale matrix is ​​easily affected by specular reflection, especially when the local highlight on the blade surface is close to the grayscale of the real edge. If edge detection is performed directly on the original pixel plane, the highlight band is easily misjudged as the leading edge contour. Therefore, this embodiment further introduces a manifold mapping function driven by both grayscale distribution and spatial neighborhood topology to enhance the expression of structural continuity.

[0090] Specifically, a two-dimensional image manifold space mapping function is constructed; this function can accept two types of inputs: one is pixel gray-level distribution information, and the other is neighborhood topological relationships; for ease of explanation, assume that a certain frame of image, after local cropping, forms 4 pixel nodes. , , , Their grayscale values ​​are 20, 25, 180, and 185, respectively. and Belongs to the dark background area. and It belongs to the highlighted boundary area;

[0091] Neighborhood connection relationship is connect , connect , connect If we only look at the difference in grayscale, and There are mutations between them; if we combine this with topological relationships, we can find... and Continuous highlighting along the same direction is more likely to form a stable boundary segment; based on this, when performing matrix transformation, the mapping function does not directly copy the grayscale, but integrates the node grayscale and neighborhood consistency into the new structural value.

[0092] The specific construction rules of the two-dimensional image manifold space mapping function are as follows: the structural value of the target pixel node is set to be the product of its normalized gray value and the neighborhood connectivity coefficient; when the target pixel node and its neighboring nodes form a connected path with a gray difference less than a preset difference threshold, the neighborhood connectivity coefficient is assigned an enhancement constant greater than 1; when the target pixel node is an isolated node or belongs to a smooth background region, the neighborhood connectivity coefficient is assigned a decay constant less than or equal to 1.

[0093] A simplified derivation can be used as follows: First, normalize the grayscale values ​​to 0 to 1, obtaining 0.10, 0.12, 0.88, and 0.90; then, based on neighborhood consistency, give correction coefficients, for example, correcting the background continuous region to 0.9 and the suspected boundary continuous region to 1.1, thus obtaining structure values ​​of 0.09, 0.11, 0.97, and 0.99 after transformation; if further written as... If the matrix is ​​given, then the initial two-dimensional image manifold space can be represented as: In this matrix, the pixel gray values ​​of the two nodes in the bottom row are greater than the preset gray value threshold, and the topological connectivity coefficient is greater than the preset connectivity threshold. Therefore, they are mapped as structurally significant regions in the manifold space.

[0094] To enable the mapping function to process the entire image frame, the transformation can be repeated using a sliding window or block-based approach; assuming an image frame is divided into 3 blocks... , , ,like If the nodes are densely connected but the grayscale is stable, the mapped value will be dominated by low-structure values; if If there are long, continuous bright edges, then a high-value ridge line will be formed after mapping; if If only isolated bright spots exist, even if the grayscale is high, the structure value will not be raised too much due to insufficient neighborhood support; this can prevent isolated reflective points from being mistakenly regarded as real edges in subsequent steps.

[0095] As an anomaly handling mechanism, if some pixel nodes are located at the image boundary and naturally lack a complete 8-neighborhood, they can be downgraded to use 4-neighborhood relationships for mapping to prevent null values ​​due to insufficient neighborhoods; if the gray levels of a local area are almost uniform, causing the normalization denominator to be close to 0, the area can be directly mapped to a low-confidence flat region and neighborhood enhancement can be skipped; if the adjacency relationship is broken due to occlusion, a break label can be added to the mapping matrix to reduce the weight of the area during subsequent energy field initialization.

[0096] In images taken at the leading edge of a blade, the true boundary of the leading edge appears as a continuous arc-shaped bright area, while the surface oil film reflection only appears as scattered bright spots. After processing with this mapping function, the arc-shaped bright area forms a continuous high-value band in the manifold space, while the oil film bright spots only appear as scattered high-value points. The subsequent energy field is more likely to form a stable convergence path in this high-value band, thereby reducing misidentification caused by oil film reflection.

[0097] The purpose of this step is to jointly encode the originally mixed grayscale information and neighborhood connectivity information in the two-dimensional image into a manifold matrix that is more suitable for structural analysis, thereby achieving differentiated representation of continuous boundaries, isolated noise and broken regions.

[0098] Furthermore, a virtual energy field distribution matrix is ​​initialized in the initial two-dimensional image manifold space based on pixel grayscale distribution information. The topological semantic decoupling degree representing edge differences is extracted according to the dynamic evolution of the initial virtual energy field in space, including:

[0099] The mapping nodes corresponding to the preset image edge regions obtained based on the initial high gradient detection are extracted from the manifold space of the initial two-dimensional image as boundary nodes. The boundary nodes are assigned preset initial energy distribution values, and a virtual energy field distribution matrix is ​​constructed by combining pixel grayscale distribution information as the initial virtual energy field.

[0100] Calculate the manifold curvature of each mapping node in the initial two-dimensional image manifold space; guide the initial virtual energy field to perform numerical diffusion iteration based on the manifold curvature of the initial two-dimensional image manifold space to obtain the steady-state distribution matrix of the virtual energy field;

[0101] The steady-state distribution matrix of the virtual energy field is processed to separate the real geometric shape edge from the optical pseudo edge, and the energy distribution difference between the real geometric shape edge and the optical pseudo edge is calculated as the topological semantic decoupling degree.

[0102] This embodiment provides a step-by-step guide for the initialization and dynamic evolution of a virtual energy field. Specifically, after completing the manifold mapping, it is still difficult to distinguish between continuous real geometric edges and local optical pseudo-edges based solely on static structural values, as both may exhibit high values ​​at a certain moment. To expose this deficiency, this embodiment further transforms the structural analysis into an energy propagation process. By observing the diffusion, convergence, and attenuation behavior of energy in the manifold space, more stable edge semantics are extracted.

[0103] Specifically, the system first extracts mapping nodes corresponding to preset edge regions from the manifold space of the initial two-dimensional image as boundary nodes; the preset edge regions can come from the initial high gradient detection, the design of regions of interest, or the projection of the results from the previous frame.

[0104] For ease of explanation, let's assume that in In the manifold matrix, the three nodes in the right column are selected as boundary nodes, and they are assigned initial energies of 0.8, 0.9, and 0.8 respectively. The remaining regions are assigned low energy values ​​between 0.1 and 0.2 based on the background gray level, thus obtaining the initial energy field distribution matrix. ;

[0105] Boundary nodes use initial energy values ​​greater than or equal to a preset initial energy threshold to construct continuous high-energy distribution paths in subsequent evolution; then, the system guides the initial energy field to perform numerical diffusion iteration based on the manifold curvature of the initial two-dimensional image manifold space;

[0106] The manifold curvature here can be understood as the trend of local structural changes: if adjacent nodes change smoothly and in the same direction in the manifold space, energy is more likely to diffuse along that direction; if the local structure changes abruptly or the connectivity is poor, energy will be blocked, diverted or attenuated.

[0107] The specific evolutionary rules of numerical diffusion iteration are as follows: In each iteration, the energy update value of the current node is the sum of its energy value in the previous round and the inflow energy from its neighboring nodes, minus the outflow energy; where the energy flow rate between nodes is negatively correlated with the manifold curvature on the connected path, the smaller the manifold curvature, the more consistent the local structure direction, and the higher the corresponding energy flow rate; in the local extrapolation example, it is assumed that... After one iteration, it was updated to:

[0108] After a second iteration, we obtain :

[0109] If the iteration continues until the average difference between two adjacent matrix iterations is less than a preset threshold, for example, less than 0.02, then the energy field can be considered to have entered a steady state, and the steady-state distribution matrix is ​​obtained. ; after obtaining Then, the system begins to separate the edges of the real geometric shape from the optical pseudo-edges; its core is not to look at a single high-value point, but to look at the continuity and dominance of the high-value region after evolution.

[0110] For example, true geometric edges often exhibit stable long-path energy convergence, while optical pseudo-edges often exhibit branch chaos or rapid attenuation during propagation due to a lack of geometric support; the system can statistically analyze the total length, number of splits, energy gradients on both sides of the path, and mean steady-state density of each high-energy path; if a path maintains its dominant channel status throughout multiple iterations, it is more likely to be identified as a true geometric edge;

[0111] Furthermore, the system calculates the topological semantic decoupling degree based on the energy distribution difference between the real geometric edge and the optical pseudo-edge; assuming the average energy of the real geometric edge region is 0.87 and the average energy of the optical pseudo-edge region is 0.46, the decoupling degree can be expressed as follows: Or it can be further normalized to 0.63; the higher the value, the more fully the real geometric shape edges in the current image are separated from optical interference, and the more reliable the subsequent skeleton extraction and local mutation identification will be.

[0112] As an anomaly handling mechanism, if the number of initial boundary nodes extracted is less than the preset lower limit of the proportion of boundary nodes, such as less than 1% of the total number of nodes in the whole image, it indicates that the distribution density of candidate edges is insufficient and the energy field is difficult to form an effective propagation. At this time, the system can automatically relax the initial gradient threshold and re-extract boundary nodes.

[0113] If the number of boundary nodes extracted is greater than the preset node proportion threshold, for example, exceeding 40% of the number of nodes, it means that the gradient change distribution range corresponding to the candidate region exceeds the preset boundary threshold. The system can first remove weak boundaries and then initialize them through non-maximum suppression.

[0114] If energy oscillations occur during the diffusion iteration process, that is, the energy of some nodes repeatedly rises and falls beyond the preset range in adjacent rounds, the diffusion step size can be reduced or the steady-state protection mode can be entered in advance to avoid noise being excessively amplified during the iteration.

[0115] In blade leading edge detection, since a coating reflection and a real knife mark coexist, both appear as bright thin lines in the original image. After initialization, the energy near the real leading edge continues to converge along the arc-shaped main direction, while the energy near the coating reflection quickly dissipates to the surrounding area after two rounds of diffusion, failing to form a stable main channel. Based on this, the system classifies the former as the edge of the real geometric shape and the latter as the optical pseudo-edge, and calculates a high degree of decoupling, providing a basis for the next step of adaptive parameter matching.

[0116] The purpose of this step is to characterize the propagation stability of the image structure through the energy field evolution process, thereby achieving algorithmic decoupling between the real geometric shape edges and optical interference, rather than relying on single-frame grayscale threshold judgment.

[0117] Furthermore, the virtual energy field steady-state distribution matrix is ​​processed to separate the real geometric shape edges from the optical pseudo-edges, and the energy distribution difference between the real geometric shape edges and the optical pseudo-edges is calculated as the topological semantic decoupling degree, including: extracting the energy convergence path and energy splitting node in the virtual energy field steady-state distribution matrix;

[0118] The edge continuity score of the energy convergence path is calculated based on the ratio of the length of the energy convergence path to the number of energy diversion nodes.

[0119] If the edge continuity score is greater than the preset continuity threshold, the edge corresponding to the energy convergence path will be judged as the true geometric shape edge.

[0120] If the edge continuity score is less than or equal to the preset continuity threshold, the edge corresponding to the energy convergence path will be identified as an optical pseudo-edge.

[0121] The ratio of the energy density of the real geometric edge to the energy density of the optical pseudo-edge is used as the topological semantic decoupling degree.

[0122] This embodiment provides an edge separation mechanism based on energy convergence paths and splitting nodes. Specifically, although a steady-state energy field has been obtained in the previous stage, if only the average energy distribution is used as the classification basis, there is still a technical defect of feature misjudgment: some optical pseudo-edges may also maintain high energy locally, especially in the continuous coating reflection area, and simple comparison of the average value is easy to fail. Therefore, this embodiment introduces a joint determination method of path length and splitting number, upgrading edge recognition from point value judgment to path behavior judgment.

[0123] Specifically, the system extracts energy convergence paths and energy diversion nodes in the steady-state energy field; the upward direction of the local energy gradient can be regarded as the energy flow direction. If multiple adjacent nodes are connected in series along approximately the same direction, they form an energy convergence path; if a point outputs a strong flow in two or more directions at the same time, it is identified as an energy diversion node.

[0124] For ease of explanation, assume that two candidate paths are extracted from the steady-state matrix. and ; There are 6 consecutive nodes, including only 1 branch node; There are 4 consecutive nodes, but 3 branching nodes are included; the system calculates the edge continuity score based on the ratio of path length to the number of branching nodes.

[0125] One simplification method is to use formulas As a rating; then The rating is , The rating is If the continuity threshold is preset to 2.0, then If the value is above the threshold, it can be considered to correspond to a stable real edge; If the edge is below the threshold, it is more likely to be an optical pseudo-edge. The advantage of this approach is that the edge of the true geometric shape usually extends along the geometric contour, with a longer main path and fewer branching.

[0126] Optical pseudo-edges are often affected by texture and reflection disturbances. Although the local gray value is abnormally high, the number of connected path nodes is less than the preset path length threshold, and the number of energy diversion nodes contained is greater than the preset diversion threshold.

[0127] After classification, the system calculates the energy density of true geometric edges and optical pseudo-edges respectively. Assuming the average energy density of all nodes identified as true geometric edges is 0.82 and the average energy density of nodes identified as optical pseudo-edges is 0.41, the topological semantic decoupling degree can be taken as... If the density of the real geometric shape edge in another frame is only 0.70, while the density of the optical pseudo edge is as high as 0.55, the decoupling degree drops to 1.27, indicating that the current optical conditions are poor and subsequent parameter configurations need to be more conservative.

[0128] As an anomaly handling mechanism, if the length of a candidate path is less than the preset path length threshold, such as less than 3 nodes, even if the score is high, it will not be directly output as the true geometric shape edge. Instead, it will be marked as a path to be confirmed and then judged in combination with the temporal consistency of adjacent frames. If the number of split nodes is 0, the denominator will be treated as 1 to avoid division by zero.

[0129] If the same node satisfies the assignment criteria of two paths, it will be assigned to the path with higher average energy and more stable direction. If the other path loses the key node and its length is insufficient, it will be automatically downgraded to an optical pseudo-edge candidate. If no path that meets the threshold is extracted in the entire image, it means that the current image structure is extremely unstable. At this time, a low decoupling degree will be returned directly and the acquisition parameters will be adjusted.

[0130] In the blade leading edge image, the true leading edge contour forms a main energy path extending along the arc surface, with a length of 12 nodes and only 2 shunting points, thus scoring higher. In contrast, the reflective stripes caused by residual cutting fluid, although locally bright, are divided into multiple short paths, and shunting occurs at the intersection of each bright spot, resulting in a lower final score. Based on this, the system retains the former as the true geometric shape edge, discards the latter as an optical pseudo-edge, and outputs a high degree of decoupling.

[0131] The purpose of this step is to change the edge determination criterion from whether it is bright or not to whether it can be stably continued on the topology, thereby improving the ability to preserve continuous real contours and reducing false detections of reflective pseudolines.

[0132] Furthermore, based on the topological semantic decoupling degree, the convergence rate of the target energy field evolution, the continuity index of the target manifold gradient, and the isomorphism preservation rate of the target structure are dynamically configured to generate an adaptive morphology analysis process, including:

[0133] Obtain historical component image analysis logs. Based on the historical component image analysis logs, configure a decoupling parameter mapping table. The decoupling parameter mapping table contains at least one mapping relationship, and each mapping relationship includes the sample topological semantic decoupling threshold range and the corresponding sample energy field evolution convergence rate, sample manifold gradient continuity index, and sample structural isomorphism preservation rate.

[0134] Using the decoupling parameter mapping table, the corresponding target energy field evolution convergence rate, target manifold gradient continuity index, and target structure isomorphism preservation rate are obtained based on topological semantic decoupling degree matching.

[0135] The convergence rate of the target energy field evolution, the continuity index of the target manifold gradient, and the isomorphism preservation rate of the target structure are combined as the adaptation morphology analysis process.

[0136] This embodiment provides a parameter adaptive configuration mechanism based on historical logs. Specifically, if the aforementioned process uses a fixed convergence rate, a fixed continuity index, and a fixed isomorphism threshold for each frame of image, it will expose obvious defects when the working conditions change: when the lighting conditions are good, the fixed parameters may be too conservative, resulting in low resolution efficiency; when the reflection is severe, the fixed parameters may be too aggressive, resulting in noise penetration. Therefore, this embodiment uses historical component image analysis logs to establish a decoupling parameter mapping table so that different image states match different resolution strategies.

[0137] Specifically, the system first obtains historical component image analysis logs; the logs record at least the decoupling degree, final skeleton integrity, accuracy of local anomaly identification, and the parameter combinations used for each analysis.

[0138] For ease of explanation, assume that three stable mapping relationships are formed in the historical logs after filtering: when the decoupling degree satisfies When the decoupling degree satisfies the following conditions, the corresponding convergence rate is 0.55, the gradient continuity exponent is 0.45, and the structural isomorphism preservation rate is 0.88; when the decoupling degree satisfies the following conditions... When the decoupling degree satisfies the following conditions, the corresponding convergence rate is 0.70, the gradient continuity exponent is 0.58, and the structural isomorphism preservation rate is 0.82; when the decoupling degree satisfies the following conditions... At that time, the corresponding convergence rate was 0.82, the gradient continuity exponent was 0.66, and the structural isomorphism preservation rate was 0.76.

[0139] The process of configuring the decoupling parameter mapping table is as follows: statistical analysis is performed on the sample data in the historical component image analysis log, and the optimal parameter combination that makes the number of connected branches of the global skeleton reach the theoretical expected value under different preset lighting conditions is extracted. The topological semantic decoupling degree interval corresponding to the optimal parameter combination is used as the threshold interval, thereby establishing the mapping relationship between the sample topological semantic decoupling degree and the sample energy field evolution convergence rate, the sample manifold gradient continuity index, and the sample structural isomorphism preservation rate.

[0140] The above relationships together form a decoupling degree parameter mapping table; after the topological semantic decoupling degree of the current image is calculated, the system performs interval matching in the mapping table; for example, if the current decoupling degree is 1.52, it falls into the middle interval, and the system selects a convergence rate of 0.70, a continuity index of 0.58, and an isomorphism preservation rate of 0.82.

[0141] If the current decoupling degree is 2.15, it means that the real edges and pseudo edges are clearly distinguished. In this case, a faster convergence rate can be adopted and the isomorphic verification threshold can be appropriately reduced to avoid excessive removal of real fine structures. In this way, the analysis process is no longer fixed, but dynamically switches according to the image state. To illustrate its role more intuitively, the adaptation morphology analysis process can be understood as a parameter combination package.

[0142] In this combination, the convergence rate determines the aggressiveness of the energy field diffusion iteration, the continuity index determines the tolerance for fractured regions during defect repair, and the isomorphism retention rate determines whether the local high-frequency structure needs to be more strictly consistent with the design topology. The three are interconnected, not isolated from each other. For example, in a low decoupling state, if only the convergence rate is reduced without increasing the isomorphism retention constraint, specular noise may still penetrate to the local anomaly identification stage. Therefore, unified matching is required.

[0143] As an anomaly handling mechanism, if the current decoupling degree happens to fall on the boundary of two intervals, for example, equal to 1.30, then the parameter combination of the more conservative interval can be used first to avoid the state switching too frequently; if the current decoupling degree is lower than the lower bound of the historical minimum interval, it indicates that an uncovered extreme image state has occurred, and at this time the lowest convergence rate and the highest isomorphism preservation rate are used as safe default values.

[0144] If the current decoupling degree is higher than the upper bound of the historical maximum interval, the highest convergence rate is adopted, but the parameter increase is limited to prevent over-reliance on a single high decoupling result. If the log sample is insufficient, for example, if the number of records in a certain decoupling interval is less than the preset number, the parameters of that interval will not be used directly in production, but will first enter the observation state, and will be officially used after the sample is sufficient.

[0145] During continuous inspection shifts of the blade leading edge, the decoupling degree is mostly above 1.9 in the morning due to the relatively clean equipment surface. The system automatically selects a high convergence rate combination to accelerate the extraction of the main skeleton. In the afternoon, the cutting fluid evaporates unevenly, and the reflection is enhanced. The decoupling degree of some images drops to about 1.2. The system then switches to a low convergence rate and high isomorphism retention rate combination to prioritize the suppression of false edge propagation and strengthen the verification of local structures, thereby maintaining the overall recognition stability.

[0146] The purpose of this mechanism is to solidify historical analysis experience into callable parameter mapping rules, thereby achieving adaptive analysis under different optical conditions and avoiding mismatch of fixed parameter schemes in complex working conditions.

[0147] Furthermore, the initial two-dimensional image manifold space is solved according to the adaptive morphology analysis process to obtain the global morphology skeleton and local structural abrupt change features, including:

[0148] According to the convergence rate of the target energy field evolution in the adaptation morphology analysis process, the number of iteration steps of the initial virtual energy field in the initial two-dimensional image manifold space is controlled to obtain the converged energy field distribution map.

[0149] Extract the incomplete features from the convergence energy field distribution map, and perform adaptive repair processing on the incomplete features according to the target manifold gradient continuity index in the adaptation morphology analysis process to extract the global morphology skeleton.

[0150] High-frequency abrupt change regions are extracted from the convergent energy field distribution map, and isomorphic mapping verification is performed on the high-frequency abrupt change regions according to the target structure isomorphic retention rate in the adaptation morphology analysis process to extract local structural abrupt change features.

[0151] This embodiment provides a shape calculation step based on adaptive parameter combination; specifically, after parameter matching is completed, if a uniform number of iteration steps and a uniform judgment standard are still used, the parameter combination cannot be effectively applied to skeleton extraction and anomaly identification; therefore, this embodiment applies the parameter combination to three actions: energy convergence control, defect repair, and local high-frequency structure verification.

[0152] Specifically, the system controls the number of iterations of the initial virtual energy field according to the convergence rate of the target energy field evolution to obtain a converged energy field distribution map. For ease of explanation, it is assumed that the current convergence rate is 0.70, which corresponds to allowing 5 rounds of diffusion iteration. If the average change of the entire map is below 0.02 in the 4th round, the system stops directly and outputs the results of the 4th round. If the system is still not stable after the 5th round, the results of the 5th round are output with an insufficient convergence mark. This ensures efficiency and avoids infinite iteration in noisy environments.

[0153] The system extracts incomplete features from the convergence energy field distribution map and performs adaptive repair by combining the target manifold gradient continuity index to obtain the global morphology skeleton. The so-called incomplete features refer to the interruption, weakening or local missingness of the main boundary that should be continuous in some areas; for example, the leading edge contour of the blade may have a 3 to 5 pixel gap under the shading of the specular highlight.

[0154] The system uses the energy flow direction, directional consistency, and surrounding gradient on both sides of the fracture as the basis for repair. When the continuity index requirement is met, a connecting line is added to restore the overall connectivity of the main skeleton. At the same time, the system also extracts high-frequency abrupt change regions from the convergent energy field distribution map and performs isomorphic mapping verification according to the target structure isomorphic retention rate to identify local structural abrupt change characteristics.

[0155] The so-called high-frequency mutation region refers to the region with drastic local energy changes and complex shape; it may be a real defect, such as chipped edges or scratches, or noise, such as liquid film reflection or stains; the system converts these regions into local topology maps and then compares them with the projected topology of the design model; if the match is high, it is retained as the real local structure; if the match is low, it is discarded as interference.

[0156] To further illustrate this process, the following local derivation example can be made: Assume that there is a fracture segment with a length of 4 nodes in the candidate region of the main framework in the convergence energy field. The energy flow angle between the two sides of the fracture is only 12 degrees, and the continuity index of the surrounding area is 0.61, which is higher than the target value of 0.58. It can be repaired and integrated into the main skeleton;

[0157] There is another high-frequency region Its local topology diagram contains 5 nodes and 2 branches, and its local structural similarity with the edge of the cooling holes in the design projection reaches 0.86, which is higher than the target retention rate of 0.82. Therefore, It was identified as a local structural mutation feature; if another high-frequency region If the similarity is only 0.49, it is judged as noise and removed;

[0158] As an anomaly handling mechanism, if there are too many breaks in the main boundary of the converged energy field distribution map, such as the number of broken segments exceeding the preset upper limit, the system will not perform large-scale automatic repair, but will only output reliable main segments to prevent noise from being mistakenly connected into a skeleton; if the area of ​​a certain high-frequency region is too small, such as containing only 1 to 2 nodes, even if the topological similarity is high, it will be cached as a weak suspicious point and wait for verification in subsequent frames; if the number of iterations caused by the convergence rate is too small, making the energy field obviously unstable, the image will be resampled first, rather than directly relying on the repair module for compensation.

[0159] In the blade leading edge detection, the convergence energy field has basically stabilized after 4 rounds of iteration, but a local fracture occurs in the middle of the leading edge due to the influence of mirror reflection. The system automatically fills in a continuous curve based on the energy flow direction at both ends of the fracture to form a complete leading edge skeleton. At the same time, a high-frequency abrupt change region is identified at the blade root. After comparing it with the local topology of the design projection, it is found that its structure does not conform to the normal chamfer relationship. Therefore, it is marked as a local edge breakage feature and enters the subsequent process alarm link.

[0160] The purpose of this step is to truly implement the adaptive parameters into executable calculation actions, thereby simultaneously ensuring the continuous recovery of the global skeleton and the reliable screening of local anomalies.

[0161] Furthermore, incomplete features are extracted from the convergence energy field distribution map, and adaptive repair processing is performed on the incomplete features according to the target manifold gradient continuity index in the adaptive morphology analysis process to extract the global morphology skeleton, including:

[0162] Calculate the manifold gradient values ​​for each local region in the convergent energy field distribution map;

[0163] Regions where the gradient value of the local manifold is less than the gradient continuity index of the target manifold are identified as gradient break regions, and these gradient break regions are used as incomplete features.

[0164] By utilizing the energy flow direction of the surrounding area of ​​the incomplete feature, path interpolation is performed on the incomplete feature to generate continuous morphological feature lines; the continuous morphological feature lines are then topologically connected to construct a global morphological skeleton.

[0165] This embodiment provides a skeleton repair step for incomplete features. Specifically, in the aforementioned solution process, if only linear connections are made based on existing high-energy regions, it is easy to misjudge neighboring noise as the main skeleton. Conversely, if no interpolation repair mechanism is configured, the real geometric boundary will continue to exhibit a broken state under optical occlusion or reflective interference. Therefore, this embodiment uses the manifold gradient continuity index to constrain the repair triggering conditions and only performs interpolation deduction on broken regions that satisfy continuous morphological logic.

[0166] Specifically, the system calculates the manifold gradient values ​​of each local region in the convergent energy field distribution map; the image can be divided into several local windows, and the magnitude of energy variation along the principal direction is calculated for each window; for ease of explanation, it is assumed that there are three adjacent windows near a certain principal boundary. , , The gradient values ​​are 0.64, 0.41, and 0.66, respectively, while the gradient continuity exponent of the target manifold is 0.58; therefore... If the value is below the target value, it is identified as a gradient break region, indicating that the missing features may be caused by highlight occlusion, grayscale distortion or local contamination.

[0167] After locating the incomplete features, the system uses the energy flow direction of the surrounding area to perform path interpolation and deduction; assuming The left main line direction angle is 32 degrees, and the right main line direction angle is 35 degrees, with the directions on both sides being basically the same; at the same time, the Euclidean distance between the left and right endpoints is only 4 nodes, so the system can insert a smooth transition path between the two.

[0168] If represented in a simplified way, the original main line might be node A—node B, node C—node D, with a missing segment in the middle. After interpolation, it can be filled into node A—node B—node X—node Y—node C—node D. Among them, the selection of nodes X and Y depends not only on their geometric positions but also on the consistency of the surrounding energy flow direction. Nodes that are consistent with the mainstream direction on both sides are preferred.

[0169] After multiple missing features are repaired, the system will connect the continuous morphological feature lines topologically to construct a global morphological skeleton. The topological connection here is not a simple splicing, but requires that adjacent line segments form a reasonable connected branch structure after connection. For example, the main skeleton of the leading edge of the blade should form a main branch, and the cooling hole boundary can form several closed branches. If the repair result of a certain segment results in an abnormal loop or a suspended branch, the repair result of that segment will be rolled back or downweighted.

[0170] As an anomaly handling mechanism, if the length of a gradient fracture region exceeds the preset upper limit, such as exceeding 20% ​​of the expected length of the main boundary, it indicates that the missing value is too large. The system will not perform automatic interpolation, but will only retain the fracture state and request re-acquisition. If the angle between the energy flow directions at both ends of the fracture is too large, such as greater than 45 degrees, it indicates that the two segments may not belong to the same topographic line. In this case, forced connection is prohibited.

[0171] If multiple candidate paths exist in the surrounding area, the system will prioritize the one with higher energy density and smoother curvature change; the remaining candidate paths will be considered as alternatives but not included in the main framework; if the number of connected branches in the framework increases abnormally after repair, it indicates that a false connection has occurred, and the system will automatically cancel the most recent interpolation.

[0172] At the middle of the leading edge of the blade, the convergent energy field shows that the main contour line is cut off by a reflective band, and the gradient value of the fracture zone drops to 0.41. The system detects that the two ends of the fracture are in the same direction and are only 4 nodes apart. The system generates a smooth interpolation transition path according to the surrounding energy flow direction to achieve topological connectivity of the leading edge contour. After interpolation repair, the global skeleton converges from the original 3 fragment branches to 1 main branch and 2 closed auxiliary branches, which meets the topological requirements of the preset 3D design model of the target component.

[0173] The purpose of this step is to restore the continuous topographic lines that have been damaged by optical factors under controlled conditions, thereby improving the integrity of the global skeleton and avoiding misconnection problems caused by unconstrained line filling.

[0174] Furthermore, high-frequency abrupt change regions are extracted from the convergence energy field distribution map, and isomorphic mapping verification is performed on these regions according to the target structural isomorphism preservation rate in the adaptation morphology analysis process. Local structural abrupt change features are then extracted, including:

[0175] In the convergent energy field distribution map, regions with energy distribution variance greater than a preset variance threshold are extracted as high-frequency mutation regions.

[0176] Extract a two-dimensional topological graph of high-frequency mutation regions, where the two-dimensional topological graph includes nodes and connected components;

[0177] Obtain a two-dimensional projection topology of the target component's preset three-dimensional design model from the current visual perspective of the visual acquisition device;

[0178] The topological similarity of nodes and connected components between a two-dimensional topological relation graph and a two-dimensional projected topological graph is used as the graph theory isomorphism matching degree.

[0179] If the graph theory isomorphism matching degree is greater than or equal to the target structure isomorphism retention rate, then the high-frequency mutation region is identified as a local structural mutation feature.

[0180] If the graph theory isomorphism matching degree is less than the target structure isomorphism retention rate, then the high-frequency abrupt change region is identified as optical interference noise and is removed.

[0181] This embodiment provides a local structural mutation extraction step based on isomorphic mapping verification. Specifically, after the global skeleton repair is completed, local high-frequency areas may still be mixed with real defects and random interference. If no further verification is performed, cutting fluid residue, local stains and noise may be mistakenly regarded as structural anomalies. Therefore, this embodiment improves the credibility of local anomalies by isomorphic comparison between the two-dimensional topology diagram and the projected topology diagram of the design model.

[0182] Specifically, the system extracts regions from the converged energy field distribution map where the energy distribution variance is greater than a preset variance threshold, designating them as high-frequency abrupt change regions. Higher variance indicates larger energy fluctuations and more complex edges within the region. For ease of explanation, we assume the threshold is set to 0.08 for a given local region. The energy variance is 0.12. It is 0.10. If it is 0.04, then , Enter the candidate set, This region was ignored as a stable area.

[0183] The system extracts a two-dimensional topological graph for each candidate region; this graph consists of nodes and connected components; nodes can be local extrema, inflection points, intersections, or key points of closed loops, and connected components reflect the connection relationships between nodes; assuming... It is transformed into a local graph containing 4 nodes and 2 branches. ; It is transformed into a local graph containing 6 nodes and 4 branches. ;

[0184] At the same time, the system acquires a two-dimensional projection topology of the target component's preset three-dimensional design model from the current visual perspective; for example, for the chamfered area at the root of a blade, a local standard diagram can be projected from the design model. It contains 4 nodes and 2 branches; the system calculates the topological similarity between the two-dimensional topological relationship graph and the two-dimensional projected topological graph as the graph theory isomorphism matching degree;

[0185] For ease of explanation, scores can be given for each of the following four aspects: node number matching, branch number matching, connection order consistency, and closed loop existence. Then, a weighted average is calculated. For each matching score, the system presets a corresponding weight coefficient that sums to 1. The node number matching score, branch number matching score, connection order consistency matching score, and closed loop existence matching score are multiplied by their respective weight coefficients and summed to calculate the final graph theory isomorphism matching degree.

[0186] Assumption and The number of nodes, branches, and connection order are all relatively consistent, with only a slight deviation at one corner position, so its matching degree is 0.85; Due to excessive branches and chaotic connections, only 0.47 is obtained; if the target structure isomorphism retention rate is 0.80, then... The identification of a local structural mutation feature indicates that although there is a mutation, it still conforms to the true structural logic at that location, and may be a normal chamfer boundary or an explainable process change. If the value is below the threshold, it is judged as optical interference noise and discarded.

[0187] It should be noted that identifying a local structural mutation feature does not equate to identifying it as an undesirable defect. Rather, it means that it is a local anomaly or change related to the real structure and is worthy of being included in the subsequent quality assessment process. The regions that are excluded will no longer be included in the morphology output in this round of analysis.

[0188] As an anomaly handling mechanism, if the current visual perspective shifts, causing inconsistencies between the pose of the design projection and the actual photograph, pose correction can be performed before calculating the isomorphic matching degree. If a candidate region has too few nodes, such as fewer than 3 nodes, its topological information is insufficient, and isomorphic comparison is not performed directly; instead, it is retained as a low-confidence observation point. If multiple candidate regions overlap spatially, region merging can be performed first, and then the topological map can be extracted uniformly to avoid repeated judgments on the same local structure. If the design model lacks a corresponding local projection template, isomorphic verification is not performed on that region; instead, a conservative strategy is adopted to mark it as a region to be re-examined.

[0189] In the blade root detection, the system extracted two high-frequency regions. The first region corresponds to a local abrupt change near the edge of the cooling hole. Its topology map is highly similar to the hole boundary at this location in the design projection, so it is retained as a true local structural feature. The second region is located in the surface cutting fluid residue area. Although the variance is high, its topology map shows a multi-branched and scattered structure, which is inconsistent with any design projection, so it is discarded as optical noise.

[0190] The purpose of this step is to add a layer of structural logic verification to the high-frequency complex region, so as to retain the mutations that are consistent with the actual part structure and exclude noise without topological basis from the abnormal candidates.

[0191] Furthermore, control instructions are generated based on the global topographic skeleton and local structural abrupt changes, and the visual acquisition parameters of the visual acquisition device are controlled in real time by a programmable logic controller, including:

[0192] The integrity score of the global topography skeleton is calculated based on the number of connected branches of the global topography skeleton, and the signal-to-noise ratio of the local structural mutation features is calculated based on the ratio of the energy density of the local structural mutation features to the preset background energy density extracted from the background region.

[0193] Among them, the visual acquisition parameters include the lighting conditions;

[0194] If the integrity score is less than the preset integrity threshold or the signal-to-noise ratio is less than the preset signal-to-noise ratio threshold, a light source adjustment command is generated and sent to the light source controller included in the vision acquisition device. The light source controller then adjusts the illumination status of the target component in real time.

[0195] If the completeness score is greater than or equal to the preset completeness threshold and the signal-to-noise ratio is greater than or equal to the preset signal-to-noise ratio threshold, then keep the current visual acquisition parameters unchanged;

[0196] This embodiment provides a closed-loop control step for adjusting acquisition parameters based on the analysis results. Specifically, if all the aforementioned analyses only remain at the image post-processing level without feeding the results back to the acquisition device, the same reflection and occlusion problems will still be repeatedly encountered under complex working conditions, making it difficult to form a stable production capacity. Therefore, this embodiment transforms the global skeleton and local structural mutation features into executable dimming criteria to achieve real-time acquisition optimization.

[0197] Specifically, the system first calculates the integrity score based on the number of connected branches in the global topography skeleton. Typically, the target component should present a finite and stable main branch structure from the current perspective. If the fragmentation is severe, it indicates that the current acquisition conditions are insufficient. For ease of explanation, it is assumed that the leading edge of the current blade should theoretically form one main skeleton and two auxiliary closed branches.

[0198] If the actual extraction result is 1 main branch plus 5 scattered short branches, then the completeness score is directly calculated according to the ratio of the total number of theoretically connected branches of the target component to the total number of actual connected branches of the global topography skeleton. The specific mathematical expression is as follows: Completeness rating =

[0199] Among them, the total number of theoretical connected branches of the target component is a dimensionless parameter, the value of which is obtained by statistically analyzing the connected branches in the two-dimensional projection topology of the target component's preset three-dimensional design model in the current visual perspective; the total number of actual connected branches of the global topology skeleton is the number of connected branches actually contained in the global topology skeleton constructed after adaptive repair processing.

[0200] For example, if the theoretical score is 3 but the actual score is 6, the completeness score is 0.50; if the preset completeness threshold is 0.75, it means that the skeleton is still incomplete; at the same time, the system calculates the signal-to-noise ratio based on the ratio of the energy density of local structural mutation features to the preset background energy density.

[0201] The preset background energy density is obtained in the following way: In the initial two-dimensional image manifold space, the connected regions with pixel gray values ​​lower than the preset background gray value threshold are selected as background regions, and the average energy density of the background regions in the converged energy field distribution map is calculated and used as the preset background energy density.

[0202] Assuming the average energy density of a local edge collapse candidate region is 0.72 and the background energy density is 0.30, the calculated signal-to-noise ratio (SNR) is 2.4. If the preset SNR threshold is 2.0, the feature has good discernibility. If the average energy density of another abnormal region is only 0.48 and the background is 0.32, the SNR is only 1.5, which is insufficient to support stable interpretation.

[0203] After completing the two assessments, the system determines whether the lighting conditions need to be adjusted. If the integrity score is lower than the integrity threshold or the signal-to-noise ratio is lower than the signal-to-noise ratio threshold, a light source adjustment instruction is generated. This instruction can include parameters such as brightness increase or decrease, incident angle switching, polarization direction rotation, and local supplementary lighting activation. For example, when the skeleton is severely fragmented, the system prioritizes reducing the positive strong light and increasing the side supplementary light ratio to reduce specular saturation.

[0204] When the signal-to-noise ratio is insufficient in a local abnormality, the illumination range can be reduced or the polarization direction can be adjusted to improve the feature contrast. The generated adjustment command is sent to the light source controller, which adjusts the illumination status of the visual acquisition device in real time. If both indicators are not lower than the threshold, the current acquisition parameters are kept unchanged to avoid frequent disturbances that cause system oscillation.

[0205] To illustrate the specific closed-loop process, the following deduction example can be made: After a certain round of analysis, the completeness score is 0.62 and the signal-to-noise ratio is 1.8, both of which fail to meet the standards; therefore, the system outputs an adjustment command. The brightness of the ring light source was reduced from 70% to 58%, the brightness of the side strip light was increased from 40% to 55%, and the polarizer angle was rotated by 10 degrees. After the next round of acquisition, the integrity score improved to 0.81 and the signal-to-noise ratio improved to 2.6. At this time, the system judgment condition was met, and the system stopped adjusting the brightness and instead maintained the current acquisition state to enter the stable detection stage.

[0206] As an anomaly handling mechanism, if the integrity score does not improve after multiple rounds of dimming, it indicates that the problem may not be with the lighting, but with lens damage, component misalignment, or mechanical vibration. In this case, the system stops dimming and issues an equipment check prompt. If the integrity score meets the standard but the signal-to-noise ratio fluctuates drastically, a sliding average method can be used to smooth the signal-to-noise ratio over the last three rounds before making a decision, avoiding repeated adjustments due to occasional noise. If the generated adjustment command exceeds the allowable range of the light source controller, such as brightness being lower than the minimum value or angle exceeding the limit, the controller executes the truncated safety value and sends the execution status back to the edge computing unit.

[0207] During the online inspection of the leading edge of the blade, a batch of blades had a lot of residual cutting fluid on the surface, which caused multiple bifurcations in the main skeleton of the initial acquisition and a low signal-to-noise ratio of the local chipping features. The system automatically reduced the front illumination and increased the oblique side illumination based on the scoring results. In the next acquisition cycle, the main line of the leading edge was obviously continuous, the energy contrast between the abnormal area at the root and the background increased, and a stable skeleton and defect features were output for the quality system to use.

[0208] The purpose of this step is to transform the morphology analysis results into real-time feedback on the acquisition environment, thereby establishing a closed-loop optimization link between visual acquisition and structural analysis, and improving the stability of continuous detection under complex surface conditions.

[0209] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A vision-based method for analyzing the morphology of precision components from images, characterized in that, The methods include: The target component's two-dimensional image sequence is acquired using a visual acquisition device according to the initial visual acquisition parameters, and pixel grayscale distribution information and spatial neighborhood topology are extracted. Based on the pixel grayscale distribution information and spatial neighborhood topology, a two-dimensional manifold space mapping is performed to obtain the initial two-dimensional image manifold space; The initial two-dimensional image manifold space is initialized with a virtual energy field distribution matrix based on the pixel grayscale distribution information, and the topological semantic decoupling degree representing edge differences is extracted according to the dynamic evolution state of the initial virtual energy field in the space. Based on the aforementioned topological semantic decoupling degree, the target energy field evolution convergence rate, the target manifold gradient continuity index, and the target structure isomorphism preservation rate are dynamically configured to generate an adaptive morphology analysis process. The initial two-dimensional image manifold space is solved according to the adaptation morphology analysis process to obtain the global morphology skeleton and local structural mutation features; Based on the global topography skeleton and local structural mutation features, control instructions are generated, and the visual acquisition parameters of the visual acquisition device are controlled in real time through a programmable logic controller. The virtual energy field is a distribution matrix constructed by combining the initial energy distribution values ​​of the boundary nodes with the pixel grayscale distribution information; the target energy field evolution convergence rate is used to control the number of iteration steps of the initial virtual energy field in the initial two-dimensional image manifold space; the target manifold gradient continuity index is used to constrain the adaptive repair of local incomplete features; and the target structural isomorphism preservation rate is used to constrain the isomorphic mapping verification of local high-frequency abrupt change regions.

2. The vision-based precision component morphology image analysis method according to claim 1, characterized in that, Based on the pixel grayscale distribution information and spatial neighborhood topology, a two-dimensional manifold space mapping is performed to obtain the initial two-dimensional image manifold space, including: Using the pixel grayscale distribution information and spatial neighborhood topology as input variables, a two-dimensional image manifold space mapping function is constructed. The initial two-dimensional image manifold space is generated by performing matrix transformation on the two-dimensional image sequence of the target component through the two-dimensional image manifold space mapping function.

3. The vision-based precision component morphology image analysis method according to claim 1, characterized in that, Based on the pixel grayscale distribution information, a virtual energy field distribution matrix is ​​initialized in the initial two-dimensional image manifold space. The topological semantic decoupling degree representing edge differences is extracted according to the dynamic evolution state of the initial virtual energy field in the space, including: The mapping nodes corresponding to the preset image edge regions obtained based on the initial high gradient detection are extracted from the initial two-dimensional image manifold space as boundary nodes. The boundary nodes are assigned preset initial energy distribution values, and a virtual energy field distribution matrix is ​​constructed by combining the pixel grayscale distribution information as the initial virtual energy field. Calculate the manifold curvature of each mapping node in the initial two-dimensional image manifold space; guide the initial virtual energy field to perform numerical diffusion iteration based on the manifold curvature of the initial two-dimensional image manifold space to obtain the steady-state distribution matrix of the virtual energy field; The virtual energy field steady-state distribution matrix is ​​processed to separate the real geometric shape edges from the optical pseudo edges, and the energy distribution difference between the real geometric shape edges and the optical pseudo edges is calculated as the topological semantic decoupling degree.

4. The vision-based precision component morphology image analysis method according to claim 3, characterized in that, The virtual energy field steady-state distribution matrix is ​​processed to separate the real geometric edges from the optical pseudo-edges. The energy distribution difference between the real geometric edges and the optical pseudo-edges is calculated as the topological semantic decoupling degree, including: Extract the energy convergence path and energy diversion node from the steady-state distribution matrix of the virtual energy field; The edge continuity score of the edge corresponding to the energy convergence path is calculated based on the ratio of the length of the energy convergence path to the number of energy diversion nodes; If the edge continuity score is greater than a preset continuity threshold, then the edge corresponding to the energy convergence path is determined to be the true geometric shape edge; If the edge continuity score is less than or equal to a preset continuity threshold, then the edge corresponding to the energy convergence path is determined to be the optical pseudo edge. The ratio of the energy density of the real geometric edge to the energy density of the optical pseudo-edge is calculated as the topological semantic decoupling degree.

5. The vision-based precision component morphology image analysis method according to claim 1, characterized in that, Based on the aforementioned topological semantic decoupling degree, the target energy field evolution convergence rate, the target manifold gradient continuity index, and the target structural isomorphism preservation rate are dynamically configured to generate an adaptive morphology analysis process, including: Obtain historical component image analysis logs, and configure a decoupling parameter mapping table based on the historical component image analysis logs. The decoupling parameter mapping table contains at least one mapping relationship, and each mapping relationship includes a sample topological semantic decoupling threshold range and the corresponding sample energy field evolution convergence rate, sample manifold gradient continuity index, and sample structural isomorphism preservation rate. Using the decoupling parameter mapping table, the corresponding target energy field evolution convergence rate, target manifold gradient continuity index, and target structure isomorphism preservation rate are obtained according to the topological semantic decoupling degree matching. The target energy field evolution convergence rate, the target manifold gradient continuity index, and the target structure isomorphism preservation rate are combined as the adaptation morphology analysis process.

6. The vision-based precision component morphology image analysis method according to claim 1, characterized in that, The initial two-dimensional image manifold space is solved according to the adapted morphology analysis process to obtain the global morphology skeleton and local structural abrupt features, including: According to the convergence rate of the target energy field evolution in the adaptation morphology analysis process, the number of iteration steps of the initial virtual energy field in the initial two-dimensional image manifold space is controlled to obtain the converged energy field distribution map. Extract the incomplete features from the convergent energy field distribution map, and perform adaptive repair processing on the incomplete features according to the target manifold gradient continuity index in the adaptive morphology analysis process to extract the global morphology skeleton. High-frequency abrupt change regions are extracted from the convergent energy field distribution map, and isomorphic mapping verification is performed on the high-frequency abrupt change regions according to the target structure isomorphic retention rate in the adaptation morphology analysis process to extract the local structural abrupt change features.

7. The vision-based precision component morphology image analysis method according to claim 6, characterized in that, Extracting incomplete features from the convergent energy field distribution map, and adaptively repairing the incomplete features according to the target manifold gradient continuity index in the adaptive topography analysis process, to extract the global topography skeleton, including: Calculate the manifold gradient values ​​for each local region in the convergent energy field distribution map; The region where the manifold gradient value is less than the target manifold gradient continuity index is identified as the gradient break region, and the gradient break region is identified as the incomplete feature. By utilizing the energy flow direction of the area surrounding the incomplete feature, path interpolation is performed on the incomplete feature to generate continuous morphological feature lines. The global shape skeleton is constructed by topologically connecting the continuous shape feature lines.

8. The vision-based precision component morphology image analysis method according to claim 6, characterized in that, High-frequency abrupt change regions are extracted from the convergent energy field distribution map, and isomorphic mapping verification is performed on the high-frequency abrupt change regions according to the target structural isomorphism retention rate in the adaptation morphology analysis process. The local structural abrupt change features are then extracted, including: In the convergent energy field distribution map, regions with energy distribution variance greater than a preset variance threshold are extracted as the high-frequency mutation regions. Extract a two-dimensional topological graph of the high-frequency mutation region, wherein the two-dimensional topological graph includes nodes and connected components; Obtain a two-dimensional projection topology of the preset three-dimensional design model of the target component from the current visual perspective of the visual acquisition device; The topological similarity of nodes and connected components between the two-dimensional topological relationship graph and the two-dimensional projected topological graph is calculated as the graph theory isomorphism matching degree. If the graph theory isomorphism matching degree is greater than or equal to the target structure isomorphism retention rate, then the high-frequency mutation region is identified as the local structural mutation feature; If the graph theory isomorphism matching degree is less than the target structure isomorphism retention rate, then the high-frequency abrupt change region is identified as optical interference noise and is removed.

9. The vision-based precision component morphology image analysis method according to claim 1, characterized in that, Based on the global topography skeleton and local structural abrupt changes, control instructions are generated, and the visual acquisition parameters of the visual acquisition device are adjusted in real time through a programmable logic controller, including: The integrity score of the global topography skeleton is calculated based on the number of connected branches of the global topography skeleton, and the signal-to-noise ratio of the local structural abrupt feature is calculated based on the ratio of the energy density of the local structural abrupt feature to the preset background energy density extracted from the background region. The visual acquisition parameters include illumination status; If the integrity score is less than a preset integrity threshold or the signal-to-noise ratio is less than a preset signal-to-noise ratio threshold, a light source adjustment command is generated and sent to the light source controller included in the vision acquisition device. The light source controller then adjusts the illumination state of the target component in real time. If the completeness score is greater than or equal to a preset completeness threshold and the signal-to-noise ratio is greater than or equal to a preset signal-to-noise ratio threshold, then the current visual acquisition parameters remain unchanged.

Images