Pipe pile inner wall welding method based on visual perception

Abstract

The application discloses a pipe pile inner wall welding method based on visual perception and belongs to the technical field of visual detection, and specifically comprises the following steps: firstly, installing an inner wall micro-texture light shielding ring and a root light guide wedge on both sides of a weld, calibrating a camera and a light field geometry, establishing and locking a local coordinate system of the weld; then, starting a ring-shaped polarized light field to light up along the weld in sections, and the light guide wedge guides the grazing light to approach the root to stabilize the incidence; then, acquiring a polarized separation image pair through a double-channel orthogonal polarization acquisition unit, and synchronously recording an azimuth marker and a sectional number; then, generating a polarized reflection map and a brightness reference map through polarization unmixing and curvature compensation, and corresponding to the local coordinate system; then, constructing a feature dictionary, extracting a weld toe profile, a root crack texture and a heat affected zone boundary therefrom, and forming a feature vector; finally, inputting the feature vector into a recognizer, combining the brightness reference map to obtain a defect mask and a label, and projecting back to the local coordinate system of the weld to complete detection.

Description

Technical Field

[0001] This invention relates to the field of visual inspection technology, and specifically to a method for welding the inner wall of pipe piles based on visual perception. Background Technology

[0002] The quality evaluation of welded inner walls of pipe piles typically relies on machine vision and endoscopic inspection. Existing solutions often employ ring light, coaxial light, dark-field side light, or area array LED light sources, combined with industrial cameras and endoscopic lenses to acquire images of the weld area. These are then combined with imaging techniques such as line laser triangulation, structured light stripes, and photometric stereo to estimate weld contour and height variations. Furthermore, filtering, edge enhancement, threshold segmentation, morphological analysis, template matching, and deep learning-based defect detection networks are used to identify defects such as cracks, porosity, and incomplete penetration.

[0003] However, the inner wall of pipe piles is often a narrow, elongated curved metal cavity, with limited space, continuously varying curvature, and high surface reflectivity. In post-weld visual inspection, traditional lighting methods such as ring lights and LED surface lights are insufficiently adapted to these geometric and material conditions. They easily produce strong specular reflections or shadows in critical areas such as the weld root and lap edges, resulting in high dynamic range scenes with both overexposure and underexposure, leading to loss of texture details and blurred boundaries. Simultaneously, due to uneven reflective distribution, the weld area exhibits pseudo-textures and bright spots, resulting in chaotic feature representation and difficulty in accurately capturing key defect information. This, in turn, misleads template-based or learning-based algorithms, reducing the accuracy and stability of defect identification and interfering with the effective judgment of welding quality. Therefore, there is an urgent need for a pipe pile inner wall welding method that is designed for curved metal cavities, possessing adaptive visual perception and lighting control, capable of stably acquiring interpretable images and reliably extracting key features under complex reflection conditions. Summary of the Invention

[0004] The purpose of this invention is to provide a visual perception-based welding method for the inner wall of pipe piles, thereby solving the problems in the background art.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A method for welding the inner wall of a pipe pile based on visual perception, comprising:

[0007] S1: Install inner wall micro-textured light-shielding rings and light-guiding wedges at the root of the weld on both sides of the weld, calibrate the camera and light field geometry, and establish and lock the local coordinate system of the weld;

[0008] S2: Activate the annular polarization light field module to illuminate segmentally along the weld seam direction. The light guide wedge guides the grazing incident light close to the root of the weld seam, limiting the range of the reflection angle and stabilizing the incident direction.

[0009] S3: The dual-channel polarization acquisition unit uses mutually orthogonal polarization to acquire polarization-separated image pairs, and simultaneously records the orientation marks and segment numbers of the light-shielding rings of the inner wall micro-textures.

[0010] S4: Based on the polarization separation image pair, perform polarization demixing and curvature compensation mapping to generate a polarization reflection map and brightness reference map of the weld area, and map the results to the local coordinate system of the weld.

[0011] S5: Construct a feature dictionary for the weld area, and extract the weld toe contour, weld root crack texture and heat-affected zone texture boundary based on the polarization reflection map to form a feature vector for the weld area.

[0012] S6: Input the feature vector into the weld area image feature recognizer, combine it with the brightness reference map to obtain the defect area mask and label, and project it back to the weld local coordinate system.

[0013] As a further aspect of the present invention: In step S1, the process of installing inner wall micro-textured light-shielding rings and light-guiding wedges at the root of the weld on both sides of the weld, calibrating the camera and light field geometry, and establishing and locking the local coordinate system of the weld is as follows:

[0014] Install inner wall micro-textured light-shielding rings on the inner walls of both sides of the weld using clamps, attach light guide wedges at the root of the weld, and affix directional marks to the outside of the light-shielding rings;

[0015] Fix the camera to the positioning bracket on the inner wall, align it with the orientation mark and the tip of the light guide wedge, place the checkerboard calibration plate to acquire images, and calculate the camera parameters and imaging geometry.

[0016] Based on camera parameters and imaging geometry, a local coordinate system for the weld is established and locked with the tip of the light guide wedge as the origin, the weld direction as the directional axis, and the pipe wall normal as the normal axis.

[0017] As a further aspect of the present invention: In step S2, the process of activating the annular polarization light field module to illuminate segmentally along the weld seam direction, and guiding the grazing incident light close to the root of the weld seam with the light guide wedge, limiting the reflection angle range and stabilizing the incident direction, is as follows:

[0018] Load the weld seam direction segment number into the controller, start the annular polarization light field module, and set the polarizer azimuth angle and segment lighting sequence.

[0019] Illuminate the current segment of the light source in segmented order, so that the light guide wedge fits against the inner wall of the weld, and adjust the position of the incident surface of the light guide wedge and the light field so that the grazing light is close to the root of the weld.

[0020] The incident direction is fixed based on the incident surface of the light guide wedge and the azimuth angle of the polarizer. Lateral reflection is blocked by the micro-textured light-shielding ring on the inner wall, thus limiting the range of the reflection angle.

[0021] As a further aspect of the present invention: In step S3, the process of the dual-channel polarization acquisition unit acquiring images with mutually orthogonal polarization to obtain polarization-separated image pairs, and simultaneously recording the orientation marks and segment numbers of the inner wall micro-texture light-shielding ring, is as follows:

[0022] Load the weld seam direction segment numbering table into the controller, configure the synchronous trigger signal and timestamp source, and set the two channel polarizers to be mutually orthogonal;

[0023] The current segment acquisition is triggered according to the segment number, and the dual-channel polarization acquisition unit is aligned with the weld area to acquire images synchronously with mutually orthogonal polarization.

[0024] Acquire two-channel raw images, pair them by channel to generate polarization-separated image pairs, and bind timestamps and segment numbers;

[0025] Identify and record the orientation markers and segment numbers of the light-shielding rings of the inner wall microtextures, and write the results into the metadata field of the polarization-separated image pair.

[0026] As a further aspect of the present invention: in step S4, the process of performing polarization demixing and curvature compensation mapping based on the polarization separation image pair to generate a polarization reflection map and a brightness reference map of the weld area, and mapping the results to the local coordinate system of the weld, is as follows:

[0027] Import the polarization separation image pairs and calibration parameters, read the segment numbers, generate the inner wall curvature grid according to the pipe wall shape, and associate it with the local coordinate system of the weld.

[0028] A polarization demixing model is established based on mutually orthogonal polarization. The polarization-separated image is demixed pixel by pixel to obtain the initial values ​​of the polarization reflection component and the brightness component.

[0029] Curvature compensation mapping is performed based on the inner wall curvature mesh and the incident direction, and the polarization reflection components are unfolded onto the local coordinate system plane of the weld.

[0030] A polarization reflection map of the weld area is generated using the expanded polarization reflection components, and a brightness reference map is generated using the initial values ​​of the brightness components, while retaining the segment number and timestamp.

[0031] The polarization reflection map and brightness reference map of the weld area are registered to the local coordinate system of the weld according to the segment number and timestamp, and the result correspondence is completed.

[0032] As a further aspect of the present invention: the specific details of establishing a polarization demixing model based on mutually orthogonal polarizations, and demixing the polarization-separated image pair pixel by pixel to obtain the initial values ​​of the polarization reflection component and the brightness component are as follows:

[0033] Read the polarization separation image pair and the polarizer azimuth angle, determine the mutually orthogonal polarization directions, and define the pixel observation vector and imaging response coefficient;

[0034] A polarization unmixing model is established based on the mutually orthogonal polarization relationship. The polarizer azimuth angle and the imaging response form a coefficient matrix, and the pixel intensity forms an observation vector.

[0035] For each pixel, the observation vector is input, and the polarization unmixing model is solved by least squares to obtain the initial values ​​of the polarization reflection component and the brightness component.

[0036] The initial values ​​of the polarization reflection component and the brightness component are used to generate corresponding images. The index is consistent with the polarization separation image pair, and the pixel positions are kept in a one-to-one correspondence.

[0037] As a further aspect of the present invention: In step S5, the process of constructing a weld area feature dictionary and extracting the weld toe contour, weld root crack texture, and heat-affected zone texture boundary based on the polarization reflection map to form a weld area feature vector is as follows:

[0038] Read the polarization reflection map of the weld area in the local coordinate system of the weld and establish the feature dictionary fields of the weld area, including weld toe contour, crack texture and heat-affected zone texture boundary;

[0039] A contour search zone is generated along the weld seam direction. The edge is tracked based on the polarization reflection gradient and intensity transition. The weld toe contour curve is extracted and registered in the weld seam region feature dictionary.

[0040] A strip window is set in the neighborhood of the weld root, the linear texture response and direction consistency are calculated, and the fine line segments are connected to obtain the crack texture at the weld root and registered.

[0041] The polarization reflection changes are scanned line by line along the pipe wall normal to locate the direction of texture abrupt change, the texture boundary of the heat-affected zone is depicted and registered in the weld area feature dictionary;

[0042] Based on the weld area feature dictionary, the weld toe contour, crack texture and heat-affected zone texture boundary are encoded into weld area feature vectors, and segment numbers and timestamp indices are retained.

[0043] As a further aspect of the present invention: In step S6, the process of inputting the feature vector into the weld area image feature recognizer, combining it with the brightness reference map to obtain the defect location mask and label, and projecting it back to the weld local coordinate system is as follows:

[0044] Load the feature vector and brightness reference map of the weld area, and set the input format and label set of the weld area image feature recognizer in the processing unit;

[0045] The feature vector is paired with the brightness reference map according to the pixel position and segment number to form a feature matrix and brightness channel, thus completing the recognition input assembly.

[0046] Start the weld area image feature recognizer, output the defect area mask and label based on the feature matrix and brightness channel, and record the index mapping table and timestamp;

[0047] Based on the segment number and timestamp, the registration parameters are called to register the defect area mask and label to the local coordinate system of the weld, ensuring that the pixel positions correspond one-to-one.

[0048] The beneficial effects of this invention are:

[0049] This invention employs an inner wall micro-textured light-shielding ring arranged on both sides of the weld and a light-guiding wedge at the weld root, locking the local coordinate system to guide grazing light close to the weld root, thus suppressing specular highlights and edge glare by limiting the reflection angle. A segmented annular polarized light field illuminates the weld, adaptively supplementing light along its direction to reduce uneven illumination caused by changes in cavity curvature. Dual-channel mutually orthogonal polarization synchronous imaging separates the specular and diffuse reflection components in hardware, significantly improving the contrast and signal-to-noise ratio of key details such as the weld toe and root. Subsequently, polarization demixing and curvature compensation mapping are performed to generate a polarization reflection map and a brightness reference map, which are then backfilled into the local coordinate system. This effectively eliminates pseudo-textures and over / underexposed areas caused by attitude and brightness drift, ensuring clear and discernible textures and boundaries at the weld root and heat-affected zone from the source.

[0050] Building upon robust imaging, a closed-loop feature extraction and recognition mechanism constrained by polarization reflection mechanisms enhances detection reliability. By constructing a weld region feature dictionary and extracting criteria such as weld toe contours, root crack textures, and heat-affected zone boundaries from the polarization reflection map, a multimodal feature vector is formed. This vector, combined with a brightness reference map, is input into an image feature recognizer to obtain defect masks and labels. This effectively suppresses false alarms and false negatives caused by highlights and shadows, achieving stable identification of defects such as incomplete penetration, cracks, and porosity. The detection results are projected back into the local coordinate system and associated with segment numbers and shading ring orientation markers, forming a traceable spatial index. This facilitates the statistical analysis of defect distribution, the location of process steps, and the optimization of welding torch posture and parameters. The overall solution requires minimal modification to existing endoscopic vision platforms, reduces power consumption and heat accumulation through segmented illumination, is compatible with pipe piles of different diameters and materials, and possesses online real-time operation capabilities. It balances high detection rates and low false alarm rates even in complex reflective environments, significantly reducing manual review and rework costs, and improving the consistency and reproducibility of welding quality assessment. Attached Figure Description

[0051] The invention will now be further described with reference to the accompanying drawings.

[0052] Figure 1 This is a schematic flowchart of a visual perception-based welding method for the inner wall of a pipe pile according to the present invention. Detailed Implementation

[0053] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0054] Please see Figure 1 As shown, the present invention is a method for welding the inner wall of a pipe pile based on visual perception, comprising:

[0055] S1: Install inner wall micro-textured light-shielding rings and light-guiding wedges at the root of the weld on both sides of the weld, calibrate the camera and light field geometry, and establish and lock the local coordinate system of the weld;

[0056] S2: Activate the annular polarization light field module to illuminate segmentally along the weld seam direction. The light guide wedge guides the grazing incident light close to the root of the weld seam, limiting the range of the reflection angle and stabilizing the incident direction.

[0057] S3: The dual-channel polarization acquisition unit uses mutually orthogonal polarization to acquire polarization-separated image pairs, and simultaneously records the orientation marks and segment numbers of the light-shielding rings of the inner wall micro-textures.

[0058] S4: Based on the polarization separation image pair, perform polarization demixing and curvature compensation mapping to generate a polarization reflection map and brightness reference map of the weld area, and map the results to the local coordinate system of the weld.

[0059] S5: Construct a feature dictionary for the weld area, and extract the weld toe contour, weld root crack texture and heat-affected zone texture boundary based on the polarization reflection map to form a feature vector for the weld area.

[0060] S6: Input the feature vector into the weld area image feature recognizer, combine it with the brightness reference map to obtain the defect area mask and label, and project it back to the weld local coordinate system.

[0061] In step S1, the process of installing inner wall micro-textured light-shielding rings and light-guiding wedges at the root of the weld on both sides of the weld, calibrating the camera and light field geometry, and establishing and locking the local coordinate system of the weld is as follows:

[0062] On the inner walls of both sides of the weld, clamps are used to secure the micro-textured light-shielding rings to their predetermined positions, ensuring the end faces of the rings are parallel to the weld edges to reduce stray reflections from the sides. Subsequently, a light guide wedge is installed at the weld root, with its incident surface facing the predetermined light source direction and its tip accurately aligned with the weld root centerline. To establish a stable directional reference, azimuth markers are affixed to the outside of the light-shielding rings. These markers use regular shapes for easy visual identification and subsequent alignment. For example, the azimuth markers may combine arrows and etched lines, with the arrows pointing towards the weld start point and the etched lines spaced at fixed intervals. The reading method is similar to clock face reading, allowing operators to quickly confirm the installation angle within the narrow tube. Through these fixing, affixing, and marking actions, an integrated installation reference system of "light-shielding ring - light guide wedge - azimuth marker" is formed, providing a stable reference for subsequent camera calibration and coordinate establishment.

[0063] The camera is fixed to the inner wall positioning bracket, ensuring its optical axis is stably aligned with the azimuth mark and the tip of the light guide wedge, guaranteeing both are simultaneously within the imaging field of view. A checkerboard calibration plate is then placed on the imaging plane between the camera and the weld. Multi-angle, multi-position image acquisition yields calibration images covering the entire working area. Using these calibration images, camera parameters and imaging geometry are calculated. Camera parameters include focal length, principal point position, and distortion coefficients, while imaging geometry describes the camera's spatial attitude relative to the azimuth mark and the tip of the light guide wedge. To improve computational stability, the checkerboard calibration plate is positioned at different tilt and deflection angles within the field of view during acquisition, creating a sufficient distribution of corner points. For example, the azimuth mark is kept on one side of the image, and the tip of the light guide wedge is positioned near the center of the image. This ensures that the solved external attitude and internal parameters are constrained within the same set of data, resulting in a consistent imaging geometry.

[0064] After obtaining the camera parameters and imaging geometry, the origin of the coordinate system is defined by the tip of the light guide wedge at the weld root. The weld orientation is used as the orientation axis, and the pipe wall normal is used as the normal axis, thus constructing a local coordinate system for the weld. The remaining axes are determined as tangential axes using a right-hand rule, ensuring that the three axes are orthogonal and consistent with the actual structural meaning. Subsequently, the transformation matrix between the camera coordinates and the local weld coordinate system is written into the system parameter area, serving as the sole spatial reference for subsequent imaging and measurement. Any image point and physical point can be mapped and queried using this transformation matrix. For ease of understanding, this coordinate system can be likened to a "transparent rectangular coordinate paper" pasted on the weld: the orientation axis extends along the length of the pile, the normal axis points perpendicularly to the inner wall, and the tangential axes are distributed circumferentially. After locking the coordinates, there is no need to re-establish the reference when re-framing; only the transformation matrix needs to be read to quickly align the orientation markers in the new image with the position of the light guide wedge tip to the predetermined coordinates, thereby ensuring that data from different batches and different perspectives maintain a consistent spatial representation within the same local weld coordinate system.

[0065] In step S2, the process of activating the annular polarization light field module to illuminate segmentally along the weld seam, and guiding the grazing incident light close to the root of the weld seam using the light guide wedge to limit the reflection angle range and stabilize the incident direction is as follows:

[0066] After loading the weld seam segment numbers into the controller, the annular polarized light field module is activated, and the polarizer azimuth angle and segment lighting sequence are set. The annular polarized light field module is arranged around the inner wall of the pipe pile, and the polarizer azimuth angle is referenced to the local coordinate system of the weld seam, maintaining a fixed relationship with the direction axis. The segment numbers divide the weld seam direction into continuous small segments, allowing the controller to schedule the corresponding light sources sequentially. For example, when the weld seam direction exhibits a curvature change, the controller lights up each small segment of the light source according to the segment number, similar to the workstation lights along a track working segment by segment, thereby providing stable polarized illumination and clear spatiotemporal identification for the current detection segment.

[0067] Following the segmented lighting sequence, the controller illuminates the current segment's light source, causing the light guide wedge to conform to the inner wall of the weld. Using the wedge's incident surface as a reference, the controller fine-tunes the light field position, bringing the grazing light close to the weld root. The wedge's incident surface and the inner wall surface form a controlled incident channel, allowing the grazing light to slide along the inner wall into the weld root region, reducing direct glare. For example, the operator observes the azimuth markers and aligns the wedge's incident surface with the tangential direction of the annular polarized light field. The grazing light then moves along the weld root like a thin band of light skimming the water's surface, thus placing the root region under continuous and uniform controlled illumination.

[0068] The incident direction is fixed based on the incident surface of the light guide wedge and the azimuth of the polarizer. Lateral reflections are blocked using a micro-textured light-shielding ring on the inner wall, limiting the reflection angle range. The polarizer azimuth determines the polarization orientation of the incident light, and the incident surface of the light guide wedge determines the incident normal and grazing path; both work together to stabilize the incident light. The micro-textured light-shielding ring on the inner wall neutralizes oblique reflections and circumferential stray light through its surface microstructure, ensuring that reflections only return to the imaging channel within a narrow angular band. For example, when metallic mirror reflections occur on the inner wall, the operator maintains the established relationship between the polarizer azimuth and the orientation axis, and positions the micro-textured surface of the light-shielding ring towards the more reflective side. This cuts off lateral reflections after the light-shielding ring, preserving an effective reflection path to the imaging surface. This configuration forms a continuous link at the control and execution level—fixed incident light, limited reflections, and clearly defined markings—providing stable input and reliable geometric reference for subsequent polarization demixing and feature extraction.

[0069] In step S3, the dual-channel polarization acquisition unit acquires images using mutually orthogonal polarization to obtain polarization-separated image pairs, and simultaneously records the orientation marks and segment numbers of the inner wall micro-texture light-shielding ring.

[0070] After loading the weld seam segment numbering table into the controller, the correspondence between the segment numbers and the illumination segments is established, and the dual-channel polarization acquisition unit is activated. Polarizers are installed at the front ends of the two imaging channels, with their azimuth angles set to be mutually orthogonal to ensure that the polarization information received by the two channels is independent. Subsequently, a synchronous trigger signal and a timestamp source are configured to enable simultaneous acquisition by both channels under the same trigger pulse, generating a unique time identifier for each trigger. This provides a stable time reference for subsequent polarization separation image synchronization. For example, the segment numbers are analogous to mileage markers arranged along the weld seam, and the timestamp source is analogous to a "timer" for each exposure; together, they fix "which segment and when" the acquisition takes place.

[0071] The controller triggers the acquisition command for the current segment according to the segment number. The optical axis of the dual-channel polarization acquisition unit is aligned with the weld area, so that the orientation mark and the root of the weld simultaneously enter the field of view. The two channels acquire images with mutually orthogonal polarization orientations under synchronous triggering, avoiding the mixing of different polarization information within a single frame. To reduce the influence of the installation environment on the alignment process, this embodiment sets the orientation mark at a fixed position outside the light-shielding ring. Before acquisition, the geometric center of the orientation mark is aligned with the field of view reference. For example, the operator aligns the orientation mark like aligning the zero mark on a ruler, and then presses the trigger button to simultaneously obtain two mutually orthogonal polarized images for that segment.

[0072] After acquisition, two channels of raw images are obtained. The timestamps and segment numbers of both channels are read, and images with the same timestamp and segment number are paired to form polarization-separated image pairs, maintaining consistent pixel indices. To ensure traceability in subsequent processing, the acquisition trigger number, exposure order, and device number are written into the record structure during pairing. For example, it's like binding two photos taken at the same time using two "filters" with different polarization orientations onto a single page, with the header indicating "segment number, when taken, and by whom," providing a basis for subsequent retrieval and comparison.

[0073] An azimuth marker recognition algorithm is invoked in the polarization-separated image pair. Based on the shape and positional relationship of the azimuth markers, azimuth information and visible numbers are extracted and matched with the corresponding segment numbers. After verification, the azimuth markers of the inner wall micro-texture shading ring and the segment numbers are written into the metadata field of the polarization-separated image pair, while retaining the timestamp and trigger sequence number to form a complete retrieval key. For example, this process is similar to filling in the "Location and Direction" entry in the image file's properties page. Subsequent coordinate registration or trajectory verification only requires reading the metadata field to locate the correct segment and orientation, avoiding confusion caused by manual recording errors. Through these four steps, a closed-loop process of segment triggering, orthogonal acquisition, channel pairing, and marker archiving is achieved, providing a clear and stable data foundation for subsequent polarization demixing, curvature compensation, and feature extraction.

[0074] In step S4, the process of performing polarization demixing and curvature compensation mapping based on the polarization separation image pair to generate a polarization reflection map and a brightness reference map of the weld area, and mapping the results to the local coordinate system of the weld, is as follows:

[0075] After importing the polarization-separated image pairs and calibration parameters, the segment numbers are read, and an inner wall curvature grid is generated based on the pipe wall shape. The inner wall curvature grid describes the pipe wall normal, tangential, and orientation relationships of each grid point, and establishes a one-to-one correspondence with the local coordinate system of the weld seam. The origin, orientation axis, and normal axis of the coordinate system remain consistent with the aforementioned calibration. For ease of understanding, the inner wall curvature grid can be viewed as "fine warp and weft lines" laid on the inner wall of the pipe. The warp lines are along the pipe wall normal, and the weft lines are along the weld seam orientation. Any image pixel mapped onto the grid can find a clear spatial position and orientation. If the pipe wall shape is slightly elliptical, the curvature grid will exhibit different curvature densities. Based on this, the correct local curvature relationship is used in subsequent steps to avoid treating the curved surface as a plane, which would cause positional deviations.

[0076] A polarization demixing model is established based on mutually orthogonal polarization. The polarization-separated image is demixed pixel by pixel to obtain the initial values ​​of the polarization reflection component and the brightness component.

[0077] Specifically, after reading the polarization-separated image pair and the polarizer azimuth angle, the orthogonal polarization directions are first determined based on the azimuth angle, and this direction pair is used as a unified reference for subsequent calculations. At the pixel level, a pixel observation vector is defined for each pixel, consisting of the grayscale values ​​of the two polarization-separated images at the same pixel location; simultaneously, an imaging response coefficient is set for each channel, which includes the influence of camera imaging response, polarizer transmission characteristics, and imaging geometry on brightness. In this way, the "observations from two polarization directions" and the "response of the camera imaging system" are mapped to the mathematical description of each pixel. For example, for a point located at the root of the weld, two sets of grayscale values ​​are read from the two images at that point, forming a two-dimensional pixel observation vector; correspondingly, a set of imaging response coefficients is associated with that point to describe the imaging relationship of that point under the two polarization directions.

[0078] Based on mutually orthogonal polarization relationships, a polarization demixing model is established at the pixel level. The model's coefficient matrix is ​​composed of the polarizer azimuth angle and the imaging response, reflecting the imaging mapping under different polarization directions; the observation vector is composed of pixel intensities, derived from the grayscale values ​​of the same position in the polarization-separated image pairs. The model divides the unknowns into two parts: the initial values ​​of the polarization reflection component and the brightness component. The former characterizes the polarization response related to the surface orientation, while the latter characterizes the brightness background independent of orientation. To ensure the ordered nature of subsequent pixel-by-pixel solutions, the observation vector is extracted one by one using the pixel row and column numbers as indices and stored in conjunction with the corresponding coefficient matrix. For example, if the polarizer azimuth angle of one channel is along the weld seam direction, and that of another channel is along the tangential direction, then the two rows of the pixel's coefficient matrix correspond to these two orientations respectively, and the two terms of the observation vector represent the grayscale values ​​of the two images at that pixel.

[0079] After the model is established, the observation vector is input for each pixel, and the polarization demixing model is solved using least squares to obtain the initial values ​​of the polarization reflection component and the brightness component of that pixel. Least squares is used to provide a consistent pixel-level estimate in the presence of imaging noise and local reflection perturbations; the solution results are cached together with the pixel index, forming a one-to-one mapping from "pixel position" to "two components". For intuitive explanation, the demixing process can be understood as distributing observations under two different polarization directions back to "the part related to direction" and "the part unrelated to direction": for pixels located near the weld toe contour, the polarization reflection component usually reflects directional details better; for pixels located in flat substrate areas, the initial value of the brightness component reflects the background brightness better. This distinction does not introduce new terms; the initial values ​​of the polarization reflection component and the brightness component are still used as a unified expression.

[0080] After pixel-by-pixel solving, the initial values ​​of the polarization reflection component and the brightness component are written into two corresponding images. To maintain spatial consistency with the polarization-separated image pair, the row and column numbers of the original image are strictly used as indices, ensuring a one-to-one correspondence between the two corresponding images and the polarization-separated image pair in terms of pixel location. Simultaneously, index mapping information is recorded in the image header to ensure that subsequent processing can directly compare and retrieve data across multiple images using the same pixel location. For example, when it is necessary to review the state of a specific pixel (e.g., row number 1000, column number 500) in various images, the operator inputs the row and column numbers to simultaneously locate the two observed values ​​in the polarization-separated image pair, the corresponding value in the polarization reflection component image, and the corresponding value in the brightness component initial value image, without any positional shift, thus avoiding errors caused by manual comparison.

[0081] After obtaining the polarized reflection components, curvature compensation mapping is performed based on the inner wall curvature mesh and the incident direction, unfolding the polarized reflection components on the curved surface onto the local coordinate system plane of the weld. The compensation mapping uses the normal and tangential information provided by the mesh to correct for changes in reflection direction and projection deformation caused by surface tilt, ensuring that the same spatial location occupies a stable position on the plane coordinate system. For example, this process can be likened to slowly flattening a label pasted on the outer wall of a cylinder along the generation line. After flattening, the text and graphics are fixed in position on the plane, and reading is no longer affected by the cylinder's curvature. Similarly, the unfolded polarized reflection components have a stable geometric relationship on the plane, facilitating subsequent image generation and retrieval.

[0082] A polarization reflection map of the weld area is generated using the unfolded polarization reflection components, and a brightness reference map is generated using the initial values ​​of the brightness components, while retaining segment numbers and timestamps. The weld area polarization reflection map is used to present the directional reflection differences between the weld toe, weld root, and heat-affected zone, while the brightness reference map is used to present the overall brightness background at the same location. Both share the same coordinate grid points and the same pixel index. For example, the polarization reflection map can be regarded as a "directional texture map," and the brightness reference map as a "non-directional light and dark background map." The two images are like two transparent sheets superimposed on the same grid paper. Directional and brightness information can be read simultaneously at any grid point, and the acquisition time and location can be traced according to the segment number and timestamp.

[0083] Finally, the registration parameters obtained from the previous calibration are called according to the segment number and timestamp to register the polarization reflectance map and brightness reference map of the weld area to the local coordinate system of the weld, completing the result correspondence. The registration process establishes a fixed relationship between the image coordinates and the actual spatial coordinates and writes them into the metadata field of the result as a unified reference for subsequent analysis and comparison. For example, the local coordinate system of the weld can be imagined as a "transparent coordinate paper" pasted on the weld. After registration, the pixel positions of the polarization reflectance map and brightness reference map strictly coincide with the grid points of this "transparent coordinate paper". When observing the same segment again at any time, only the segment number and timestamp need to be read to put the new data back into the same grid position, achieving stable alignment across batches and across viewpoints.

[0084] In step S5, the process of constructing a weld area feature dictionary and extracting the weld toe contour, weld root crack texture, and heat-affected zone texture boundary based on the polarization reflection image to form a weld area feature vector is as follows:

[0085] In the local coordinate system of the weld, the polarization reflection map of the weld area is first read, and the image grid is then mapped point by point to the orientation axis, normal axis, and tangential axis of the local coordinate system of the weld. Subsequently, feature dictionary fields for the weld area are established, with fixed names for "weld toe contour," "crack texture," and "heat-affected zone texture boundary." For ease of retrieval, the feature dictionary uses segment number as the primary key and timestamp as the secondary key. Each entry in the dictionary includes reference coordinates, pixel index, and data source marker. This ensures that any subsequent processing can access the polarization reflection information at the same location based on the same coordinates and index. For example, when the operator inputs a segment number and timestamp index, the polarization reflection map of that segment is directly located, and three corresponding blank fields are automatically created in the feature dictionary, awaiting subsequent writing of the weld toe contour, crack texture, and heat-affected zone texture boundary.

[0086] A contour search band is generated on the polarization reflectance map along the weld seam direction. The search band is located in a narrow region where the weld toe may appear, and its width is determined by the estimated position of the weld toe in the local coordinate system and the geometric relationship with the pipe wall. Subsequently, the polarization reflectance gradient and intensity transition are calculated within the search band. A stable edge trajectory is obtained by continuous tracking along the direction, and the trajectories of adjacent frames are stitched together in the coordinate system to form a complete weld toe contour curve. The curve is registered in the "weld toe contour" field of the weld region feature dictionary in an ordered point column form, and the segment number and timestamp index at the time of generation are recorded. For example, the search band can be understood as a narrow channel spreading along the weld seam. Within the channel, it moves along the point of maximum gradient, and the turning position is indicated by the intensity transition, thus obtaining a continuous weld toe contour line.

[0087] In the vicinity of the weld root, a strip window aligned with the weld direction is set. One end of the strip window is close to the weld root, and the other end points towards the substrate region. The linear texture response and directional consistency are calculated column by column within the window, and fine line segments with prominent responses are connected to obtain the crack texture at the weld root. The connection strategy prioritizes directional consistency, connecting segments with similar and mutually pointing directions to avoid irrelevant textures from being mixed in. The obtained crack texture is stored as an ordered set of fine lines in the "Crack Texture" field of the weld region feature dictionary, along with the pixel index, segment number, and timestamp index. For example, the strip window can be imagined as a narrow comb, with its teeth combing the image along the weld direction, separating the fine and long crack fibers from the background texture and sequentially connecting these fibers into a traceable fine line structure.

[0088] The polarization reflection variation is scanned line by line along the pipe wall normal. Each scan line advances from the weld area towards the substrate area, recording the rate of change of polarization reflection and the abrupt turning points of the texture direction. The turning points of consecutive scan lines within the same segment are connected to form the texture boundary of the heat-affected zone. This boundary is represented in the coordinate system as a continuous curve distributed around the weld, used to mark the texture transition zone of the heat-affected zone. After the depiction is completed, the curve is registered in the "Heat-Affected Zone Texture Boundary" field of the weld area feature dictionary as a control point sequence, and the segment number and timestamp index are saved simultaneously. For example, line-by-line scanning can be likened to repeatedly slicing horizontally at the same location, marking a point at the abrupt texture change; by gradually slicing along the direction, a series of connected points are obtained, and the curve formed by this series of points is the texture boundary of the heat-affected zone.

[0089] Once the three items—"weld toe contour," "crack texture," and "heat-affected zone texture boundary"—have been registered, they are encoded according to the weld area feature dictionary to generate a weld area feature vector. The encoding follows a fixed order and fixed fields: first, the geometric parameters and orientation description of the weld toe contour are written; second, the line segment distribution and direction statistics of the crack texture are written; and finally, the curve shape and normal position sequence of the heat-affected zone texture boundary are written. The feature vector, along with the segment number, timestamp index, and pixel index mapping table, is stored together to form a standard input that can be directly called by subsequent recognition modules. For example, the feature vector can be understood as a fully indexed summary card, with the key quantities and positional references of "contour," "thin line," and "boundary" recorded in a uniform order. When comparing the same segment under different working conditions or at different times, only two cards need to be retrieved under the same index, allowing alignment of the three types of features within the same coordinate system for item-by-item comparison and subsequent recognition.

[0090] In step S6, the process of inputting the feature vector into the weld area image feature recognizer, combining it with the brightness reference map to obtain the defect location mask and label, and projecting it back to the weld local coordinate system is as follows:

[0091] After loading the weld area feature vector and brightness reference map, the input format and label set of the weld area image feature recognizer are set in the processing unit. The input format specifies the column order of the feature matrix, the encoding method of the brightness channel, and the recording field of the pixel index. The label set provides the defect categories to be identified and their numbering rules. The feature vector is derived from the encoding of the weld toe contour, crack texture, and heat-affected zone texture boundaries in the previous steps, and the brightness reference map provides the overall light and dark background at the same location. For ease of understanding, the feature vector can be regarded as a "description card" for each pixel, containing geometric direction, texture intensity, and relative position; the brightness reference map can be regarded as a "background color map" to supplement the overall illumination information. Through this setting, the recognizer obtains a unified field meaning and label number before receiving data, avoiding field ambiguity during subsequent parsing.

[0092] The feature vector and the brightness reference are linked according to pixel position and segment number. Figure 1 Each pairing is used to form a feature matrix and a brightness channel, completing the recognition input assembly. Pairing uses the pixel row and column numbers as indices and the segment number as the primary key, ensuring that the same pixel occupies the same position in both the feature matrix and the brightness channel. During assembly, the acquisition timestamp and device identifier are simultaneously written, forming a three-way association of "pixel index—segment number—timestamp". For example, when processing the k-th frame of a segment, the feature vector of that pixel is written to the r-th row and c-th column of the feature matrix, and the corresponding brightness value is written to the r-th row and c-th column of the brightness channel. The segment number and timestamp of that pixel are recorded in the same index record. In this way, the data received by the recognizer remains consistent in both space and time, allowing subsequent outputs to accurately trace back to the original position.

[0093] After activating the weld area image feature recognizer, it outputs defect area masks and labels based on the feature matrix and brightness channel, and records the index mapping table and timestamps. The defect area mask represents whether a pixel belongs to a defect area in a binary or multi-valued manner, with different values ​​corresponding to different label numbers. Labels are assigned values ​​based on the label set to distinguish between crack types, incomplete fusion types, porosity types, or other preset categories. The index mapping table stores the row and column numbers, segment numbers, and timestamps of each mask pixel, ensuring that any defect area can be accurately found in the original data. For example, when the recognizer forms a continuous mask near the weld root, it assigns the entire area the label "crack texture," and records the row and column numbers of the four corner points of the area's bounding box and their corresponding timestamps in the index mapping table. Subsequent viewing by entering this index will directly access the pixel set and label information for that area.

[0094] Based on the segment number and timestamp, the registration parameters obtained from the previous calibration are used to register the defect area mask and label to the weld local coordinate system, maintaining a one-to-one correspondence between pixel positions. The registration process maps the row and column numbers of each pixel in the mask and label images to the direction and normal axes in the weld local coordinate system, aligning them with the existing coordinate grid. After registration, the defect area mask and label are "projected back" to their exact positions within the weld local coordinate system, facilitating subsequent statistics and verification. For example, when checking the crack start and end positions of a segment, the operator can read the registration result corresponding to that segment to obtain the start and end coordinates of the crack on the direction axis and its span on the normal axis. If the same segment is identified again at different times, the new mask and label, using the same set of registration parameters, will also be projected to the same coordinate grid points, facilitating comparison of defect morphology changes under the same coordinate reference.

[0095] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for welding the inner wall of a pipe pile based on visual perception, characterized in that, Includes the following steps: S1: Install inner wall micro-textured light-shielding rings and weld root light-guiding wedges on both sides of the weld, calibrate the camera and light field geometry, and establish and lock the local coordinate system of the weld; the specific process is as follows: Install inner wall micro-textured light-shielding rings on the inner walls of both sides of the weld using clamps, attach light guide wedges at the root of the weld, and affix directional marks to the outside of the light-shielding rings; Fix the camera to the positioning bracket on the inner wall, align it with the orientation mark and the tip of the light guide wedge, place the checkerboard calibration plate to acquire images, and calculate the camera parameters and imaging geometry. Based on camera parameters and imaging geometry, a local coordinate system for the weld is established and locked with the tip of the light guide wedge as the origin, the weld direction as the direction axis, and the pipe wall normal as the normal axis. S2: Activate the annular polarization light field module to illuminate segmentally along the weld seam direction. The light guide wedge guides the grazing incident light close to the root of the weld seam, limiting the range of the reflection angle and stabilizing the incident direction. S3: The dual-channel polarization acquisition unit uses mutually orthogonal polarization to acquire polarization-separated image pairs, and simultaneously records the orientation marks and segment numbers of the light-shielding rings of the inner wall micro-textures. S4: Based on the polarization separation image pair, perform polarization demixing and curvature compensation mapping to generate a polarization reflection map and brightness reference map of the weld area, and map the results to the local coordinate system of the weld. S5: Construct a feature dictionary for the weld area, and extract the weld toe contour, weld root crack texture and heat-affected zone texture boundary based on the polarization reflection map to form a feature vector for the weld area. S6: Input the feature vector into the weld area image feature recognizer, combine it with the brightness reference map to obtain the defect area mask and label, and project it back to the weld local coordinate system.

2. The method for welding the inner wall of a pipe pile based on visual perception according to claim 1, characterized in that, In step S2, the process of activating the annular polarization light field module to illuminate segmentally along the weld seam, and guiding the grazing incident light close to the root of the weld seam with the light guide wedge, limiting the reflection angle range and stabilizing the incident direction, is as follows: Load the weld seam direction segment number into the controller, start the annular polarization light field module, and set the polarizer azimuth angle and segment lighting sequence. Illuminate the current segment of the light source in segmented order, so that the light guide wedge fits against the inner wall of the weld, and adjust the position of the incident surface of the light guide wedge and the light field so that the grazing light is close to the root of the weld. The incident direction is fixed based on the incident surface of the light guide wedge and the azimuth angle of the polarizer. Lateral reflection is blocked by the micro-textured light-shielding ring on the inner wall, thus limiting the range of the reflection angle.

3. The method for welding the inner wall of a pipe pile based on visual perception according to claim 1, characterized in that, In step S3, the process of the dual-channel polarization acquisition unit acquiring images with mutually orthogonal polarization to obtain polarization-separated image pairs, and simultaneously recording the orientation marks and segment numbers of the inner wall micro-texture light-shielding ring, is as follows: Load the weld seam direction segment numbering table into the controller, configure the synchronous trigger signal and timestamp source, and set the two channel polarizers to be mutually orthogonal; The current segment acquisition is triggered according to the segment number, and the dual-channel polarization acquisition unit is aligned with the weld area to acquire images synchronously with mutually orthogonal polarization. Acquire two-channel raw images, pair them by channel to generate polarization-separated image pairs, and bind timestamps and segment numbers; Identify and record the orientation markers and segment numbers of the light-shielding rings of the inner wall microtextures, and write the results into the metadata field of the polarization-separated image pair.

4. The method for welding the inner wall of a pipe pile based on visual perception according to claim 1, characterized in that, In step S4, the process of performing polarization demixing and curvature compensation mapping based on the polarization separation image pair to generate a polarization reflection map and a brightness reference map of the weld area, and mapping the results to the local coordinate system of the weld, is as follows: Import the polarization separation image pairs and calibration parameters, read the segment numbers, generate the inner wall curvature grid according to the pipe wall shape, and associate it with the local coordinate system of the weld. A polarization demixing model is established based on mutually orthogonal polarization. The polarization-separated image is demixed pixel by pixel to obtain the initial values ​​of the polarization reflection component and the brightness component. Curvature compensation mapping is performed based on the inner wall curvature mesh and the incident direction, and the polarization reflection components are unfolded onto the local coordinate system plane of the weld. A polarization reflection map of the weld area is generated using the expanded polarization reflection components, and a brightness reference map is generated using the initial values ​​of the brightness components, while retaining the segment number and timestamp. The polarization reflection map and brightness reference map of the weld area are registered to the local coordinate system of the weld according to the segment number and timestamp, and the result correspondence is completed.

5. The method for welding the inner wall of a pipe pile based on visual perception according to claim 4, characterized in that, The specific details of establishing a polarization demixing model based on mutually orthogonal polarization, and demixing the polarization-separated image pair pixel by pixel to obtain the initial values ​​of the polarization reflection component and the brightness component are as follows: Read the polarization separation image pair and the polarizer azimuth angle, determine the mutually orthogonal polarization directions, and define the pixel observation vector and imaging response coefficient; A polarization unmixing model is established based on the mutually orthogonal polarization relationship. The polarizer azimuth angle and the imaging response form a coefficient matrix, and the pixel intensity forms an observation vector. For each pixel, the observation vector is input, and the polarization unmixing model is solved by least squares to obtain the initial values ​​of the polarization reflection component and the brightness component. The initial values ​​of the polarization reflection component and the brightness component are used to generate corresponding images. The index is consistent with the polarization separation image pair, and the pixel positions are kept in a one-to-one correspondence.

6. The method for welding the inner wall of a pipe pile based on visual perception according to claim 1, characterized in that, In step S5, the process of constructing a weld area feature dictionary and extracting the weld toe contour, weld root crack texture, and heat-affected zone texture boundary based on the polarization reflection map to form a weld area feature vector is as follows: Read the polarization reflection map of the weld area in the local coordinate system of the weld and establish the feature dictionary fields of the weld area, including weld toe contour, crack texture and heat-affected zone texture boundary; A contour search zone is generated along the weld seam direction. The edge is tracked based on the polarization reflection gradient and intensity transition. The weld toe contour curve is extracted and registered in the weld seam region feature dictionary. A strip window is set in the neighborhood of the weld root, the linear texture response and direction consistency are calculated, and the fine line segments are connected to obtain the crack texture at the weld root and registered. The polarization reflection changes are scanned line by line along the pipe wall normal to locate the direction of texture abrupt change, the texture boundary of the heat-affected zone is depicted and registered in the weld area feature dictionary; Based on the weld area feature dictionary, the weld toe contour, crack texture and heat-affected zone texture boundary are encoded into weld area feature vectors, and segment numbers and timestamp indices are retained.

7. The method for welding the inner wall of a pipe pile based on visual perception according to claim 1, characterized in that, In step S6, the process of inputting the feature vector into the weld area image feature recognizer, combining it with the brightness reference map to obtain the defect location mask and label, and projecting it back to the weld local coordinate system is as follows: Load the feature vector and brightness reference map of the weld area, and set the input format and label set of the weld area image feature recognizer in the processing unit; The feature vector is paired with the brightness reference map according to the pixel position and segment number to form a feature matrix and brightness channel, thus completing the recognition input assembly. Start the weld area image feature recognizer, output the defect area mask and label based on the feature matrix and brightness channel, and record the index mapping table and timestamp; Based on the segment number and timestamp, the registration parameters are called to register the defect area mask and label to the local coordinate system of the weld, ensuring that the pixel positions correspond one-to-one.

Images