Method for constructing robust evaluation data model of three-dimensional intelligent biological identification instrument

Abstract

The application is suitable for the field of three-dimensional intelligent biological identification technology, and provides a method for constructing a robust evaluation data model of a three-dimensional intelligent biological identification instrument, which comprises the following steps: dynamically calculating a scaling ratio according to the geometric characteristics of an original three-dimensional biological sample model to generate a scale change model; identifying surface feature areas based on a Gaussian curvature model to construct a two-term loss function containing perturbation smoothness and curvature change constraints, and generating a natural noise perturbation model by using an optimization algorithm; generating structural defect and surface defect models by combining Poisson reconstruction with proportional clipping or curvature-based point cloud removal; and constructing a multi-grade interference parameter system to quantitatively evaluate the identification error rate and average interference error of the instrument. The application realizes systematic, standardized and automated generation of test samples, and provides a scientific and quantitative evaluation basis for the robustness of three-dimensional biological identification instruments.

Description

Technical Field

[0001] This invention relates to the field of three-dimensional intelligent biometrics technology, specifically a method for constructing a robustness evaluation data model for three-dimensional intelligent biometrics instruments. Background Technology

[0002] With the rapid development of artificial intelligence and 3D imaging technology, 3D intelligent biometric instruments (such as plant seed identifiers based on 3D point clouds or meshes, and facial recognition terminals) have been widely used in fields such as agricultural quarantine and security monitoring. Robustness is a key indicator that measures how well such instruments can maintain their recognition performance and accuracy when faced with uncertainties and interference factors.

[0003] Existing robustness evaluation technologies for 3D intelligent biometric instruments have the following main drawbacks: 1. Uncontrollable construction of test model categories and insufficient coverage: Existing methods cannot systematically and controllably modify data models for various complex situations such as scale changes, noise interference, and structural defects, resulting in biased evaluation results; 2. Lack of standardized procedures: The test samples used by different laboratories or research institutions lack uniform standards, resulting in poor comparability of evaluation results; 3. Low level of automation: Existing methods rely on manual creation of robustness evaluation test samples, which is inefficient and difficult to guarantee consistency; 4. Insufficient quantitative analysis: Lack of a precise quantitative indicator system to assess the robustness level; 5. High cost: Traditional methods require a large number of real samples and have a high loss rate, resulting in huge testing costs.

[0004] Therefore, in view of the above situation, there is an urgent need to provide a method for constructing a robustness evaluation data model for three-dimensional intelligent biological identification instruments in order to overcome the shortcomings in current practical applications. Summary of the Invention

[0005] The purpose of this invention is to provide a method for constructing a robustness evaluation data model for a three-dimensional intelligent biological identification instrument, effectively solving the problems mentioned in the background art.

[0006] This invention is implemented as follows: a method for constructing a robustness evaluation data model for a three-dimensional intelligent biological identification instrument, comprising the following steps: Step S1: Generation of scale-changing 3D data model: Obtain the original 3D biological sample model, dynamically calculate the scaling factor based on the geometric features of the original 3D biological sample model, and generate a scale-changing 3D data model. Step S2: Noise perturbation 3D data model generation: Calculate the Gaussian curvature of the 3D model surface, introduce controlled geometric perturbations in the high curvature region, and balance the perturbation intensity and surface smoothness through optimization algorithm to generate a noise perturbation 3D data model. Step S3: Generation of defective 3D data model: Generate a 3D data model with structural or surface defects by proportional cropping or curvature-based point cloud removal and reconstruction. Step S4, Robustness Evaluation: Construct a multi-level test set based on the scale-varying 3D data model, the noise-perturbed 3D data model, and the 3D data model with structural or surface defects. Test the 3D intelligent biometric instrument under test and quantify its robustness index.

[0007] As a further aspect of the present invention: in step S1, the specific steps for dynamically calculating the scaling factor are as follows: The maximum bounding box length of the original model is calculated as the baseline size. The minimum and maximum scaling ratios are calculated based on the preset target length range. A specified number of scaled version models are generated at equal intervals within the minimum and maximum scaling ratio ranges.

[0008] As a further aspect of the present invention: in step S2, the introduction of controlled geometric disturbance specifically includes: A bivariate loss function is constructed, and the loss function is minimized using a boundary-constrained quasi-Newton optimization algorithm (L-BFGS-B) to determine the optimal perturbation vector; wherein the loss function is: In the formula, The L-BFGS-B solver represents the feature extractor of the target recognition system. Iterative minimization under constraints , This is the perturbed 3D point set, which is the optimal set of vertex positions that the L-BFGS-B algorithm aims to solve for. This represents the total number of vertices in a 3D model or point cloud. For the original first vertex coordinates For the perturbation of the first The coordinates of each vertex; λ is the weighting balance coefficient; The calculation is performed on the vertices after all perturbations. With the original vertex The square of the Euclidean distance between them.

[0009] As a further aspect of the present invention: in step S2, after the noise-perturbed three-dimensional data model is generated, an anisotropic Laplacian smoothing process is also included to eliminate high-frequency noise artifacts and maintain the visual naturalness of the biological sample.

[0010] As a further aspect of the present invention: in step S3, the generation step of the three-dimensional data model of the structural defect is as follows: The bounding box of the computational model determines the longest dimension, and the longest dimension is pruned forward or backward according to a preset percentage range to generate a three-dimensional data model with structural defects.

[0011] As a further aspect of the present invention: in step S3, the step of generating the three-dimensional data model of the surface defect is as follows: Calculate the curvature value of each point in the model point cloud; The kd-tree algorithm is used to identify high curvature regions where the curvature value exceeds a preset threshold. Delete the point cloud data in the high curvature region and its neighborhood according to the preset removal radius; The remaining discrete point cloud was reconstructed into a three-dimensional data model of surface defects using the Poisson reconstruction algorithm.

[0012] As a further aspect of the present invention: in step S4, the multi-level test set divides the robustness evaluation into at least four levels, with higher levels representing greater interference intensity.

[0013] As a further aspect of the present invention: the quantification of its robustness index includes: Calculate the recognition error rate, normalized error rate, and mean interference error (mCE) ​​of the instrument under test on each level of test set.

[0014] As a further aspect of the present invention: the original three-dimensional biological sample model is in STL format.

[0015] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the method for constructing a robust evaluation data model for a three-dimensional intelligent biometric identification instrument as described above.

[0016] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention proposes an automatic generation method for three-dimensional data models of scale changes based on dynamic calculation, which solves the problem that existing technologies cannot systematically simulate samples of different scales. It realizes standardized and automated scale change testing. The batch processing of the Trimesh library enables 100% automation, which greatly improves testing efficiency and sample consistency. This invention employs a noise perturbation 3D data model generation algorithm based on Gaussian curvature, which accurately applies perturbation in high curvature regions, solving the problem of lack of specificity and controllability in noise testing. Combined with loss function optimization, Bayesian optimization, and smoothing processing, it can accurately apply perturbation in key regions while maintaining the visual realism of the samples, thus improving the realism of the test. This invention solves the problem of balancing perturbation intensity and visual naturalness by constructing a bi-term loss function and optimizing it with the L-BFGS-B algorithm through perturbation refinement techniques of constrained multi-objective optimization and Bayesian optimization. This makes the generated noise perturbation model both meet the test validity and conform to the visual characteristics of real samples. This invention proposes a method for generating three-dimensional data models of defects based on proportional clipping and curvature analysis. This method can accurately control the degree, location, and type of defects, solve the problem of lack of standardization and controllability in defect testing, realize standardized testing of structural and surface defects, and ensure the integrity of the defect model through Poisson reconstruction. This invention establishes a four-level standardized multi-level robustness evaluation system, sets multi-dimensional quantitative test indicators and evaluation indicators, solves the problem of lack of quantitative standards for robustness evaluation, realizes accurate quantitative evaluation of instrument robustness, and makes the evaluation results of different institutions comparable. Attached Figure Description

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

[0018] Figure 1 This is a schematic diagram illustrating the generation of the scale-changing sample model in this invention, showing the morphology of the same biological sample at different scaling ratios.

[0019] Figure 2 The image shows a comparison of the three-dimensional models generated under different noise parameters in this invention, demonstrating the effect of Gaussian curvature perturbation.

[0020] Figure 3 This is a schematic diagram of the structural defect generation model at different proportions in this invention.

[0021] Figure 4 This is a comparison diagram of the point cloud model after defect processing and the corresponding Poisson reconstructed 3D model in this invention.

[0022] Figure 5 This is a schematic diagram illustrating the principle of using Hausdorff distance for model difference quantization in this invention.

[0023] Figure 6 This is a comparison diagram of surface defect models generated under the influence of different parameters in this invention. Detailed Implementation

[0024] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0025] The present invention will be further explained below with reference to specific embodiments.

[0026] Please see Figures 1-6 The present invention provides a method for constructing a robustness evaluation data model for a three-dimensional intelligent biological identification instrument, the method comprising the following steps: Step S1, Scale-Change 3D Data Model Generation Module: Input the original 3D model in STL format; The maximum bounding box length of the calculated model is used as the baseline dimension; The scaling factor range is dynamically calculated based on the target length range; Generate a specified number of scaled versions at equal intervals within a magnification range; Use the trimesh library for batch processing to generate 3D data models with scale variations that meet the requirements.

[0027] In this step, the target length range is 6.0-10.0mm (configurable), the number of magnifications is 5 (adjustable), and the level of automation is 100% batch processing.

[0028] Step S2, Noise Perturbation 3D Data Model Generation Module: Gaussian curvature calculation: The Gaussian curvature of each vertex in the STL model is calculated using the angle defect method; Perturbation application: Introduce small perturbations that follow a Gaussian distribution in regions of high curvature; The optimization process is as follows: A bivariate loss function is constructed, and the perturbation is optimized under constraints using a quasi-Newton optimization algorithm with boundary constraints (L-BFGS-B), where the loss function is: In the formula, The L-BFGS-B solver represents the feature extractor of the target recognition system. Iterative minimization under constraints , This refers to the perturbed set of 3D points (optimization variables), which is the set of "optimal vertex positions" that the L-BFGS-B algorithm aims to solve for. This represents the total number of vertices in a 3D model or point cloud. For the original first vertex coordinates (e.g.) Throughout the optimization process, it serves as a fixed reference benchmark. For the perturbation of the first The vertex coordinates contain the original coordinates and the adversarial perturbation quantity (i.e., ...). ), which are the parameters that the algorithm is actually adjusting; λ is the weight balance coefficient, used to adjust the ratio between "geometric fidelity" and "feature distortion"; The calculation is performed on the vertices after all perturbations. With the original vertex The square of the Euclidean distance between them (L2 norm).

[0029] Bayesian optimization is used to further refine the perturbation parameters; Smoothing: Apply anisotropic Laplacian smoothing to maintain visual naturalness.

[0030] In this step, the perturbation strength ε is 0.2-0.7 (adjustable at four levels), the smoothing factor λ is 0.5-1.0, and the number of iterations is 10-50.

[0031] Step S3: Generation of the defective 3D data model, specifically a 3D data model with structural or surface defects: The steps for generating a three-dimensional data model of structural defects are as follows: Calculate the model bounding box and determine the longest dimension; Cut proportionally according to the set percentage (10%-40%); Supports both forward and reverse cropping modes.

[0032] The steps for generating a 3D data model of surface defects are as follows: Calculate the curvature of each point in the point cloud; Identify regions with high curvature (using the kd-tree algorithm); Remove high curvature areas by a set radius (0.5-2.0mm); The Poisson Reconstruction algorithm is used to generate a complete defect model.

[0033] Step S4, Robustness Evaluation: Grading system: Establish a four-level evaluation system (levels 1-4, with increasing severity); Test metrics include: Scale variation: magnification parameters 1-5; Noise variation: ε parameter 0.2-0.7; Overall defect: 10%-40% of the defect was removed; Surface defects: Remove defects with a radius of 0.5-2.0 mm; Evaluation metrics include: recognition error rate, normalized error rate, and mCE value.

[0034] This invention also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of the method for constructing a robust evaluation data model for a three-dimensional intelligent biometric identification instrument as described above.

[0035] Example 1: Robustness test of the three-lobed ragweed seed model Original model: 3D scan model of three-lobed ragweed seed (STL format, length 8.76mm); Test configuration: Scale variation: 5 versions were generated (0.685-1.141x); Noise perturbation: ε=0.5, λ=0.7, 30 iterations; Structural defects: 30% resection; Surface defects: Remove 3 areas with a radius of 1.5 mm.

[0036] Test results: PointNet error rate increased from 11.9% (Level 1) to 25.6% (Level 4). The PointNet++ error rate increased from 12.9% (Level 1) to 27.6% (Level 4). The effectiveness of the evaluation system has been verified.

[0037] Example 2: Large-scale sample batch processing Processing scale: 60 seed models × 4 variations × 4 levels = 960 test samples; Processing time: Average processing time per model <30 seconds; Consistency verification: For 10 random samples generated with the same parameters, the standard deviation of the Hausdorff distance is <0.01 mm.

[0038] Example 3: Practical Application Verification Application scenario: On-site plant quarantine at ports; Test sample: 50 defective seeds collected on-site; Verification results: The difference in the identification error rate between the simulated defect samples generated by this invention and the actual defect seeds is less than 5%, proving that the simulation effect is realistic and reliable.

[0039] The embodiments of the present invention also include the following alternative implementations: In the scaling step, dynamic calculation can be omitted, and a preset fixed scaling list (such as [0.5,0.7,1.3,1.5]) can be used directly for scaling.

[0040] In the noise perturbation step, if fine perturbation based on curvature is not desired, uniform random perturbation or perturbation based on the normal vector direction can be used as a fast alternative.

[0041] In the defect generation step, structural defects can be replaced by geometric methods such as sphere cutting or plane cutting instead of proportional trimming; surface defects can be replaced by randomly removing a specified proportion of point cloud instead of curvature-based removal.

[0042] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a robustness evaluation data model for a three-dimensional intelligent biological identification instrument, characterized in that, Includes the following steps: Step S1: Generation of scale-changing 3D data model: Obtain the original 3D biological sample model, dynamically calculate the scaling factor based on the geometric features of the original 3D biological sample model, and generate a scale-changing 3D data model. Step S2: Noise perturbation 3D data model generation: Calculate the Gaussian curvature of the 3D model surface, introduce controlled geometric perturbations in the high curvature region, and balance the perturbation intensity and surface smoothness through optimization algorithm to generate a noise perturbation 3D data model. Step S3: Generation of defective 3D data model: Generate a 3D data model with structural or surface defects by proportional cropping or curvature-based point cloud removal and reconstruction. Step S4, Robustness Evaluation: Construct a multi-level test set based on the scale-varying 3D data model, the noise-perturbed 3D data model, and the 3D data model with structural or surface defects. Test the 3D intelligent biometric instrument under test and quantify its robustness index.

2. The method for constructing a robustness evaluation data model for a three-dimensional intelligent biological identification instrument according to claim 1, characterized in that, In step S1, the specific steps for dynamically calculating the scaling factor are as follows: The maximum bounding box length of the original model is calculated as the baseline size. The minimum and maximum scaling ratios are calculated based on the preset target length range. A specified number of scaled version models are generated at equal intervals within the minimum and maximum scaling ratio ranges.

3. The method for constructing a robustness evaluation data model for a three-dimensional intelligent biological identification instrument according to claim 1, characterized in that, In step S2, the introduction of controlled geometric perturbation specifically includes: Construct a bivariate loss function and minimize it using the boundary-constrained quasi-Newton optimization algorithm L-BFGS-B to determine the optimal perturbation vector; where the loss function is: In the formula, The L-BFGS-B solver represents the feature extractor of the target recognition system. Iterative minimization under constraints , This is the perturbed 3D point set, which is the optimal set of vertex positions that the L-BFGS-B algorithm aims to solve for. This represents the total number of vertices in a 3D model or point cloud. For the original first vertex coordinates For the perturbation of the first The coordinates of each vertex; λ is the weighting balance coefficient; The calculation is performed on the vertices after all perturbations. With the original vertex The square of the Euclidean distance between them.

4. The method for constructing a robustness evaluation data model for a three-dimensional intelligent biological identification instrument according to claim 3, characterized in that, In step S2, after the noise-perturbed three-dimensional data model is generated, an anisotropic Laplacian smoothing process is also included to eliminate high-frequency noise artifacts and maintain the visual naturalness of the biological sample.

5. The method for constructing a robustness evaluation data model for a three-dimensional intelligent biological identification instrument according to claim 1, characterized in that, In step S3, the steps for generating the three-dimensional data model of the structural defect are as follows: The bounding box of the computational model determines the longest dimension, and the longest dimension is pruned forward or backward according to a preset percentage range to generate a three-dimensional data model with structural defects.

6. The method for constructing a robustness evaluation data model for a three-dimensional intelligent biological identification instrument according to claim 1, characterized in that, In step S3, the steps for generating the three-dimensional data model of the surface defect are as follows: Calculate the curvature value of each point in the model point cloud; The kd-tree algorithm is used to identify high curvature regions where the curvature value exceeds a preset threshold. Delete the point cloud data in the high curvature region and its neighborhood according to the preset removal radius; The remaining discrete point cloud was reconstructed into a three-dimensional data model of surface defects using the Poisson reconstruction algorithm.

7. The method for constructing a robustness evaluation data model for a three-dimensional intelligent biological identification instrument according to claim 1, characterized in that, In step S4, the multi-level test set divides the robustness evaluation into at least four levels, with higher levels representing greater interference intensity.

8. The method for constructing a robustness evaluation data model for a three-dimensional intelligent biological identification instrument according to any one of claims 1 to 7, characterized in that, The metrics used to quantify its robustness include: Calculate the recognition error rate, normalized error rate, and average interference error of the instrument under test on each level of test set.

9. The method for constructing a robustness evaluation data model for a three-dimensional intelligent biological identification instrument according to claim 1, characterized in that, The original three-dimensional biological sample model is in STL format.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method for constructing a robustness evaluation data model for a three-dimensional intelligent biometric instrument as described in any one of claims 1 to 9.

Images