A method and device for cross-grain sanding based on a texture direction field

Abstract

The application provides a method and device for polishing along the grain based on a texture direction field, comprising obtaining wood surface image data and wood three-dimensional surface data; determining texture direction field data based on the wood surface image data; identifying the region type of a plurality of local regions based on the texture direction field data to generate region type label data; obtaining three-dimensional polishing direction constraint data based on the wood three-dimensional surface data to generate polishing path data along the grain, adjusting to obtain target polishing path data and polishing process parameter data corresponding to the target polishing path data. The application can provide a more fine direction basis for polishing along the grain, thereby improving the adaptability, stability and surface treatment quality of polishing.

Description

Technical Field

[0001] This invention relates to the technical field of wood workpiece sanding methods, and in particular to a method and apparatus for sanding along the grain based on the grain direction field. Background Technology

[0002] During the sanding process of wood and furniture surfaces, it is generally necessary to sand along the grain as much as possible. If the sanding direction forms a large angle with the wood grain, it is easy to form horizontal or diagonal scratches on the wood surface. These scratches are more noticeable after subsequent painting, applying a transparent coating, or polishing, affecting the surface quality of the wood products.

[0003] Existing robotic arm sanding methods typically follow preset straight lines, reciprocating paths, or generate sanding trajectories based on the overall main direction, making it difficult to adapt to the complex texture variations on real wood surfaces. Actual wood surfaces may exhibit curved textures, wavy textures, swirling textures around knots, texture bifurcation, abrupt changes in texture direction at joints, and chaotic texture direction on end faces. Relying solely on a single main direction or a fixed sanding path for the entire surface cannot accurately constrain the sanding direction of the robotic arm in different local areas.

[0004] Furthermore, for curved wood or irregularly shaped furniture surfaces, the texture direction in the 2D image needs to be combined with the 3D surface geometry to form a 3D sanding direction constraint that the robotic arm can execute. If the path is planned solely based on the 2D image direction, directional deviations are easily generated in curved areas; if special areas such as knots, panel joints, low-confidence regions, and end-face texture areas are not distinguished, problems such as abrupt path changes, local over-sanding, missed sanding, or inconsistent sanding quality are likely to occur. Therefore, there is an urgent need for a method that can adaptively generate the robotic arm sanding path and sanding process parameters based on the texture direction field of the wood surface. Summary of the Invention

[0005] In view of the shortcomings of the prior art described above, the purpose of this invention is to provide a method and apparatus for parallel sanding based on the texture direction field. This invention can provide a more precise directional basis for parallel sanding, thereby improving the adaptability, stability and surface treatment quality of sanding.

[0006] To achieve the above and other related objectives, this invention provides a method for sanding along the grain based on a texture direction field, comprising: acquiring wood surface image data and wood three-dimensional surface data; determining texture direction field data based on the wood surface image data, wherein the texture direction field data is used to characterize the local texture direction corresponding to multiple local regions on the wood surface; identifying the region type of the multiple local regions based on the texture direction field data, and generating region type label data; performing three-dimensional direction transformation on the texture direction field data based on the wood three-dimensional surface data to obtain three-dimensional sanding direction constraint data corresponding to the multiple local regions; generating sanding path data along the grain based on the three-dimensional sanding direction constraint data, and performing regional adjustments on the sanding path data along the grain and the corresponding sanding process parameters according to the region type label data to obtain target sanding path data and sanding process parameter data corresponding to the target sanding path data.

[0007] Preferably, determining texture orientation field data based on wood surface image data includes: inputting wood surface image data into a pre-trained orientation field estimation model, the orientation field estimation model including a multi-scale texture feature extraction unit and an output unit connected to the multi-scale texture feature extraction unit; extracting multi-scale texture features from the wood surface image data through the multi-scale texture feature extraction unit; outputting an orientation angle map and a confidence map based on the multi-scale texture features through the output unit; determining local texture orientation data corresponding to multiple local regions based on the orientation angle values ​​corresponding to multiple local regions in the orientation angle map, and determining texture orientation confidence data corresponding to multiple local regions based on the confidence values ​​corresponding to multiple local regions in the confidence map; and generating texture orientation field data based on the local texture orientation data and texture orientation confidence data corresponding to multiple local regions.

[0008] Preferably, region type identification is performed on multiple local regions based on texture direction field data to generate region type label data, including: obtaining local texture direction data and texture direction confidence data corresponding to multiple local regions from the texture direction field data; performing at least one of the following region type identifications on the multiple local regions: identifying panel joint regions based on the directional abrupt change features of local texture direction data of adjacent local regions, and generating panel joint region label data; identifying knot regions based on the directional wrapping change features of local texture direction data, and generating knot region label data; identifying end face texture regions based on the directional disorder features and texture direction confidence data of local texture direction data, and generating end face texture region label data; identifying texture bifurcation point regions based on the directional bifurcation features of local texture direction data, and generating texture bifurcation region label data; identifying low-confidence regions based on texture direction confidence data, and generating low-confidence region label data; and generating region type label data based on at least one of the panel joint region label data, knot region label data, end face texture region label data, texture bifurcation region label data, and low-confidence region label data.

[0009] Preferably, based on the three-dimensional surface data of wood, the texture direction field data is transformed into three-dimensional directions to obtain three-dimensional sanding direction constraint data corresponding to multiple local regions. This includes: obtaining local texture direction data corresponding to multiple local regions from the texture direction field data; determining the three-dimensional surface regions corresponding to multiple local regions according to the correspondence between the wood surface image data and the three-dimensional surface data of wood; determining the local tangent plane corresponding to the three-dimensional surface region, mapping the local texture direction data corresponding to the local region to the corresponding local tangent plane to obtain the three-dimensional wood grain direction vector in the local tangent plane; and generating three-dimensional sanding direction constraint data corresponding to the local region based on the three-dimensional wood grain direction vector.

[0010] Preferably, the region type label data includes knot region label data. Based on the region type label data, the parallel grinding path data and corresponding grinding process parameters are regionally adjusted to obtain target grinding path data and corresponding grinding process parameter data. This includes: determining knot regions based on the knot region label data and identifying the path segments to be adjusted corresponding to the knot regions from the parallel grinding path data; generating multiple concentric ring path segments with intervals based on the directional circumferential variation characteristics corresponding to the knot regions; converting the multiple concentric ring path segments into target grinding path segments corresponding to the knot regions based on the three-dimensional grinding direction constraint data corresponding to the knot regions; updating the path segments to be adjusted based on the target grinding path segments, and adjusting the grinding process parameters corresponding to the knot regions to obtain target grinding path data and corresponding grinding process parameter data.

[0011] Preferably, the region type label data includes panel joint region label data. Based on the region type label data, the grain-following sanding path data and corresponding sanding process parameters are adjusted regionally, including: determining the panel joint region based on the panel joint region label data, and determining the path segment to be adjusted corresponding to the panel joint region from the grain-following sanding path data; determining the joint transition region based on the panel joint region, and generating a direction-gradient path segment within the joint transition region based on the three-dimensional sanding direction constraint data corresponding to both sides of the panel joint region; updating the path segment to be adjusted based on the direction-gradient path segment, and adjusting the sanding process parameters corresponding to the panel joint region to obtain the target sanding path data and the sanding process parameter data corresponding to the target sanding path data.

[0012] Preferably, the region type label data includes at least one of low-confidence region label data and end-face texture region label data. The region-based adjustment of the grain-following polishing path data and corresponding polishing process parameters is performed based on the region type label data, including: determining the region to be adjusted based on at least one of the low-confidence region label data and end-face texture region label data, and determining the path segment to be adjusted corresponding to the region to be adjusted from the grain-following polishing path data; when the region to be adjusted is determined to be a low-confidence region or an end-face texture region, switching the path adjustment strategy corresponding to the region to be adjusted to a conservative coverage path strategy, and switching the polishing process parameter adjustment strategy corresponding to the region to be adjusted to a conservative polishing parameter strategy; updating the path segment to be adjusted based on the conservative coverage path strategy, and adjusting the polishing process parameters corresponding to the region to be adjusted based on the conservative polishing parameter strategy, to obtain the target polishing path data and the corresponding polishing process parameter data.

[0013] The present invention also provides a parallel sanding device based on texture direction field, comprising: The system comprises the following modules: a data acquisition module for acquiring wood surface image data and wood 3D surface data; an orientation field determination module for determining texture orientation field data based on wood surface image data, which characterizes the local texture orientation corresponding to multiple local regions on the wood surface; a region type recognition module for identifying the region type of multiple local regions based on texture orientation field data and generating region type label data; an orientation constraint generation module for performing 3D orientation transformation on texture orientation field data based on wood 3D surface data to obtain 3D sanding orientation constraint data corresponding to multiple local regions; and a path parameter generation module for generating parallel sanding path data based on 3D sanding orientation constraint data, and regionalizing the parallel sanding path data and corresponding sanding process parameters according to the region type label data to obtain target sanding path data and corresponding sanding process parameter data.

[0014] The present invention also provides an electronic device, comprising: a processor; a memory, communicatively connected to the processor, the memory storing a computer program, and when the processor executes the computer program, implementing the above-described method for parallel grinding based on texture direction field.

[0015] The present invention also provides a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the above-described method for parallel grinding based on texture direction field.

[0016] As described above, this invention provides a method and apparatus for grain-side sanding based on texture direction field. This application determines texture direction field data covering multiple local areas based on wood surface image data, rather than relying solely on a single principal direction of the entire wood surface, thus providing a more refined directional basis for grain-side sanding. Furthermore, this application identifies the region types of different local areas based on texture direction field data, enabling subsequent path planning to employ different sanding strategies for different regions, which helps reduce the risk of transverse scratches, path abrupt changes, or localized over-sanding in special texture areas. Moreover, this application combines three-dimensional wood surface data to convert two-dimensional local texture directions into three-dimensional sanding direction constraint data within the three-dimensional surface local tangent plane, allowing the robotic arm sanding path to simultaneously satisfy both wood grain direction constraints and surface geometry constraints, thereby improving the accuracy of grain-side sanding on curved wood surfaces. Finally, this application also regionalizes the grain-side sanding path data and sanding process parameters based on region type label data, improving the adaptability, stability, and surface treatment quality of the robotic arm's automatic sanding. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 The diagram shows the overall process of a parallel sanding method based on texture direction field provided in an embodiment of the present invention. Figure 2 The diagram shown is a schematic diagram of a sub-process for determining texture orientation field data provided in an embodiment of the present invention. Figure 3 The diagram shown is a sub-process diagram for generating region type label data according to an embodiment of the present invention. Figure 4 The diagram shown is a sub-process diagram of generating three-dimensional grinding direction constraint data according to an embodiment of the present invention. Figure 5The diagram shown is a sub-process flow chart of adjusting the path and process parameters of the nodule region according to an embodiment of the present invention. Figure 6 The diagram shown is a sub-process flow diagram of adjusting the splicing joint path and process parameters according to an embodiment of the present invention. Figure 7 The diagram shown is a schematic diagram of a sub-process for adjusting the path and process parameters of a low-confidence region or end-face texture region according to an embodiment of the present invention. Figure 8 The diagram shown is a structural schematic of a parallel-texture polishing device based on a texture direction field provided in an embodiment of the present invention. Figure 9 The diagram shown is a structural schematic of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0019] The embodiments of this application are described below with reference to the accompanying drawings. It should be understood that the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, other embodiments obtained by those skilled in the art without creative effort should all fall within the protection scope of this application.

[0020] This application's embodiments can be applied to automated sanding scenarios for various wooden workpieces. When using traditional robotic arms for sanding, problems such as horizontal scratches, diagonal scratches, localized over-sanding, abrupt path changes, repeated sanding at boundaries, or inconsistent sanding quality can easily occur in localized areas when the wood surface has curved textures, wavy textures, texture branching and merging, swirling textures around knots, abrupt changes in texture direction between adjacent boards in panel furniture, chaotic texture direction on end faces, or changes in the three-dimensional texture direction of curved surfaces. Especially after subsequent painting, polishing, or coating with a transparent coating, horizontal scratches become more pronounced, thus affecting the surface quality of the wood products.

[0021] Figure 1 This is a schematic diagram illustrating the overall process of the method provided in the embodiments of this application. Figure 1 As shown, this method can be executed by a server, industrial computer, robotic arm controller, edge computing device, or other electronic device with data processing capabilities. The method may include steps S110 to S150.

[0022] Step S110: Obtain wood surface image data and wood three-dimensional surface data.

[0023] In this embodiment, the wood surface image data can be acquired by an image acquisition device located at the end of a robotic arm, a fixed support, or an independent scanning station. The image acquisition device may include an industrial area array color camera, a line scan camera, a color-depth (RGB-D) camera, a structured light camera, or an industrial camera with a polarized lens. In a preferred embodiment, an industrial area array color camera is used in conjunction with low-angle side lighting to acquire wood surface images, and a structured light scanner or laser line scanner is used to acquire three-dimensional wood surface data. The resolution of the industrial area array color camera can be no less than 20 megapixels. Each pixel can correspond to an actual scale of 0.05 mm to 0.5 mm, or each image can cover a wood surface area of ​​approximately 200 mm × 200 mm. The angle between the low-angle side lighting source and the surface to be measured can be 5 degrees to 25 degrees to enhance the contrast of the wood grain texture.

[0024] In this embodiment, the three-dimensional surface data of the wood can be obtained from structured light scanning, laser line scanning, color-depth (RGB-D) reconstruction, multi-frame depth fusion, import of computer-aided design (CAD) models, or modeling of furniture design drawings. The three-dimensional surface data of the wood can be acquired synchronously with the wood surface image data at the same acquisition station, or it can be acquired in advance at a separate three-dimensional scanning station.

[0025] It should be noted that simultaneously acquiring image texture information and 3D geometric information through the above methods provides a unified data foundation for subsequent longitudinal sanding of curved wood surfaces by the robotic arm. Wood surface image data is used to characterize the local direction of the wood grain, while wood 3D surface data is used to characterize the spatial position, normal, and surface morphology of the wood surface. Combining the two can avoid the problem of ignoring surface variations when planning paths only on a two-dimensional plane.

[0026] Step S120: Determine texture orientation field data based on wood surface image data. Texture orientation field data is used to characterize the local texture orientation corresponding to multiple local areas on the wood surface.

[0027] In this embodiment, the texture orientation field data is not a single principal direction calculated for the entire wood surface, but rather a set of local texture orientations determined for multiple local regions. The output texture orientation field data may include local texture orientation data and texture orientation confidence data for each local region. This texture orientation field data can be used to identify special texture regions and can also be used to generate three-dimensional sanding direction constraints in combination with three-dimensional wood surface data. By replacing a single principal direction with a local orientation field, the subsequent sanding path can adapt to the bending, bifurcation, abrupt changes, and local disorder of the real wood grain.

[0028] Step S130: Based on the texture orientation field data, perform region type identification on multiple local regions and generate region type label data.

[0029] In this embodiment, local texture direction data and texture direction confidence data can be extracted from the texture direction field data, and region type identification can be performed on multiple local regions, such as identifying panel joint regions, knot regions, end face texture regions, texture bifurcation regions, and low-confidence regions. Region type label data is used to identify the texture state and polishing processing strategy corresponding to different local regions. Region type label data may include at least one of panel joint region label data, knot region label data, end face texture region label data, texture bifurcation region label data, and low-confidence region label data. Different region types can correspond to different path adjustment strategies and different polishing process parameter adjustment strategies.

[0030] The output region type label data can be used together with texture orientation field data and 3D grinding orientation constraint data in step S150 to avoid the robotic arm using the same strategy for all regions according to the normal grain path, thereby reducing the risk of local horizontal scratches, path abrupt changes and over-grinding.

[0031] Step S140: Based on the three-dimensional surface data of wood, perform three-dimensional orientation transformation on the texture orientation field data to obtain three-dimensional sanding orientation constraint data corresponding to multiple local areas.

[0032] In this embodiment, the local texture direction in the wood surface image data is typically a direction angle expressed in a two-dimensional image plane, while the robotic arm sanding process occurs on a three-dimensional curved surface. Therefore, it is necessary to combine the three-dimensional wood surface data to convert the local texture direction in the two-dimensional image into a three-dimensional wood grain direction vector in the local tangent plane of the three-dimensional curved surface.

[0033] In one embodiment, the three-dimensional surface region corresponding to each local region can be determined based on the correspondence between wood surface image data and wood three-dimensional surface data; a local cutting plane can be determined based on three-dimensional surface points or three-dimensional patches, and the local texture direction can be mapped to the local cutting plane to obtain a three-dimensional wood grain direction vector.

[0034] The output 3D sanding direction constraint data can include 3D surface point coordinates, surface normal vectors, 3D wood grain direction vectors, tool feed direction constraints, and tool posture constraints. This data can constrain the generation of the sanding path along the grain and constrain the posture of the robotic arm's end-effector tool, ensuring that the tool feed direction is consistent with or approximately consistent with the wood grain direction, and that the tool axis maintains a preset angle with the normal of the wood surface.

[0035] Step S150: Generate parallel grinding path data based on three-dimensional grinding direction constraint data, and perform regional adjustments on parallel grinding path data and corresponding grinding process parameters according to region type label data to obtain target grinding path data and grinding process parameter data corresponding to the target grinding path data.

[0036] In this embodiment, the sanding path data can be a sequence of path points generated along the three-dimensional wood grain direction. For conventional texture areas, the streamline integral method can be used to generate the sanding path along the three-dimensional direction field. Specifically, a path seed point can be selected at the boundary of the area to be sanded or within the area. Starting from the seed point, the path is integrated along the three-dimensional wood grain direction vector of the current local area, and the direction vector is updated when the path enters an adjacent local area. Alternatively, a basic coverage path can be generated first, such as a reciprocating straight line or a zigzag path, and then the basic coverage path can be corrected segment by segment according to the three-dimensional sanding direction constraint data to make the local feed direction of the path consistent with or approximately consistent with the local wood grain direction.

[0037] In one embodiment, when the path reaches the boundary of the area to be sanded, enters a low-confidence area, enters a knot area, approaches the panel joint area, enters a preset avoidance area, or the path length reaches a preset upper limit, the current path can be terminated, and a new path seed point can be selected to generate the next path. By setting path direction update rules and path termination conditions, the sanding path along the grain can be extended continuously in the normal area, and handled by the corresponding regionalization strategy in special areas.

[0038] In this embodiment, the target sanding path data may include a path point sequence, path point 3D coordinates, surface normal vector, 3D wood grain direction vector, tool posture, path segment type, and region label. Sanding process parameter data may include feed rate, downforce, sandpaper grit, path spacing, path overlap rate, and tool head type. For knot areas, panel joint areas, low-confidence areas, or end-face texture areas, regional adjustments can be performed according to the region type label data.

[0039] In one embodiment, the path spacing can characterize the lateral distance between adjacent parallel-grind sanding paths, and the path overlap rate can characterize the degree of overlap between the effective sanding coverage areas corresponding to adjacent parallel-grind sanding paths. The path spacing can be determined based on the effective sanding width of the sanding tool and the path overlap rate. Specifically, if the effective sanding width of the sanding tool is w and the path overlap rate is k, then the path spacing d can satisfy d = w × (1 - k). By setting the path overlap rate, a preset overlapping coverage area can be formed between adjacent parallel-grind sanding paths, thereby reducing the risk of missed sanding and improving the uniformity of sanding coverage.

[0040] It should be noted that by combining path direction constraints with regional adjustments to process parameters, the robotic arm can maintain its efficiency along the grain in conventional areas and adopt a more reliable local strategy in complex texture areas, thereby improving the consistency of overall polishing quality.

[0041] Figure 2 This is a flowchart illustrating the process of determining texture orientation field data provided in an embodiment of this application. Figure 2 As shown, step S120 may include steps S121 to S125.

[0042] Step S121: Input the wood surface image data into the pre-trained orientation field estimation model. The orientation field estimation model includes a multi-scale texture feature extraction unit and an output unit connected to the multi-scale texture feature extraction unit.

[0043] In this embodiment, the wood surface image data can be a preprocessed RGB image, a grayscale image, a low-angle side-lit enhanced image, a polarized image, or a combination of these images. This wood surface image data can be cropped, scaled, or normalized according to a preset size before being input into the orientation field estimation model. Converting the wood surface image data into an input tensor that the orientation field estimation model can process allows the model to uniformly handle wood grain structures at different scales, in different directions, and under different lighting conditions. The orientation field estimation model is used to identify the local orientation of the wood grain from the wood surface image data and outputs an orientation angle map and a confidence map corresponding to the image space.

[0044] In some embodiments, the orientation field estimation model can be pre-trained in the following manner. The training dataset may include wood surface images of different wood species such as oak, walnut, cherry, pine, and rubberwood; it may also include different texture states such as straight grain, curved grain, wavy grain, knot grain, splicing seams, and end face texture; and it may include wood surface images with different surface shapes such as planar, single-curved, and hyperboloid surfaces, as well as under different lighting conditions. The training data can be collected using low-angle side lighting, polarized light, diffuse light, or a combination of multiple light sources to improve the robustness of the model to different actual collection scenarios.

[0045] In one embodiment, training labels can be generated by manually annotating key wood grain lines. The operator can annotate several key wood grain lines on the image, and then an algorithm generates dense orientation field ground values ​​between the annotated lines using Laplace interpolation, thin-plate spline interpolation, or other interpolation methods. In another embodiment, orientation angles can be directly annotated in the grid by the operator, or an initial orientation field value can be generated using a semi-automatic annotation tool, followed by manual correction of high-error areas. Confidence labels can be generated based on the reliability of manual annotation, orientation angle prediction error, neighborhood orientation consistency, or annotation results for specific regions.

[0046] In one embodiment, the training loss may include at least one of orientation angle loss, confidence loss, orientation field smoothing loss, and hard example sample weight loss. The total training loss can be a weighted sum of the orientation angle loss, confidence loss, orientation field smoothing loss, and hard example sample weight loss, and the weights can be adjusted based on the validation results on the training set. After training, the orientation field estimation model receives the current wood surface image data during the inference phase and outputs an orientation angle map and a confidence map. The training phase is used to update the model parameters, and the inference phase is used to estimate the orientation field of the workpiece to be polished; these are different data processing phases.

[0047] In another embodiment, a multi-scale Gabor filter bank can be used to determine the texture orientation field data without employing a pre-trained orientation field estimation model. Multiple Gabor filters can be set up, for example, one filter every 10 degrees from 0 to 170 degrees, forming 18 directional filters; or one filter every 5 degrees, forming 36 directional filters. Multiple scales of Gabor filters can also be set up, for example, small scales for capturing local curved texture details, medium scales for capturing regular texture directions, and large scales for capturing macroscopic texture directions. For each local window, the Gabor filter response intensity in different directions can be calculated, and the filter direction with the highest response intensity can be taken as the local wood grain direction; alternatively, interpolation can be performed on the maximum response direction and its adjacent directions, or a weighted average of multiple high-response directions can be taken by multiple angle vectors based on their response intensity to obtain the local texture direction. The confidence level under the Gabor filtering method can be determined based on the peak-to-peak ratio between the maximum and second-largest responses, the proportion of the maximum response to the total response, or the consistency of neighborhood directions. Both this Gabor filtering method and the pre-trained orientation field estimation model method can output texture orientation field data in a unified format.

[0048] In one embodiment, the multi-scale texture feature extraction unit of the orientation field estimation model can employ a fully convolutional network (FCN) with an encoder-decoder structure. The encoder can use a Residual Network (ResNet), a U-Net encoder, a High Resolution Network (HRNet) backbone, or other convolutional feature extraction structures to perform multi-level downsampling on the input wood surface image data and extract multi-scale texture features. The output unit can include an orientation angle output head and a confidence output head, where the orientation angle output head outputs orientation angle-related features, and the confidence output head outputs the orientation estimation reliability for each pixel or each local region. In another embodiment, a network structure with multi-scale feature extraction capabilities, such as DeepLab v3+, can also be used.

[0049] Step S122: Extract multi-scale texture features from wood surface image data using a multi-scale texture feature extraction unit.

[0050] In this embodiment, multi-scale texture features may include low-level gradient features for characterizing fine texture edges, mid-level texture structure features for characterizing local texture orientation, and high-level context features for characterizing macroscopic wood grain variation trends. Local wood grain on a wood surface may simultaneously contain both fine texture and macroscopic wavy texture; therefore, single-scale features can easily lead to unstable local orientation estimation.

[0051] In one embodiment, the multi-scale texture feature extraction unit may specifically include multiple convolutional layers, residual modules, dilated convolutional modules, or feature pyramid modules to extract multi-scale texture features under different receptive fields. For smaller-scale texture features, local curved texture details can be captured; for medium-scale features, regular texture directions can be captured; and for larger-scale features, macroscopic texture directions and the trend of changes in the directions of adjacent regions can be captured. During the decoding stage, texture features of different resolutions can be fused through skip connections to avoid loss of detail. Multi-scale texture feature extraction avoids the problem of incomplete direction estimation caused by relying solely on a single window or scale, ensuring that the direction field estimation results reflect both local texture details and maintain macroscopic direction continuity.

[0052] Step S123: Output orientation angle map and confidence map based on multi-scale texture features through the output unit.

[0053] In this embodiment, the orientation angle map is used to represent the wood grain orientation angle corresponding to each pixel or each local region in the image, and the confidence map is used to represent the reliability of the corresponding orientation angle. Since the wood grain direction has 180-degree equivalence, that is, the orientation angle θ and θ+180 degrees represent the same wood grain direction, the orientation angle map does not need to directly output a single angle value, but can represent the direction through a double angle transformation.

[0054] In one embodiment, the orientation angle θ can be represented as the cosine and sine of 2θ through a double-angle transformation, and a dual-channel orientation feature map is output by the orientation field estimation model. The orientation angle can then be obtained by inverse solving the bivariate arctangent function (atan2). This method avoids discontinuities in the orientation angle at the 0-degree and 180-degree boundaries, thereby reducing the instability of model training and inference. In another embodiment, the orientation angle output head can also directly output the orientation angle θ, and the influence of boundary discontinuities can be eliminated through a cyclic loss function or angle wrapping processing.

[0055] By simultaneously outputting the direction angle and confidence level, subsequent path planning can not only know "which direction to polish", but also "whether the direction is reliable".

[0056] Step S124: Based on the orientation angle values ​​corresponding to multiple local regions in the orientation angle map, determine the local texture orientation data corresponding to each of the multiple local regions, and based on the confidence values ​​corresponding to multiple local regions in the confidence map, determine the texture orientation confidence data corresponding to each of the multiple local regions.

[0057] In this embodiment, the local regions can preferably be divided into two-dimensional image meshes based on the wood surface image data, for example, dividing the wood surface image into multiple two-dimensional local regions at 32×32 pixels or other sizes. For scenarios requiring correspondence with three-dimensional wood surface data, the two-dimensional local regions can be further mapped to surface regions on the three-dimensional wood surface based on the mapping relationship between the wood surface image data and the three-dimensional wood surface data. Optionally, after completing the registration of the image and the three-dimensional surface data, the local regions can also be determined directly according to the physical size mesh or three-dimensional triangular mesh facets of the three-dimensional surface, for example, dividing the surface into physical sizes of 5 mm × 5 mm to 20 mm × 20 mm, or using triangular facets or adjacent merged facets of the three-dimensional model as local regions. All of the above local regions are used to carry the corresponding local texture direction data and texture direction confidence data.

[0058] In one embodiment, when the orientation angle map is output pixel-by-pixel, the orientation angles of all pixels within a local region can be cyclically counted to obtain the local texture orientation data for that region. Specifically, the orientation angles of each pixel within the region can be converted into double-angle vectors, and then the double-angle vectors can be averaged or weighted averaged. Finally, the orientation angles of the local region can be obtained by inverse solving. For the confidence score, the mean, quantiles, or weighted average of the confidence scores within the local region can be calculated to obtain the texture orientation confidence score data for that local region.

[0059] Step S125: Generate texture orientation field data based on the local texture orientation data and texture orientation confidence data corresponding to multiple local regions respectively.

[0060] In this embodiment, texture orientation field data can be stored according to local region numbers. Each local region's data record can include a region number, two-dimensional image coordinates, three-dimensional surface point coordinates, surface normal vector, orientation angle θ, texture orientation confidence level c, and region type label. The orientation angle θ can range from 0 to 180 degrees, and the confidence level c can range from 0 to 1. Before region type recognition is performed, the region type label can be empty or a regular region label; after subsequent step S130 is completed, the corresponding region type label can be written into this data structure.

[0061] The generated texture orientation field data can be used for region type recognition, 3D orientation transformation, and path planning for grinding along the grain. Through a unified data structure, the outputs of deep learning models, Gabor filter banks, or other orientation estimation methods can all be converted into a consistent orientation field representation, thus facilitating subsequent use of robotic arm path constraints.

[0062] Figure 3 This is a flowchart illustrating the process of generating region type label data as provided in an embodiment of this application. Figure 3 As shown, step S130 may include steps S131 to S137. Steps S132 to S136 may be parallel identification branches, and one or more of them may be executed, without being required to be executed in a fixed order.

[0063] Step S131: Obtain local texture direction data and texture direction confidence data corresponding to multiple local regions from the texture direction field data.

[0064] In this embodiment, local texture orientation data may include the orientation angle, double-angle orientation vector, or orientation histogram of each local region. Texture orientation confidence data can indicate whether the orientation estimation of the local region is reliable. For data obtained through an orientation field estimation model, the confidence level can be derived from the confidence output header or post-processing calculation; for data obtained through a Gabor filter bank, the confidence level can be derived from orientation response concentration, peak-to-peak ratio, or neighborhood consistency.

[0065] In one embodiment, the region number, two-dimensional coordinates, local orientation angle θ, confidence level c, and neighboring region index for each local region can be read from the texture orientation field data. The neighboring region index can be determined based on the 4-neighborhood or 8-neighborhood of the two-dimensional mesh, or based on the adjacent patch relationships of the three-dimensional surface mesh. To handle the 180-degree equivalence of orientation angles, subsequent calculations of orientation difference, curl, divergence, or orientation interpolation can be performed in the double-angle vector domain.

[0066] By using both orientation angle and confidence level, different phenomena such as orientation abrupt change, orientation circling, orientation confusion, orientation bifurcation, and low orientation confidence can be identified, thus providing clear technical basis for region type label data.

[0067] Step S132: Identify the panel joint area based on the directional change features of the local texture direction data of adjacent local areas, and generate panel joint area label data.

[0068] In this embodiment, the splicing seam area typically appears at the junction of adjacent boards. Due to the different origins of the wood grain in the adjacent boards, the local grain direction on both sides of the seam may undergo a step-like abrupt change. The splicing seam area can be characterized by a direction angle difference greater than a preset abrupt change threshold, and the high-difference areas are continuously distributed along a line.

[0069] In one embodiment, the angular difference between each local region and its 4- or 8-neighbor regions can be calculated, and the minimum angular difference is taken considering 180-degree equivalence. When the angular difference is greater than a preset abrupt change threshold of 30 to 45 degrees, and the highly disparate local region forms a linear region extending along a one-dimensional direction after connected component analysis, this linear region can be used as a candidate region for panel joints. The length threshold for the linear region can be 20 to 50 millimeters; when the length of the candidate region for panel joints is greater than this linear region length threshold, the candidate region for panel joints can be determined as the panel joint region.

[0070] The label data for the panel joint area can include the joint line location, joint line direction, local texture direction on both sides of the joint, joint area boundary, and joint recognition confidence.

[0071] Step S133: Identify the scar region based on the orientation wrap-around change features of local texture orientation data, and generate scar region label data.

[0072] In this embodiment, the knot region typically has wood grain that varies in a swirling or spiral pattern around the center of the knot. Unlike ordinary curved textures, the orientation field of the knot region may exhibit singularity in its local neighborhood, with the orientation changing around the center.

[0073] In one embodiment, the orientation angle θ of each local region can be converted into a double-angle orientation vector, and the curl, divergence, or circumference index of the orientation field can be calculated within a preset neighborhood of each local region. The preset neighborhood can be a 5×5 or 7×7 local region. When the absolute value of the neighborhood curl exceeds a preset threshold, and the neighborhood orientation exhibits a circumference distribution, the center of that neighborhood can be determined as the node center. Subsequently, a radial search can be performed outward from the node center, and the node boundary can be determined based on the confidence decrease region, the curl attenuation range, or the image texture morphology. For example, when the curl amplitude in the radial direction attenuates to less than 20% to 30% of the center value, the node boundary can be determined.

[0074] Scar region labeling data can include scar center, scar boundary, scar radius, orientation variation characteristics, and scar region confidence.

[0075] Step S134: Identify the end face texture region based on the directional disorder features of local texture direction data and texture direction confidence data, and generate end face texture region label data.

[0076] In this embodiment, the end-face texture area can be a cross-section of the wood end or a region approximating the end face. Annual rings, vessels, or truncated wood fibers in this area may result in dense, directional, or unstable grain patterns. Forcing a single grain-oriented direction into the end-face texture area may actually cause path instability or localized scratches.

[0077] In one embodiment, end-face texture regions can be jointly identified based on local texture density, directional response dispersion, and confidence level. When the average local gradient magnitude exceeds a preset threshold, multiple directional responses are simultaneously high, the directional response entropy exceeds a preset threshold, or the proportion of the main direction peak is lower than a preset threshold, the local region can be considered directionally disordered. Simultaneously, if the texture direction confidence level of the local region and its neighborhood is lower than a preset threshold of 0.4 to 0.5, and the area of ​​the region continuously satisfying the above conditions exceeds a preset end-face region area threshold, for example, exceeding the area of ​​a corresponding 10 mm × 10 mm physical size region, or being comparable to the effective area of ​​a grinding tool, then it can be determined as an end-face texture region.

[0078] The end-face texture region label data may include the end-face region boundary, end-face region center, degree of orientation disorder, average confidence level, and conservative polishing strategy number.

[0079] Step S135: Identify texture bifurcation point regions based on the directional bifurcation features of local texture direction data, and generate texture bifurcation region label data.

[0080] In this embodiment, the texture bifurcation point region can represent the area where the wood grain flow lines split or merge. This region may appear at natural grain bifurcation points, at localized changes in the wood grain growth, or at the transition area around knots. If the texture bifurcation region is directly integrated along a single direction to generate a path, it may result in sudden path bifurcation, looping, or insufficient local coverage.

[0081] In one embodiment, texture bifurcation regions can be identified based on the topology of directional field streamlines. When directional field streamlines near a local region result in two or more subsequent paths after passing through that region, or when multiple streamlines converge in that region, it can be identified as a texture bifurcation point region. In another embodiment, the directional response distribution within a local region can be detected. When two or more non-adjacent high-response directions exist simultaneously in the Gabor filter response or directional probability distribution, and the angular interval between different high-response directions is greater than 30 degrees, it can be identified as a texture bifurcation candidate region.

[0082] Texture bifurcation region label data can include bifurcation point location, number of bifurcation directions, main direction of each branch, bifurcation region boundary, and bifurcation identification confidence.

[0083] Step S136: Identify low-confidence regions based on texture orientation confidence data and generate low-confidence region label data.

[0084] In this embodiment, the low-confidence region can be a region where the reliability of orientation estimation is low. Low-confidence regions may be caused by wood surface reflection, contamination, weathering, weak grain, image noise, end-face texture, knot centers, or scan shadows. Forcing the orientation field output as the grain-following path orientation in low-confidence regions may lead to incorrect path orientation.

[0085] In one embodiment, local regions with a confidence level (c) below a preset threshold can be marked as low-confidence candidate regions. The confidence level threshold can be between 0.4 and 0.6. If the area of ​​a region composed of consecutive low-confidence local regions exceeds a preset area threshold, for example, exceeding the area of ​​a corresponding 20 mm × 20 mm physical size region, or being comparable to the contact area of ​​a sand table, then low-confidence region label data can be generated. In another embodiment, low-confidence regions can also be determined based on both the neighborhood orientation standard deviation and the texture response intensity, for example, if the neighborhood orientation standard deviation exceeds 20 to 30 degrees, or the total intensity of the local texture response is below a preset threshold.

[0086] Low-confidence region label data may include the low-confidence region boundary, region area, average confidence level, lowest confidence level, and conservative polishing strategy number.

[0087] Step S137: Generate region type label data based on at least one of the following: panel joint area label data, knot area label data, end face texture area label data, texture bifurcation area label data, and low confidence area label data.

[0088] In this embodiment, the region type label data can be a collection of multiple special region labels, or it can be a label map covering all local regions. A local region can correspond to only one region type, or the final region type can be determined according to a preset priority when multiple labels overlap. For example, when the center region of a knot overlaps with a low-confidence region, the knot region strategy or conservative strategy can be executed first; when the panel seam region overlaps with a low-confidence region, the speed can be reduced and the direction can be gradually changed simultaneously.

[0089] In one embodiment, the region type label data may include at least one of the following: region number, region type, center location, boundary profile, confidence level, sanding strategy number, and associated process parameter set. The region type may include a regular region, a knot region, a panel joint region, an end-face texture region, a texture bifurcation region, and a low-confidence region. The boundary profile may be represented by a two-dimensional vertex sequence, a three-dimensional boundary point sequence, or a closed curve. The associated process parameter set may record the recommended sandpaper grit, feed rate range, downforce range, path spacing, and tool head type for that region.

[0090] The region type label data will be incorporated into the subsequent path parameter generation process. By labeling different region types, the wood texture perception results can be directly converted into the basis for robotic arm path planning and sanding process parameter adjustment.

[0091] Figure 4 This is a flowchart illustrating the process of generating three-dimensional grinding direction constraint data as provided in an embodiment of this application. Figure 4 As shown, step S140 may include steps S141 to S144.

[0092] Step S141: Obtain local texture direction data corresponding to multiple local regions from the texture direction field data.

[0093] In this embodiment, the local texture direction data can be a direction angle θ in a two-dimensional image plane, or a double-angle direction vector or a direction histogram. The local texture direction data of each local region corresponds to the wood grain direction of the wood surface within that local region.

[0094] In one embodiment, the local region number, two-dimensional image coordinates, orientation angle θ, and texture orientation confidence can be read from the texture orientation field data. For each local region, the orientation angle of the center point of the region can be selected, or the average result of the double-angle vector of the orientation angles within the region can be selected as the representative orientation. For local regions with low confidence, the orientation angle data can be retained, but the weight of the orientation constraint can be reduced or a conservative strategy can be triggered when generating the polishing orientation constraint later.

[0095] The output of step S141 is used for subsequent 2D to 3D orientation conversion. By extracting local orientations from the texture orientation field data, a corresponding wood grain orientation input can be established for each 3D surface region.

[0096] Step S142: Based on the correspondence between the wood surface image data and the wood three-dimensional surface data, determine the three-dimensional surface regions corresponding to multiple local regions.

[0097] In this embodiment, the correspondence between wood surface image data and wood 3D surface data can be established through camera intrinsic parameter calibration, camera extrinsic parameter calibration, hand-eye calibration, feature point matching, or texture mapping. Camera intrinsic parameter calibration can obtain the camera intrinsic parameter matrix and lens distortion coefficients; camera extrinsic parameter calibration can establish the rotation matrix and translation vector between the camera coordinate system and the workpiece coordinate system; for a camera mounted at the end of a robotic arm, a fixed transformation matrix between the camera coordinate system and the end-of-arm coordinate system can be established through hand-eye calibration; for a workpiece with an existing 3D mesh model, a mapping table between 2D image pixels and 3D surface points can be established through texture coordinate mapping or feature point matching.

[0098] In one embodiment, for the center pixel or representative pixel of each local region, the coordinates of the corresponding 3D surface point can be looked up based on the camera's intrinsic and extrinsic parameters and 3D surface data. If a precise 3D point does not exist at the location corresponding to a certain 2D pixel, the point closest to the projection ray on the 3D surface can be taken, or interpolation can be performed within adjacent triangular facets. The 3D surface region can be a 3D surface point, a triangular facet, or a local surface patch. When a local region corresponds to multiple 3D surface points, the set of these points or the fitted surface patch can be used as the 3D surface region corresponding to that local region.

[0099] A correspondence is established between local regions of a 2D image and 3D curved surface regions, enabling 2D orientation angles to be mapped onto the actual grinding surface. This correspondence ensures that subsequent robot arm path points have clear 3D positions in both the workpiece coordinate system and the robot arm base coordinate system.

[0100] Step S143: Determine the local tangent plane corresponding to the three-dimensional surface region, map the local texture direction data corresponding to the local region to the corresponding local tangent plane, and obtain the three-dimensional wood grain direction vector in the local tangent plane.

[0101] In this embodiment, the local tangent plane can be determined by three-dimensional surface points and surface normal vectors. The surface normal vector can be obtained through point cloud neighborhood plane fitting, triangular mesh normal interpolation, or surface fitting. For example, for point cloud data, a predetermined number of neighboring points of each three-dimensional point can be subjected to least-squares plane fitting, and the normal vector of the fitted plane can be used as the normal vector of that point; for triangular mesh data, the normal vectors of the triangular facets can be weighted by area or angle and interpolated to the vertices, and then interpolated to any point within the facets by centroid coordinates; for non-uniform rational spline (NURBS) surfaces or quadratic surfaces, the normal vector can be calculated using the surface analytical expression. Furthermore, two orthogonal tangent basis vectors can be constructed within the tangent plane, using the three-dimensional surface points as the origin and the normal vectors as the normals of the local tangent plane. Subsequently, based on the rotation relationship between the two-dimensional image coordinate axes and the tangent plane basis vectors, the two-dimensional direction angle can be converted into a local tangent plane direction angle, and the three-dimensional wood grain direction vector can be obtained based on this local tangent plane direction angle. The three-dimensional wood grain direction vector lies in the local tangent plane, is orthogonal to the surface normal vector, and is consistent with or approximately consistent with the wood grain direction at that point.

[0102] In another embodiment, the center pixel of the local area and the auxiliary pixel offset by a preset pixel distance along the local texture direction can be taken, and the corresponding three-dimensional surface points can be determined according to the correspondence between the wood surface image data and the wood three-dimensional surface data. The difference vector between the two three-dimensional surface points is projected onto the local tangent plane and normalized to obtain the three-dimensional wood grain direction vector in the local tangent plane.

[0103] Even if the image acquisition involves perspective projection or the wood surface is curved, the two-dimensional texture direction can be accurately converted into directional constraints on the three-dimensional curved surface. The output three-dimensional wood grain direction vector can be used to constrain the feed direction of the robotic arm tool. For curved wood surfaces, the three-dimensional wood grain direction vector will change continuously or abruptly locally with the change of surface position, allowing the robotic arm end effector to adjust its movement direction in three-dimensional space to follow the change of wood grain direction.

[0104] Step S144: Generate three-dimensional sanding direction constraint data corresponding to the local area based on the three-dimensional wood grain direction vector.

[0105] In this embodiment, the 3D sanding direction constraint data may include 3D surface point coordinates, surface normal vector, 3D wood grain direction vector, tool feed direction constraint, tool posture constraint, and confidence information. The tool feed direction constraint can represent the robotic arm end effector feeding along the 3D wood grain direction vector or its opposite direction. Since the wood grain direction has 180-degree equivalence, the 3D wood grain direction vector and its opposite direction can both be used as the grain direction. The specific feed direction can be determined based on path continuity, robotic arm reachability, and the path start and end points.

[0106] In one embodiment, the tool axis of the end-effector sanding tool can maintain a preset angle with the surface normal vector. For example, the tool axis can be perpendicular to the wood surface, i.e., parallel or antiparallel to the surface normal vector; or it can maintain an angle of inclination of 5 to 15 degrees with the surface normal vector to adapt to specific sanding processes. The tool feed direction can be consistent with or approximately consistent with the three-dimensional wood grain direction vector, for example, with an angle not exceeding 5 degrees. The tool lateral axis can be determined by the cross product of the tool axis and the tool feed direction, thereby forming the attitude constraint of the end-effector sanding tool. The robotic arm controller can obtain the joint angles that satisfy the above position and attitude constraints through inverse kinematics. The three-dimensional sanding direction constraint data is entered into the path parameter generation process. By simultaneously constraining the tool feed direction and tool attitude, the robotic arm can not only sand along the grain in planar areas but also sand along the three-dimensional wood grain direction in curved areas, thereby reducing the directional deviation caused by directly applying two-dimensional directions to curved wood surfaces.

[0107] Figure 5 This is a flowchart illustrating the process of regionalizing the path and process parameters for the nodule region provided in an embodiment of this application. Figure 5 As shown, when the region type label data includes scar region label data, the regionalization adjustment in step S150 may include steps S151 to S154.

[0108] In this embodiment, regional adjustment is used to locally update path segments and process parameters related to special texture regions in the grain-following polishing path data. Path segments generated based on region type label data and adapted to the texture features of corresponding special regions can be called region-specific path segments; path segments formed by associating region-specific path segments with 3D polishing direction constraint data, surface normals, and tool poses are called target polishing path segments. Target polishing path segments can be used to replace, insert, or smoothly connect path segments to be adjusted in the grain-following polishing path data.

[0109] Step S151: Determine the knot area based on the knot area label data, and determine the path segment to be adjusted corresponding to the knot area from the grinding path data along the grain.

[0110] In this embodiment, the knot region label data may include at least one of the following: knot center, knot boundary, knot radius, orientation variation characteristics, knot region confidence, and knot region identifier. The boundary of the knot region may be represented by a two-dimensional image boundary contour, a three-dimensional surface boundary point sequence, or a local curved surface closed boundary. The grain-following grinding path data may include multiple path segments, each path segment including multiple path points and the corresponding three-dimensional coordinates, surface normal vector, path tangential direction, tool posture, and grinding process parameters of each path point.

[0111] In one embodiment, the path segment to be adjusted can be determined based on the spatial overlap between the boundary of the knot region and the grinding path data along the grain. Specifically, when one or more path points in the grinding path segment along the grain are located within the boundary of the knot region, or when the path segment intersects with the boundary of the knot region, the path segment can be determined as the path segment to be adjusted corresponding to the knot region.

[0112] Step S151 allows for locating the area of ​​knots that needs regional adjustment within the existing grain-following polishing path data. This ensures that subsequent processing only affects the knot area and its adjacent transition area, avoiding unnecessary changes to the grain-following path in the regular area.

[0113] Step S152: Based on the directional circumferential change characteristics corresponding to the scar region, generate multiple concentric ring path segments with set intervals.

[0114] In this embodiment, dedicated path segments for the node region can be generated based on the node center, node boundary, and directional variation features. These dedicated path segments can include multiple concentric ring path segments spaced apart, such that the tangential direction of the path extends along the surrounding texture direction of the node region.

[0115] In one embodiment, the ring spacing between adjacent concentric ring path segments can be determined with the node center as the center and the outer radius corresponding to the node boundary as the maximum radius, based on the effective width of the sand disc and the path overlap rate.

[0116] In another embodiment, when the node boundary is not a regular circle, multiple closed path segments offset inward or outward can be generated based on the node boundary, so that the multiple closed path segments are distributed around the center of the node and approximately form concentric ring path segments.

[0117] Step S153: Based on the three-dimensional grinding direction constraint data corresponding to the nodule region, convert multiple concentric ring path segments into target grinding path segments corresponding to the nodule region.

[0118] In this embodiment, sampling points for concentric ring path segments can be generated first in the two-dimensional image coordinate system, the local parameter domain of the workpiece, or the local plane of the knot region. Then, based on the correspondence between the wood surface image data and the three-dimensional surface data of the wood, each sampling point is mapped to the three-dimensional surface of the wood to obtain the three-dimensional coordinates of the path points located on the three-dimensional surface of the knot region. For each path point, the tool axis direction can be determined based on the surface normal vector of the three-dimensional surface region where the path point is located, and the tool feed direction can be determined based on the path tangential direction of the concentric ring path segment and the three-dimensional grinding direction constraint data.

[0119] In one embodiment, the target grinding path segment may include multiple path points and the corresponding three-dimensional coordinates, surface normal vector, path tangential direction, tool axis direction, tool feed direction, path point connection order, path segment type, and associated knot region labels for each path point. Thus, the knot region path not only adapts to the surrounding texture in two-dimensional form but also satisfies surface tangential motion and tool posture constraints on the three-dimensional surface.

[0120] Step S154: Update the path segment to be adjusted based on the target grinding path segment, and adjust the grinding process parameters corresponding to the nodule area to obtain the target grinding path data and the corresponding grinding process parameter data.

[0121] In this embodiment, the path segment to be adjusted corresponding to the knot area in the grain-following polishing path data can be replaced with the target polishing path segment, or the target polishing path segment can be inserted into the path segment to be adjusted, and the part of the original path that is not suitable for the texture direction of the knot area can be deleted.

[0122] In one embodiment, the polishing process parameters corresponding to the knot area can be adjusted to lower downforce, lower feed rate, finer sandpaper grit, or a flexible tool head. For example, the downforce in the knot area can be reduced to 50% to 90% of the conventional downforce, the feed rate reduced to 60% to 90% of the conventional rate, and the sandpaper grit increased by one or two levels, for example, from 240 grit to 320 grit or 400 grit. If necessary, the rigid sanding disc can be replaced with a flexible grinding wheel or a flexible polishing wheel to reduce the risk of localized over-polishing caused by changes in hardness and abrupt texture changes in the knot area. This reduces transverse damage, localized burns, or over-polishing in the knot area and improves the polishing uniformity of the surface near the knot.

[0123] Figure 6 This is a flowchart illustrating the process of regionalizing the adjustment of the splicing joint area path and process parameters provided in the embodiments of this application. Figure 6 As shown, when the area type label data includes panel joint area label data, the area adjustment in step S150 may include steps S161 to S164.

[0124] Step S161: Determine the panel joint area based on the panel joint area label data, and determine the path segment to be adjusted corresponding to the panel joint area from the grain sanding path data.

[0125] In this embodiment, the panel seam area label data may include at least one of the following: seam line position, seam line direction, seam area boundary, local texture direction on both sides of the seam, direction difference on both sides of the seam, and panel seam area confidence level. The panel seam area may be a linear area extending along the seam line, or a strip-shaped area extending to both sides from the seam line as the center.

[0126] In one embodiment, the path segment to be adjusted can be determined based on the overlap between the boundary of the panel joint area and the grain-following sanding path data. Specifically, when a path segment crosses the joint line, enters the boundary of the panel joint area, or when the difference in local texture direction on both sides of the path segment is greater than a preset direction change threshold, the path segment can be identified as the path segment to be adjusted corresponding to the panel joint area. Thus, the path range requiring direction transition and process parameter adjustment at the joint can be located in the grain-following sanding path data, preventing the robotic arm from suddenly crossing two boards with different texture directions along a single path direction.

[0127] Step S162: Determine the joint transition area based on the panel joint area.

[0128] In this embodiment, the joint transition area provides space for the gradual change of path direction between different wood grain directions on both sides of the joint. The joint transition area can be a strip-shaped area extending a predetermined distance to both sides from the joint line of the panels. The path direction within this strip-shaped area no longer directly adopts a single wood grain direction on either side of the joint, but rather transitions continuously between the wood grain directions on both sides.

[0129] In one embodiment, a transition area for the joint can be formed by extending 5 mm to 30 mm to both sides of the joint line. The extension distance can be determined based on the directional differences between the two sides of the joint, the effective width of the sanding disc, the curvature of the wood surface, and the sanding quality requirements. When the directional differences between the two sides of the joint are large or the effective width of the sanding disc is large, the width of the transition area can be increased; when the directional differences between the two sides of the joint are small or the space around the joint is limited, the width of the transition area can be decreased. Thus, the robotic arm end effector gradually changes its feed direction and tool posture as it enters the joint area, passes through the center of the joint, and leaves the joint area, rather than switching instantaneously along the joint line.

[0130] Step S163: Based on the three-dimensional grinding direction constraint data corresponding to both sides of the splicing joint area, generate the direction gradient path segment in the joint transition area.

[0131] In this embodiment, the three-dimensional sanding direction constraint data corresponding to both sides of the panel joint area may include the three-dimensional wood grain direction vector of the board on one side of the joint and the three-dimensional wood grain direction vector of the board on the other side of the joint. The direction gradient path segment is a region-specific path segment corresponding to the panel joint area, used to smoothly transition from the three-dimensional wood grain direction on one side to the three-dimensional wood grain direction on the other side within the joint transition area.

[0132] In one embodiment, the two sides outside the joint transition area can each generate a grain-following path segment based on their respective three-dimensional sanding direction constraint data. Specifically, the grain-following path segment on one side of the joint is generated according to the three-dimensional wood grain direction vector of that side of the board, and the grain-following path segment on the other side of the joint is generated according to the three-dimensional wood grain direction vector of the other side of the board. Subsequently, a direction-gradient path segment is generated within the joint transition area based on the three-dimensional wood grain direction vectors on both sides to achieve a continuous connection between the grain-following path segments on both sides.

[0133] In one embodiment, for any path point within the joint transition area, the directional interpolation weight can be determined based on the normalized distance of the path point relative to the two boundaries of the joint transition area. When the path point is close to the first side boundary, the transition direction is close to the three-dimensional wood grain direction of the first side panel; when the path point is close to the second side boundary, the transition direction is close to the three-dimensional wood grain direction of the second side panel; when the path point is located near the center of the joint, the transition direction is determined based on the three-dimensional wood grain directions on both sides and the interpolation weight.

[0134] The direction gradient path segment can include the three-dimensional coordinates of each path point within the seam transition area, the surface normal vector, the transition three-dimensional direction vector, the tool axis direction, the tool feed direction, and the path point connection sequence.

[0135] Step S164: Update the path segment to be adjusted based on the direction gradient path segment, and adjust the grinding process parameters corresponding to the splicing joint area to obtain the target grinding path data and the grinding process parameter data corresponding to the target grinding path data.

[0136] In this embodiment, similarly, a direction-gradient path segment can be used to replace the path segment to be adjusted that passes through the seam transition area in the grain-following grinding path data, or the direction-gradient path segment can be spliced ​​with the grain-following path segments on both sides of the seam. The grinding process parameters corresponding to the splicing seam area can include the feed speed, downforce, path overlap rate, and path smoothing constraint near the seam center. For example, the robotic arm can begin to decelerate linearly before entering the seam transition area, reduce the feed speed to 50% to 70% of the normal speed near the seam center, and then linearly restore the normal speed after leaving the seam transition area. The downforce can be reduced or kept to change smoothly near the seam center to avoid over-grinding caused by local height differences at the seam.

[0137] By following the steps above, we can reduce the polishing marks caused by abrupt changes in the texture direction on both sides of the seam, and reduce the impact of abrupt changes in the trajectory direction of the robotic arm on the stability of polishing.

[0138] Figure 7 This is a flowchart illustrating the process of adjusting the path and process parameters of the low-confidence region and / or end-face texture region provided in the embodiments of this application. Figure 7 As shown, when the region type label data includes at least one of low confidence region label data and end face texture region label data, the regionalization adjustment in step S150 may include steps S171 to S173.

[0139] Step S171: Determine the region to be adjusted based on at least one of the low confidence region label data and the end face texture region label data, and determine the path segment to be adjusted corresponding to the region to be adjusted from the grain grinding path data.

[0140] In this embodiment, the low-confidence region is a region with low reliability in texture direction estimation, and the end-face texture region is a region with dense texture direction, disordered direction, or where it is difficult to determine a stable longitudinal direction. The low-confidence region label data may include at least one of the following: low-confidence region boundary, region area, average confidence level, lowest confidence level, and low-confidence region identifier. The end-face texture region label data may include at least one of the following: end-face region boundary, degree of direction disorder, average confidence level, end-face region identification confidence level, and end-face region identifier.

[0141] In one embodiment, the path segment to be adjusted can be determined based on the overlap between the boundary of the area to be adjusted and the grain-following polishing path data. Specifically, when a path point is located within a low-confidence region or an end-face texture region, or when a path segment passes through a low-confidence region or an end-face texture region, the path segment or a portion of the path segment located within the area to be adjusted can be identified as the path segment to be adjusted. This allows the path segments corresponding to low-confidence regions and end-face texture regions to be separated from ordinary grain-following paths, avoiding the forced adoption of a grain-following direction when texture direction estimation is unreliable or the texture direction is chaotic, thus preventing incorrect paths.

[0142] Step S172: When it is determined that the area to be adjusted is a low-confidence area or an end-face texture area, the path adjustment strategy corresponding to the area to be adjusted is switched to a conservative coverage path strategy, and the grinding process parameter adjustment strategy corresponding to the area to be adjusted is switched to a conservative grinding parameter strategy.

[0143] In this embodiment, a conservative coverage path strategy is used to prioritize coverage integrity, path smoothness, and low-damage polishing when the reliability of the orientation field is insufficient or it is difficult to determine a stable grain direction. The conservative coverage path strategy can include at least one of the following: contour-following path, uniform coverage path, low-speed dense coverage path, short-stroke reciprocating coverage path segment, and flexible polishing path. For low-confidence regions, uniform coverage path segments or low-speed dense coverage path segments can be generated based on the boundaries of the low-confidence regions, ensuring that the path primarily meets the requirements of coverage integrity and posture smoothness, rather than directly using local directions with low confidence in the orientation field. For end-face texture regions, contour-following path segments or short-stroke reciprocating coverage path segments can be generated based on the outer contour of the end-face region, and the direction changes between adjacent path segments can be controlled to be gradual, thereby reducing the risk of scratches caused by chaotic end-face texture orientation.

[0144] In one embodiment, a conservative polishing parameter strategy may include reducing downforce, reducing feed rate, increasing sandpaper grit, increasing path overlap, reducing path spacing, or using a flexible grinding wheel or flexible polishing wheel. For example, downforce in low-confidence areas or end-face texture areas may be reduced to 50% to 70% of the conventional downforce, feed rate may be reduced to 50% to 70% of the conventional rate, and finer sandpaper or a flexible polishing wheel may be used for polishing.

[0145] Step S173: Update the path segment to be adjusted based on the conservative coverage path strategy, and adjust the grinding process parameters corresponding to the area to be adjusted based on the conservative grinding parameter strategy to obtain the target grinding path data and the corresponding grinding process parameter data.

[0146] In this embodiment, a conservative coverage path segment corresponding to the area to be adjusted can be generated based on a conservative coverage path strategy, and this conservative coverage path segment can be used as a region-specific path segment for the area to be adjusted. Subsequently, the conservative coverage path segment can be mapped onto the three-dimensional surface based on the three-dimensional surface data of the wood, and combined with the surface normal vector of the corresponding path point, the path tangential direction, and the tool posture constraint to form the target sanding path segment corresponding to the area to be adjusted.

[0147] In one embodiment, the target polishing path segment can be used to replace the path segment to be adjusted located in the low confidence region or end face texture region in the parallel polishing path data. Alternatively, the target polishing path segment can be inserted into the path segment to be adjusted, and the part of the original path that depends on the low confidence direction field generation can be deleted.

[0148] In one embodiment, the updated target sanding path data may include normal grain path segments, conservative overlay path segments, and transition path segments. The sanding process parameters corresponding to the conservative overlay path segments may employ low pressure, low speed, finer sandpaper, higher path overlap, or a flexible tool head. This reduces the risk of sanding damage in areas with low reliability in direction estimation or chaotic end-face texture, and improves the overall consistency of the surface finish.

[0149] In some embodiments, when the region type label data includes texture bifurcation region label data, the texture bifurcation region can be determined based on the texture bifurcation region label data, and the path segment to be adjusted corresponding to the texture bifurcation region can be determined from the grain-following grinding path data. The texture bifurcation region can be divided into multiple sub-regions according to the multiple main branch directions within the texture bifurcation region, and local grain-following path segments can be generated for each sub-region based on the corresponding 3D grinding direction constraint data. Near the bifurcation center, short path transition segments, low-speed coverage path segments, or conservative coverage path segments can be generated, and the downforce and feed rate can be reduced to minimize directional abrupt changes caused by path bifurcation. When the texture bifurcation region overlaps with knot regions, panel joint regions, low-confidence regions, or end-face texture regions, a path adjustment strategy and grinding process parameter adjustment strategy can be selected based on preset region priorities.

[0150] In some embodiments, a strategy mapping relationship can be formed between region type label data, path adjustment strategy, and grinding process parameter adjustment strategy. Specifically, regular regions can use a smooth path with consistent grain, and employ conventional feed speed, conventional downforce, and conventional path spacing; knot regions can use concentric ring path segments, and reduce downforce, reduce feed speed, increase sandpaper grit, or use a flexible tool head; panel joint regions can use a path segment with gradually changing direction, and reduce feed speed and smoothly adjust downforce near the joint center; low-confidence regions can use a conservative coverage path segment, and reduce downforce, reduce feed speed, and increase path overlap rate; end-face texture regions can use a uniform coverage path segment, a contour-following path segment, or a flexible polishing path segment, and employ finer sandpaper or a flexible polishing wheel. Through this strategy mapping relationship, region type label data can be transformed into a basis for differentiated control of robotic arm paths and grinding parameters.

[0151] In some embodiments, when the same local area corresponds to multiple area type labels, the final path adjustment strategy and grinding process parameter adjustment strategy can be determined based on preset area priorities. For example, when a knot area overlaps with a low-confidence area, a concentric ring path segment corresponding to the knot area can be used, superimposed with a conservative grinding parameter strategy corresponding to the low-confidence area; when a panel joint area overlaps with a low-confidence area, a direction-gradient path segment corresponding to the panel joint area can be used, with reduced feed speed and downward pressure; when an end-face texture area overlaps with a low-confidence area, a conservative coverage path strategy and a conservative grinding parameter strategy can be preferentially adopted. By setting area priorities, conflicts between multiple strategies can be avoided.

[0152] In one specific application example, the workpiece to be sanded is a panel tabletop. The surface of this panel tabletop has a panel joint, and a knot area exists on one side of the joint. For the regular area, the system can use streamline integration to generate sanding path segments along the grain. For the knot area, the system generates multiple concentric annular path segments centered on the knot center, converting these concentric annular path segments into the target sanding path segments corresponding to the knot area. For the panel joint area, the system generates along-grain path segments on both sides of the joint according to their respective three-dimensional wood grain directions, and forms a joint transition area centered on the joint line. Within the joint transition area, path segments with gradually changing directions are generated according to the three-dimensional wood grain directions on both sides of the joint.

[0153] Finally, the system merges the regular grain path segments, the target sanding path segments in the knot area, the directional gradient path segments in the panel joint area, and the transition path segments between each area into target sanding path data, and configures corresponding sanding process parameter data for each path segment for sanding. This allows for simultaneous processing of regular wood grain, knot texture, and panel joint texture on the same wood workpiece, thereby reducing transverse scratches and path abrupt changes and improving the quality of automated sanding.

[0154] Figure 8 This is a schematic diagram of the structure of a parallel-texture polishing device 800 based on a texture direction field, provided in an embodiment of this application. Figure 8 As shown, the device 800 may include a data acquisition module 810, a direction field determination module 820, a region type identification module 830, a direction constraint generation module 840, and a path parameter generation module 850.

[0155] The data acquisition module 810 is used to acquire wood surface image data and wood three-dimensional surface data. In some embodiments, the wood surface image data output by the data acquisition module 810 can be sent to the orientation field determination module 820, and the output wood three-dimensional surface data can be sent to the orientation constraint generation module 840.

[0156] The orientation field determination module 820 is used to determine texture orientation field data based on wood surface image data. The texture orientation field data output by the orientation field determination module 820 can be sent to the region type recognition module 830 and the orientation constraint generation module 840.

[0157] The region type recognition module 830 is used to identify the region type of multiple local regions based on texture orientation field data and generate region type label data. The region type label data output by the region type recognition module 830 can be sent to the path parameter generation module 850.

[0158] The orientation constraint generation module 840 is used to perform three-dimensional orientation transformation on the texture orientation field data based on the three-dimensional surface data of wood, to obtain three-dimensional sanding orientation constraint data corresponding to multiple local regions. The three-dimensional sanding orientation constraint data output by the orientation constraint generation module 840 can be sent to the path parameter generation module 850.

[0159] The path parameter generation module 850 is used to generate parallel grinding path data based on the three-dimensional grinding direction constraint data, and to perform regional adjustments on the parallel grinding path data and the corresponding grinding process parameters according to the region type label data, so as to obtain the target grinding path data and the grinding process parameter data corresponding to the target grinding path data.

[0160] In some embodiments, the device 800 may further include a control output module 860. The control output module 860 can be used to convert target grinding path data and grinding process parameter data into robotic arm control commands and output them to the robotic arm controller. The robotic arm control commands may include path point pose sequences, feed speed commands, downforce control commands, tool head switching commands, and sandpaper grit number prompts.

[0161] It should be noted that the device 800 in this embodiment can be used to execute any of the above method embodiments. The data acquisition module 810 can be used to execute step S110, the direction field determination module 820 can be used to execute step S120, the region type identification module 830 can be used to execute step S130, the direction constraint generation module 840 can be used to execute step S140, and the path parameter generation module 850 can be used to execute step S150. Each of the above modules can be implemented as a software program, a hardware circuit, or a combination of software and hardware.

[0162] Figure 9 This is a schematic diagram of the structure of an electronic device 900 provided in an embodiment of this application. Figure 9 As shown, the electronic device 900 may include a processor 910 and a memory 920. The memory 920 is communicatively connected to the processor 910 and stores a computer program. When the processor 910 executes the computer program, it can implement the texture direction field-based parallel grinding method in any of the above method embodiments.

[0163] In this embodiment, the electronic device 900 can be a server, industrial computer, edge computing device, robotic arm controller, workstation, or other device with data processing capabilities. The processor 910 can include one or more of a central processing unit (CPU), graphics processing unit (GPU), neural network processor (NPU), or industrial control processor. The memory 920 can store orientation field estimation models, calibration parameters, process parameter libraries, path planning programs, wood surface image data, wood three-dimensional surface data, and sanding process records. The electronic device 900 may also include a communication interface 930 and a bus 940; the processor 910, memory 920, and communication interface 930 can be connected via the bus 940. The communication interface 930 can be used to communicate with image acquisition devices, three-dimensional scanning devices, robotic arm controllers, force sensors, databases, or host computer systems. It should be noted that... Figure 9 The structure shown is merely an example and does not constitute a limitation on the hardware structure of the electronic device 900.

[0164] This application also provides a computer-readable storage medium. The computer-readable storage medium stores a computer program. When executed by a processor, the computer program can implement the texture direction field-based parallel-texture polishing method in any of the above method embodiments. The computer-readable storage medium may include a read-only memory, random access memory, flash memory, solid-state drive, hard disk, portable hard drive, optical disk, portable storage disk, memory card, or other media capable of storing a computer program.

[0165] The above embodiments can be implemented individually or combined with each other without conflict. Without departing from the technical concept of this application, any embodiment that still generates target grinding path data and grinding process parameter data based on texture direction field data, region type label data, and three-dimensional grinding direction constraint data should be understood as an optional embodiment of this application.

Claims

1. A method for parallel sanding based on texture direction field, characterized in that, include: Acquire image data and three-dimensional surface data of wood; Based on the wood surface image data, texture direction field data is determined, and the texture direction field data is used to characterize the local texture direction corresponding to multiple local areas of the wood surface. Based on the texture orientation field data, region type identification is performed on multiple local regions to generate region type label data; Based on the three-dimensional surface data of the wood, the texture direction field data is transformed into three-dimensional directions to obtain three-dimensional sanding direction constraint data corresponding to multiple local regions; Based on the three-dimensional grinding direction constraint data, parallel grinding path data is generated, and the parallel grinding path data and corresponding grinding process parameters are regionally adjusted according to the region type label data to obtain target grinding path data and grinding process parameter data corresponding to the target grinding path data.

2. The parallel-texture polishing method based on texture direction field according to claim 1, characterized in that, The determination of texture orientation field data based on the wood surface image data includes: The wood surface image data is input into a pre-trained orientation field estimation model, which includes a multi-scale texture feature extraction unit and an output unit connected to the multi-scale texture feature extraction unit. The multi-scale texture feature extraction unit extracts multi-scale texture features from the wood surface image data. The output unit outputs an orientation angle map and a confidence map based on the multi-scale texture features. Based on the orientation angle values ​​corresponding to the multiple local regions in the orientation angle map, local texture orientation data corresponding to the multiple local regions are determined, and based on the confidence values ​​corresponding to the multiple local regions in the confidence map, texture orientation confidence data corresponding to the multiple local regions are determined. The texture direction field data is generated based on the local texture direction data and texture direction confidence data corresponding to the multiple local regions respectively.

3. The parallel-texture polishing method based on texture direction field according to claim 1, characterized in that, The step of identifying the region type of multiple local regions based on the texture orientation field data and generating region type label data includes: Obtain local texture direction data and texture direction confidence data corresponding to multiple local regions from the texture direction field data; For the multiple local regions, at least one of the following region types is identified: Based on the directional abrupt change features of the local texture direction data of adjacent local regions, the panel seam area is identified, and panel seam area label data is generated. Based on the directional wrap-around variation features of the local texture direction data, the nodule region is identified, and nodule region label data is generated; Based on the directional chaos features and texture direction confidence data of the local texture direction data, the end face texture region is identified, and end face texture region label data is generated. Based on the directional bifurcation features of the local texture direction data, identify texture bifurcation point regions and generate texture bifurcation region label data; Based on the texture direction confidence data, low-confidence regions are identified, and low-confidence region label data is generated; The region type label data is generated based on at least one of the panel joint area label data, the knot area label data, the end face texture area label data, the texture bifurcation area label data, and the low confidence area label data.

4. The parallel-texture polishing method based on texture direction field according to claim 1, characterized in that, The step of performing a three-dimensional orientation transformation on the texture direction field data based on the three-dimensional surface data of the wood to obtain three-dimensional sanding direction constraint data corresponding to multiple local regions includes: Obtain local texture direction data corresponding to multiple local regions from the texture direction field data; Based on the correspondence between the wood surface image data and the wood three-dimensional surface data, the three-dimensional surface regions corresponding to the multiple local regions are determined respectively; Determine the local cutting plane corresponding to the three-dimensional surface region, and map the local texture direction data corresponding to the local region to the corresponding local cutting plane to obtain the three-dimensional wood grain direction vector in the local cutting plane; Based on the three-dimensional wood grain direction vector, generate three-dimensional sanding direction constraint data corresponding to the local area.

5. The parallel-texture polishing method based on texture direction field according to claim 3, characterized in that, The region type label data includes the knot / scar region label data. The step of regionally adjusting the grain-following polishing path data and corresponding polishing process parameters based on the region type label data to obtain target polishing path data and polishing process parameter data corresponding to the target polishing path data includes: Based on the knot area label data, the knot area is determined, and the path segment to be adjusted corresponding to the knot area is determined from the grain grinding path data; Based on the directional circumferential change characteristics corresponding to the nodular region, multiple concentric ring path segments with intervals are generated. Based on the three-dimensional grinding direction constraint data corresponding to the nodule region, the multiple concentric ring path segments are converted into target grinding path segments corresponding to the nodule region. The path segment to be adjusted is updated based on the target grinding path segment, and the grinding process parameters corresponding to the knot area are adjusted to obtain the target grinding path data and the corresponding grinding process parameter data.

6. The parallel-texture polishing method based on texture direction field according to claim 3, characterized in that, The region type label data includes the panel joint region label data. The step of regionally adjusting the grain-following sanding path data and corresponding sanding process parameters based on the region type label data includes: The panel joint area is determined based on the panel joint area label data, and the path segment to be adjusted corresponding to the panel joint area is determined from the grain sanding path data. Based on the panel joint area, a joint transition area is determined, and based on the three-dimensional grinding direction constraint data corresponding to both sides of the panel joint area, a direction gradient path segment is generated within the joint transition area. The path segment to be adjusted is updated based on the directional gradient path segment, and the grinding process parameters corresponding to the panel joint area are adjusted to obtain the target grinding path data and the grinding process parameter data corresponding to the target grinding path data.

7. The parallel-texture polishing method based on texture direction field according to claim 3, characterized in that, The region type label data includes at least one of low-confidence region label data and end-face texture region label data. The step of regionalizing the grain-following polishing path data and corresponding polishing process parameters based on the region type label data includes: The region to be adjusted is determined based on at least one of the low confidence region label data and the end face texture region label data, and the path segment to be adjusted corresponding to the region to be adjusted is determined from the grain-following polishing path data. When the area to be adjusted is determined to be a low-confidence area or an end-face texture area, the path adjustment strategy corresponding to the area to be adjusted is switched to a conservative coverage path strategy, and the polishing process parameter adjustment strategy corresponding to the area to be adjusted is switched to a conservative polishing parameter strategy. The path segment to be adjusted is updated based on the conservative coverage path strategy, and the grinding process parameters corresponding to the area to be adjusted are adjusted based on the conservative grinding parameter strategy to obtain the target grinding path data and the corresponding grinding process parameter data.

8. A parallel-texture polishing device based on a texture direction field, characterized in that, include: The data acquisition module is used to acquire image data of the wood surface and three-dimensional surface data of the wood; The orientation field determination module is used to determine texture orientation field data based on the wood surface image data, and the texture orientation field data is used to characterize the local texture orientation corresponding to multiple local areas of the wood surface. The region type recognition module is used to perform region type recognition on multiple local regions based on the texture orientation field data, and generate region type label data. The orientation constraint generation module is used to perform three-dimensional orientation transformation on the texture orientation field data based on the three-dimensional surface data of the wood to obtain three-dimensional sanding orientation constraint data corresponding to multiple local regions. The path parameter generation module is used to generate parallel grinding path data based on the three-dimensional grinding direction constraint data, and to perform regional adjustments on the parallel grinding path data and the corresponding grinding process parameters according to the region type label data to obtain target grinding path data and grinding process parameter data corresponding to the target grinding path data.

9. An electronic device, characterized in that, include: processor; A memory, communicatively connected to the processor, stores a computer program, and when the processor executes the computer program, implements the texture direction field-based parallel grinding method as described in any one of claims 1 to 7.

10. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program, which, when executed by a processor, implements the grain-following polishing method based on the texture direction field as described in any one of claims 1 to 7.

Images