High-speed Marking Recognition Method For Welded Chips Based On Improved Deep Convolutional Neural Networks

In the wave of global digital intelligent manufacturing,machine vision-based inspection technology acts a significant role to ensure the quality of chips in electronics manufacturing.The chip marking recognition technology has a high demand and widespread applications.As the Optical Character Recognition(OCR)gradually penetrates into the industrial field,designing and implementing an OCR method that is suitable for chip-related manufacturing scenarios has become a critical way to solve the chip marking recognition problem.To date,a majority of the chip marking recognition methods used in the industry are traditional morphology-based methods.No matter whether it is the traditional OCR methods or the recent deep learning-based methods,there are many challenges in dealing with chip marking recognition in the production environment.On the one hand,image quality problems such as low brightness,high noise,and poor resolution extensively exist in the real applications;on the other hand,some issues such as the blurred chip markings,high diversity of chip appearance,and interfering welding materials also hinder the performance of the existing methods.To overcome these challenges,this thesis devotes to carrying out a high-speed marking recognition method for welded chips based on deep convolutional neural networks.The main contributions of the thesis are as follows:(1)Proposing a chip image preprocessing method based on deep convolutional networks.Preprocessing can improve the quality of the chip images,reduce the elements that affect the marking recognition,and make the images more suitable for applying the OCR algorithm than before the processing.In this thesis,the problems of low contrast and high noise of the chip images are solved by an improved illumination enhancement algorithm.Meanwhile,an improved image segmentation algorithm is used to remove the interference information generated by chip welding,and a residual network is used to uniformly identify the typesetting direction.Finally,the algorithms are integrated as a preprocessing module based on deep convolutional networks.Experiments prove that this module can effectively suppress the image quality problems and reduce the difficulty of the OCR algorithm caused by the chip appearances.The images processed by this module can produce better text detection and recognition performance in the subsequent OCR algorithm.(2)Proposing a marking character extraction algorithm based on weakly supervised learning.The marking character extraction can organize the markings in the chip image into a character image sequence.This thesis designs a text detection algorithm based on weakly supervised learning,which can effectively reduce the actual cost of labeling production data.At the same time,this thesis aims to solve the problem of compact printing and tiny marking size,and an adaptive padding algorithm and a dynamic label update strategy are proposed to improve the precision of training label generation.In addition,this thesis proposes a character boundary correction algorithm to further improve the accuracy of the character extraction.Experiments prove that the algorithm has a high recall rate of 96.71% for the character detection in the chip markings.(3)Proposing a chip marking character recognition algorithm based on deep convolutional network.Aiming at the problems of poor resolution,unclearness,and unbalanced sample distribution,this thesis proposes two data enhancement methods for chip single characters through category balance and adaptive color scale enhancement to improve the recognizability and learning effect of character images.Through the optimization of the down-sampling process and loss function of the network,a marking character recognition algorithm based on convolutional neural network is proposed,and experiments verify that the marking character recognition algorithm has remarkable recognition accuracy of 99.52%.(4)Developing a chip marking character recognition system based on deep learning.This thesis implements the above-mentioned design and introduction of various algorithms,and integrates them into a marking recognition system with good portability,scalability,and simplicity.The overall test of the chip marking recognition system proves that the chip marking recognition system has a recall rate of 93.98%,and it has better performance on the characteristics of the chip image than the conventional deep learningbased OCR algorithms,as well as can adapt to multiple chip types.Under the premise of image hardware acceleration,the running speed of the algorithm is 20.81 fps,being greater than the status quo of the existing process,and it meets the requirements of the production environment.