Research On Recognition Method Of Batch Wafer Identifications Based On Machine Vision

With the rapid development of the electronic information industry,semiconductor devices have been widely used.Wafers are the basic materials for the production of semiconductor devices.Typically,wafers are manufactured with monocrystalline silicon rods produced by the Czochralski method and further processed by rolling,slicing,polishing,cleaning,and other processes.During the manufacturing of wafers,the laser marker is printed on the wafer surface as a unique identification(ID)for processing and circulation in producing departments,and the wafer ID is used to index the production process data in the database for product life cycle management.Because the wafer surface is smooth and the marking characters are tiny and shallow,the general character recognition equipment cannot be directly applied.At present,most domestic wafer production process still adopts manual recognition and recording method,leading to low efficiency,high error rate,and labor intensity.The requirements of automatic,intelligent,and large-scale production have not been satisfied.Facing the problems of current wafer-ID recognition methods,this paper proposes a batch wafer-ID recognition method based on machine vision.Considering the high reflection of the wafer surface,the small character size,and shallow marks on wafers,this paper designs a specific image acquisition scheme and character recognition algorithms,enabling non-contact recognition of batch wafer-IDs with high speed and high accuracy.In this paper,a machine vision solution based on linear light sources and a highresolution industrial camera is proposed with a specifically designed wafer cassette base,so that all the wafer IDs in a cassette can be clearly imaged with a single snap.With the visual imaging scheme,a character rectification algorithm based on 3D model registration is proposed to recover the perspective deformation of characters and overcome the difficulty of character recognition.The maximum stable extremum region method and the non-maximum suppression method are further combined to achieve accurate character segmentation.And then,a lightweight convolutional neural network is designed to classify characters with high efficiency and accuracy.Multiple data sets are constructed by collecting wafer ID images on the production line for comparative experiments.The experimental results show that the proposed method has significant advantages in accuracy and efficiency compared with the stateof-the-art algorithms and existing products on the market.In addition,the proposed system also provides a human-machine interface software,which can be operated conveniently and quickly.This system can seamlessly connect to the production execution system and the enterprise resource management system,which is helpful to realize the interconnection of production data and improve the automatic level of wafer manufacturing.