臺灣博碩士論文加值系統

近年來隨著科技不斷進步，城市化發展與都市建設日漸成形，汽機車早已成為人們主要的交通與運輸工具，因此如何有效管理汽機車已成為社會上一個重要的問題，隨著使用汽機車的人數增加，汽機車管理需要在各種情境和場合下都能適用，因此「一車一號」的車牌辨識方法應運而生，透過這種方式，只需了解車牌即可對車輛進行管理。因此車牌辨識技術已經成為交通管制和數位執法的重要方法之一，且隨著影像辨識技術的進步，車牌辨識系統已廣泛應用於生活中，包括門禁出入管理、區間測速與身分查詢等，如今如何更迅速且正確地自動辨識車牌已成為該技術的主要研究方向，而AI技術中，神經網路能夠運用現有資料進行學習，從而實現自動辨識，以減少人力缺乏和人為失誤導致的管理問題，此方式不僅適用於一般環境，還能夠應對複雜多變的情境，如不同天候、日夜環境與拍攝角度等問題。故本研究針對車牌辨識開發了車牌與字元辨識系統，能夠提供操作者即時得知當前車牌偵測狀況。本論文主要可分為三大部分進行探討，第一部分為建置視覺車牌辨識系統，透過網路攝影機對監控空間進行拍攝，並架設光電感測器進行感測，運用資料擷取卡讀取光電感測器的即時訊號，再通過建立的車牌與字元辨識程式，對字元辨識流程進行設計;第二部分為利用神經網路的學習特性，以及在圖像辨識中的出色表現，建立車牌與字元的辨識模型，本研究使用YOLOv5(You Only Look Once Version 5)將收集的汽車圖像作為模型的訓練資料進行輸入，汽車車牌特徵作為輸出進行車牌辨識的模型建置，並對車牌辨識模型所辨識的車牌位置進行提取，以提取圖像作為模型的訓練資料進行二次輸入，車牌字元特徵作為輸出進行車牌字元辨識模型的建置，以實現辨識車牌等性能，並克服了光線條件、拍攝角度與遮擋變化而帶來的挑戰;第三部分為神經網路檢測車牌與影像辨識的方法進行整合，運用將網路攝影機當前畫面進行擷取並作為輸入，使用YOLOv5進行車牌辨識，並將汽車車牌辨識結果作為輸出，隨後對車牌辨識結果進行擷取並以此資料作為輸入，輸出車牌字元辨識結果，整體實驗得知YOLOv5預測模型的平均準確度為99%。本研究的目的在於不同環境下皆能對車牌與字元進行辨識，因此將所收集的車牌數據依照日夜進行分類，並分別驗證，結果可發現無論是日間或夜間進行車牌與字元辨識的準確度皆高於98%，因此本系統沒有使用時間限制，同時可全自動運作，以節省人事成本並提高車輛管理效率。

In recent years, with the continuous advancement of technology and the ongoing development of urbanization and urban construction, automobiles and motorcycles have become the primary means of transportation for people. Effectively managing vehicles has become a crucial societal issue. As the number of vehicle users increases, vehicle management needs to be applicable in various situations and scenarios. Hence, the "one vehicle, one number" license plate recognition method has emerged. Through this method, vehicle management can be achieved by understanding the license plate alone. License plate recognition technology has become one of the essential methods in traffic control and digital law enforcement. With the progress of image recognition technology, license plate recognition systems are widely used in various aspects of life, including access control, speed measurement, and identity verification. The focus of current research is on how to rapidly and accurately automatically recognize license plates. In the field of AI technology, neural networks can learn from existing data to achieve automatic recognition, reducing issues related to manpower shortages and human errors in management. This approach is not only applicable in general environments but also capable of handling complex and changing situations, such as different weather conditions, day and night environments, and various shooting angles. Therefore, this paper developed a license plate and character recognition system, providing operators with real-time information on the current license plate detection status. The thesis is divided into three main parts for exploration. The first part involves building a visual license plate recognition system using network cameras to capture surveillance space. Light sensors are deployed for detection, and real-time signals from the light sensor are extracted using a data acquisition card. The license plate and character recognition program is then designed for the character recognition process. In the second part, leveraging the learning characteristics of neural networks and their outstanding performance in image recognition, a recognition model for license plates and characters is established. The paper utilizes YOLOv5 (You Only Look Once Version 5), using collected car images as training data to input into the model. Car license plate features serve as output for constructing the license plate recognition model. The recognized license plate positions are extracted, and the extracted images are used as training data for a second input. Car license character features serve as output for building the license plate character recognition model, achieving the recognition performance and overcoming challenges posed by changes in lighting conditions, shooting angles, and obstructions. In the third part, the method of neural network detection of license plates and image recognition is integrated. The current view of the network camera is captured and used as input. YOLOv5 is employed for license plate recognition, and the recognized results are output. Subsequently, the recognized license plate results are extracted, and this data is used as input. The output is the recognition result for license plate characters. The overall experiment reveals that the average accuracy of the YOLOv5 prediction model is 99%. The aim of this paper is to achieve recognition of license plates and characters in different environments. Therefore, the collected license plate data is classified based on day and night, and separate validations are performed. The results show that whether during the day or night, the accuracy of license plate and character recognition is above 98%. Hence, the system operates without time constraints, is fully automatic, saves personnel costs, and enhances vehicle management efficiency.

摘要...i
Abstract...ii
誌謝...iv
目錄...v
表目錄...viii
圖目錄...ix
符號說明...xii
第一章 緒論...1
1.1 研究背景與動機...1
1.2 研究目的...2
1.3 論文架構...2
第二章 文獻回顧...3
2.1 CNN辨識分類物體之應用國內外相關研究...3
2.2 車牌辨識之應用國內外相關研究...7
第三章 研究架構與方法...13
3.1 研究架構與流程...13
3.2 車牌辨識系統架構及硬體...14
3.2.1 車牌辨識系統架構...14
3.2.2 資料擷取卡...14
3.2.3 光電感測器...15
3.2.4 網路攝影機...15
3.2.5 電腦規格與模型訓練環境...16
3.3 神經網路辨識車牌基本概論...17
3.4 卷積神經網路...18
3.4.1 卷積神經網路基本概論...18
3.4.2 輸入層(Input Layer)...20
3.4.3 卷積層(Convolution Layer)...21
3.4.4 池化層(Pooling Layer)...23
3.4.5 全連接層(Fully Connected Layer)...24
3.4.6 輸出層(Output Layer)...24
3.5 模型演算法...25
3.5.1 模型版本與結構...25
3.5.2 車牌物件偵測...27
3.5.3 物件的交集與聯集比...28
3.5.4 平均精度與均值平均精度...29
第四章 實驗結果...34
4.1 視覺車牌辨識系統建置...34
4.1.1 車牌辨識系統之環境架設...34
4.1.2 HMI車牌識別系統與運作流程...35
4.2 車牌與字元資料庫圖像數據建構及模型訓練...36
4.2.1 建構車牌圖像數據...37
4.2.2 車牌數據標註...37
4.2.3 建構車牌數據...38
4.2.4 車牌字元數據標註...39
4.2.5 車牌圖像與字元模型訓練...41
4.2.6 模型辨識...47
4.3 車牌字元識別...49
4.3.1 車牌字元樣式...49
4.3.2 字元辨識模型...50
4.4 選用模型辨識準確度分析...51
4.4.1 車牌圖像辨識選擇模型...51
4.4.2 車牌字元辨識選擇模型...52
4.5 辨識車牌圖像及字元系統實驗...53
4.5.1 白天辨識情境...53
4.5.2 夜晚辨識情境...54
4.5.3 辨識錯誤原因調查...55
第 五 章 結論與未來展望...56
參考文獻...57

[1].Keyu Wu, Mahdi Abolfazli Esfahani, Shenghai Yuan, Han Wang, 2019, "TDPP-Net: Achieving three-dimensional path planning via a deep neural network architecture", Neurocomputing Volume 357, Pages 151-162, September
[2].Mehrad Aria, Mahdi Hashemzadeh, Nacer Farajzadeh, 2022, "QDL-CMFD:A Quality-independent and deep Learning-based Copy-Move image forgery detection method", Neurocomputing Volume 511, Pages 213-236, October
[3].Xinchen Lina, Yang Tanga, Huaglory Tianfieldb, Feng Qiana, Weimin Zhonga, 2019, "A novel approach to reconstruction based saliency detection via convolutional neural network stacked with auto-encoder", Neurocomputing Volume 349, Pages 145-155, July
[4].Rubina Akter, Van-Sang Doan, Jae-Min Lee, Dong-Seong Kim, 2021, "CNN-SSDI: Convolution neural network inspired surveillance system for UAVs detection and identification", Computer Networks Volume 201, 108519, December
[5].Joris Guérin, Stéphane Thiery, Eric Nyiri, Olivier Gibaru, Byron Boots, 2021, "Combining pretrained CNN feature extractors to enhance clustering of complex natural images", Neurocomputing Volume 423, Pages 551-571, January
[6].Ruifan Li, Haoyu Liang, Yihui Shi, Fangxiang Feng, Xiaojie Wang, 2020, "Dual-CNN: A Convolutional language decoder for paragraph image captioning", Neurocomputing Volume 396, Pages 92-101, July
[7].Cong Zhang, Qi Wang, Xuelong Li, 2021, "V-LPDR: Towards a unified framework for license plate detection, tracking, and recognition in real-world traffic videos", Neurocomputing Volume 449, Pages 189-206, August
[8].Salah Alghyaline, 2022, "Real-time Jordanian license plate recognition using deep learning", Journal of King Saud University - Computer and Information Sciences Volume 34, Issue 6, Part A, Pages 2601-2609, June
[9].Naaman Omar, Abdulkadir Sengur, Salim Ganim Saeed Al-Ali, 2020, "Cascaded deep learning-based efficient approach for license plate detection and recognition", Expert Systems with Applications Volume 149, 113280, July
[10].Zied Selmi, Mohamed Ben Halimaa, Umapada Pal, M. Adel Alimi, 2020, "DELP-DAR system for license plate detection and recognition", Pattern Recognition Letters Volume 129, Pages 213-223, January
[11].Songhua Zhang, Xiuling Lu, Zhangjie Lu, 2023, "Improved CNN-based CatBoost model for license plate remote sensing image classification", Signal Processing Volume 213, 109196, December
[12].Di Wang, Yumin Tiana, Wenhui Geng, Lin Zhao, Chen Gong, 2020, "LPR-Net: Recognizing Chinese license plate in complex environments", Pattern Recognition Letters Volume 130, Pages 148-156, February
[13].YOLOv5. Available: https://github.com/ultralytics/yolov5
[14].LabelImg. Available: https://github.com/tzutalin/labelImg
[15].中華民國公路總局，號牌型式及編碼規則表。參考網址： https://www.thb.gov.tw/page?node=92d4a6e2-9afb-464d-a2eb-2be8d26d8d89
[16].汽車模型. https://grabcad.com/library/super-sport-car-1
[17].Ian Goodfellow，2019，"深度學習 Deep Learning"，陳仁和譯，台北：碁峰。
[18].陳慶峯，2022，"基於神經網路智能化辨識刀具影像特徵及預測刀具壽命"，國立虎尾科技大學機械與電腦輔助工程系碩士學位論文
[19].湯仁愷，2023，"應用YOLOv5和CNN深度學習技術於車牌辨識研究"，國立臺北科技大學車輛工程系碩士班碩士學位論文
[20].張文奕，2020，"以深度學習網路做即時車輛及車牌辨識"，國立雲林科技大學電機工程系碩士學位論文
[21].林語揚，2022，"基於卷積神經網路的傾斜校正之車牌辨識模型"，國立虎尾科技大學自動化工程系碩士學位論文