Method for extracting strong reflective v-shaped weld features based on semantic topological segmentation and cross-polarity

Abstract

This invention discloses a method for extracting features of highly reflective V-shaped welds based on semantic topological segmentation and cross-product polarity. Addressing the issues of strong secondary reflection and easily confused topological order of feature points in V-shaped grooves, the method extracts key points using an attitude network and adaptively sorts them according to image coordinates to construct a skeleton. It then defines the region of interest and extracts sub-pixel scattered points within a controlled Gaussian gravitational field. Robust fitting of the segmented point set and solving for analytical intersections are performed, followed by circular domain cross-validation to reconstruct a noise-free absolute geometric skeleton. Directed vectors are constructed sequentially, and the root points of the weld are adaptively stripped using two-dimensional cross-product polarity. Finally, a spatiotemporal Kalman filter mechanism is introduced to combat prediction jitter and spatter occlusion, and the groove cross-sectional area is calculated using outer product modulus quantization. This invention completely eliminates morphological distortion, achieving high-fidelity extraction and smooth, stable visual servo control under extreme conditions.

Description

Technical Field

[0001] This invention relates to the fields of image processing and intelligent welding vision technology, and in particular to a method for high-fidelity feature extraction and cross-sectional quantization of structured light weld morphology under strong reflection and strong spatter interference. Background Technology

[0002] In industrial robot vision weld seam tracking, compared with ordinary open fillet welds, the physical structure of V-groove welds is prone to forming a severe "optical cavity effect".

[0003] When the laser is projected into the V-groove, severe secondary or even multiple specular reflections occur between the side walls and the bottom, causing the actual weld root to be completely submerged by a large, false, bright halo.

[0004] Existing methods based on the traditional gray-scale centroid method are local greedy algorithms. They are not only easily affected by the secondary reflection at the bottom of the V-shaped bevel, causing "corner distortion" and resulting in the loss of the true acute angle shape of the root features, but also cannot separate the real light stripe from the highlight noise.

[0005] Pure deep learning end-to-end prediction methods lack rigorous geometric and physical constraints, making it impossible to meet sub-pixel level precision measurement requirements. Furthermore, the inherent pixel-level jitter in network predictions can easily cause high-frequency oscillations at the robotic arm's end effector.

[0006] Furthermore, existing vision algorithms rely heavily on the absolute horizontal mounting position of the camera when separating the edge points and the bottom root point of a V-groove. If the camera tilts mechanically or the workpiece deflects in an industrial setting, the feature classification logic based on a single absolute coordinate comparison will directly fail. Summary of the Invention

[0007] The technical problem to be solved by this invention is to provide a method for extracting features of highly reflective V-shaped welds based on semantic topological segmentation and cross product polarity, so as to solve the problems of low extraction accuracy, easy distortion, extreme dependence on absolute coordinate system, and poor anti-jitter interference ability of existing technologies under severe secondary reflection of V-shaped grooves.

[0008] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: a method for extracting features of highly reflective V-shaped welds based on semantic topological segmentation and cross product polarity, comprising the following steps:

[0009] The V-shaped weld image is acquired and input into the attitude estimation network to extract 5 topological key points. The range is calculated and sorted in an adaptive coordinate system. The local region of interest is defined by connecting the key points to enhance the image.

[0010] A controlled Gaussian gravitational field is constructed using the ordered connection as the mathematical expectation, and the Gaussian weighted gray-level centroid is calculated to extract local sub-pixel scattered points.

[0011] Robust line fitting is performed independently on the segmented point set to obtain a main arm straight line model that is resistant to secondary reflection distortion.

[0012] The intersection point is solved analytically by combining the equations of adjacent main arm lines, and the original key points are used for adaptive circular domain cross-validation.

[0013] The high-fidelity intersections are connected in sorted order to output a noise-free absolute geometric skeleton light stripe image. Then, the unique root feature point of the weld is extracted by adaptively removing the sign polarity of the two-dimensional vector outer product. Spatiotemporal Kalman filtering is introduced to combat pixel jitter and spatter occlusion. Finally, the two-dimensional apparent cross-sectional area of ​​the V-groove is calculated by using vector outer product modulus quantization to achieve intelligent control output.

[0014] The beneficial effects of the present invention are as follows: First, the present invention eliminates the "cutting angle distortion" and optical cavity effect interference at the bottom of the V-shaped bevel, transforms the connection of points into line into segmented intersection, filters out the background and reconstructs an absolutely pure mathematical skeleton light stripe image.

[0015] Second, this invention introduces the right-hand screw rule into visual feature recognition, which can accurately isolate the unique root feature point by relying solely on the polarity change of the cross product of continuous paths, thus realizing the physical and semantic decoupling of the root point and the wall point.

[0016] Third, this invention cleverly utilizes the same topological vector to not only solve the camera tilt interference, but also provides a smooth tracking trajectory and accurate adaptive wire feeding section basis for the welding expert system through spatiotemporal filtering and cross product area derivation, without the need for complex three-dimensional calibration. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of skeleton generation and local region of interest constraint in this invention;

[0018] Figure 2 This is a schematic diagram of the controlled Gaussian extraction and piecewise line reconstruction of the present invention;

[0019] Figure 3 This is a schematic diagram of the adaptive circular domain cross-validation of the present invention;

[0020] Figure 4 This is an image showing the noise-free absolute geometric skeleton light stripe extraction effect of the present invention;

[0021] Figure 5 This is a schematic diagram of the ultimate vector cross product polarity and V-shaped root adaptive recognition principle of the present invention;

[0022] Figure 6 This is a comparison chart of the anti-jitter smoothing tracking trajectory after the introduction of spatiotemporal Kalman filtering in this invention. Detailed Implementation

[0023] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments.

[0024] It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0025] This embodiment takes a structured light image of a V-groove weld that is severely affected by secondary reflection interference as an example. The specific implementation steps are as follows:

[0026] Step 1: Obtain an image of the V-shaped weld seam containing structured light stripes. Input this image into a pre-trained pose estimation network to extract five topological key points of the continuous light stripes. ,in .

[0027] Calculate the range of all keypoints in the horizontal and vertical directions of the image to determine the spatial distribution of the light stripes. If the X-axis range is greater than the Y-axis range, sort the keypoints monotonically from smallest to largest X-axis coordinate; otherwise, sort them by Y-axis coordinate, thus completely resolving the topological chaos caused by camera tilt. Based on the sorted keypoints, define the physical region of interest (ROI) mask and perform CLAHE local contrast enhancement on the original image to obtain the enhanced image. .

[0028] Step 2: Within the corridor formed by connecting adjacent key points in the sorting process, connect the points according to the theoretical x-coordinates. Construct a controlled Gaussian gravitational field.

[0029] Local sub-pixel scattered points are extracted using the Gaussian weighted gray-level centroid method, where the Gaussian weighted centroid is... The calculation formulas are as follows:

[0030]

[0031] In the formula, The x-coordinate of the image pixels; To enhance the local image in coordinates The pixel grayscale value at that location; Connect the lines for sorting. The theoretical mathematical expectation of the line, x-axis; This is the set Gaussian tolerance coefficient.

[0032] This formula, combined with ROI masking, attenuates the weight of secondary spatter noise points that deviate significantly from the theoretical centerline within the V-groove optical cavity to zero, thereby extracting a clean local sub-pixel scattered point set.

[0033] Step 3: For the above independent line segment sets, use a robust algorithm with Huber loss for initial fitting, and remove residual halo ionized noise by setting a tolerance threshold.

[0034] Then, the least squares method was used for high-precision reconstruction to obtain the equation of the noise-resistant main arm straight line, which is expressed as follows:

[0035]

[0036] Step 4: Solve the equations of the adjacent main arm lines simultaneously, and obtain the mathematical intersection point analytically. .

[0037] Predict key points using the corresponding native network Set the effective verification radius with the center as the center. Construct an adaptive verification circle domain.

[0038] Calculate the Euclidean distance between the mathematical intersection point and the corresponding center of the circle. The formula for verifying the circular domain is independent as follows:

[0039]

[0040] In the formula, The coordinates of the mathematical intersection point obtained by analytically solving the equations of the main arm straight lines; The pose estimation network predicts the coordinates of the corresponding native topological key points. This is the set effective verification radius threshold. If the above formula is satisfied, it is determined that the network prediction and geometric reconstruction are consistent, and the intersection point is confirmed as a high-fidelity analytical intersection point.

[0041] Step 5: Geometric skeleton reconstruction and root stripping using cross product polarity adaptive method. Connect the first and last key points to the verified high-fidelity analytical intersection points in sequence to construct a continuous sequence of directed spatial feature trajectories.

[0042] After acquiring extremely high-precision feature path points, the algorithm connects these path points on a pure black background and outputs the result, directly generating an absolute geometric skeleton light stripe image (e.g., [image of a black background]) that filters out all impurities, splashes, and halos. Figure 4 As shown in the figure, this greatly improves the anti-interference capability of subsequent processing.

[0043] Then, iterate through the internal intersections and construct the inflow vectors accordingly. With outflow vector Using the right-hand screw rule, we can calculate the Z-axis component of the two-dimensional cross product of adjacent vectors. Its calculation formula is as follows:

[0044]

[0045] In the formula, For the inflow vector, This is the outflow vector; Inflow vectors Coordinate components on the X and Y axes; These are the outflow vectors. The coordinate components on the X and Y axes. The algorithm counts and finds the outer product. The characteristic point where the polarity of the sign is reversed (inward).

[0046] Because the physical structure of a V-groove dictates that only the bottom geometric inflection point is in the opposite direction to the bending direction of the two side edges, the algorithm determines that the only feature point where polarity is reversed is the absolute weld root feature point.

[0047] Spatiotemporal continuous tracking and bevel section quantization to resist geometric jitter. In real-world continuous industrial welding scenarios, the inherent pixel-level prediction jitter of neural networks, as well as occasional large-particle spatter occlusion, can cause high-frequency oscillations or brief loss of extracted feature points over time. To ensure absolute smoothness and continuity of feature extraction in the time dimension, this invention introduces a spatiotemporal Kalman filtering mechanism.

[0048] Constructing system state vectors for feature points at the weld root ,in The root pixel coordinates, This represents the velocity of the root in pixel coordinates. When the video stream enters the... At frame rate, the state prediction is first performed using the optimal estimate from the previous frame, and the state prediction equations are independent as follows:

[0049]

[0050] In the formula, For the first The state prediction vector of the frame; For the first The optimal state estimation vector of the frame; This is the state transition matrix constructed based on the uniform motion model.

[0051] If the current number is If the frame image is obscured by splashing, causing extraction failure, or if there are drastic pixel jumps in feature points, the algorithm adaptively degrades, directly converting the state prediction vector. The coordinate components in the image are used as the root output position of the current frame to achieve anti-interference feedforward smoothing compensation.

[0052] If the current number is The frame image is normal, and the coordinate vectors of the root feature points actually observed have been successfully extracted. Then, the tracker is optimally corrected using the Kalman state update equation, and the calculation formula is as follows:

[0053]

[0054] In the formula, For the first time after merging observation data Frame-optimal state estimation vector; This is a dynamically calculated Kalman gain matrix used to balance the confidence weights of the prediction model and the actual observations. This is the observation matrix.

[0055] Furthermore, to provide a geometric basis for subsequent welding process parameters, this invention performs feature quantization on the morphology of the V-groove. Using the optimal root feature point from the tracking output as the spatial geometric origin, a spatial direction vector pointing towards the feature point on the left side wall edge is constructed. Spatial direction vector pointing to the feature point on the right side wall edge .

[0056] Based on spatial vector subtraction and the geometric norm theorem, calculate the two-dimensional apparent opening width of the current V-groove. The formulas are independent as follows:

[0057]

[0058] Based on the geometric and physical meaning of the cross product of two vectors (the magnitude of the cross product of two vectors is equal to the area of ​​the parallelogram they span), the two-dimensional apparent cross-sectional filling area of ​​the V-groove is derived using the calculated vector data. The formulas are independent as follows:

[0059]

[0060] In the above two formulas, the symbols This represents the magnitude of the vector (Euclidean norm). Ultimately, the control system uses the smooth and continuous weld root trajectory as the target path for welding torch position tracking; simultaneously, it uses the dynamically quantized opening width... With cross-sectional filling area As feedback input to the welding expert system, it enables smooth tracking under jitter interference and adaptive intelligent adjustment of wire feed.

[0061] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A strong reflective light V-shaped weld seam feature extraction method based on semantic topological segmentation and cross-multiplying polarity, characterized in that, Includes the following steps: Step 1: Obtain an image of a V-shaped weld seam containing structured light stripes, input it into a pre-trained pose estimation network, extract 5 topological key points; calculate the range of each point in the horizontal and vertical directions of the image, adaptively select coordinate axes with larger spans for monotonic sorting, delineate local regions of interest, and perform local contrast enhancement. Step 2: Within the corridor formed by connecting adjacent sorting key points, construct a controlled Gaussian gravitational field with the mathematical expectation of the connecting line, calculate the local Gaussian weighted gray-level centroid, and extract the initial sub-pixel scattered point set of each segment. Step 3: Perform robust line fitting independently on the sub-pixel scattered point set in each segment to obtain a segmented main arm straight line model that is resistant to secondary reflection distortion. Step 4: Solve the equations of adjacent segmented main arm lines simultaneously to obtain the absolute geometric intersection point analytically; construct an adaptive verification circle with the corresponding original topological key point as the center, and calculate the spatial distance between the intersection point and the center of the circle to perform cross-validation of the validity. Step 5: Connect the first and last key points with the verified absolute geometric intersection points in sequence, filter out the background and reconstruct a noise-free absolute geometric skeleton light stripe; construct a continuous spatial directed tracking vector, calculate the two-dimensional outer product of adjacent directed tracking vectors, and realize the adaptive recognition and stripping of the feature points at the root of the tested V-shaped weld according to the topological polarity law of the outer product sign.

2. The strong reflective V-shaped weld seam feature extraction method based on semantic topology segmentation and cross-multiplication polarity according to claim 1, characterized in that, In step 2, the dynamic weight function of the controlled Gaussian attractive field is: wherein, is the pixel gray value of the local enhanced image coordinate at the local enhanced image coordinate, is the theoretical mathematical expected horizontal coordinate of the line corridor at the line, is the set Gaussian tolerance coefficient.

3. The method of claim 1, wherein, In step 3, a linear fitting algorithm with Huber loss function is used to initially fit each segment point set. After setting a tolerance threshold to remove local points, the least squares method is used for secondary high-precision reconstruction.

4. The method of claim 1, wherein, The specific method for determining the two-dimensional cross product and topological polarity law in step 5 is as follows: calculate the two-dimensional cross product of all internal corner feature points of the tested V-shaped weld. Define the inflow vector as , the outflow vector as , and the Z component of the two-dimensional cross product as : wherein, the feature point ranked in the th position points to a two-dimensional spatial vector of the feature point ranked in the th position; the feature point ranked in the th position points to a two-dimensional spatial vector of the feature point ranked in the th position; Statistically the Z component Sign polarity, due to the physical bending direction of the V-shaped groove edge corner is opposite to the bottom root corner, the sign polarity is the only feature point determination and locking as the absolute root feature point, so as to accurately decouple and separate the root feature point from the remaining four groove wall feature points.

5. The method according to claim 1 or 4, characterized in that, Step 5 is followed by anti-jitter smoothing and bevel geometry feature quantization steps based on spatiotemporal Kalman filtering, specifically: Construct a Kalman spatiotemporal tracker with the pixel coordinates and pixel velocity of feature points at the weld root as state variables; When the current image frame is not splashed by occlusion and the root feature point observation is successfully stripped The tracker is corrected by the Kalman state update equation: When the current image frame is occluded by splashing or experiences severe pixel jitter, the predicted position of the root feature point is output using the Kalman state prediction equation: wherein, is the optimal state estimation vector; is the state prediction vector; is the Kalman gain matrix; is the observation matrix; is the state transition matrix; Further, taking the optimal root feature point of the tracking output as the origin, a spatial direction vector pointing to the left side wall edge feature point is constructed With the spatial direction vector pointing to the right side wall edge feature point The two-dimensional groove apparent opening width And the two-dimensional apparent cross-sectional filling area : The calculated root continuous motion trajectory and apparent cross-sectional fill area Sequence as visual control input for adaptive adjustment of welding speed and wire feed by a welding expert system.

Images