A machine learning-based three-dimensional model object recognition method, device, and medium
By generating a manufacturing syntax dual-code description package and a manufacturing conformity certificate, and combining machine learning methods, the problem of weak geometric mapping accuracy and semantic correlation in 3D model object recognition is solved, achieving high-precision and interpretable 3D object recognition.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHANGYANG HANGRUN (XIAN) SYSTEM TECHNOLOGY CO LTD
- Filing Date
- 2026-04-02
- Publication Date
- 2026-06-26
AI Technical Summary
Existing methods for recognizing 3D model objects suffer from insufficient accuracy in 3D geometric mapping and weak semantic correlation in manufacturing, which affects the reliability of applications in industrial scenarios.
By generating a manufacturing syntax dual-code description package and combining it with machine learning methods, the topological constraint code and shape syntax code are integrated to generate a manufacturing consistency certificate, ensuring the overlap between semantic segmentation and geometric features, and achieving accurate recognition of 3D objects.
It enhances the interpretability and industrial applicability of 3D model recognition, and improves the geometric consistency and accuracy of the recognition results.
Smart Images
Figure CN122289659A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of 3D model recognition technology, and in particular to a method, device and medium for recognizing 3D model objects based on machine learning. Background Technology
[0002] In the field of 3D model object recognition, traditional methods typically rely on geometric feature extraction, multi-view analysis, or deep learning techniques to achieve object identification. Geometric feature-based methods perform matching by extracting local features from point clouds or meshes, while multi-view analysis methods utilize multi-view 2D projections of the 3D model, combined with image processing techniques, to identify object features. Furthermore, deep learning methods directly learn feature representations from 3D data using convolutional neural networks or graph neural networks. These methods have wide applications in industrial design, intelligent manufacturing, and other fields, playing a particularly important role in computer-aided design model retrieval, reverse engineering, and quality inspection.
[0003] Although existing methods have achieved some success in 3D object recognition, they still have limitations: traditional multi-view analysis methods usually rely only on 2D projection information for recognition, lacking a precise mapping mechanism with 3D geometric structures, which may lead to strong viewpoint dependence and insufficient spatial positioning accuracy; although deep learning-based recognition methods can automatically extract features, their interpretability is weak, making it difficult to effectively combine manufacturing semantics for high-precision object recognition, thus affecting the reliability of applications in industrial scenarios. Summary of the Invention
[0004] In view of the aforementioned existing problems, the present invention is proposed.
[0005] Therefore, this invention provides a machine learning-based method for recognizing 3D model objects to address the problems of insufficient accuracy in 3D geometric mapping and weak semantic relevance in manufacturing.
[0006] To solve the above-mentioned technical problems, the present invention provides the following technical solution: In a first aspect, the present invention provides a method for recognizing 3D model objects based on machine learning, comprising, Collect CAD 3D models, render them to generate standard CAD views, and write viewpoint identifiers. Then, summarize the CAD 3D models, standard CAD views, and viewpoint identifiers to generate a multi-view evidence set. Based on the multi-view evidence set, candidate regions are detected in the standard CAD view. The candidate regions are located to the CAD 3D model by the view backtracking localization method and geometric intersection is performed to generate a set of 3D ROI positioning volumes. Extract topological features from the three-dimensional ROI location volume set to generate topological constraint codes. Simultaneously, perform shape decomposition to generate shape syntax codes. Integrate the topological constraint codes and shape syntax codes to generate a manufacturing syntax dual-code description package. The manufacturing syntax dual-code description package is classified and interpreted to generate a three-dimensional semantic segment. The three-dimensional semantic segment is projected onto the CAD standard view and the overlap with the candidate region is calculated to generate a manufacturing conformity certificate. Extract the 3D semantic segment from the manufacturing conformity certificate, map the 3D semantic segment to the CAD 3D model, generate a direction annotation layer, and associate it with the manufacturing conformity certificate to generate a manufacturing semantic object recognition list.
[0007] As a preferred embodiment of the machine learning-based 3D model object recognition method of the present invention, the steps for generating the multi-view evidence set are as follows: Align the coordinates of the CAD 3D model, generate a view reference system, and collect a list of manufacturing feature categories; Based on the viewpoint reference frame, set the rendering directions for the main view direction, the right view direction, and the top view direction; Render the CAD 3D model, generate a standard CAD view, and write viewpoint identifiers; The viewpoint labels of the CAD standard views are organized to generate viewpoint coverage curves, and then summarized with the CAD 3D model, CAD standard views and viewpoint labels to generate a multi-view evidence set.
[0008] As a preferred embodiment of the machine learning-based 3D model object recognition method of the present invention, the steps for generating the 3D ROI location set are as follows: Convert standard CAD views to grayscale to generate grayscale images; Read the grayscale values from the pixels of the grayscale image, and set the horizontal direction of the grayscale image as the grayscale row direction and the vertical direction of the grayscale image as the grayscale column direction; Discrete the grayscale values and perform vector synthesis to generate the grayscale change amplitude and direction. The pixels of the standard CAD view are traversed to generate edge pixels, which are then summarized to generate an edge set. The edge set is scanned along the gray-level row direction and gray-level column direction, and consecutive adjacent edge pixels are merged into edge segments. The connecting edge segments are then connected to generate candidate regions. Read the viewpoint identifier, cut the CAD 3D model along the candidate region, and generate a single-view positioning volume group; Perform geometric intersection on the single-view positioning volume group to generate a fine positioning volume, and summarize the fine positioning volumes to generate a three-dimensional ROI positioning volume set.
[0009] As a preferred embodiment of the machine learning-based 3D model object recognition method of the present invention, the steps for generating the topological constraint code are as follows: According to the rendering direction, the finely positioned volumes of the 3D ROI positioning volume set are divided into three-dimensional mesh blocks according to the front view direction, right view direction, and top view direction. Define the adjacency relationship of the 3D mesh blocks, and extract the connectivity, number of loops, and number of cavities of the 3D mesh blocks based on the adjacency relationship; The connectivity, number of loops, and number of cavities are transcoded to generate connectivity code, number of loops code, and number of cavities code, which are then spliced and encapsulated to generate topology constraint code.
[0010] As a preferred embodiment of the machine learning-based 3D model object recognition method of the present invention, the steps for generating the manufacturing syntax dual-code description package are as follows: Arrange the external dimensions of the 3D mesh blocks to generate geometric principal axis feature vectors; Encode the geometric principal axis feature vectors to generate principal axis codes; The three-dimensional mesh block is projected and clipped along its projection outline. Simultaneously, the grayscale values of the projection outline of the three-dimensional mesh block are differentially calculated and accumulated item by item to generate shape spectrum coefficients. The shape spectral coefficients are encoded to generate spectral codes; The spindle code and the spectrum code are concatenated to generate the shape syntax code, which is then integrated with the topology constraint code to generate a manufacturing syntax dual-code description package.
[0011] As a preferred embodiment of the machine learning-based 3D model object recognition method of the present invention, the steps of classifying and interpreting the manufacturing syntax dual-code description package to generate 3D semantic segmentation are as follows: The topological constraint code and shape syntax code of the manufacturing syntax dual-code description package are concatenated to generate the interpretation input vector. The interpretation input vector is then judged by the topological constraint code to generate a category determination. The input vector is interpreted according to the front view direction, right view direction, and top view direction to generate front view direction interpretation, right view direction interpretation, and top view direction interpretation. The front view direction interpretation, right view direction interpretation, and top view direction interpretation are aligned according to the viewpoint identifier and merged according to the order of viewpoint coverage curve to generate three-dimensional semantic segmentation.
[0012] In a preferred embodiment of the machine learning-based 3D model object recognition method of the present invention, the steps for generating a manufacturing conformity certificate are as follows: The 3D semantic segmentation is projected onto the CAD standard view according to the viewpoint identifier to generate a semantic segmentation projection. The overlap between the semantic segmentation projection and the candidate region is calculated to generate overlap in the main view direction, right view direction, and top view direction. This overlap is then combined with the category determination and 3D semantic segmentation to generate a manufacturing consistency certificate.
[0013] As a preferred embodiment of the machine learning-based 3D model object recognition method of the present invention, the steps for generating the manufacturing semantic object recognition list are as follows: The single-view localizers of the single-view localizer group are located based on semantic segmentation projection, and the directional entity index is generated by counting the number of 3D mesh blocks covered by the single-view localizers. Numerical comparison and vector synthesis are performed on the directional entity indexes to generate CAD entity indexes. The CAD entity indexes are then bound to 3D mesh blocks and written into the CAD 3D model to generate the directional annotation layer. The directional labeling layer and manufacturing conformity certificate are sorted according to the view coverage curve to generate a list of manufacturing semantic objects.
[0014] In a second aspect, the present invention provides a computer device including a memory and a processor, wherein the memory stores a computer program, wherein when the computer program is executed by the processor, it implements any step of the machine learning-based 3D model object recognition method described in the first aspect of the present invention.
[0015] Thirdly, the present invention provides a computer-readable storage medium having a computer program stored thereon, wherein: when the computer program is executed by a processor, it implements any step of the machine learning-based 3D model object recognition method described in the first aspect of the present invention.
[0016] The beneficial effects of this invention are as follows: by generating a manufacturing syntax dual-code description package, the structured encoding of manufacturing semantics is realized, which enhances the interpretability and industrial applicability of 3D model recognition; by generating a manufacturing conformity certificate, the overlap between semantic segmentation and geometric features is verified, ensuring the geometric consistency of the recognition results and improving the accuracy and reliability of 3D object recognition. Attached Figure Description
[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. 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 This is a flowchart of a machine learning-based method for recognizing 3D model objects.
[0019] Figure 2A flowchart for generating a syntax double-code description package.
[0020] Figure 3 A flowchart for generating conformance certificates.
[0021] Figure 4 A flowchart for generating a list of semantic object recognition. Detailed Implementation
[0022] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings.
[0023] Many specific details are set forth in the following description in order to provide a full understanding of the invention. However, the invention may also be practiced in other ways different from those described herein, and those skilled in the art can make similar extensions without departing from the spirit of the invention. Therefore, the invention is not limited to the specific embodiments disclosed below.
[0024] Secondly, the term "one embodiment" or "embodiment" as used herein refers to a specific feature, structure, or characteristic that may be included in at least one implementation of the present invention. The phrase "in one embodiment" appearing in different places in this specification does not necessarily refer to the same embodiment, nor is it a single or selective embodiment that is mutually exclusive with other embodiments.
[0025] Reference Figures 1-4 This is one embodiment of the present invention, which provides a machine learning-based method for recognizing 3D model objects, including the following steps: S1. Collect the CAD 3D model, render it to generate a CAD standard view, and write viewpoint identifiers. Then, summarize the CAD 3D model, CAD standard view, and viewpoint identifiers to generate a multi-view evidence set.
[0026] Align the coordinates of the CAD 3D model, generate a viewpoint reference system, and collect a list of manufacturing feature categories.
[0027] Furthermore, the three-dimensional coordinate axes of the CAD 3D model are read, and the three principal axes and the origin of the three-dimensional coordinate axes are used as the viewpoint reference system.
[0028] Collect through holes and slots, blind holes and slots, closed cavities and stepped cavities, pipe channels and annular channels, assembly interfaces and adjacent feature sets from the CAD 3D model, and integrate them to generate a list of manufacturing feature categories.
[0029] Based on the viewpoint reference frame, the rendering direction is set, the CAD 3D model is rendered, a standard CAD view is generated, and viewpoint identifiers are written.
[0030] Furthermore, the three principal axes of the view reference system are extracted, and the rendering directions are set as the main view direction, the right view direction, and the top view direction, respectively. The CAD 3D model is rendered in the order of main view direction, right view direction, and top view direction to generate three standard CAD views.
[0031] It should be noted that the steps to generate three standard CAD views are as follows: Align the geometric center of the CAD 3D model with the origin of the view reference system; align the principal axis of the CAD 3D model with the current rendering direction; using the geometric center of the CAD 3D model as the center of a sphere, take the maximum distance from the center of the sphere to the vertex of the CAD 3D model as the radius of the sphere, and generate a framing reference sphere containing the CAD 3D model; orthogonally project the framing reference sphere onto the current rendering direction to generate a framing reference circle, and use the boundary of the framing reference circle as the framing range; based on the framing range, orthogonally project the CAD 3D model along the main view direction, the right view direction, and the top view direction as the current rendering direction to generate three CAD standard views, use the current rendering direction as the viewpoint identifier, record the time of generating the CAD standard views as the time stamp, and write the viewpoint identifier and time stamp into the CAD standard views.
[0032] The viewpoint labels of the CAD standard views are organized to generate viewpoint coverage curves, and then summarized with the CAD 3D model, CAD standard views and viewpoint labels to generate a multi-view evidence set.
[0033] Furthermore, the viewpoint identifiers and time markers of the three standard CAD views are recorded sequentially according to the rendering order, and then connected by polylines to generate a viewpoint coverage curve.
[0034] The view coverage curve, CAD 3D model, three standard CAD views, view markers, and time stamps are uniformly summarized and encapsulated to generate a multi-view evidence set.
[0035] S2. Based on the multi-view evidence set, candidate regions are detected in the standard CAD view. The candidate regions are located in the CAD 3D model by the view backtracking localization method and geometric intersection is performed to generate a set of 3D ROI positioning volumes.
[0036] The standard CAD view is converted to grayscale to generate a grayscale image. The grayscale value is read from the pixels of the grayscale image, and the horizontal direction of the grayscale image is set as the grayscale row direction and the vertical direction of the grayscale image is set as the grayscale column direction. The grayscale values are discretized and vectorized to generate the grayscale change amplitude and grayscale change direction.
[0037] Furthermore, the standard CAD view is converted to grayscale to generate a grayscale image. The grayscale value of each pixel in the grayscale image is read. The horizontal direction of the grayscale image is set as the grayscale row direction, and the vertical direction is set as the grayscale column direction. Discrete difference is performed on the grayscale value of each pixel in the grayscale row direction and the grayscale column direction to generate horizontal grayscale variation and vertical grayscale variation. The horizontal grayscale variation and vertical grayscale variation are vector synthesized to generate grayscale variation amplitude and grayscale variation direction.
[0038] The pixels of the standard CAD view are traversed to generate edge pixels, which are then summarized to create an edge set.
[0039] Furthermore, the grayscale change amplitude of the current pixel is extracted, and the grayscale change amplitudes of the adjacent pixels on both sides are extracted along the grayscale change direction of the current pixel. The grayscale change amplitude of the current pixel is compared with the grayscale change amplitudes of the adjacent pixels on both sides. If the grayscale change amplitude of the current pixel is less than the grayscale change amplitude of any adjacent pixel, then the grayscale change amplitude of the current pixel is set to zero.
[0040] The pixels in the standard CAD view are traversed, and pixels with non-zero grayscale variation are marked as edge pixels. The edge pixels are then aggregated to generate an edge set.
[0041] It should be noted that performing discrete difference analysis on the grayscale value of each pixel in the grayscale row direction and the grayscale column direction to generate horizontal and vertical grayscale changes means: extracting the grayscale value of the current pixel and the grayscale value of its right neighbor, and calculating the difference between the grayscale value of the current pixel and the grayscale value of the right neighbor to generate horizontal grayscale changes; extracting the grayscale value of the top neighbor pixel, and calculating the difference between the grayscale value of the current pixel and the grayscale value of the top neighbor to generate vertical grayscale changes.
[0042] It should be noted that extracting the grayscale change amplitude of adjacent pixels on both sides along the grayscale change direction of the current pixel means: moving the current pixel by one pixel along the grayscale change direction and in the opposite direction of the grayscale change direction, respectively, and taking them as adjacent pixels.
[0043] The edge set is scanned to generate edge segments, and the adjacent edge segments are connected to generate candidate regions.
[0044] Furthermore, the edge set is scanned along the grayscale row direction and grayscale column direction, and consecutive adjacent edge pixels are merged into edge segments; the ends of all intersecting or end-connected edge segments are connected to generate candidate regions, and all candidate regions are summarized to generate a candidate region set.
[0045] Read the viewpoint identifier, cut the CAD 3D model along the candidate region, and generate a single-view positioning volume group.
[0046] Furthermore, the viewpoint identifier of the CAD standard view to which the candidate region belongs is read. Using the outline of the candidate region as the boundary, the viewpoint reference sphere is penetrated along the rendering direction indicated by the viewpoint identifier to generate a viewpoint extrusion. The viewpoint extrusion and the CAD 3D model are placed in the same coordinate system, and the non-overlapping parts of the CAD 3D model are trimmed to generate a single-view positioning volume. The process of generating single-view positioning volumes for the same candidate region is repeated in three CAD standard views according to the order of the viewpoint coverage curve to generate a single-view positioning volume group.
[0047] Perform geometric intersection on the single-view positioning volume group to generate a fine positioning volume, and summarize the fine positioning volumes to generate a three-dimensional ROI positioning volume set.
[0048] The three single-view positioning bodies of the single-view positioning body group are placed in the same coordinate system, and the intersecting part of the three single-view positioning bodies is extracted as the fine positioning body.
[0049] All precise positioning objects are compiled and encapsulated to generate a 3D ROI positioning object set.
[0050] It should be noted that when three single-view positioning volumes have no intersecting parts, or only two single-view positioning volumes intersect, a fine positioning volume will not be generated.
[0051] It should be noted that ROI (Region of Interest) refers to a precisely located area.
[0052] S3. Extract topological features from the three-dimensional ROI positioning volume set, generate topological constraint codes, and simultaneously perform shape decomposition to generate shape syntax codes. Integrate the topological constraint codes and shape syntax codes to generate a manufacturing syntax dual-code description package.
[0053] According to the rendering direction, the fine positioning volume of the 3D ROI positioning volume set is divided into three parts to generate 3D mesh blocks.
[0054] Furthermore, each fine-positioning object in the 3D ROI positioning object set is divided according to the main view direction, right view direction, and top view direction to generate a division direction. First, the main view direction is used as the main view division direction to divide the fine-positioning object: the diameter of the view reference sphere is read, and the number of pixels of the view reference circle diameter in the main view division direction is read; then, the span pixel count of the candidate region is counted, and the division spacing is calculated based on the diameter of the view reference sphere, the number of pixels of the view reference circle diameter, and the span pixel count. The expression is: ; in, To divide the spacing, The diameter of the reference sphere for framing. The number of pixels equal to the diameter of the reference circle for framing. This represents the number of pixels across the span.
[0055] Based on the main view division direction, the precision positioning body is divided according to the division interval. At the same time, the right view direction is used as the right view division direction to divide the precision positioning body, and the top view direction is used as the top view division direction to divide the precision positioning body.
[0056] A three-dimensional mesh block is generated by dividing each precise positioning volume three times.
[0057] It should be noted that the span pixel count refers to the number of left and right pixels between the leftmost and rightmost edge pixels of the candidate region outline along the grayscale row direction; and the number of top and bottom pixels between the topmost and bottommost edge pixels of the candidate region outline along the grayscale column direction; the smaller of the left and right pixel count and the top and bottom pixel count is taken as the span pixel count.
[0058] Define adjacency relationships and extract the connectivity, number of loops, and number of cavities of the 3D mesh blocks based on these relationships.
[0059] Furthermore, adjacency refers to the fact that three-dimensional mesh blocks share edges or faces.
[0060] Arbitrarily select a 3D mesh block as the starting 3D mesh block, mark the starting 3D mesh block as connected and set a connected set. Merge 3D mesh blocks that share edges or faces with the starting 3D mesh block into the connected set and mark them as connected. Using the newly merged 3D mesh blocks as the starting point, continue to expand based on shared edges or faces until no more mesh blocks sharing edges or faces can be found. Continue to select unmarked 3D mesh blocks and set connected sets until all 3D mesh blocks are marked as connected. Calculate the connected sets as the connectivity.
[0061] In each connected set, the three-dimensional mesh blocks are visited sequentially along common edges or common surfaces, and the arrival order is recorded. When a visited three-dimensional mesh block that is not the current source is reached again during the visit, it is determined that a closed channel is generated, and the closed channels are accumulated according to the principle of non-repetition, and the number of loops is generated.
[0062] For the precisely positioned body that has been divided, a continuous grid frame is added using the same method as dividing the 3D grid blocks. Cells not occupied by the 3D grid blocks are defined as empty blocks. All empty blocks connected to the boundary are removed. The remaining empty blocks are grouped according to their coplanar or edge-sharing relationships to generate empty set. The number of empty set is recorded to generate the number of cavities.
[0063] The connectivity, number of loops, and number of cavities are transcoded to generate connectivity code, number of loops code, and number of cavities code, which are then spliced and encapsulated to generate topology constraint code.
[0064] Furthermore, the connectivity is transcoded according to the rule mapping to generate connectivity code elements: when the connected set is 0, the connectivity code element is marked as 0; when the connected set is 1, the connectivity code element is marked as 1; when there are more than 1 connected sets, the connectivity code element is marked as 2+. The loop quantity is transcoded according to the rule mapping to generate loop quantity code elements: when the closed channel is 0, the loop quantity code element is marked as 0; when the closed channel is 1, the loop quantity code element is marked as 1; when the closed channel is more than 1, the loop quantity code element is marked as 2+. The cavity quantity is encoded according to the rule mapping to generate cavity quantity code: when the empty space set is 0, the cavity quantity code is marked as 0; when the empty space set is 1, the cavity quantity code is marked as 1; when the empty space set is more than 1, the cavity quantity code is marked as 2+. The topology constraint code is generated by concatenating and encapsulating the connectivity code, loop quantity code, and cavity quantity code in that order.
[0065] Simultaneously, the external dimensions of the three-dimensional mesh blocks are arranged to generate geometric principal axis feature vectors.
[0066] Furthermore, the external dimensions of the 3D mesh block in the three directions of the 3D coordinate axis are statistically analyzed. The largest external dimension is taken as the principal axis external dimension, and the direction of the principal axis external dimension is taken as the principal axis. The second largest external dimension is taken as the secondary axis external dimension, and the direction of the secondary axis external dimension is taken as the secondary axis. The smallest external dimension is taken as the minor axis external dimension, and the direction of the minor axis external dimension is taken as the minor axis. The principal axis external dimensions, secondary axis external dimensions, and minor axis external dimensions are arranged in the order of principal axis, secondary axis, and minor axis to generate the geometric principal axis feature vector.
[0067] Encode the geometric principal axis feature vectors to generate principal axis codes.
[0068] Furthermore, the maximum and minimum values of the geometric principal axis feature vectors are extracted according to the division direction. The maximum value of the geometric principal axis feature vector is mapped to 1, and the minimum value is mapped to 0. The geometric principal axis feature vectors between the maximum and minimum values are mapped to a range of 0 to 1, generating normalized geometric principal axis feature vector values. The normalized geometric principal axis feature vector values are multiplied by the span pixel number according to the division direction and rounded to generate feature indexes (if the feature index value is not greater than 1, it is recorded as 1; if the feature index value is greater than the span pixel number, it is recorded as the span pixel number; if the feature index value is between 1 and the span pixel number, it is directly recorded as the feature index). The feature indexes are sorted according to the front view direction, right view direction, and top view direction to generate the principal axis code.
[0069] The three-dimensional mesh block is projected and clipped along its projection outline. Simultaneously, the grayscale values of the projection outline of the three-dimensional mesh block are differentially calculated and accumulated to generate shape spectrum coefficients.
[0070] Furthermore, based on the viewpoint identifier, the 3D mesh block is projected onto the CAD standard view, the projection outline of the 3D mesh block is extracted, and the CAD standard view is cropped according to the projection outline of the 3D mesh block to generate a mesh image fragment.
[0071] In a grid image segment, the gray values of the projected contour of the 3D grid block are differentially expressed along the gray-level row direction or gray-level column direction to generate a gray-level difference sequence. The gray-level difference sequence is then accumulated item by item to generate a cumulative gray-level change sequence. The cumulative gray-level change sequence is recorded as shape spectrum coefficients.
[0072] The shape spectral coefficients are encoded to generate spectral codes.
[0073] Furthermore, the maximum and minimum values of the shape spectrum coefficients are extracted according to the division direction. The maximum value of the shape spectrum coefficient is mapped to 1, and the minimum value is mapped to 0. Shape spectrum coefficients between the maximum and minimum values are mapped to a range of 0 to 1, generating normalized shape spectrum coefficient values. The normalized shape spectrum coefficient values are multiplied by the span pixel number according to the division direction and rounded to generate a spectrum index (if the spectrum index value is not greater than 1, it is recorded as 1; if the spectrum index value is greater than the span pixel number, it is recorded as the span pixel number; if the spectrum index value is between 1 and the span pixel number, it is directly recorded as the spectrum index). The spectrum indexes are sorted according to the front view direction, right view direction, and top view direction to generate spectrum codes.
[0074] The spindle code and the spectrum code are concatenated to generate the shape syntax code, which is then integrated with the topology constraint code to generate a manufacturing syntax dual-code description package.
[0075] Furthermore, the main axis code and spectrum code are read sequentially according to the main view direction, right view direction, and top view direction, and then spliced together in the order of main axis code first and spectrum code last to generate shape syntax code.
[0076] The viewpoint identifier, time stamp, topological constraint code, and shape syntax code for each direction are sequentially concatenated to generate a direction segment. The direction segments are then concatenated in the order of main view, right view, and top view to generate a double-code record. All double-code records are then integrated to generate a manufacturing syntax double-code description package.
[0077] S4. Classify and interpret the manufacturing syntax dual-code description package, generate three-dimensional semantic segmentation, project the three-dimensional semantic segmentation onto the CAD standard view, calculate the overlap with the candidate region, and generate a manufacturing conformity certificate.
[0078] The topological constraint code and shape syntax code are concatenated to generate the decoding input vector. The topological constraint code is then used to determine the decoding input vector to generate a category determination.
[0079] Furthermore, the fields are concatenated in the order of connectivity code, loop quantity code, cavity quantity code, spindle code, and spectrum code to generate the interpretation input vector. The interpretation input vector is then checked for name consistency and value consistency while maintaining the concatenation order. The corresponding entries in the manufacturing feature category directory are used as labels. The interpretation input vector is judged by the topological constraint code and classified into the manufacturing feature category directory to generate the topological constraint judgment. At the same time, the viewpoint identifier and time stamp are recorded to generate the training sample set. The training sample set is divided into K folds to generate grouped training sample sets. The fold index is written according to the division order, and K is fixed as the fold number.
[0080] It should be noted that the steps for generating topological constraint determinations by judging the decoded input vector using topological constraint codes and classifying it into the manufacturing feature category list are as follows: When the connectivity symbol is 0, the interpretation input vector is directly discarded; when the connectivity symbol is 1 and the cavity number symbol is 0 and the loop number symbol is 1, it is classified as a through hole or through slot; when the connectivity symbol is 1 and the cavity number symbol is 0 and the loop number symbol is 0, it is classified as a blind hole or blind slot; when the connectivity symbol is 1 and the cavity number symbol is 1 and the loop number symbol is 0, it is classified as a closed cavity or stepped cavity; when the connectivity symbol is 1 and the cavity number symbol is 2+ and the loop number symbol is 1, it is classified as a pipe channel or annular channel; when the connectivity symbol is 2+, it is preferentially classified as an assembly interface or adjacent feature set; the interpretation input vectors that have been classified are summarized to generate a topological constraint determination.
[0081] Machine learning is used to determine the consistency of topological constraint decisions generated by topological constraint codes. K-fold cross-validation is used to train gradient boosting decision trees: a parameter list is established in the gradient boosting decision tree training configuration, and the parameter list is initialized to generate initial values for the number of trees, the maximum depth, and the learning rate. First, the number of trees parameter is trained. A list of tree number values is established in the parameter list. The total number of entries in the interpreted input vector is read as the upper limit of the number of trees, and the initial value of the number of trees is read as the lower limit of the number of trees. The initial value of the number of trees is multiplied by multiples of K to generate the number of trees parameter until the upper limit of the number of trees is reached. The initial value of the number of trees is written as the first tree number parameter in the list of tree number value, and the other tree number parameters are written in sequence. K-fold cross-validation loop is performed on the grouped training sample sets: the grouped training sample set with fold index 1 is used as the verification data, and the grouped training sample sets with fold indices 2 to K are used as the training data. The initial value of the maximum depth is set as the maximum depth parameter setting, and the initial value of the learning rate is set as the learning rate parameter.
[0082] Take the first tree quantity parameter and generate the same number of trees as the tree quantity parameter. For the first tree, perform hierarchical generation according to the maximum depth parameter, with the maximum depth parameter as the highest level. In the first level, group the sample training set according to connectivity code elements, and generate connectivity sample training sets. Then, continue to divide the connectivity sample training set in the order of loop quantity code elements, cavity quantity code elements, spindle code, and spectrum code, generating sample training subsets. Count the number of corresponding entries in the manufacturing feature category directory of each sample training subset, and take the entry with the most entries as the classification label of the training sample subset. In the second level, group the sample training subsets generated in the first level and assign classification labels, until the sample training subsets generated in the previous level are grouped and assigned classification labels at the maximum depth parameter, then stop the hierarchical generation of this tree. Perform hierarchical generation for each tree according to the maximum depth parameter.
[0083] Read each category label one by one, extract the number of entries for each category label, generate a tree count sequence, use the learning rate parameter as a weight, perform a weighted summation with the tree count sequence, record the weighted summation process, and generate an interpretation decision function.
[0084] The verification data is input into the interpretation decision function for verification, a verification category decision is generated, the number of entries that are consistent with the topological constraint decision is counted, the consistency ratio is calculated, and the tree number parameter, the fold of the verification data and the consistency ratio are recorded.
[0085] Continue using the training sample set with fold index 2 as the verification data, and use the training sample set with other fold indices as the training data to generate hierarchies, set the interpretation decision function, and obtain the consistency ratio through verification; until the training sample set with fold index K is used as the verification data, and the training sample set with other fold indices is used as the training data to generate hierarchies, set the interpretation decision function, and obtain the consistency ratio through verification to complete the tree number adjustment, and set the tree number parameter with the highest consistency ratio as the fixed tree number parameter.
[0086] The same training method as the tree number parameter is used to train the maximum depth parameter to generate a fixed maximum depth parameter, in which the fixed tree number parameter is used as the tree number parameter.
[0087] Finally, the same training method as the tree number parameter is used to train the learning rate parameter to generate a fixed learning rate parameter, wherein the fixed tree number parameter is used as the tree number parameter, and the fixed maximum depth parameter is used as the maximum depth parameter.
[0088] The trained gradient boosting decision tree is used to interpret the input vector and determine the class classification and class consistency ratio.
[0089] The input vector is interpreted according to the front view direction, right view direction, and top view direction to generate front view direction interpretation, right view direction interpretation, and top view direction interpretation. The front view direction interpretation, right view direction interpretation, and top view direction interpretation are aligned according to the viewpoint identifier and merged according to the order of viewpoint coverage curve to generate three-dimensional semantic segmentation.
[0090] The 3D semantic segmentation is projected onto the CAD standard view according to the viewpoint identifier to generate a semantic segmentation projection. The number of pixels jointly covered by the semantic segmentation projection and the candidate region and the number of pixels merged are counted. The ratio of the number of jointly covered pixels to the number of merged covered pixels is used as the directional coincidence. The directional coincidence in the main view direction, the right view direction and the top view direction are calculated in sequence to generate the main view direction coincidence, the right view direction coincidence and the top view direction coincidence.
[0091] The system integrates the overlap of the main view direction, the overlap of the right view direction, the overlap of the top view direction, the category determination, the category consistency ratio, and the 3D semantic segmentation to generate a manufacturing consistency certificate.
[0092] S5. Extract the 3D semantic segmentation of the manufacturing conformity certificate, map the 3D semantic segmentation to the CAD 3D model, generate the orientation annotation layer, and associate it with the manufacturing conformity certificate to generate a manufacturing semantic object recognition list.
[0093] The single-view localization volume is located based on semantic segmentation projection, and the directional entity index is generated by counting the number of 3D mesh blocks covered by the single-view localization volume.
[0094] Furthermore, in the standard CAD view, the semantic segmentation projection is aligned with the candidate region, and the candidate region is located to a single-view positioning volume. The number of 3D mesh blocks covered by the single-view positioning volume is counted, and directional entity indexes are generated: positioning is performed in the main view direction, right view direction, and top view direction, and the number of covered 3D mesh blocks is counted to generate entity indexes in the main view direction, right view direction, and top view direction.
[0095] Numerical comparison and vector synthesis are performed on the directional entity indexes to generate CAD entity indexes. The CAD entity indexes are then bound to 3D mesh blocks and written into the CAD 3D model to generate the directional annotation layer.
[0096] Furthermore, the entity indices in the main view direction, right view direction, and top view direction are numerically compared. The entity index with the largest value is selected as the retained direction. The entity indices in the other two directions are then merged into the entity index of the retained direction through vector merging to generate the CAD entity index.
[0097] The CAD entity index is bound to the 3D mesh block to generate directional annotation fragments. By writing the directional annotation fragments into the CAD 3D model, a directional annotation layer is generated.
[0098] The directional labeling layer and manufacturing conformity certificate are sorted according to the view coverage curve to generate a list of manufacturing semantic objects.
[0099] Furthermore, at the direction annotation layer, the manufacturing consistency certificate is connected to the CAD entity index. The viewpoint identifier, the number of 3D mesh blocks, the overlap of the main view direction, the overlap of the right view direction, the overlap of the top view direction, and the category determination are uniformly encapsulated to generate a manufacturing semantic object recognition list.
[0100] This embodiment also provides a computer device applicable to the machine learning-based 3D model object recognition method, including: a memory and a processor; the memory is used to store computer-executable instructions, and the processor is used to execute the computer-executable instructions to implement the machine learning-based 3D model object recognition method proposed in the above embodiment.
[0101] The computer device can be a terminal, comprising a processor, memory, communication interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The communication interface is used for wired or wireless communication with external terminals; wireless communication can be achieved through Wi-Fi, carrier networks, NFC (Near Field Communication), or other technologies. The display screen can be an LCD screen or an e-ink screen. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad on the computer device's casing, or an external keyboard, touchpad, or mouse.
[0102] This embodiment also provides a storage medium storing a computer program that, when executed by a processor, implements the machine learning-based 3D model object recognition method proposed in the above embodiments. The storage medium can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as Static Random Access Memory (SRAM), Electrically Erasable Programmable Read-Only Memory (EEPROM), Erasable Programmable Read Only Memory (EPROM), Programmable Red-Only Memory (PROM), Read-Only Memory (ROM), magnetic storage, flash memory, magnetic disk, or optical disk.
[0103] In summary, this invention achieves structured encoding of manufacturing semantics by generating a dual-code description package of manufacturing syntax, thereby enhancing the interpretability and industrial applicability of 3D model recognition; and by generating a manufacturing conformity certificate, it verifies the overlap between semantic segmentation and geometric features, ensuring the geometric consistency of the recognition results and improving the accuracy and reliability of 3D object recognition.
[0104] It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.
Claims
1. A method for recognizing 3D model objects based on machine learning, characterized in that: include, Collect CAD 3D models, render them to generate standard CAD views, and write viewpoint identifiers. Then, summarize the CAD 3D models, standard CAD views, and viewpoint identifiers to generate a multi-view evidence set. Based on the multi-view evidence set, candidate regions are detected in the standard CAD view. The candidate regions are located to the CAD 3D model by the view backtracking localization method and geometric intersection is performed to generate a set of 3D ROI positioning volumes. Extract topological features from the three-dimensional ROI location volume set to generate topological constraint codes. Simultaneously, perform shape decomposition to generate shape syntax codes. Integrate the topological constraint codes and shape syntax codes to generate a manufacturing syntax dual-code description package. The manufacturing syntax dual-code description package is classified and interpreted to generate a three-dimensional semantic segment. The three-dimensional semantic segment is projected onto the CAD standard view and the overlap with the candidate region is calculated to generate a manufacturing conformity certificate. Extract the 3D semantic segmentation of the manufacturing conformity certificate, generate a directional annotation layer, and associate it with the manufacturing conformity certificate to generate a manufacturing semantic object recognition list.
2. The machine learning-based 3D model object recognition method as described in claim 1, characterized in that: The steps for generating the multi-view evidence set are as follows: Align the coordinates of the CAD 3D model and generate a view reference frame; Based on the viewpoint reference frame, set the rendering directions for the main view, right view, and top view; Render the CAD 3D model, generate a standard CAD view, and write viewpoint identifiers; The viewpoint labels of the CAD standard views are organized to generate viewpoint coverage curves, and then summarized with the CAD 3D model, CAD standard views and viewpoint labels to generate a multi-view evidence set.
3. The machine learning-based 3D model object recognition method as described in claim 2, characterized in that: The steps for generating the 3D ROI location set are as follows. Convert standard CAD views to grayscale to generate grayscale images; Read the grayscale values from the pixels of the grayscale image, and set the horizontal direction of the grayscale image as the grayscale row direction and the vertical direction of the grayscale image as the grayscale column direction; Discrete the grayscale values and perform vector synthesis to generate the grayscale change amplitude and direction. The pixels of the standard CAD view are traversed to generate edge pixels, which are then summarized to generate an edge set. The edge set is scanned along the gray-level row direction and gray-level column direction, and consecutive adjacent edge pixels are merged into edge segments. The connecting edge segments are then connected to generate candidate regions. Read the viewpoint identifier, cut the CAD 3D model along the candidate region, and generate a single-view positioning volume group; Perform geometric intersection on the single-view positioning volume group to generate a fine positioning volume, and summarize the fine positioning volumes to generate a three-dimensional ROI positioning volume set.
4. The machine learning-based 3D model object recognition method as described in claim 3, characterized in that: The steps for generating the topological constraint code are as follows: According to the rendering direction, the finely positioned volumes of the 3D ROI positioning volume set are divided into three-dimensional mesh blocks according to the front view direction, right view direction, and top view direction. Define the adjacency relationship of the 3D mesh blocks, and extract the connectivity, number of loops, and number of cavities of the 3D mesh blocks based on the adjacency relationship; The connectivity, number of loops, and number of cavities are transcoded to generate connectivity code, number of loops code, and number of cavities code, which are then spliced and encapsulated to generate topology constraint code.
5. The machine learning-based 3D model object recognition method as described in claim 4, characterized in that: The steps for generating the manufacturing syntax dual-code description package are as follows: Arrange the external dimensions of the 3D mesh blocks to generate geometric principal axis feature vectors; Encode the geometric principal axis feature vectors to generate principal axis codes; The three-dimensional mesh block is projected and clipped along its projection outline. Simultaneously, the grayscale values of the projection outline of the three-dimensional mesh block are differentially calculated and accumulated item by item to generate shape spectrum coefficients. The shape spectral coefficients are encoded to generate spectral codes; The spindle code and the spectrum code are concatenated to generate the shape syntax code, which is then integrated with the topology constraint code to generate a manufacturing syntax dual-code description package.
6. The machine learning-based 3D model object recognition method as described in claim 5, characterized in that: The steps for classifying and interpreting the dual-code description package of the manufacturing syntax to generate three-dimensional semantic segmentation are as follows. The topological constraint code and shape syntax code of the manufacturing syntax dual-code description package are concatenated to generate the interpretation input vector. The interpretation input vector is then judged by the topological constraint code to generate a category determination. The input vector is interpreted according to the front view direction, right view direction, and top view direction to generate front view direction interpretation, right view direction interpretation, and top view direction interpretation. The front view direction interpretation, right view direction interpretation, and top view direction interpretation are aligned according to the viewpoint identifier and merged according to the order of viewpoint coverage curve to generate three-dimensional semantic segmentation.
7. The machine learning-based 3D model object recognition method as described in claim 6, characterized in that: The steps for generating the manufacturing conformity certificate are as follows: The 3D semantic segmentation is projected onto the CAD standard view according to the viewpoint identifier to generate a semantic segmentation projection. The overlap between the semantic segmentation projection and the candidate region is calculated to generate overlap in the main view direction, right view direction, and top view direction. This overlap is then combined with the category determination and 3D semantic segmentation to generate a manufacturing consistency certificate.
8. The machine learning-based 3D model object recognition method as described in claim 7, characterized in that: The steps for generating the manufacturing semantic object identification list are as follows: The single-view localizers of the single-view localizer group are located based on semantic segmentation projection, and the directional entity index is generated by counting the number of 3D mesh blocks covered by the single-view localizers. Numerical comparison and vector synthesis are performed on the directional entity indexes to generate CAD entity indexes. The CAD entity indexes are then bound to 3D mesh blocks and written into the CAD 3D model to generate the directional annotation layer. The directional labeling layer and manufacturing conformity certificate are sorted according to the view coverage curve to generate a list of manufacturing semantic objects.
9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that: When the processor executes the computer program, it implements the steps of the machine learning-based 3D model object recognition method according to any one of claims 1 to 8.
10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by the processor, it implements the steps of the machine learning-based 3D model object recognition method according to any one of claims 1 to 8.