目前車牌辨識技術越來越成熟,應用的方面也變得更廣泛,車牌辨識之流程,基本上是先進行車牌偵測,接著再進行車牌辨識,許多研究都是用單純深度學習或者圖形辨識技術來達成,很少結合兩種技術的優缺點,因此,本論文利用工業用電腦視覺函式庫-Euresys Open eVision與深度學習做結合,主要透過深度學習來克服複雜背景下的車牌偵測,並且運用工業用電腦視覺函式庫來提升辨識效率。本論文提出的技術主要分為三個階段:第一階段為使用深度學習來偵測車牌在影像上的位置,第二階段進行車牌視角轉換,將歪斜的車牌轉為正面視角,接著在第三階段進行字元辨識。經過實驗的分析統計後,本研究之車牌辨識準確率達96.7%以上,單一車牌的辨識時間為77.2ms,在實際環境下達成一定的實用性。
At present, the license plate recognition (LPR) technique is gradually maturing and widely used in daily life. License plate detection and optical character recognition are two basic steps. Many studies simply utilize deep learning technique, or traditional computer vision technique to realize LPR. Therefore, deep learning technique and industrial computer vision library (Euresys Open eVision) are attempted to be integrated in this study. The proposed approach can be divide into three stage: the first stage is to apply deep learning to locate the license plate in the image. The second stage is license plate correction. It will transform license plate image to the image with expected angle. Finally, it will recognize the license plate by splitting the individual character. In the experimental study, the accuracy of the implemented system can reach 96.7% with an average execution time of 77.2ms. It shows the proposed approach can achieve a high accuracy in the practical environment.