臺灣博碩士論文加值系統

本論文研製一個汽車車牌辨識系統，使用雷登轉換矯正傾斜車牌，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。
車牌定位使用灰階化將影像通道數量由三減少至一，中值濾波過濾雜訊，邊緣偵測
以凸顯影像中的垂直邊緣和使用大津二值化。接著使用形態學運算，經由區塊面積大小、
寬高比例、區塊面積中的輪廓數量作為篩選，取得車牌候選位置。
拍攝車牌可能形成車牌圖像傾斜情形，須做傾斜矯正以利後續辨識。車牌傾斜矯正
包含水平傾斜矯正與垂直傾斜矯正，使用雷登轉換計算出車牌的傾斜角度，對傾斜車牌
進行水平矯正，使用仿射轉換對水平矯正後的車牌做垂直修正，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，先將定位後車牌進行二值化，對其執
行水平投影，切除中間車牌字元以外的區域，接著使用垂直投影，將同行的黑色像素數
目做累加，依照字元間的間隔依序進行切割，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，最後將切割後的字元進行特徵
比對完成字元辨識。

This thesis develops a tilt vehicle license plate recognition system and uses Radon
Transform method for tilted license plate. The architecture is mainly divided into four parts:
license plate detection, license plate tilt correction, characters segmentation and characters
recognition. License plate detection grayscales the image to reduce the number of image
channels from three to one. The median filter filters out noise, and the edge detection method
accentuates vertical edges in images. Otsu's method separates the image into two classes,
foreground and background, further simplifying the image information. Then, morphology is
used to connect specific blocks. The license plate candidate positions are obtained through
block area size, width-to-height ratio, and the number of contours in the block area as filters.

License plate tilt correction includes horizontal tilt correction and vertical tilt correction,
The Radon Transform method is used to calculate the tilt angle of the tilted license plate,
followed by performing horizontal correction. The affine transformation algorithm is used to
perform vertical tilt correction on the horizontally corrected license plate, completing the tilt
correction process.

The character segmentation process includes horizontal projection segmentation and
vertical projection segmentation. First, binarize the positioned license plate and perform
horizontal projection on it to cut off the area other than the middle license plate characters. Then,
use vertical projection to accumulate the number of black pixels in each column and cut them
sequentially according to the intervals between characters to complete the character cutting.

Character recognition uses convolutional neural networks to train character samples, and
finally compares features of the segmented characters to complete the character recognition
process.

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 研究方向 ..................................................................................................................... 1
1.3 文獻探討 ..................................................................................................................... 2
1.4 論文架構 ..................................................................................................................... 3
第二章 車牌定位 ...................................................................................................................... 4
2.1 車牌樣式 ..................................................................................................................... 4
2.2 灰階化 ......................................................................................................................... 5
2.3 中值濾波 ..................................................................................................................... 5
2.4 邊緣偵測 ..................................................................................................................... 6
2.5 大津二值化 ................................................................................................................. 7
2.6 影像形態學 ................................................................................................................. 9
第三章 車牌傾斜矯正與字元切割 ........................................................................................ 11
3.1 雷登轉換 ................................................................................................................... 11
3.2 仿射轉換 ................................................................................................................... 13
3.3 字元切割 ................................................................................................................... 14
第四章 車牌字元辨識 ............................................................................................................ 16
4.1 卷積層(Convolution Layer) ...................................................................................... 16
4.2 池化層(Pooling Layer) .............................................................................................. 17
4.3 平坦層(Flatten) ......................................................................................................... 17
4.4 全連接層(Fully Connected Layer) ............................................................................ 18
4.5 激活函數(Activation Function) ................................................................................ 19
4.6 隨機失活(Dropout) ................................................................................................... 19
4.7 卷積神經網路架構 ................................................................................................... 20
第五章 實驗結果 .................................................................................................................... 22
5.1 實驗環境 ................................................................................................................... 22
5.2 實驗結果 ................................................................................................................... 23
5.2.1 車牌定位結果 ........................................................................................................ 23
5.2.2 車牌傾斜矯正結果 ................................................................................................ 24
5.2.3 字元切割結果 ........................................................................................................ 26
5.2.4 字元辨識結果 ........................................................................................................ 26
5.3 車牌辨識失敗情形 ................................................................................................... 27
第六章 結論 ............................................................................................................................ 29
參考文獻 .................................................................................................................................. 30
Extended Abstract ..................................................................................................................... 31

[1] 李孟學，2017，“車牌模擬與車牌定位實作”，中華科技大學電子工程系研究所碩士
論文。

[2] 趙洋，閔升鋒，李大舟，2022，“基於HSV空間顏色和紋理特徵的車牌定位算法研
究”，瀋陽化工大學學報，36卷，1期，pp. 90-95。

[3] 咸志杰，方芹芹，季偉偉，李海霞，2021，“基於車牌混和特徵的車牌定位算法研究”，
現代信息科技，5卷，2期，pp. 76-79, 84。

[4] 林語揚，2022，“基於卷積神經網路的傾斜矯正之車牌辨識模型”，國立虎尾科技大
學自動化工程系碩士倫文。

[5] 陳韋榤，2009，“應用快速廣義霍夫轉換於歪斜車牌定位與角度估算”，國立臺灣科
技大學高分子工程系碩士論文。

[6] Hitesh Rajput, Tanmoy Som and Soumitra Kar, 2016, “Using Radon Transform to
Recognize Skewed Images of Vehicular License Plates”, IEE Computer Society, Vol. 49, Issue:1, pp. 59-65, 14 January.

[7] N.N. Kamal and E. Tariq, 2021, “License Plate Tilt Correction: A Review”, Engineering and Technology Journal, Vol. 39, Part B, No. 01, pp. 101-116, 25 March.

[8] Bingyan Lin, Fengping He, et al, 2023, “License Plate Character Recognition with Lightweight Convolutional Neural Networks”, IEEE 2nd International Conference on
Electrical Engineering, Big Data and Algorithms (EEBDA), Changchun, China, 24-26
February.

[9] Qinglin Wu, Caiyun Xu, et al, 2022, “Research on License Plate Recognition System Based on OpenCV”, International Symposium on Computational Intelligence and Design (ISCID), Hangzhou, China, 01 February.

[10] Nobuyuki Otsu, 1979, “A Threshold Selection Method from Gray-Level Histograms”, IEEE Transactions on Systems, Man, and Cybernetics, Vol 9, Issue:1, pp. 62-66, January.

[11] A. Jain and S. Ansari, 1984, “Radon Transform Theory for Random Fields and Optimum Image Reconstruction from Noisy Projections”, ICASSP '84. IEEE International Conference on Acoustics, Speech, and Signal Processing, San Diego, CA, USA, 19-21
March.

[12] Kunihiko Fukushima, 1980, “Neocognitron: A self-organizing Neural Network Model for A Mechanism of Pattern Recognition Unaffected by Shift in Position”, Biological Cybernetics, Vol. 36, Issue:4, pp. 193-202, April.