利用卷積神經網路進行印刷電路板上點陣式序號的字元辨識

Character Recognition for Dot-Printed Series Numbers on PCB Using Convolutional Neural Networks

許智閔(CHIH-MIN HSU)

指導教授 ： 繆紹綱

中原大學/電機資訊學院/電子工程研究所/碩士(2018年)

https://doi.org/10.6840/cycu201800639

摘要

近年來，工業4.0的風潮使自動化設備的需求越來越高。經過與一自動化生產公司的接觸，了解該公司在自動化機械上會遇到的問題以及建議。本研究是根據該公司近期的研發需求所提出，希望能以電腦視覺技術自動辨認打印在PCB上的序號，這個需求在類似的公司都有。本研究的核心是字元辨識(character recognition)的問題，屬於固定字型的印刷字體，與傳統字元辨識問題較不同之處是要處理由雷射打點所形成的點陣式印刷(dot matrix printing)字元，而不是常見的視覺連續性筆畫(continuous stroke)的字元。此外，點陣式印刷字元更容易因外力因素導致歪斜及不完整，進而增加辨識難度。
本研究要實現功能如下的系統：輸入含有序號的影像，並輸出準確的序號。此系統含有三個子系統：(1)利用霍夫找線找出影像中感興趣區域(region of interest, ROI)(即含有序號的可能區域)；(2)藉由水平與垂直投影法進行影像分割(image segmentation)，自動將PCB上的序號切割出來；(3)利用卷積神經網路(convolutional neural network, CNN)來進行序號上字元的辨識。
實驗結果顯示，在具有完整性較高的序號影像或是具有破損的序號影像上都得到非常高的辨識率，經過卷積神經網路的訓練結果，在具有破損字體的序號影像中辨識率也達到95%以上，部分破損字體甚至連人眼也無法準確辨識，這展現出深度學習的強大能力。整體而言，系統能成功辨識點陣式印刷字體的字元，解決自動化機械在字元辨識上的問題。

關鍵字

PCB ； 字元辨識 ； 影像分割 ； 感興趣區域 ； 深度學習 ； CNN

並列摘要

In recent years, the trend of Industry 4.0 has made the demand for automation equipment higher and higher. After contacting an automation production company, we realized the company’s problems and suggestions on automated machinery. This study is based on the company’s recent research and development needs. Specifically, they hoped to use computer vision technology to automatically identify the series number printed on a PCB. This is a common demand from similar companies. The core of this research is the problem of optical character recognition (OCR), which falls into the type of fixed font. Compared with the traditional OCR problem, this study deals with dot matrix printing characters formed by laser dots instead of common continuous stroke characters. Furthermore, dot matrix printing characters are more likely to be skewed and incomplete due to external forces. Therefore, the difficulty of doing this type of OCR will increase.
This study will implement a system that can have the following function: Given an image containing a series number, the system shall produce the exact number. This system contains three subsystems: (1) Use the Hough line detection to find the region of interest (ROI) (i.e., the possible region containing the series number on the PCB.). (2) Use the horizontal and vertical projection method to perform image segmentation and cut out the serial number from the PCB automatically. (3) Use the convolutional neural network (CNN) to recognize characters in the series number.

並列關鍵字

PCB ； OCR ； Image Segmentation ； ROI ； Deep Learning ； CNN

參考文獻

[1] A. Farhat, A. Al-Zawqari, and A. Al-Qahtani, “OCR Based Feature Extraction and Template Matching Algorithms for Qatari Number Plate,”

2016 International Conference on Industrial Informatics and Computer Systems (CIICS), March 13-15, 2016.

[2] 支持向量機器 (Support Vector Machine)，

https://cg2010studio.com/2012/05/20/%E6%94%AF%E6%8C%81%E5%90%91%E9%87%8F%E6%A9%9F%E5%99%A8-support-vector-machine/，2018/06。

[3] 深度學習模型，https://medium.com/@syshen/%E7%89%A9%E9%AB%94%E5%81%B5%E6%B8%AC-object-detection-740096ec4540，2018/6