Research On Multi-objective Lens Position And Concave-convex Surface Recognition System Based On Machine Vision

In the production of lenses,multiple processes are required.The most effective way to improve efficiency is to reduce the product transfer time between processes in the case where the time required for each process is constant.For some small and medium-sized factories in China,this step is generally done manually,with lower efficiency and higher labor costs.Machine vision technology is a method that uses digital image processing to analyze the acquired images and extract the information of interest.This technology has been widely used in industrial production due to its high efficiency,high reliability and low cost.Introducing machine vision technology into lens production can improve the overall production efficiency and reduce the cost.In this paper,based on machine vision,a method for positioning and concave-convex surface recognition of multi-target lenses was designed,which mainly included the following parts:Aiming at the shortcomings of the traditional connected domain labeling algorithm that could not distinguish the targets that were close together,an improved connected domain labeling algorithm was proposed.The algorithm only needed to scan the image once,to extract the internal contour of each individual target area,the corresponding area and external rectangle information to complete the image segmentation.The coordinates of the center point of the circumscribed rectangle of each extracted target area were taken as the coordinates of the corresponding lens.After verification,the positioning accuracy could reach 1 mm.Using the Convolutional Neural Network(CNN)was as a classifier to realize the recognition of the concave and convex surface of the lens.First,the parameters of a single model were determined through parameter comparison experiment,and then the CNN integrated model was constructed by combining integrated learning and Convolutional Neural Network,which further improved the system's recognition accuracy.After testing,the recognition rate of this integrated model could reach 99%.The system was realized by FPGA+PC,the algorithms of image acquisition,preprocessing and connected domain labeling were easily implemented in parallel onthe FPGA side.The computational complex Convolutional Neural Network was implemented on the PC side,and the data was transmitted through the Gigabit Ethernet.After verification,the scheme was feasible and can meet the actual needs.