Fast Image Stitching System For Irregular Bottle Based On Multi-view Stereo Vision

In the beverage,food and pharmaceutical industries,more and more machine vision inspection equipment is being used.The inspection of product labels such as plastic bottles,cans and pharmaceuticals is an essential part of the production process.Traditional machine vision inspection methods employ several cameras to acquire images,and then inspect the acquired images separately.However,for the following reasons:firstly,because the captured bottle labels are non-planar,the cylindrical label images will have distortion;secondly,in order to detect the complete and clear parts of the labels,it is often necessary to capture images with partial overlapping areas between them.The former may affect the accuracy of the detection results,while the latter limits the further improvement of the detection speed.To address the above situation,this paper designs a high-speed stitching method for the cylindrical bottle label images.The method stitches the images of labels collected from multiple directions to generate an unfolding image at high speed,which create a basis for subsequent label detection.This paper’s research content and innovations are summarized as follows.(1)In this paper,the system hardware architecture is designed and a four-camera hardware experimental platform is built,which can capture the side images of the bottle from four different orientations respectively.Then,the camera calibration and known geometric information are used to establish an accurate four-camera coordinate system for the subsequent study of the stitching system.(2)For the splicing problem of the cylindrical bottle labels with constant radius and linearly varying radius,this paper first establishes the bottle model,and then obtains the pose of the bottle label model in the actual 3D space by fitting the actual captured bottle images,and establishes the mapping relationship between the 3D points of the spatial label and the pixel points of the plane image captured by the camera,so as to obtain the bottle label unfolding image.(3)To address the problem of the stitching cylindrical bottle labels with changing radius and non-linear variation,this paper adopts the 3D point cloud model of the bottle with multi-view modeling,and obtains the accurate 3D model of the bottle through 3D point cloud data processing,and then also establishes the mapping relationship between the 3D points of the spatial label and the pixel points of the plane image taken by the camera,so as to obtain the bottle label unfolding image.(4)A software system for high-speed stitching of the bottle labels was constructed,and finally the purpose of obtaining the side unfolding image of bottle labels from the captured bottle label images was achieved.The experiment shows that the stitching system proposed in this paper can rapidly and accurately stitch the captured bottle label images to generate an unfolding image.The average stitching speed of each label can reach 143.5ms when using a 1.6 million color camera.