CN-121994800-A - Industrial design appearance defect intelligent detection system based on computer vision

CN121994800ACN 121994800 ACN121994800 ACN 121994800ACN-121994800-A

Abstract

The invention discloses an intelligent detection system for industrial design appearance defects based on computer vision, which relates to the technical field of computer vision and industrial intelligent detection, and in the operation of the system, multi-angle space information and multispectral imaging data of the surface of a product to be detected are obtained through a preset structured light source module and an array type industrial camera cluster so as to reconstruct high-dimensional vision characteristic distribution of the surface of the product, noise suppression and fusion are performed, generating an enhanced feature map with high signal-to-noise ratio, receiving the enhanced feature map, identifying defects by using an attention mechanism, generating a defect candidate region and preliminary classification probability, reducing calculated amount by knowledge distillation and realizing less sample migration, deploying an optimized detection model to a high-parallelism hardware acceleration platform, calculating a detection result in real time, and sending a sorting instruction, an alarm signal or a process parameter correction instruction to a production execution system according to a preset quality judgment threshold value.

Inventors

YAO MINGGANG
YAO XUETING
CHEN YAN

Assignees

杭州淇璟科技有限公司

Dates

Publication Date: 20260508
Application Date: 20260127

Claims (10)

1. The industrial design appearance defect intelligent detection system based on computer vision is characterized by comprising a multidimensional light field sensing unit, a self-adaptive image enhancement unit, a multiscale defect characterization neural network, a knowledge migration and model compression unit and an edge calculation and closed-loop control unit; The multi-dimensional light field sensing unit is used for acquiring multi-angle space information and multi-spectrum imaging data of the surface of a product to be detected through a preset structured light source module and an array type industrial camera cluster so as to reconstruct high-dimensional visual characteristic distribution of the surface of the product; the self-adaptive image enhancement unit is used for performing background noise suppression, illumination normalization and high dynamic range fusion processing on the original image acquired by the multidimensional light field sensing unit to generate an enhanced feature map with high signal-to-noise ratio; The multi-scale defect characterization neural network is used for receiving the enhanced feature map, identifying and positioning geometric distortion, surface scratch, chromatic aberration abnormality and structural damage of the surface of the product under different spatial scales through a deep feature extraction operator and a global attention mechanism, and generating a defect candidate region and a preliminary classification probability; the knowledge migration and model compression unit is used for carrying out knowledge distillation on the multi-scale defect characterization neural network by utilizing a preset teacher network, reducing the calculated amount of the model on the premise of keeping the detection precision, and realizing rapid domain self-adaption on the defect characteristics of the new-specification product by combining a few-sample learning algorithm; The edge calculation and closed-loop control unit is used for deploying the optimized detection model to the high-parallelism hardware acceleration platform, calculating a detection result in real time, and sending a sorting instruction, an alarm signal or a process parameter correction instruction to the production execution system according to a preset quality judgment threshold value.
2. The intelligent detection system for the appearance defects of the industrial design based on the computer vision, which is characterized in that the multidimensional light field sensing unit further comprises 1 group of annular array light sources consisting of multi-color temperature light emitting diodes and 1 group of polarized illumination assemblies; the annular array light source is integrated with monochromatic light sources with wavelengths of 460 nanometers, 525 nanometers and 630 nanometers respectively; The polarized illumination assembly eliminates mirror glare on the surface of the product due to fresnel reflection by adjusting the relative angle between the polarizer and the analyzer.
3. The intelligent detection system for visual appearance defects of industrial design according to claim 1, wherein the adaptive image enhancement unit is further used for performing texture rejection operation based on Fourier transform when processing industrial products with periodic textures, and comprises the following steps: the method includes converting an image into a spectral space by a fast Fourier transform, identifying and blocking periodic frequency components representing normal textures in the spectral space, and recovering a residual image containing only non-periodic defects by an inverse Fourier transform.
4. The intelligent detection system for the appearance defects of the industrial design based on computer vision according to claim 1, wherein the multi-scale defects represent the specific logic of an attention mechanism integrated in the neural network, and the intelligent detection system is characterized in that the statistical description quantity of channel levels is obtained through a global average pooling layer, and a nonlinear mapping relation between channels is constructed by utilizing two fully connected layers so as to generate a channel attention vector; Carrying out maximum pooling and average pooling along the channel dimension on the feature map, and generating a space attention diagram by using 1 convolution kernel of 7 times 7 after splicing; The input feature map is subjected to pixel-by-pixel product with the channel attention vector and the spatial attention in an effort to achieve focusing on critical defect areas.
5. The intelligent detection system for appearance defects of industrial design based on computer vision according to claim 1, wherein the multi-scale defect characterization neural network performs uncertainty estimation before outputting detection results, and the method comprises the following steps: Forward reasoning with random inactivation attribute is carried out on an input image for a plurality of times through a Monte Carlo sampling technology, so that the confidence level of a detection conclusion is evaluated; And feeding the calibration result back to the model updating flow in real time as new evidence of online learning.
6. The intelligent detection system for the appearance defects of the industrial design based on computer vision according to claim 1, wherein the knowledge migration and model compression unit simulates probability distribution of the teacher network at the defect edges by introducing soft label cross entropy and intermediate feature mapping similarity loss when performing the knowledge distillation processing, so that a lightweight student network obtains sub-pixel level positioning capability; wherein the knowledge distillation process is performed asynchronously on separate computing clusters.
7. The intelligent detection system for appearance defects of industrial design based on computer vision according to claim 1, wherein the specific process of executing the process trend prediction algorithm by the edge calculation and closed-loop control unit is as follows: establishing a state transition matrix comprising normal state, metastable state, drift state and fault state; Calculating probability values of different states of the current process in real time according to detection results in the continuous time sequence; And triggering the process parameter correction instruction when the probability value of the drift state exceeds 0.6.
8. The intelligent detection system for the appearance defects of the industrial design based on the computer vision, which is disclosed in claim 1, is characterized in that the edge calculation and closed-loop control unit adopts a double-redundancy architecture, and the heartbeat monitoring mechanism ensures that when a main calculation node has hardware faults, the backup node is instructed to complete taking over within 50 milliseconds; and a real-time operating system is deployed at the bottom layer of the edge computing and closed-loop control unit, and the visual acquisition task and the defect judgment task are ensured to obtain the highest execution authority through a priority preemption scheduling mechanism.
9. The intelligent detection system for visual appearance defects of industrial design based on computer vision according to claim 1, wherein the system further comprises a centralized defect characteristic fingerprint database for storing defect image characteristic vectors subjected to encryption processing; the defect characteristic fingerprint database is communicated with a plurality of production bases through an industrial Internet platform, and random noise is introduced before the characteristic vector is uploaded by utilizing a differential privacy algorithm, so that cross-region collaborative evolution of a defect mode is realized.
10. The intelligent detection system for the appearance defects of the industrial design based on the computer vision according to claim 1, wherein the multidimensional light field sensing unit further comprises 1 group of stripe generating modules based on structured light projection, and the stripe generating modules are used for projecting sine stripes with phase movement to the surface of a product; The system is also used for collecting stripe deformation images modulated by the surface morphology of the product through an oblique observation camera and restoring 3-dimensional point cloud data of the surface of the product by utilizing a phase unwrapping algorithm.

Description

Industrial design appearance defect intelligent detection system based on computer vision Technical Field The invention relates to the technical field of computer vision and industrial intelligent detection, in particular to an industrial design appearance defect intelligent detection system based on computer vision. Background In the development process of intelligent manufacturing and industrial automation, the appearance quality control of industrial products is a key component of a production and manufacturing link. With the evolution of computer vision technology, an automatic detection means based on image processing gradually replaces the traditional manual naked eye identification, and becomes a core technical support for guaranteeing the delivery qualification rate of products. The field covers a plurality of technical branches such as image acquisition, feature extraction, mode identification, automatic execution and the like, and has a vital effect on improving the automation degree of a production line, enhancing the consistency of products and reducing the labor cost of manufacturing links. The intelligent detection technology of the industrial design appearance defects based on computer vision focuses on utilizing a high-resolution imaging system and a deep learning algorithm to realize automatic identification and refined classification of the tiny defects on the surface of the product. The basic principle of the technology is that the multidimensional visual information of the product surface is captured by an industrial camera, and the abnormal states such as foreign matters, chromatic aberration, deformation, structural damage and the like in the image are rapidly positioned and quantitatively analyzed by combining a predefined defect characteristic model or an end-to-end deep neural network architecture. The method aims at replacing the traditional manual sampling inspection mode through high-frequency non-contact visual scanning, so that an intelligent quality monitoring system covering the whole production flow is established. The prior art still presents significant limitations in dealing with visual inspection in complex industrial environments. The strong fluctuation of the illumination condition of the industrial site and the high reflection and high interference characteristics generated by complex materials on the surface of the product often lead to low signal-to-noise ratio of the acquired image, so that the weak defect characteristics are extremely easy to submerge in background noise. Although the deep learning technology introduces stronger feature expression capability, the existing detection framework has obvious omission phenomenon when dealing with extremely small size defects, and has insufficient flexibility adaptability to multi-specification and multi-form product lines, so that an algorithm model needs to be input with extremely high re-labeling and migration cost when facing product substitution. In addition, the high computational complexity of the large-scale depth algorithm and the extremely short single beat requirement of the industrial pipeline have outstanding contradictions, and the extremely unbalanced sample distribution caused by the difficulty in acquiring the defect sample in the production process makes the over-detection rate and the false-detection rate of the system difficult to be cooperatively optimized under the complex working condition, so that the actual deployment effect of the defect detection system in the precision manufacturing field is severely restricted. Disclosure of Invention Aiming at the defects of the prior art, the invention provides an intelligent detection system for the appearance defects of industrial design based on computer vision, and solves the problems in the background art. The intelligent detection system for the appearance defects of the industrial design based on computer vision comprises a multidimensional light field sensing unit, a self-adaptive image enhancement unit, a multiscale defect characterization neural network, a knowledge migration and model compression unit and an edge calculation and closed-loop control unit; The multi-dimensional light field sensing unit is used for acquiring multi-angle space information and multi-spectrum imaging data of the surface of a product to be detected through a preset structured light source module and an array type industrial camera cluster, wherein the structured light source module utilizes a pulse width modulation technology to perform microsecond-level dynamic adjustment on light intensities of different spectral bands and is matched with a synchronous trigger controller to drive the array type industrial camera cluster to perform nanosecond-level synchronous acquisition so as to reconstruct high-dimensional visual feature distribution of the surface of the product; The self-adaptive image enhancement unit is used for performing joint transformation of a space domain