Research On Multi View 3D Reconstruction Method For PCB Component Defect Recognition

With China’s comprehensive into the new era of 5G network,the Internet of Things and other electronic information industry is booming,printed circuit board(Printed Circuit Board,PCB)as the realization of the electrical interconnection of the components of the interface substrate,its surface mount(Surface Mounted Technology,SMT)after the electronic components of the pin warping,component inequality,and the detection of three-dimensional defects is particularly important.The detection and identification of three-dimensional defects is particularly important.Raster projection technology as a key link based on three-dimensional data for defect identification,with non-contact,high measurement accuracy advantages,but the application of traditional monocular single raster structure for three-dimensional reconstruction,PCB component height difference and projection angle will lead to the existence of shadows,making it more difficult to obtain high-precision three-dimensional information of components.Therefore,the traditional raster projection model can not carry out all-round three-dimensional reconstruction,and there is a shortage in the measurement and defect identification of PCB components,so the study of multi-view three-dimensional reconstruction method has important practical significance.This topic intends to study the multi-view 3D reconstruction method for PCB component defect identification,taking PCB components as the research object,realizing the highprecision 3D reconstruction of PCB components by multi-view 360 degree projection,focusing on solving the problems of component shadow segmentation,multi-view phase fusion and 3D defect identification in the reconstruction process.The research content of this paper mainly includes the following 4 points:(1)Overall system solution design.Firstly,we introduce the technical requirements of 3D defect recognition of this topic.We analyze in detail the shading and reconstruction blind area difficulties of this topic,design the scheme of multi-view 3D reconstruction system based on multi-grating projection of this topic,introduce the components of the system,and build the hardware system of image acquisition and raster projection.At the same time,we design the overall architecture and algorithm scheme of the software according to the multi-view projection characteristics of the system.(2)A multi-scale self-attention-based component shadow segmentation method.The shadow segmentation method based on streak projection characteristics is proposed to address the problem of component shadows in single-grating 3D reconstruction by using the pixel change characteristics of the shadow region within the projection period.To address the shortage of adaptive segmentation performance,a multi-scale shadow segmentation network is designed to integrate convolution and self-attention to achieve robust shadow segmentation by combining multi-scale feature extraction and self-attention mechanism.For the case of inaccurate edges of the segmented shadow region,the phase-shifted streak information is introduced to guide the network to learn the shadow edges and achieve accurate shadow segmentation.(3)3D reconstruction method based on multi-view phase fusion.Aiming at the reconstruction blindness problem of monocular single-grating 3D reconstruction,a multi-view3 D reconstruction method based on multi-grating projection is designed.The calibration method of raster projection 3D reconstruction system is introduced in detail,including camera calibration,phase solution and system calibration derivation to complete the conversion calculation from 2D phase to 3D point cloud.For the problem of different light intensities in the calibration field of view,a dynamic threshold segmentation algorithm is proposed to achieve robust extraction of calibration points.For multi-view phase data,a joint Gaussian weight phase fusion method based on channel filtering is proposed to realize multi-view phase fusion.For the case of poor generalization,the phase fusion method based on Generate Adversise Network(GAN)is proposed to improve the phase fusion accuracy.(4)Component defect identification method based on point cloud alignment and classification.To design the defect recognition algorithm combining 2D and 3D methods for the 3D defects studied in this project.For the 3D defect localization problem of components,we design the point cloud alignment method based on Iterative Closest Point(ICP)optimization,and realize the defect localization segmentation by aligning the component point cloud with the template point cloud and combining with the distance threshold.The defect recognition algorithm based on Support Vector Machine(SVM)classification model is designed for the segmented set of defect points.For the problem that feature extraction is not robust enough in the recognition process,the Point Net++ point cloud classification network is improved to integrate multi-scale features to assist classification and improve the recognition accuracy of target defects.This paper mainly studies the multi-view 3D reconstruction algorithm for PCB component defect identification,and based on this,writes the software for PCB component 3D defect identification system,which mainly includes the functional modules of system calibration,3D reconstruction and defect identification.To verify the effectiveness of the multi-view 3D reconstruction algorithm,this paper completes the shadow segmentation accuracy test,multiview 3D measurement accuracy test and defect recognition accuracy test based on the self-built data set.The test results show that the multi-view 3D reconstruction system built in this paper has good accuracy and stability,the average intersection ratio of shadow segmentation is 0.87,the absolute accuracy of 3D measurement is 0.04 mm,the repetition accuracy is 0.01 mm,the accuracy rate of defect recognition running test is above 98%,the over-check rate is less than1.5%,the leakage rate is less than 0.5%,which meets the requirements of the subject.