A Research On Paper Counting System Based On Machine Vision

A variety of industries,such as printing industry and publishing industry,play significant roles in the development of national economy,in which paper counting is one of the most important problems that need to be resolved immediately.In recent years,paper counting systems based on machine vision were proposed by many researchers,but they cannot adapt to the thinner paper.In order to count the amount of paper whose average thickness is about 0.08 mm,this dissertation proposed a novel method,which contains two stages: image mosaic and paper counting.In this article,the following work is finished:(1)Imaging system was designed and the core components of it were selected appropriately.The imaging system is the foundation of paper counting,whose main function is to take a series of pictures of paper's profile.In this system,we use a camera to shoot,and it can be operated to take a series of images from top to bottom.After this,we studied the selection of the industrial camera,lens and light source according to the thickness and other features of paper.Then,the characteristics of the images are analyzed further to make it easier to stitch images and count paper.(2)This dissertation also proposed a novel corner detector and a method to extract feature points and improve the repeatability of them,then,the RANSAC algorithm was improved using DBSCAN clustering algorithm.To detect corners,the image is segmented into binary images by dynamic threshold algorithm firstly;then,mathematical morphology is applied to remove noise and bump,fill tiny holes,and connect narrow gaps;after this,skeletons are extracted by thinning algorithm,and endpoints of skeletons was extracted as potential feature points finally.In the second part,to improve repeatability of feature points,the length of skeletons corresponding to inliers and outliers in potential feature points are analyzed separately.Then,according to distribution of lengths,those potential feature points with small skeleton length are removed,so,the proportion of inliers in remaining points is increased.In the last part,the matching feature points in two images having a relation of translation only are analyzed,and the feature vectors of matching feature points are constructed.After this,those vectors are processed by DBSCAN clustering algorithm,and the RANSAC algorithm is used to completing image mosaicing.(3)In the stage of paper counting,the paper thickness model describing the relationship between thicknesses and pressures is constructed,and then we combined the model with gray projection method to count the number of sheets.Firstly,the images are preprocessed by Hough transform and image enhancement,in which the Hough transform is used to correct the linear tilt of the image and image enhancement could highlight the stripes to make them easier to count.After preprocessing,this article established a paper thickness model to describe the relationship between paper thickness and the amount of paper,and then combined this model with gray projection method to count paper.Results show that the proposed corner detector could extract the feature points effectively in paper overlapping images and the repeatability of extracted feature points is about 60%,which obviously overcome those traditional algorithms,including Harris,SUSAN and FAST.By comparing the combination of DBSCAN and RANSAC with the RANSAC,the quality of the stitched images was improved remarkably.Besides,the accuracy of algorithm for paper with 0.08 mm thickness is slightly more than 98%.