夜間影像的車牌定位與辨識

根據交通部的統計顯示，夜間交通事故的發生率大約是白天的1.5倍。由於普通相機有限的動態範圍或場景內光線不明的情況，夜晚所拍攝到的影像品質通常較低；所以如何準確提取夜間影像中有價值的線索一直是個值得探討的問題。在這篇論文中，我們提出了一個針對夜晚環境下的汽車車牌辨識系統。車牌字元辨識系統一般可分為五個處理階段：彩色圖像的前處理、車牌定位、車牌校正、字元切割和字元辨識。我們著重在前四個階段。由於在非均勻照明、低對比度、非預期的色偏和複雜背景下做出車牌的定位和辨識並非容易，首先，必須在前處理時進行我們提出的利用改良暗通道增強影像方法以解決這些現象。接著，為了檢測車牌區域，也就是車牌定位與校正步驟的總稱，藉由結合車牌多種特性和數學形態學操作可以用來取得僅含車牌的區域。此外，我們提出了一種改良的Bernsen二元化方法，對於像是低對比、灰塵、汙漬和非均勻光線較能與之抗衡，更重要的是能簡化後續的傾斜校正。在補償了車牌由於拍攝角度所造成的歪斜後，最後，藉由我們的新方法車牌會被分割成數個字元。這些字元再被正規化，並使用以卷積神經網路(Convolutional Neural Networks)實現的分類器完成字元識別。因此，我們所提出的系統可應用於移動式成像設備，有效改善夜間車輛圖像的視覺感受，並提供自動車牌辨識。

關鍵字

彩色圖像增強；高動態範圍； retinex演算法；白平衡；車牌檢測；車牌定位；二元化；傾斜校正；字元分割；字元辨識

並列摘要

According to the Ministry of Transportation, it has indicated that the incidence of road traffic accidents is about 1.5 times higher during the nighttime than during the day. Since the quality of the acquired images are often low at night time owing to the limited dynamics of current cameras and the changeable lighting environment conditions, how the valuable clues can be accurately extracted from nighttime images has been an issue worth exploring. In this thesis, we present a vehicle license plate recognition system especially for vehicle images under a nighttime environment. License plate character recognition system has been generally divided into five processing steps: color image preprocessing, license plate localization, license plate correction, character segmentation and character recognition. We focused on the first four steps. Since non-uniform illumination, low contrast, unwanted color cast and complex backgrounds made localization and recognition challenging, for one thing, an image enhancement method using improved dark channel prior was proposed for image preprocessing to resolve these phenomena. Next, for detecting license plate regions, which was the general term for license plate localization and correction steps, the combination of the various features of license plate and mathematical morphology operations is used to obtain the region contained license plate only. Besides, the proposed improved Bernsen binarization method is robust to degradations such as low contrast, dust, smear and uneven illumination and, more importantly, can simplify the subsequent tilt correction processing. After the skew compensation for license plates due to various camera angle, lastly, the extracted license plate was split into a number of characters from each other by our new segmentation method. These characters were normalized and recognized using the classifier which has been performed using Convolutional Neural Networks (CNNs). Therefore, the proposed system can be applied to mobile imaging devices to effectively improve visual perception for nighttime vehicle images and provide automatic license plate recognition.

並列關鍵字

Color image enhancement ； high dynamic range ； retinex ； white balance ； license plate detection ； license plate localization ； binarization ； tilt correction ； character segmentation ； character recognition

參考文獻

[1] C. Tsai and Z. Yeh, “Contrast compensation by fuzzy classification and image illumination analysis for back-lit and front-lit color face images,” IEEE Trans. on Consumer Electronics, vol. 56, no. 3, pp. 1570-1578, August 2010.

[2] S. Wu, S. Xie, S. Rahardja, and Z. Li, “A robust and fast anti-ghosting algorithm for high dynamic range imaging,” IEEE Conf. Int. Conf. on Image Processing, pp. 397-400, September 2010.

[6] Y. Rao, Y. Zhang, and J. Gou, “Gradient fusion method for night video enhancement,” ETRI Journal, vol. 35, no. 5, pp. 923, October 2013.

[7] E. Reinhard, M. Ashikhmin, B. Gooch, and P. Shirley, “Color transfer between images,” IEEE Computer Graphics and Applications, vol. 21, no. 5, pp. 34-41, Sep/Oct. 2001.

[8] Alexander Toet, “Natural colour mapping for multiband nightvision imagery,” Information Fusion, vol. 4, no. 3, pp. 155-166, September 2003.

國際替代計量

夜間影像的車牌定位與辨識

主題瀏覽