CN-122023241-A - Industrial visual inspection model-oriented countermeasure sample generation and robustness automatic verification system

CN122023241ACN 122023241 ACN122023241 ACN 122023241ACN-122023241-A

Abstract

The invention discloses an automatic challenge sample generation and robustness verification system for an industrial visual detection model, which comprises four core modules and an optional closed loop iteration optimization module, wherein the four core modules comprise challenge sample generation, three-dimensional test matrix construction, robustness automatic assessment and improved report generation, a high-fidelity challenge sample with SSIM (solid-state image) of more than or equal to 0.85 and PSNR (solid-state image) of more than or equal to 25dB is generated through cooperation of a generated challenge network (CGAN/DCGAN) and a physical imaging model with illumination, noise and shielding simulation, a full-dimensional test matrix is constructed by defect type-imaging condition-interference intensity, mAP (mean time point) and omission ratio) indexes are automatically calculated and visually presented, and an improved report with weak link analysis, parameter adjustment and sample supplement suggestions is output to support 'generate-evaluate-optimize' closed loop iteration. The system is compatible with a main flow frame model, is suitable for various industrial detection scenes, can shorten the model verification period by more than 70%, improves mAP by 8% -12%, obviously reduces the omission rate/false detection rate, and ensures that the model industrial scenes are reliably landed.

Inventors

ZHANG XINGXING
TENG YIFEI
YE XUANCHEN
GUO MINGMING
WANG XIAOJUAN
CHEN SHUANG
JIANG XIN
ZHAO JUNLI
ZHANG YANXUE
MENG FANDA

Assignees

中国电子技术标准化研究院华东分院

Dates

Publication Date: 20260512
Application Date: 20251219

Claims (10)

1. An industrial visual inspection model-oriented challenge sample generation and robustness automated verification system, comprising: The countermeasure sample generation module is used for synthesizing a high-fidelity industrial countermeasure sample based on the cooperative work of the generated countermeasure network and the physical imaging model; the three-dimensional test matrix construction module is used for constructing a test matrix by taking the defect type-imaging condition-interference intensity as a three-dimensional dimension; The robustness automatic evaluation module is used for inputting the countermeasure sample into a model to be evaluated, automatically calculating performance indexes and generating a robustness evaluation result; And the improvement report generation module is used for automatically generating an improvement report containing the model optimization direction according to the evaluation result.
2. The industrial visual inspection model oriented challenge sample generation and robustness automated verification system of claim 1, wherein the generation of challenge networks in the challenge sample generation module generates challenge networks on condition of generation of challenge networks or deep convolution generates challenge networks, by pre-training of industrial real defect sample data sets, corresponding types of challenge sample base features can be generated according to input defect type label orientations.
3. The industrial vision inspection model-oriented challenge sample generation and robustness automated verification system of claim 1, wherein the physical imaging model comprises: The illumination simulation sub-module is used for simulating adjustable parameters of illumination angle, illumination intensity and color temperature; the noise simulation submodule is used for simulating adjustable parameters of Gaussian noise, impulse noise and noise intensity; and the shielding simulation sub-module is used for simulating adjustable parameters of shielding area occupation ratio, shielding shape and shielding transparency.
4. The industrial visual inspection model-oriented challenge sample generation and robustness automatic verification system of claim 1, wherein the illumination simulation sub-module has adjustable parameters ranging from 0-180 degrees of illumination angle, 50-1000 lux of illumination intensity and 3000-6500K of color temperature; the noise intensity of the noise simulation submodule is measured by a signal to noise ratio, and the range is 10dB-40dB; the adjustable parameter range of the shielding simulation sub-module is that the shielding area accounts for 5% -50% and the shielding transparency is 0.1-0.9, and the shielding shape supports regular and irregular shapes; the three-dimensional test matrix generates a specific test case in a Cartesian integration or layered sampling mode.
5. The automated industrial visual inspection model-oriented challenge sample generation and robustness verification system of claim 1, wherein the three-dimensional test matrix construction module: the "defect type" dimension includes surface scratches, depressions, protrusions, stains, cracks, dimensional deviations, and the like; the dimension of the imaging condition corresponds to different combinations of illumination, noise and shielding parameters in the physical imaging model; The dimension of interference intensity is divided into three grades of low, medium and high, and the three grades correspond to the parameter ranges of slight, medium and serious interference in an industrial scene respectively.
6. The automated industrial visual inspection model-oriented challenge sample generation and robustness verification system of claim 1, wherein the performance metrics computed by the automated robustness assessment module include average precision mean, miss rate, false miss rate, the module being capable of automatically counting the model's predictions for each test case and visualizing the model's performance distribution through thermodynamic diagrams under different combinations of test dimensions.
7. The automated industrial visual inspection model-oriented challenge sample generation and robustness verification system of claim 1, wherein the challenge sample generation module further comprises a sample fidelity verification sub-module for performing quality verification by calculating a structural similarity index and a peak signal-to-noise ratio of the synthesized challenge sample and the true defect sample, and determining that the challenge sample is a high fidelity sample and is included in the test flow when SSIM is greater than or equal to 0.85 and PSNR is greater than or equal to 25 dB.
8. The industrial visual inspection model-oriented challenge sample generation and robustness automated verification system of claim 1, wherein the improvement report generation module generates an improvement report comprising: Model weak link analysis specifically indicates performance decay data under specific defect types, imaging conditions or interference intensities; parameter adjustment suggestions comprise training regularization parameters of the model to be evaluated and adjustment directions of a data enhancement strategy; sample supplementation advice clearly requires a new challenge sample type and corresponding physical imaging parameter range.
9. The automated challenge sample generation and robustness verification system for an industrial visual inspection model according to claim 1, further comprising an automated iteration optimization module for feeding back the robustness assessment result to the challenge sample generation module, dynamically adjusting and generating a training strategy of a challenge network or a parameter combination of a physical imaging model, regenerating a targeted challenge sample and supplementing the targeted challenge sample to a three-dimensional test matrix, and realizing "generation-evaluation-optimization" closed loop iteration until model robustness meets a preset threshold.
10. The automated industrial visual inspection model-oriented challenge sample generation and robustness verification system of claim 1, wherein the system is adapted for industrial visual inspection scenarios including automotive part appearance inspection, electronic component pin defect inspection, food packaging sealing performance inspection, or photovoltaic module spallation inspection.

Description

Industrial visual inspection model-oriented countermeasure sample generation and robustness automatic verification system Technical Field The invention relates to the technical field of industrial visual inspection and artificial intelligent model verification intersection, in particular to an automatic verification system for generating an countermeasure sample and robustness for an industrial visual inspection model. Background Along with the deep penetration of artificial intelligence technology in the field of industrial manufacturing, industrial visual inspection models are widely applied to key links such as product defect detection, size measurement, appearance screening and the like, and become core supports for improving production efficiency and guaranteeing product quality. However, the landing application of the existing industrial vision detection model still faces a plurality of technical bottlenecks, which lead to insufficient reliability and stability in actual production scenes: In an industrial scene, the probability of occurrence of product defects (such as micro cracks and recessive pits) is low, the types are various, and the collection of partial defect samples can damage the product integrity (such as hidden cracks of photovoltaic modules and cracks of precise electronic elements), so that the coverage range of a real defect sample library is limited and the quantity of the real defect sample library is scarce. Traditional model test relies on a real sample collected manually, and is difficult to cover all possible defect types and application scenes, so that the model test is insufficient. Physical scene interference simulation is missing, namely in an industrial production environment, physical factors such as illumination change (such as workshop dome lamp angle offset and natural light interference), environmental noise (such as camera sensor noise and transmission link interference), random shielding (such as dust coverage and transmission belt dirt) and the like can directly influence the quality of a detection image. The existing model training and testing are mostly based on image data in a standardized laboratory environment, and lack of accurate simulation on the industrial real physical interference leads to excellent performance of the model in the laboratory scene, but serious performance attenuation occurs due to environmental fluctuation after online, and the omission rate and the false detection rate are remarkably increased. The test case design lacks systematicness, the current model robustness test mostly adopts a scattered test case design mode, only a single defect type or a single interference factor is verified, and a full-dimension test system covering defect type-imaging condition-interference intensity is not formed. Such a fragmentation test fails to fully expose the performance shortboards of the model in complex combined scenarios, resulting in undiscovered model potential risk i.e. online application. The robustness assessment has low automation degree, and the existing model robustness assessment is dependent on manual execution of test cases, statistics of test results and calculation of performance indexes, and has complicated flow and low efficiency. For complex scenes containing hundreds or even thousands of test cases, the traditional manual evaluation method needs to take a plurality of weeks, has long verification period and seriously affects the iteration efficiency of the model. Meanwhile, errors are prone to occur in manual statistics, and the evaluation result is distorted. The model optimization lacks a targeted guide that the existing assessment tool can only output performance indexes (such as accuracy and omission factor) of the model, cannot locate core reasons (such as insufficient identification capability of specific defect types and feature extraction failure in a strong noise environment) of the performance attenuation of the model, and does not provide a specific optimization direction. The research personnel need to spend a large amount of time to analyze data and debug parameters, which results in long iterative optimization period and poor effect of the model. In the prior art, three links of anti-sample generation, model evaluation and parameter optimization are mutually independent, and effective feedback linkage is lacked. The generated countermeasure sample cannot be adjusted in a targeted manner according to the model evaluation result, so that the countermeasure effectiveness of the sample is insufficient, the evaluation result cannot directly reverse sample generation and model parameter optimization, a closed loop of generation-evaluation-optimization is formed, and the efficiency of improving the model robustness is further reduced. In order to solve the above problems, there is a need for an integrated verification system capable of accurately generating high-fidelity industrial countermeasure samples, covering a multi-