CN-121999241-A - Automatic micro-assembly micro-part multi-interference complex image positioning method

CN121999241ACN 121999241 ACN121999241 ACN 121999241ACN-121999241-A

Abstract

The invention relates to the technical field of automatic micro-assembly, in particular to a method for positioning a multi-interference complex image of a micro part of automatic micro-assembly. The method comprises the steps of S1, obtaining a block b containing complete features to be detected through template matching or target detection deep learning algorithm on an original picture a of a part acquired by a vision system, S2, obtaining a block c with texture interference removed and clear fuzzy features through image preprocessing by using a super-resolution reconstruction network and a texture removing network on the basis of the block b, S3, obtaining edge information d of the features to be detected through binarization and a self-adaptive Canny algorithm on the basis of the block c, S4, removing interference points by using a contour moment algorithm, a RANSAC algorithm and an iterative least squares fitting algorithm, and obtaining the pose of the part through fitting lines or circles, wherein the obtained edge information d is a point set containing partial interference information.

Inventors

HE JIAHANG
WANG FULIN
GUO XUETAO
CHEN LIANG

Assignees

西安精微传感科技有限公司

Dates

Publication Date: 20260508
Application Date: 20251227

Claims (10)

1. An automated micro-assembly micro-part multi-interference complex image positioning method is characterized by comprising the following steps: s1, obtaining a block b containing complete features to be detected by using a template matching or target detection depth learning algorithm on an original picture a of a part acquired by a vision system; s2, on the basis of the block b, performing image preprocessing by using a super-resolution reconstruction network and a texture removing network to obtain a block c with clear texture interference and fuzzy characteristics; s3, based on the block c, acquiring edge information d of the feature to be detected by using a binarization and self-adaptive Canny algorithm; And S4, the obtained edge information d is a point set containing partial interference information, interference points are removed by using a contour moment algorithm, a RANSAC algorithm and an iterative least squares fitting algorithm, and the pose of the part is obtained through a fitting line or a circle.
2. The method according to claim 1, wherein S1 comprises: Selecting according to the actual part performance, if the part structure is clear, the surface is clean, the same feature to be detected of different parts has good consistency of imaging effect under a vision system, and template matching with simple algorithm is used; If the surface processing quality of the part is slightly poor, the imaging consistency of the same feature to be detected of different parts under a vision system is slightly poor, and a deep learning target detection algorithm with better robustness is used.
3. The method according to claim 1, wherein S2 comprises: Selecting according to the imaging effect of the actual part, and if the part features are not on the same horizontal plane and exceed the depth of field of the vision system, recovering a fuzzy region by using a super-resolution reconstruction technology at the moment to make the features clear again; if the texture of the surface of the part is too heavy or the texture interferes with the edge feature recognition, a texture removing deep learning network is needed to remove the texture area of the picture.
4. The method according to claim 1, wherein S3 comprises: Features of the part are generally represented as areas with stronger edges, namely areas with larger gradients, if the surfaces of the part are clean, the processing precision is high, the features under the vision system and the background edges are clear, and the binarization of a fixed threshold value and a Canny algorithm are directly used to obtain edge information of the part; If the consistency of the surface states of the parts is slightly poor, the processing precision is low, the edges of the features and the background under the vision system are not obvious or the edge intensities caused by the difference of the parts are different, and the self-adaptive Canny algorithm is required to be used for extracting the edge information.
5. The method of claim 4, wherein the adaptive Canny algorithm extracts edge information comprising: The Canny algorithm needs to determine a fixed low threshold c1 and a fixed high threshold c2 when extracting edge information; The adaptive Canny algorithm firstly uses an Otsu binarization algorithm to process a block c to obtain an oxford binarization threshold u, at the moment, the Canny binarization threshold c1=u×α1 and c2=u×α2 can be used for realizing the extraction of adaptive edges of different parts and under different illumination, wherein a1 is a low threshold coefficient and a2 is a high threshold coefficient.
6. The method of claim 1, wherein S4 comprises: Firstly, carrying out preliminary interference point removal by using a RANSAC algorithm to obtain a point set e, and then carrying out further interference information removal and final pose determination by using iterative least square fitting; Or directly determining the pose by using a contour moment algorithm.
7. The method of claim 6, wherein iterating a least squares fitting algorithm comprises: Step 1, performing least square fitting on all data points in a total data set e to estimate a target geometric model M, wherein the model M is a basic geometric shape of a straight line and a circle; Step 2, calculating residual errors between each point in the total data set e and the model M, wherein the distance from the point to the straight line is adopted if the model is the straight line, and the distance from the point to the circumference is adopted if the model is the circle; Step 3, eliminating all abnormal points with residual errors exceeding sigma according to a set maximum allowable error threshold sigma, and forming a new data set e by the residual points; Step 4, reducing the current allowable error threshold sigma according to a preset error decreasing step delta sigma, and updating to sigma=sigma Δσ; And 5, repeatedly executing the steps 1 to 4 until the allowable error sigma is reduced below a preset minimum threshold value, and ending the algorithm.
8. The method of claim 6, wherein the use of the RANSAC algorithm in combination with iterative least squares comprises: in actual use, the RANSAC method is selected to be used for carrying out preliminary outlier removal by utilizing a larger threshold value, and then the iterative least squares fitting method is used for fitting.
9. The method of claim 6, wherein the contour moment algorithm comprises: firstly, carrying out closed operation treatment on a Canny image to ensure that the contour of a part to be identified is closed; On the basis, the prior information screening of the closed contour is carried out according to the characteristics of the part, wherein the prior information screening comprises the area or the perimeter; And calculating the mass center and the pose of the part by using the image moment of the feature to be identified.
10. The method of claim 6, wherein the selection of the RANSAC and the use of the iterative least squares algorithm and the contour moment algorithm comprises: when the feature to be identified is a non-closed contour, using a RANSAC and iterative least squares algorithm; When the feature to be identified is a closed contour, a contour moment algorithm is used.

Description

Automatic micro-assembly micro-part multi-interference complex image positioning method Technical Field The invention relates to the technical field of automatic micro-assembly, in particular to a method for positioning a multi-interference complex image of a micro part of automatic micro-assembly. Background With the development of modern manufacturing technology to microminiaturization and high integration, automatic micro-assembly technology plays an increasingly important role in the fields of aerospace, micro-electromechanical system precision optics, semiconductor packaging and the like. Micro-assembly has significant technical specificity compared to traditional macro-scale assembly. Firstly, the object of the micro assembly operation is usually a micro part with millimeter level or even sub millimeter level, the micro part has small volume, light weight and fine structure, is extremely easy to be influenced by small external force or static electricity and other environmental factors to displace or damage in the operation process, and has higher requirements on positioning precision and operation stability. Second, the assembly accuracy of micro-assembly systems is typically on the order of microns or even sub-microns, which is far higher than the millimeter accuracy required by conventional assembly systems. A micro-assembly system based on microscopic visual image localization is therefore typically used. Unlike conventional assembly vision systems, which only need to perform coarse positioning or target recognition, micro-assembly requires higher requirements for the vision system, which not only needs to achieve high-precision positioning in the micron level or even in the submicron level, but also needs to maintain good robustness and real-time processing capability under multiple image interference. The image positioning method for automatic micro-assembly generally comprises the steps of image acquisition, feature extraction, geometric fitting and the like, and aims to provide accurate target position and attitude information for an assembly system. However, due to factors such as image virtual focus, complex surface texture of parts, large background interference, uneven illumination and the like caused by depth of field limitation in a micro-assembly environment, the conventional visual positioning method (such as a standard Canny algorithm, least square fitting and the like) often has the problems of large fitting deviation, high false recognition rate and the like in practical application, and seriously influences the precision and efficiency of an assembly system. Meanwhile, the micro-assembly parts are various in variety and obvious in geometric characteristic difference, and challenges such as local shielding, contact deformation and uncertain spatial attitude between the clamp holder and the parts are also faced in the assembly process, so that the difficulty of image processing is further increased. In summary, the existing visual positioning method has the problems of insufficient adaptability, unstable positioning precision and the like in a complex image environment, and is difficult to meet the comprehensive requirements of a micro-assembly system on high precision, high robustness and real-time performance. Therefore, it is highly desirable to provide a micro part positioning method for an automated micro assembly system, which can still maintain stable, accurate and efficient feature recognition and model fitting capability under the existence of multiple image interference factors such as virtual focus, texture aliasing, part batch difference, illumination variation and the like, so as to improve the intelligent level and industrial application value of the micro assembly system. Disclosure of Invention The invention aims to provide an automatic micro-part multi-interference complex image positioning method, which aims to solve the problems of low positioning precision and poor robustness caused by factors such as image virtual focus, texture interference, edge blurring, illumination non-uniformity and the like in the micro-assembly process. The technical scheme is as follows: An automated micro-assembly micro-part multi-interference complex image positioning method, comprising: s1, obtaining a block b containing complete features to be detected by using a template matching or target detection depth learning algorithm on an original picture a of a part acquired by a vision system; s2, on the basis of the block b, performing image preprocessing by using a super-resolution reconstruction network and a texture removing network to obtain a block c with clear texture interference and fuzzy characteristics; s3, based on the block c, acquiring edge information d of the feature to be detected by using a binarization and self-adaptive Canny algorithm; And S4, the obtained edge information d is a point set containing partial interference information, interference points are removed by u