Object simulation modeling method and device for AI visual recognition

Abstract

The present invention relates to the technical field of data recognition, and provides a method and device for modeling simulation of an object for AI visual recognition, including scanning an object, obtaining point cloud data, performing noise reduction, inputting the data into a preset point cloud segmentation model, obtaining segmented point cloud data and feature point cloud data of the object, extracting the geometric features of the object based on the segmented point cloud data, obtaining a set of geometric feature parameters of the object, performing feature matching on the set of geometric feature parameters and the feature point cloud data, obtaining a simulated three-dimensional model of the object and a simulated texture model of the object, and inputting the simulated texture model into a preset AI visual recognition system to perform recognition and classification of the simulated object. By generating a simulated three-dimensional model and a simulated texture model and inputting the model into an AI visual recognition system, the problem that the quality of point cloud data is easily affected by environmental factors in actual operation and there is a lot of data noise, which affects the modeling effect and leads to insufficient precision or inaccurate matching, is improved.

Description

Technical Field

[0001] The present application relates to the technical field of data recognition, and in particular to an object simulation modeling method and device for AI visual recognition. Background Art

[0002] With the rapid development of artificial intelligence technology, AI visual recognition has been widely used in many fields such as autonomous driving, security monitoring, industrial automation, and smart retail. In these applications, accurate identification and classification of objects are crucial. As an important part of AI visual recognition, the 3D modeling technology of objects can help the system better understand the shape, size, texture and other information of objects, thereby improving the accuracy and efficiency of recognition. With the continuous advancement of 3D scanning technology, computer vision and machine learning algorithms, object simulation modeling has become one of the key technologies to promote industry innovation.

[0003] Among the relevant technical means, the visual recognition object simulation modeling technology mainly relies on technical means such as 3D scanning, point cloud data processing and deep learning. The three-dimensional point cloud data of the object is obtained through methods such as laser scanning or structured light, and the geometric features of the object are extracted by combining point cloud noise reduction and segmentation algorithms, and three-dimensional reconstruction is completed through feature matching. Furthermore, the appearance information of the object is integrated into the model through technologies such as texture mapping to facilitate AI system recognition and classification. This process can better reflect the shape, size and surface details of the object, helping to improve the performance and robustness of the recognition system.

[0004] Regarding the above technical solution, although relatively accurate geometric information of objects can be obtained through three-dimensional scanning and point cloud data processing, in actual operation, the quality of point cloud data is easily affected by changes in lighting and the reflectivity of the object surface. There is a lot of data noise, which affects the modeling effect and leads to problems such as insufficient precision or inaccurate matching. Summary of the invention

[0005] In order to improve the problem that the quality of point cloud data is easily affected by environmental factors in actual operations, there is a lot of data noise, which affects the modeling effect and leads to insufficient precision or inaccurate matching, the present application provides an object simulation modeling method and device for AI visual recognition.

[0006] The present invention provides an object simulation modeling method for AI visual recognition, comprising: performing 3D scanning on an object to obtain three-dimensional point cloud data of the object, and performing noise reduction processing on the three-dimensional point cloud data to obtain noise-reduced point cloud data; inputting the noise-reduced point cloud data into a preset point cloud segmentation model to obtain segmentation point cloud data and feature point cloud data of the object, extracting geometric features of the object based on the segmentation point cloud data to obtain a set of geometric feature parameters of the object, performing feature matching on the set of geometric feature parameters and the feature point cloud data to obtain a target matching point set, a feature vector set and a matching accuracy value; performing three-dimensional reconstruction of the object according to the target matching point set, the feature vector set and the matching accuracy value to obtain a simulated three-dimensional model of the object, generating a texture mapping model of the object using the simulated three-dimensional model to obtain a simulated texture model of the object; and inputting the simulated three-dimensional model and the simulated texture model into a preset AI visual recognition system to perform recognition and classification of simulated objects.

[0007] As a preferred solution, the steps of performing 3D scanning on the object to obtain 3D point cloud data of the object, and performing denoising processing on the 3D point cloud data to obtain denoised point cloud data include: performing 3D scanning on the object using a laser scanning device to obtain original point cloud data, and performing denoising processing on the original point cloud data to obtain preliminary denoised point cloud data; performing point cloud density homogenization processing on the preliminary denoised point cloud data using a mean filtering algorithm to obtain homogenized point cloud data, performing fine denoising processing on the homogenized point cloud data using a region growing algorithm to obtain fine denoised point cloud data, performing normal estimation on the fine denoised point cloud data according to the Boltzmann method to obtain a set of normal vectors; processing the set of normal vectors according to non-local mean filtering to obtain smoothed point cloud data, performing edge-preserving denoising processing on the smoothed point cloud data according to non-local mean filtering to obtain edge-preserving denoised point cloud data, and performing data fusion on the edge-preserving denoised point cloud data to obtain denoised point cloud data.

[0008] As a preferred solution, the step of inputting the denoised point cloud data into a preset point cloud segmentation model to obtain segmented point cloud data and feature point cloud data of the object, extracting geometric features of the object based on the segmented point cloud data, and obtaining a set of geometric feature parameters of the object includes: using a gradient-based normal vector estimation algorithm to perform normal vector estimation on the denoised point cloud data to obtain normal vector data, inputting the normal vector data into a point cloud segmentation model, and in the point cloud segmentation model, using a clustering algorithm to process the normal vector data to obtain preliminary segmented point cloud data and surface feature point cloud data; According to the region growing segmentation algorithm, the preliminary segmentation point cloud data is subjected to region growing segmentation to obtain refined segmentation point cloud data, and the refined segmentation point cloud data is subjected to boundary detection by using the boundary detection algorithm to obtain boundary point cloud data; polynomial surface fitting is applied to perform surface fitting on the boundary point cloud data and the surface feature point cloud data to obtain surface feature point cloud data, and the surface feature point cloud data is subjected to feature extraction by using the SIFT algorithm to obtain geometric feature point cloud data; the geometric feature parameter set is calculated on the geometric feature point cloud data by using the HOG feature algorithm to obtain a geometric feature parameter set.

[0009] As a preferred scheme, the step of performing feature matching on the geometric feature parameter set and the feature point cloud data to obtain a target matching point set, a feature vector set and a matching accuracy value includes: normalizing the geometric feature parameter set by minimum-maximum normalization to obtain a normalized geometric feature parameter set, scaling the feature point cloud data based on mean variance normalization to obtain scaled feature point cloud data; describing the normalized geometric feature parameter set and the scaled feature point cloud data by a feature descriptor algorithm to obtain a feature descriptor set; screening the feature descriptor set by a RANSAC algorithm to obtain a target matching point set, calculating feature vectors of the target matching point set by a K nearest neighbor algorithm to obtain a feature vector set, and calculating matching accuracy of the feature vector set according to the root mean square error to obtain a matching accuracy value.

[0010] As a preferred scheme, the steps of performing three-dimensional reconstruction of the object according to the target matching point set, the feature vector set and the matching accuracy value to obtain a simulated three-dimensional model of the object, and using the simulated three-dimensional model to generate a texture mapping model of the object to obtain a simulated texture model of the object include: performing surface reconstruction on the matching point set by Poisson reconstruction to obtain a preliminary simulated three-dimensional model, performing surface smoothing on the preliminary simulated three-dimensional model based on the feature vector set by a global optimization algorithm to obtain an optimized surface, performing shape optimization on the optimized surface based on the matching accuracy value according to an iterative closest point algorithm to obtain a simulated three-dimensional model of the object; performing texture coordinate calculation on the simulated three-dimensional model to obtain texture coordinate data, and performing preliminary texture mapping on the texture coordinate data by a texture mapping algorithm to obtain a simulated texture model of the object.

[0011] As a preferred solution, the steps of calculating texture coordinates of the simulated three-dimensional model to obtain texture coordinate data, performing preliminary texture mapping on the texture coordinate data using a texture mapping algorithm, and obtaining a simulated texture model of the object include: calculating texture coordinates of the simulated three-dimensional model using a UV unfolding algorithm to obtain texture coordinate data, normal vectors, and texture illumination, performing preliminary texture mapping on the texture coordinate data using a Laplace smoothing algorithm to obtain a preliminary simulated texture model; performing texture detail enhancement on the preliminary simulated texture model based on the normal vector using a high dynamic range imaging algorithm to obtain a detail enhanced simulated texture model, performing illumination correction on the detail enhanced simulated texture model using a Hough transform illumination correction algorithm to obtain an illumination corrected simulated texture model; and performing texture optimization on the illumination corrected simulated texture model using a texture optimization algorithm based on a multi-scale convolutional neural network to obtain a simulated texture model of the object.

[0012] As a preferred solution, the step of inputting the simulated three-dimensional model and the simulated texture model into a preset AI visual recognition system to identify and classify the simulated object includes: applying principal component analysis to extract features from the simulated three-dimensional model and the simulated texture model to obtain feature extraction data, and performing feature fusion on the feature extraction data by weighted average method to obtain fused feature data; inputting the fused feature data into a preset neural network model for preliminary classification to obtain preliminary classification results, and performing confidence calculation on the preliminary classification results by Bayesian inference to obtain confidence data; optimizing the preliminary classification results by using an optimization algorithm to obtain optimized classification results, and performing confidence calculation on the optimized classification results according to cross-validation. Perform precision verification to obtain precision verification data, and perform category correction on the precision verification data through post-processing technology to obtain a corrected classification result; after inputting the simulated three-dimensional model, the simulated texture model and the corrected classification result into the AI ​​visual recognition system, use a convolutional neural network to perform deep learning processing on the input simulated three-dimensional model and the simulated texture model to extract important image features therefrom; use a target detection algorithm to locate and select the image features to determine the object area in the image features; combined with the corrected classification result, classify the objects through a classification layer preset in the AI ​​visual recognition system, identify the type and attribute of the simulated object, and output the recognition result and classification result.

[0013] The present application also provides an object simulation modeling device for AI visual recognition, including: an acquisition module, used to perform 3D scanning on the object to obtain three-dimensional point cloud data of the object, and perform noise reduction processing on the three-dimensional point cloud data to obtain noise-reduced point cloud data; an extraction module, used to input the noise-reduced point cloud data into a preset point cloud segmentation model to obtain segmentation point cloud data and feature point cloud data of the object, extract the geometric features of the object based on the segmentation point cloud data to obtain a set of geometric feature parameters of the object, perform feature matching on the set of geometric feature parameters and the feature point cloud data to obtain a target matching point set, a feature vector set and a matching accuracy value; a mapping module, used to perform three-dimensional reconstruction of the object according to the target matching point set, the feature vector set and the matching accuracy value to obtain a simulated three-dimensional model of the object, generate a texture mapping model of the object using the simulated three-dimensional model, and obtain a simulated texture model of the object; and a recognition module, used to input the simulated three-dimensional model and the simulated texture model into a preset AI visual recognition system to recognize and classify simulated objects.

[0014] Compared with the prior art, the present application has the following beneficial effects: good modeling effect and high recognition accuracy. By performing three-dimensional scanning of the object and denoising of the point cloud data, combined with point cloud segmentation, geometric feature extraction and matching, a high-precision simulated three-dimensional model is generated, and a simulated texture model is further generated. The simulated three-dimensional model and the simulated texture model are input into the AI ​​visual recognition system, so that the recognition process can fully combine the geometric features and appearance texture features of the object, improve the accuracy and robustness of recognition and classification, and improve the actual operation. The quality of point cloud data is easily affected by environmental factors, and there is a lot of data noise, which affects the modeling effect, resulting in insufficient accuracy or inaccurate matching. BRIEF DESCRIPTION OF THE DRAWINGS

[0015] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the drawings required for use in the embodiments or the description of the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying creative labor.

[0016] The structures, proportions, sizes, etc. illustrated in the drawings of this specification are only used to match the contents disclosed in the specification so as to facilitate understanding and reading by persons familiar with this technology. They are not used to limit the conditions under which the present invention can be implemented, and therefore have no substantive technical significance. Any structural modification, change in proportion or adjustment of size, without affecting the effects and purposes that can be achieved by the present invention, should still fall within the scope of the technical contents disclosed by the present invention.

[0017] Figure 1 It is a flowchart of an object simulation modeling method for AI visual recognition provided by an embodiment of the present invention;

[0018] Figure 2 It is a schematic block diagram of the structure of an object simulation modeling device for AI visual recognition provided by an embodiment of the present invention.

[0019] Description of reference numerals:

[0020] 10. Object simulation modeling device for AI visual recognition; 11. Acquisition module; 12. Extraction module; 13. Mapping module; 14. Recognition module. DETAILED DESCRIPTION

[0021] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0022] The flowcharts shown in the accompanying drawings are only examples and do not necessarily include all the contents and operations / steps, nor must they be executed in the order described. For example, some operations / steps may also be decomposed, combined or partially merged, so the actual execution order may change according to actual conditions.

[0023] It should also be understood that the terms used in this application specification are only for the purpose of describing specific embodiments and are not intended to limit the application. As used in this application specification and the appended claims, unless the context clearly indicates otherwise, the singular forms "a", "an" and "the" are intended to include plural forms.

[0024] It should be further understood that the term “and / or” used in the specification and appended claims refers to any combination and all possible combinations of one or more of the associated listed items, and includes these combinations.

[0025] The technical solution of the present invention is further described below with reference to the accompanying drawings and through specific implementation methods.

[0026] Embodiment 1:

[0027] like Figure 1 As shown, the object simulation modeling method for AI visual recognition provided in the embodiment of the present application includes steps S100 to S400.

[0028] Step S100: 3D scan the object to obtain three-dimensional point cloud data of the object, and perform noise reduction processing on the three-dimensional point cloud data to obtain noise-reduced point cloud data.

[0029] In this step, a laser scanning or structured light scanning device is used to perform a three-dimensional scan of the object to generate three-dimensional point cloud data containing the surface morphology information of the object. In order to remove the noise introduced by factors such as lighting and equipment errors in the point cloud data, a noise reduction algorithm based on Gaussian filtering or bilateral filtering is used to process the point cloud data to obtain the noise-reduced point cloud data; specifically, the three-dimensional scanning completes the point cloud collection of the complete surface of the object through point-by-point scanning and splicing, and then the collected point cloud data is checked for geometric consistency to remove outliers. During the noise reduction process, the point cloud surface area is optimized in combination with a local smoothing algorithm to ensure that the point cloud data is both complete and clear, and has high-quality subsequent processing capabilities.

[0030] For example, a high-precision 3D scanner was used to scan an irregular geometric object and obtain 3D point cloud data containing 500,000 points. After Gaussian noise reduction processing, about 10% of the noise points were removed, making the point cloud data more uniform and accurate, providing high-quality input data for subsequent operations.

[0031] Step S200: input the denoised point cloud data into a preset point cloud segmentation model to obtain segmentation point cloud data and feature point cloud data of the object, extract the geometric features of the object based on the segmentation point cloud data to obtain a set of geometric feature parameters of the object, perform feature matching on the set of geometric feature parameters and the feature point cloud data to obtain a target matching point set, a feature vector set and a matching accuracy value.

[0032] In this step, the denoised three-dimensional point cloud data is segmented using a deep learning-based point cloud segmentation model (such as the PointNet++ model), and the segmented point cloud data of different structural areas are extracted, and the key feature point cloud data are marked. Subsequently, the geometric feature extraction algorithm based on the segmented point cloud data calculates the geometric feature parameter set of the object, including parameters such as curvature, planarity, and boundary features. A feature matching algorithm based on distance and direction is used for the geometric feature parameter set and the feature point cloud data, and finally the target matching point set, feature vector set, and matching accuracy value are obtained; specifically, during the segmentation process, iterative optimization is used to ensure that the segmented point cloud data has accurate regional divisions, and the geometric feature extraction uses shape descriptor technology (such as SIFT or FPFH descriptors) to quantify the surface features of the object. Feature matching ensures a balance between matching accuracy and speed by combining nearest neighbor search with matching thresholds.

[0033] For example, a complex object composed of multiple materials was segmented into points, and the main plane area and edge feature area were successfully extracted, and the feature parameters of these areas were annotated using the FPFH descriptor. Finally, a target matching point set and feature vector set with an accuracy of 99.8% were obtained during the matching process.

[0034] Step S300: reconstruct the object in three dimensions according to the target matching point set, the feature vector set and the matching accuracy value to obtain a simulated three-dimensional model of the object, and generate a texture mapping model of the object using the simulated three-dimensional model to obtain a simulated texture model of the object.

[0035] In this step, based on the target matching point set and feature vector set, the three-dimensional reconstruction of the object is completed through the surface fitting algorithm (such as Poisson reconstruction or triangulation) to generate a simulated three-dimensional model of the object. Subsequently, the surface texture data of the object is collected through a high-resolution camera, and the texture mapping algorithm (such as UV mapping) is used to fuse the texture data with the simulated three-dimensional model to obtain a simulated texture model of the object; specifically, during the three-dimensional reconstruction process, the surface smoothing technology is combined to eliminate the sharp feature points on the model surface to ensure the smoothness and realism of the model. During texture mapping, the multi-view image stitching technology is used to correct the texture data so that the texture is perfectly aligned with the model surface.

[0036] For example, a complex mechanical part is reconstructed in three dimensions to generate a simulated three-dimensional model with a precision of up to 1mm. Texture data is collected by shooting at multiple angles, and the surface metallic gloss characteristics of the object are completely mapped onto the model, ultimately resulting in a high-fidelity simulated texture model.

[0037] Step S400: input the simulated three-dimensional model and the simulated texture model into a preset AI visual recognition system to identify and classify the simulated object.

[0038] In this step, the AI ​​visual recognition system based on the deep learning algorithm receives the simulated 3D model and the simulated texture model, and inputs them into the pre-trained classification network for feature extraction and classification operations. This process uses the appearance features of the simulated texture model and the geometric features of the simulated 3D model to improve the accuracy and reliability of recognition; specifically, during the classification process, the model's shape features (such as surface area, volume, etc.) and texture features (such as color distribution, texture direction, etc.) are jointly analyzed to optimize the classification performance. At the same time, the input model can be quickly recognized through real-time reasoning.

[0039] For example, a group of objects made of the same material but with different shapes were identified and successfully classified into five categories with a classification accuracy of up to 98.7%.

[0040] In this embodiment, the three-dimensional point cloud data of the object is obtained by three-dimensional scanning of the object, and the three-dimensional point cloud data is subjected to noise reduction processing to improve the quality of the data. Subsequently, the three-dimensional point cloud data after noise reduction is input into the preset point cloud segmentation model, the segmentation point cloud data and feature point cloud data of the object are extracted, and the geometric features of the object are extracted in detail based on the segmentation point cloud data to form a set of geometric feature parameters of the object. Next, the geometric feature parameter set and the feature point cloud data are subjected to feature matching operations to obtain the target matching point set, feature vector set and matching accuracy value. Subsequently, the object is three-dimensionally reconstructed according to the target matching point set, feature vector set and matching accuracy value to generate a simulated three-dimensional model of the object, and a texture mapping model of the object is further generated based on the model, thereby finally obtaining a simulated texture model of the object. Finally, the simulated three-dimensional model and the simulated texture model are input into the preset AI visual recognition system to realize the accurate recognition and classification of simulated objects. Through the steps of denoising, segmentation, geometric feature extraction and feature matching of 3D point cloud data, the accuracy and efficiency of 3D reconstruction of objects are greatly improved, and the problem of insufficient modeling accuracy caused by noise or environmental interference in the existing technology is effectively solved. In actual operation, the quality of point cloud data is easily affected by environmental factors, and there is a lot of data noise, which affects the modeling effect and leads to insufficient accuracy or inaccurate matching.

[0041] Embodiment 2:

[0042] In step S100, a laser scanning device is used to perform 3D scanning on the object to obtain original point cloud data, and the original point cloud data is denoised to obtain preliminary noise-reduced point cloud data.

[0043] By using a laser scanner to scan the surface of an object with high precision, the spatial geometric data of the object is obtained, and the original three-dimensional point cloud data is obtained; specifically, the laser scanning device emits a laser beam and records the time information of the reflection back, so as to accurately calculate the position of each point on the surface of the object and generate three-dimensional point cloud data. The original point cloud data obtained by scanning is denoised by applying a Gaussian filter to remove errors caused by the noise in the scanning environment.

[0044] For example, when scanning a complex object, some noise points may be generated due to uneven reflectivity or environmental interference. The Gaussian filtering algorithm can effectively remove these noises to obtain cleaner preliminary point cloud data.

[0045] The point cloud density of the preliminary denoised point cloud data is homogenized by the mean filtering algorithm to obtain homogenized point cloud data. The homogenized point cloud data is finely denoised by the region growing algorithm to obtain fine denoised point cloud data. The normal of the fine denoised point cloud data is estimated according to the Boltzmann method to obtain a set of normal vectors.

[0046] The mean filter algorithm is used to homogenize the density of the point cloud data and optimize the point cloud distribution, making the distance between each point more uniform. Specifically, the mean filter algorithm adjusts the point cloud data to a more uniform distribution by calculating the mean position of the neighborhood points around each point. Then, the region growing algorithm is used to perform fine noise reduction on the processed uniform point cloud to further reduce irregular points or isolated points.

[0047] For example, when dealing with complex objects (such as mechanical parts with complex surfaces), the region growing algorithm can remove noise points that do not conform to the overall shape by growing the local area.

[0048] The normal vector set is processed according to non-local mean filtering to obtain smooth point cloud data, the smooth point cloud data is subjected to edge-preserving denoising processing according to non-local mean filtering to obtain edge-preserving denoised point cloud data, and the edge-preserving denoised point cloud data is subjected to data fusion to obtain denoised point cloud data.

[0049] The normal vector is smoothed by the non-local mean filtering algorithm, thereby reducing the impact of noise on the normal direction and making the normal data more stable. Specifically, the non-local mean filtering algorithm optimizes the normal vector estimation of each point by comprehensively considering the information of the neighboring areas in the point cloud, making the surface of the point cloud smoother. Then, the non-local mean filtering algorithm is used to perform edge-preserving denoising on the smoothed point cloud data, removing excess noise while maintaining important edge information.

[0050] For example, when processing point cloud data of complex objects (such as sculptures), the non-local mean filtering algorithm can avoid blurring the edges of the objects while removing small noise points on the surface.

[0051] In step S200, the denoised point cloud data is input into a preset point cloud segmentation model to obtain segmentation point cloud data and feature point cloud data of the object, and the geometric features of the object are extracted based on the segmentation point cloud data to obtain a set of geometric feature parameters of the object, including: using a gradient-based normal vector estimation algorithm to estimate the normal vector of the denoised point cloud data to obtain normal vector data, inputting the normal vector data into the point cloud segmentation model, and in the point cloud segmentation model, using a clustering algorithm to process the normal vector data to obtain preliminary segmentation point cloud data and surface feature point cloud data.

[0052] By using a gradient-based normal vector estimation algorithm to estimate the normal vector of the denoised point cloud data, the normal information of each point is obtained, providing a basis for subsequent point cloud segmentation; specifically, the gradient-based normal vector estimation method calculates the normal vector by analyzing the local curvature of each point in the point cloud, and then inputs the normal vector data into the point cloud segmentation model, and uses a clustering algorithm to process these normal vectors, thereby realizing automatic segmentation of objects.

[0053] For example, when processing complex surfaces of objects (such as the shell of a car), the gradient-based normal vector estimation algorithm can accurately capture surface details and segment the sub-regions of the object through the clustering algorithm.

[0054] The preliminary segmented point cloud data is subjected to region growing segmentation according to the region growing segmentation algorithm to obtain refined segmented point cloud data, and the refined segmented point cloud data is subjected to boundary detection using the boundary detection algorithm to obtain boundary point cloud data.

[0055] The preliminary segmented point cloud data is further processed through the region growing algorithm to ensure that each area on the surface of the object can be segmented according to the actual shape; specifically, the region growing algorithm further segments the data according to the density and similarity of the points in the point cloud data, separates the different areas of the object, and accurately detects the boundaries of these areas through the boundary detection algorithm.

[0056] For example, for objects with obvious edge features (such as building materials), region growing algorithms can help refine the segmentation results, while boundary detection can clarify the outer contours of the object.

[0057] Polynomial surface fitting is applied to perform surface fitting on the boundary point cloud data and the surface feature point cloud data to obtain the surface feature point cloud data. The SIFT algorithm is used to extract features from the surface feature point cloud data to obtain the geometric feature point cloud data.

[0058] The boundary point cloud and surface feature point cloud data are fitted by the polynomial surface fitting algorithm to generate a smooth surface that conforms to the actual geometric shape of the object; specifically, the polynomial surface fitting can accurately fit the surface features in the point cloud data, eliminate the influence of noise points, and ensure the continuity of the surface. Subsequently, the SIFT (Scale Invariant Feature Transform) algorithm is used to extract features from the surface feature point cloud data to extract representative geometric feature points.

[0059] For example, when reconstructing the surface of a building, polynomial surface fitting can make the surface smoother, while the SIFT algorithm can extract iconic feature points from it, such as the edges of doors and windows.

[0060] The HOG feature algorithm is used to calculate the geometric feature parameter set of the geometric feature point cloud data to obtain the geometric feature parameter set.

[0061] The geometric feature parameters of the geometric feature point cloud data are calculated through the HOG (Histogram of Directed Gradients) feature algorithm to obtain the description information of each feature point; specifically, the HOG algorithm captures the shape characteristics of the object surface by calculating the gradient direction distribution in the local area, and the obtained set of geometric feature parameters can effectively describe the geometric structure of the object.

[0062] For example, when extracting the geometric features of the human skeleton model, the HOG feature can extract the edge information of the joints, which is of great significance for human recognition.

[0063] In step S200, feature matching is performed on the geometric feature parameter set and the feature point cloud data to obtain a target matching point set, a feature vector set and a matching accuracy value, including: normalizing the geometric feature parameter set by minimum-maximum normalization to obtain a normalized geometric feature parameter set, and scaling the feature point cloud data based on mean-variance normalization to obtain scale-transformed feature point cloud data.

[0064] The geometric feature parameter set is normalized through minimum-maximum normalization, and each feature value is unified to the same scale interval to eliminate the differences between different feature dimensions; specifically, during the normalization process, each feature value is mapped to a specific range (such as [0,1]) to ensure that all features are processed at the same scale. At the same time, the feature point cloud data is scaled using mean-variance normalization to ensure that the scales of different feature point cloud data are consistent and eliminate the impact caused by differences in object size.

[0065] For example, when dealing with objects of various shapes, after normalization, the feature values ​​are no longer affected by the object size or measurement unit, making them more suitable for subsequent feature matching and analysis.

[0066] The normalized geometric feature parameter set and the feature point cloud data after scale transformation are described by the feature descriptor algorithm to obtain a feature descriptor set.

[0067] Through the feature descriptor algorithm, the data after normalization and scale transformation is described, and the unique descriptor of each feature point is extracted; specifically, the feature descriptor algorithm can extract feature information with strong invariance, such as corner points, edges, etc., from geometric features and point cloud data, and generate feature descriptors for subsequent matching and classification.

[0068] For example, in face recognition, the feature descriptor algorithm can extract descriptors of key parts of the face, and can still effectively identify the face even if the facial posture changes.

[0069] The feature descriptor set is screened by the RANSAC algorithm to obtain the target matching point set, the feature vector of the target matching point set is calculated by the K nearest neighbor algorithm to obtain the feature vector set, and the matching accuracy of the feature vector set is calculated according to the root mean square error to obtain the matching accuracy value.

[0070] The RANSAC algorithm is used to screen feature descriptors and remove point pairs that do not meet the matching requirements to obtain a high-quality target matching point set. Specifically, the RANSAC algorithm uses random sampling and iterative verification to exclude matching data that does not meet expectations and ensure the accuracy of the remaining point set. Then, the K nearest neighbor algorithm is used to further calculate the feature vector of the target matching point set to generate a feature vector set of matching points. Finally, the matching accuracy of these feature vectors is calculated using the root mean square error to obtain the matching accuracy.

[0071] For example, when recognizing objects, the RANSAC algorithm can eliminate incorrect matching points, while the K nearest neighbor algorithm can further optimize the matching accuracy based on known proximity relationships.

[0072] In step S300, the matching point set is reconstructed by Poisson reconstruction to obtain a preliminary simulated three-dimensional model, the preliminary simulated three-dimensional model is smoothed by a global optimization algorithm based on a feature vector set to obtain an optimized surface, and the shape of the optimized surface is optimized based on the matching accuracy value according to an iterative closest point algorithm to obtain a simulated three-dimensional model of the object.

[0073] The matching point set is reconstructed through the Poisson reconstruction algorithm to generate a preliminary 3D model and restore the geometric shape of the object. Specifically, the Poisson reconstruction algorithm can generate a continuous surface based on the matching point set and normal information to obtain a rough 3D model. Then, the global optimization algorithm is used to smooth the 3D model based on the feature vector set to reduce discontinuous or erroneous areas and finally generate an optimized 3D surface.

[0074] For example, when reconstructing the facade of a building, Poisson reconstruction can restore the wall curves, while the global optimization algorithm can make the wall smoother and more realistic.

[0075] The texture coordinates of the simulated three-dimensional model are calculated to obtain texture coordinate data, and the texture coordinate data is preliminarily texture mapped using a texture mapping algorithm to obtain a simulated texture model of the object.

[0076] The texture coordinates of the simulated 3D model are calculated by the UV unfolding algorithm, and the surface of the 3D model is unfolded into a 2D plane to obtain the texture coordinate data; specifically, the UV unfolding algorithm unfolds the 3D surface of the object into a 2D plane area, thereby generating the texture coordinates corresponding to each surface point. Then, the texture mapping algorithm is used to perform preliminary texture mapping on these coordinates to generate a preliminary simulated texture model of the object.

[0077] For example, when generating textures for furniture items, the UV unfolding algorithm can unfold the surface of the furniture into a two-dimensional texture map, and the texture mapping algorithm can match it with the texture map to create a realistic visual effect.

[0078] Among them, the steps of calculating the texture coordinates of the simulated three-dimensional model to obtain texture coordinate data, performing preliminary texture mapping on the texture coordinate data using a texture mapping algorithm, and obtaining a simulated texture model of the object include: using a UV unfolding algorithm to calculate the texture coordinates of the simulated three-dimensional model to obtain texture coordinate data, normal vectors, and texture lighting, and using a Laplace smoothing algorithm to perform preliminary texture mapping on the texture coordinate data to obtain a preliminary simulated texture model.

[0079] The texture coordinates of the simulated 3D model are calculated by the UV unfolding algorithm, and the surface of the 3D model is unfolded into a 2D plane to obtain the texture coordinate data; specifically, the UV unfolding algorithm unfolds the 3D surface of the object into a 2D plane area, thereby generating the texture coordinates corresponding to each surface point. Then, the texture mapping algorithm is used to perform preliminary texture mapping on these coordinates to generate a preliminary simulated texture model of the object.

[0080] For example, when generating textures for furniture items, the UV unfolding algorithm can unfold the surface of the furniture into a two-dimensional texture map, and the texture mapping algorithm can match it with the texture map to create a realistic visual effect.

[0081] The texture details of the preliminary simulated texture model are enhanced based on the normal vector by using a high dynamic range imaging algorithm to obtain a detail enhanced simulated texture model. The detail enhanced simulated texture model is corrected for illumination by using a Hough transform illumination correction algorithm to obtain a illumination corrected simulated texture model.

[0082] The high dynamic range imaging (HDRI) algorithm is used to enhance the texture details of the preliminary simulated texture model based on the normal vector, improving the surface illumination and detail performance of the model. Specifically, the HDRI algorithm enhances the brightness and contrast of the image through multiple exposure synthesis, making the surface details of the object richer. Then, the Hough transform algorithm is used to perform illumination correction on the texture model, adjust the illumination effect of the texture, and ensure that the surface illumination is natural and uniform.

[0083] For example, for a metal object, HDRI can enhance the reflection details of its surface, while Hough transform can correct the position of the light source to make the surface of the object appear more realistic.

[0084] The illumination correction simulation texture model is optimized by a texture optimization algorithm based on a multi-scale convolutional neural network to obtain a simulation texture model of the object.

[0085] Through the texture optimization algorithm based on the multi-scale convolutional neural network (CNN), the simulated texture model after illumination correction is optimized to further improve the accuracy and detail of the texture. Specifically, the multi-scale convolutional neural network optimizes the texture in detail through convolution operations at multiple scales, and can process texture details of different sizes and complexities, making the model texture more realistic.

[0086] For example, when optimizing furniture textures, a convolutional neural network can make the texture’s wood grain details more distinct while removing unnecessary noise.

[0087] In step S400, principal component analysis is applied to extract features from the simulated three-dimensional model and the simulated texture model to obtain feature extraction data, and feature fusion is performed on the feature extraction data by weighted average method to obtain fused feature data.

[0088] Principal component analysis (PCA) is used to extract features from the simulated 3D model and texture model, reducing the dimension of the data while retaining key feature information, thereby obtaining feature extraction data; specifically, principal component analysis uses linear transformation to convert high-dimensional feature data into feature vectors with the largest energy in low-dimensional space, extracting the shape features of the 3D model and the texture details of the texture model. Subsequently, the weighted average method is used to fuse the geometric features of the 3D model and the visual features of the texture model to generate fused feature data containing comprehensive information.

[0089] For example, when identifying complex objects (such as mechanical parts), principal component analysis can extract their main geometric shape features, while the weighted average method can combine texture and geometric information to achieve a more comprehensive object description.

[0090] The fused feature data is input into the preset neural network model for preliminary classification to obtain preliminary classification results, and the confidence of the preliminary classification results is calculated through Bayesian inference to obtain confidence data.

[0091] The fused feature data is preliminarily classified through the preset deep neural network model to generate prediction results for each category. Subsequently, the confidence of the preliminary classification results is calculated using Bayesian inference to evaluate the reliability of each classification result. Specifically, the neural network extracts important information from the fused feature data through multi-layer convolution and full connection operations and generates preliminary classification results. Bayesian inference combines classification probability and prior knowledge to calculate the confidence of each category to help screen out the most reliable results.

[0092] For example, when identifying a set of electronic components, the neural network model can quickly classify different types, while Bayesian inference provides confidence for each category, ensuring the reliability of the final classification results.

[0093] The optimization algorithm is used to optimize the preliminary classification results to obtain optimized classification results. The optimized classification results are precision-checked based on cross-validation to obtain precision-checked data. The precision-checked data are categorized through post-processing technology to obtain corrected classification results.

[0094] The preliminary classification results are adjusted through the optimization algorithm to further improve the accuracy and stability of the classification. Subsequently, the accuracy of the optimized classification results is evaluated by cross-validation to ensure the robustness and reliability of the classification model. Finally, the accuracy verification data is corrected by post-processing technology to eliminate the existing misclassification phenomenon and output the final corrected classification results; specifically, the optimization algorithm improves the classification performance by adjusting the model parameters or increasing the feature weights, and the cross-validation verifies the performance of the model through multiple groupings, thereby discovering and correcting the classification deviation. Post-processing technology corrects the classification results by analyzing contextual information or specific rules.

[0095] For example, when identifying industrial parts, optimization algorithms can correct classification errors of marginal samples, while cross-validation can improve the robustness of the overall classification, and post-processing techniques can correct individual misclassifications based on contextual information.

[0096] After the simulated three-dimensional model, simulated texture model and corrected classification results are input into the AI ​​visual recognition system, the convolutional neural network is used to perform deep learning processing on the input simulated three-dimensional model and simulated texture model to extract important image features.

[0097] The convolutional neural network is used to jointly extract features of the simulated three-dimensional model and the simulated texture model to analyze the geometric shape and surface texture of the object. Specifically, the convolutional neural network extracts the shape features of the three-dimensional model and the visual features of the texture model through multi-layer convolution kernel operations, and combines the pooling layer and activation function to improve the compactness and nonlinear ability of the feature expression.

[0098] For example, when identifying a polygonal object, a convolutional neural network is able to extract its main edge features while analyzing the color and pattern information in the texture to provide key data for classification.

[0099] The target detection algorithm is used to locate and select the image features to determine the object area in the image features.

[0100] The extracted image features are located and analyzed through the target detection algorithm to accurately mark the position and boundary area of ​​the object in the image. Specifically, the target detection algorithm analyzes the feature map generated by the convolutional neural network, uses the bounding box prediction technology to select the precise position of the object, and further refines the target area based on the classification results.

[0101] For example, when detecting multiple stacked objects, the object detection algorithm can separately select the location of each object and label its category information.

[0102] Combined with the corrected classification results, the objects are classified through the classification layer preset in the AI ​​visual recognition system. After identifying the type and attributes of the simulated objects, the recognition results and classification results are output.

[0103] By combining the classification layer with the corrected classification results, the target area is finally classified and the type and attribute information of the object is output; specifically, the classification layer performs a comprehensive analysis of the fused geometric and texture features, combines the corrected classification results, outputs the final classification label, and generates the corresponding object attribute information.

[0104] For example, when identifying a set of mechanical parts, it is possible to output the specific model and specific properties (such as material or function) of each part.

[0105] In this embodiment, by integrating multiple technologies such as point cloud data processing, geometric feature extraction, 3D reconstruction, texture optimization, and AI visual recognition, high-precision recognition and classification of complex objects can be achieved. Laser scanning and multi-level noise reduction technology ensure the high quality of point cloud data; Poisson reconstruction and multi-scale convolutional networks are used to optimize 3D models and textures to ensure the authenticity and richness of details of the model; combined with deep learning and classification algorithms, accurate recognition and classification output of objects are finally achieved. The implementation of this solution significantly improves the efficiency and accuracy of recognition, providing strong technical support for industrial automation, intelligent manufacturing and logistics.

[0106] Embodiment 3:

[0107] like Figure 2 As shown, the present application also provides an object simulation modeling device 10 for AI visual recognition, including an acquisition module 11, an extraction module 12, a mapping module 13 and a recognition module 14.

[0108] The acquisition module 11 is mainly used to perform 3D scanning on the object to obtain the three-dimensional point cloud data of the object, and perform noise reduction processing on the three-dimensional point cloud data to obtain the noise-reduced point cloud data.

[0109] The acquisition module 11 completes the 3D scanning of the object, obtains the 3D point cloud data of the object, and performs multi-stage noise reduction processing on the original data. The acquisition module 11 uses high-precision laser scanning equipment and advanced noise reduction algorithms, such as Gaussian filtering and non-local mean filtering, to ensure that the generated point cloud data has high quality and low noise, providing a reliable basis for subsequent feature extraction.

[0110] The extraction module 12 is mainly used to input the denoised point cloud data into a preset point cloud segmentation model to obtain the segmentation point cloud data and feature point cloud data of the object, extract the geometric features of the object based on the segmentation point cloud data to obtain the geometric feature parameter set of the object, perform feature matching on the geometric feature parameter set and the feature point cloud data to obtain the target matching point set, feature vector set and matching accuracy value.

[0111] The segmentation and feature extraction of the noise-reduced point cloud data are achieved through the extraction module 12. The extraction module 12 combines the deep learning point cloud segmentation model and the traditional geometric algorithm to accurately segment the segmentation point cloud data and feature point cloud data of the object, and extracts the geometric feature parameter set based on these data. Subsequently, the module uses feature matching algorithms (such as the RANSAC algorithm and the K nearest neighbor algorithm) to calculate the target matching point set, feature vector set and matching accuracy value, further enhancing the accuracy of object modeling.

[0112] The mapping module 13 is mainly used to reconstruct the object in three dimensions according to the target matching point set, the feature vector set and the matching accuracy value to obtain a simulated three-dimensional model of the object, and to generate a texture mapping model of the object using the simulated three-dimensional model to obtain a simulated texture model of the object.

[0113] The three-dimensional reconstruction and texture mapping of the object are completed through the mapping module 13. The mapping module 13 generates a simulated three-dimensional model based on the Poisson reconstruction algorithm and global optimization technology, and carefully optimizes the texture model through UV expansion, texture mapping and multi-scale convolutional neural network to generate a high-fidelity and realistic simulated texture model. The design of the mapping module 13 ensures that the geometry and texture details of the model can still be accurately presented in complex scenes.

[0114] The recognition module 14 is mainly used to input the simulated three-dimensional model and the simulated texture model into a preset AI visual recognition system to recognize and classify the simulated objects.

[0115] The simulated 3D model and simulated texture model are input into the preset AI visual recognition system through the recognition module 14, and the geometric and texture features are extracted using a deep convolutional neural network, and the simulated objects are accurately recognized and classified in combination with the target detection and classification algorithms. The recognition module 14 further optimizes the accuracy and robustness of recognition through principal component analysis, feature fusion and Bayesian inference.

[0116] In this embodiment, through modular design and functional division of labor, the simulation modeling and AI visual recognition process of objects are made more systematic, intelligent and efficient. The acquisition module 11 ensures the high quality of the data source, the extraction module 12 realizes accurate feature segmentation and matching, the mapping module 13 ensures the realism and accuracy of the model, and the recognition module 14 improves the accuracy and stability of classification and recognition.

[0117] It should be noted that technicians in the relevant technical field can clearly understand that for the convenience and conciseness of description, the specific working process of the above-described device and each module can refer to the corresponding process in the aforementioned embodiment of the object simulation modeling method for AI visual recognition, and will not be repeated here.

[0118] As described above, the above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that the technical solutions described in the aforementioned embodiments may still be modified, or some of the technical features thereof may be replaced by equivalents. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for object simulation modeling for AI visual recognition, characterized in that: include: Performing 3D scanning on the object to obtain 3D point cloud data of the object, and performing noise reduction processing on the 3D point cloud data to obtain noise-reduced point cloud data; The denoised point cloud data is input into a preset point cloud segmentation model to obtain segmented point cloud data and feature point cloud data of the object, geometric features of the object are extracted based on the segmented point cloud data to obtain a set of geometric feature parameters of the object, and feature matching is performed on the set of geometric feature parameters and the feature point cloud data to obtain a target matching point set, a feature vector set and a matching accuracy value; The target matching point set is reconstructed through Poisson reconstruction to obtain a preliminary simulated three-dimensional model, the preliminary simulated three-dimensional model is smoothed based on the feature vector set using a global optimization algorithm to obtain an optimized surface, the optimized surface is shape optimized based on the matching accuracy value according to an iterative closest point algorithm to obtain a simulated three-dimensional model of the object; the texture coordinates of the simulated three-dimensional model are calculated using a UV unfolding algorithm to obtain texture coordinate data, normal vectors and texture illumination, the texture coordinate data is preliminarily texture mapped using a Laplace smoothing algorithm to obtain a preliminary simulated texture model; the preliminary simulated texture model is texture-detailedly enhanced based on the normal vector using a high dynamic range imaging algorithm to obtain a detail-enhanced simulated texture model, the detail-enhanced simulated texture model is illumination-corrected using a Hough transform illumination correction algorithm to obtain an illumination-corrected simulated texture model; the illumination-corrected simulated texture model is texture-optimized using a texture optimization algorithm based on a multi-scale convolutional neural network to obtain a simulated texture model of the object; The principal component analysis is applied to extract features of the simulated three-dimensional model and the simulated texture model to obtain feature extraction data, and the feature extraction data is fused by weighted average method to obtain fused feature data; the fused feature data is input into a preset neural network model for preliminary classification to obtain a preliminary classification result, and the confidence of the preliminary classification result is calculated by Bayesian inference to obtain confidence data; the preliminary classification result is optimized by an optimization algorithm to obtain an optimized classification result, and the precision of the optimized classification result is verified by cross-validation to obtain precision verification data, and the precision verification data is corrected by post-processing technology to obtain a corrected classification result; after the simulated three-dimensional model, the simulated texture model and the corrected classification result are input into the AI ​​visual recognition system, the convolutional neural network is used to perform deep learning processing on the input simulated three-dimensional model and the simulated texture model to extract important image features therefrom; the image features are positioned and framed by a target detection algorithm to determine the object area in the image features; combined with the corrected classification result, the objects are classified by a classification layer preset in the AI ​​visual recognition system, and after identifying the type and attribute of the simulated object, the recognition result and the classification result are output.

2. The object simulation modeling method for AI visual recognition according to claim 1 is characterized in that: The step of performing 3D scanning on the object to obtain three-dimensional point cloud data of the object, and performing noise reduction processing on the three-dimensional point cloud data to obtain noise-reduced point cloud data includes: Using a laser scanning device to perform 3D scanning on the object to obtain original point cloud data, and performing denoising processing on the original point cloud data to obtain preliminary denoised point cloud data; Performing point cloud density homogenization processing on the preliminary denoised point cloud data by using a mean filtering algorithm to obtain homogenized point cloud data, performing fine denoising processing on the homogenized point cloud data by using a region growing algorithm to obtain fine denoised point cloud data, and performing normal estimation on the fine denoised point cloud data according to a Boltzmann method to obtain a normal vector set; The normal vector set is processed according to non-local mean filtering to obtain smooth point cloud data, the smooth point cloud data is subjected to edge-preserving denoising processing according to non-local mean filtering to obtain edge-preserving denoised point cloud data, and data fusion is performed on the edge-preserving denoised point cloud data to obtain denoised point cloud data.

3. The object simulation modeling method for AI visual recognition according to claim 1 is characterized in that: The step of inputting the denoised point cloud data into a preset point cloud segmentation model to obtain segmented point cloud data and feature point cloud data of the object, extracting geometric features of the object based on the segmented point cloud data, and obtaining a set of geometric feature parameters of the object comprises: Using a gradient-based normal vector estimation algorithm to perform normal vector estimation on the denoised point cloud data to obtain normal vector data, inputting the normal vector data into a point cloud segmentation model, and in the point cloud segmentation model, processing the normal vector data using a clustering algorithm to obtain preliminary segmentation point cloud data and surface feature point cloud data; Performing region growing segmentation on the preliminary segmented point cloud data according to a region growing segmentation algorithm to obtain refined segmented point cloud data, and performing boundary detection on the refined segmented point cloud data using a boundary detection algorithm to obtain boundary point cloud data; Polynomial surface fitting is applied to perform surface fitting on the boundary point cloud data and the surface feature point cloud data to obtain surface feature point cloud data; SIFT algorithm is used to perform feature extraction on the surface feature point cloud data to obtain geometric feature point cloud data; HOG feature algorithm is used to calculate a set of geometric feature parameters on the geometric feature point cloud data to obtain a set of geometric feature parameters.

4. The object simulation modeling method for AI visual recognition according to claim 1, characterized in that: The step of performing feature matching on the geometric feature parameter set and the feature point cloud data to obtain a target matching point set, a feature vector set and a matching accuracy value comprises: Normalizing the geometric feature parameter set by minimum-maximum normalization to obtain a normalized geometric feature parameter set, and scaling the feature point cloud data based on mean-variance normalization to obtain scale-transformed feature point cloud data; Describing the normalized geometric feature parameter set and the feature point cloud data after the scale transformation by a feature descriptor algorithm to obtain a feature descriptor set; The feature descriptor set is screened by the RANSAC algorithm to obtain a target matching point set, feature vectors are calculated on the target matching point set using the K nearest neighbor algorithm to obtain a feature vector set, matching accuracy is calculated on the feature vector set according to the root mean square error to obtain a matching accuracy value.

5. An object simulation modeling device for AI visual recognition, characterized in that: include: An acquisition module is used to perform 3D scanning on an object to obtain three-dimensional point cloud data of the object, and to perform noise reduction processing on the three-dimensional point cloud data to obtain noise-reduced point cloud data; An extraction module is used to input the denoised point cloud data into a preset point cloud segmentation model to obtain segmented point cloud data and feature point cloud data of the object, extract the geometric features of the object based on the segmented point cloud data to obtain a set of geometric feature parameters of the object, perform feature matching on the set of geometric feature parameters and the feature point cloud data to obtain a target matching point set, a feature vector set and a matching accuracy value; A mapping module is used to perform surface reconstruction on the target matching point set through Poisson reconstruction to obtain a preliminary simulated three-dimensional model, perform surface smoothing on the preliminary simulated three-dimensional model based on the feature vector set using a global optimization algorithm to obtain an optimized surface, perform shape optimization on the optimized surface based on the matching accuracy value according to an iterative closest point algorithm to obtain a simulated three-dimensional model of the object; perform texture coordinate calculation on the simulated three-dimensional model using a UV unfolding algorithm to obtain texture coordinate data, normal vector and texture illumination, perform preliminary texture mapping on the texture coordinate data using a Laplace smoothing algorithm to obtain a preliminary simulated texture model; perform texture detail enhancement on the preliminary simulated texture model based on the normal vector using a high dynamic range imaging algorithm to obtain a detail enhanced simulated texture model, perform illumination correction on the detail enhanced simulated texture model using a Hough transform illumination correction algorithm to obtain an illumination corrected simulated texture model; perform texture optimization on the illumination corrected simulated texture model using a texture optimization algorithm based on a multi-scale convolutional neural network to obtain a simulated texture model of the object; The recognition module is used to extract features from the simulated three-dimensional model and the simulated texture model by applying principal component analysis to obtain feature extraction data, perform feature fusion on the feature extraction data by weighted average method to obtain fused feature data; input the fused feature data into a preset neural network model for preliminary classification to obtain preliminary classification results, perform confidence calculation on the preliminary classification results by Bayesian inference to obtain confidence data; optimize the preliminary classification results by using an optimization algorithm to obtain optimized classification results, perform accuracy verification on the optimized classification results according to cross-validation to obtain accuracy verification data, and verify the accuracy of the precision by post-processing technology. The degree verification data is corrected for category to obtain a corrected classification result; after the simulated three-dimensional model, the simulated texture model and the corrected classification result are input into the AI ​​visual recognition system, a convolutional neural network is used to perform deep learning processing on the input simulated three-dimensional model and the simulated texture model to extract important image features therefrom; the image features are positioned and framed by a target detection algorithm to determine the object area in the image features; combined with the corrected classification result, the objects are classified by a classification layer preset in the AI ​​visual recognition system, and after identifying the type and attribute of the simulated object, the recognition result and the classification result are output.

Images