| With the vigorous development of the printed circuit board(PCB)industry,the in-spection efficiency of traditional PCB quality inspection methods has been difficult to adapt to the current PCB production needs.In recent years,automatic optical inspection systems(AOI systems)based on machine vision technology have gradually become the main PCB quality inspection methods.There are two types of common PCB defects: the first category is defects in the 2D range such as scratches,foreign objects,and compo-nent reverse? the second category is defects in the 3D range such as trace bending and offset,and trace crossing.In order to detect the first type of defects,the AOI system uses a high-resolution industrial camera with a very limited field of view to take pictures to obtain partial images of the PCB,and then stitch the collected images,and use machine vision theory to locate and detect the stitched panorama.For the second type of defects,3D models of PCB components need to be reconstructed in real time and inspected.And point cloud stitching is the most critical step in 3D reconstruction.In view of the par-ticularity of the AOI system,this paper proposes a new method for real-time stitching of high-resolution PCB images(RTStitch)and a new method for real-time stitching of large point clouds(PC-RTStitch).The main research contents and innovations of the thesis are as follows:(1)This paper proposes to classify PCB images and uses different methods to process different types of images.This paper also proposes to obtain the maximum overlap position of the images to be stitched through cosine similarity calculation to prepare for subsequent accurate registration.There are few feature points in the overlap area of some PCB images.Therefore,this paper proposes to first classify images according to the degree of gray level change in the overlapped area of the image,and further process the images with fewer feature points in the overlapped area.After using the cosine similarity calculation to obtain the maximum overlap position,the subsequent improved template matching algorithm is concentrated in the maximum overlap area to improve the speed and accuracy of registration.(2)This paper improves the template matching algorithm,and further improves the accuracy and efficiency of registration to realize real-time stitching of high-resolution PCB images.The selection of the template image must include the column with the largest gray change in the maximum overlap area so that the template image will contain as many feature points as possible.Experiments show that for the first type of image,RTStitch improves the registration accuracy by more than 2% compared with the other three rep-resentative algorithms.While for the second type of image,the registration accuracy is improved by more than 10%.In addition,RTStitch has more than doubled the speed of reg-istration stitching.The most important thing is that RTStitch can meet the high-precision and real-time requirements of the AOI system.(3)This paper proposes to detect SIFT key points of the point cloud,and only calculate the normals of the key points,and use multi-core/multi-thread to calculate FPFH features to save calculation time.The amount of point cloud data in the AOI system is large,so the Voxel Grid filter is used to filter the input large point cloud,and only the normals of the detected SIFT key points are calculated to save calculation time.Then the SAC-IA initial registration is performed according to the FPFH features calculated by multi-threading to provide an initial value for accurate registration.(4)This paper proposes to use the LM algorithm to improve the Point-to-Plane Iterative Closest Point algorithm to achieve real-time accurate splicing of point clouds.Experiments show that compared with the classic ICP algorithm and its improved algo-rithm,PC-RTStitch maintains high registration accuracy while splicing large point clouds in real time,and can be applied to real-time splicing large point clouds in AOI systems. |
PDF Full Text Request