本論文製作一個汽車車牌的辨識系統，對於傾斜車牌提出矯正的方法，架構主要分為車牌定位、車牌傾斜矯正、字元切割以及字元辨識四個部分。
車牌定位使用索貝爾邊緣偵測以凸顯影像中的垂直邊緣，利用中值濾波過濾雜訊，接著使用形態學對影像作處理，經由區塊面積大小、寬高比例、車牌字元間的變化規律作為篩選，取得車牌候選位置。
車牌傾斜矯正包含水平傾斜矯正與垂直傾斜矯正，先使用霍夫轉換計算出傾斜車牌的傾斜角度，對傾斜車牌進行水平矯正，使用剪切轉換對水平矯正後的車牌做垂直傾斜修正，完成車牌傾斜矯正。
字元切割包含水平投影切割與垂直投影切割，先將車牌候選位置進行二值化，對其進行水平投影，取得中間車牌字元的區域，接著使用垂直投影，將同行的黑色像素數目做累加，依照字元間的最低點依序進行切割，完成字元切割。
字元辨識是使用卷積神經網路對字元樣本進行訓練，並將字元切割後的字元進行特徵比對完成字元辨識。

This thesis presents a vehicle license plate recognition system, which proposes methods for tilt angle detection and correction of license plates. There are four subtasks of the recognition system, including license plate detection, license plate tilt correction, character segmentation and character recognition.
The license plate location method uses the Sobel edge detection technique to enhance the vertical edges in the vehicle image. Subsequently, the median filter is applied for noise removal from the image, and morphological operations are performed on the edge images. There are several criteria for determining the locations of license plate candidates, including the size of the area, aspect ratio, and variations among characters on the license plate.
The license plate tilt correction process involves both horizontal and vertical tilt correction. In the horizontal tilt correction process, the Hough transform is employed to detect the angle of tilt plates. Horizontal correction is then applied to rectify the horizontally tilted license plate. The shear transformation algorithm is used to correct the vertical tilt problem of the license plate, thus the license plate tilt correction process is completed.
The character segmentation process consists of horizontal projection segmentation and vertical projection segmentation. Horizontal projection is applied to obtain the region of the central characters on the license plate candidate regions, after being thresholded. Accumulating the black pixels of the characters in each column of the plate, vertical projection is used to segment plate characters.
Convolutional Neural Network (CNN) algorithm is used to character recognition. The character dataset is input into the built neural network for training. The character recognition is done in the license plate using convolutional neural network classifier.

摘要……………i
Abstract……………ii
誌謝……………iv
目錄……………v
表目錄……………vii
圖目錄……………viii
符號說明……………x
第一章 緒論……………1
1.1 研究動機……………1
1.2 研究方法……………1
1.3 文獻探討……………1
1.4 論文架構……………2
第二章 研究方法……………3
2.1 車牌定位……………3
2.1.1 灰階化……………4
2.1.2 邊緣偵測……………5
2.1.3 中值濾波……………8
2.1.4 大津二值化演算法……………9
2.1.5 形態學……………10
2.2 車牌傾斜矯正……………12
2.2.1 霍夫轉換……………13
2.2.2 水平傾斜矯正……………15
2.2.3 垂直傾斜矯正……………16
2.3 字元切割……………17
2.3.1 車牌影像處理……………18
2.3.2 水平投影切割……………18
2.3.3 垂直投影切割……………19
2.4 CNN車牌字元辨識模型……………19
第三章 實驗結果與討論……………25
3.1 實驗環境……………25
3.1.1 實驗設備……………25
3.1.2 汽車車牌拍攝……………25
3.2 實驗結果……………26
3.2.1 車牌定位結果……………26
3.2.2 車牌傾斜矯正結果……………27
3.2.3 字元切割結果……………29
3.2.4 字元辨識結果……………29
3.3 車牌辨識失敗的情形……………30
第四章 結論與未來展望……………32
4.1 結論……………32
4.2 未來展望……………32
參考文獻……………33
Extended Abstract……………34

[1]江忠錦，“斜側方影像之車牌定位與辨識研究”，中華大學機械工程學系研究所，碩士論文，2010。
[2]藍正皓，“雙規格車牌辨識系統之實作”，國立臺灣海洋大學電機工程學系，碩士論文，2013。
[3]孫正豪，“新式汽車車牌辨識之研究”，國立聯合大學機械工程學系碩士班，碩士論文，2014。
[4]陳韋榤，“應用快速廣義霍夫轉換於歪斜車牌定位與角度估算”，國立台灣科技大學高分子工程系，碩士論文。
[5] Lee, C.-H., K.-L. You, and Y.-P. Lin, “Dynamic real-time license plate recognition” Journal of Technology, Vol. 25, No. 2, pp. 151-165, Jun. 2010.
[6] Wang Y., Huag L.-C., and Li Qin, “License Plate Tilt Correction Based on Region Partition and Hough Transform” Journal of Chongqing University of Technology, Vol. 31, No. 3, pp. 97-104, Mar. 2017.
[7]李曉莎，林森，“基於Python+OpenCV車牌識別技術的研究”，數學技術與應用，37卷，6期，頁95-97，6月，2019。
[8] N.N. Kamal and E. Tariq “License Plate Tilt Correction: A Review,” Engineering and Technology Journal, Vol. 39, Part B, No. 01, pp. 101-116, 2021.
[9] Y. Lecun, L. Bottou, Y. Bengio and P. Haffner, “Gradient-based learning applied to document recognition”, Proceedings of the IEEE, Vol. 86, Issue. 11, pp. 2278-2324, Nov. 1998.
[10] J. Li, M. Yang and L. Huang, “Tilt Angle Detection and Correction of License Plate Based on Convolutional Neural Network”, International Journal of Mechatronics and Applied Mechanics, Issue. 11, pp. 101-116, 2022.
[11] Nobuyuki Otsu, “A threshold selection method from gray-level histograms”, IEEE Trans. Sys. Man. Cyber, Vol. 9, Issue. 1, pp. 62-66, Jan. 1979.
[12] R.O. Duda, and R.E. Hart. Use of the Hough Transform to Detect Lines and Curves in Pictures, Communications of the ACM, Vol. 15, Issue. 1, pp. 11-15. Jan. 1972.
[13] Deep Learning by Josh Patterson, Adam Gibson. https://www.oreilly.com/library/view/deep-learning/9781491924570/ch01.html