應用於車牌辨識之雙行車牌偵測與字元切割

Abstract

隨著自動化牌照定位與辨識系統日益普及，其準確率的要求也越來越高。大部分的牌照定位系統可分為兩種，ㄧ種是使用影像處理與電腦視覺技術，基於梯度(gradient)計算的索貝爾(Sobel)邊緣偵測方法，尋找紋理密度高的區域當作可能的牌照位置；另一種則是基於哈爾(Haar-like)特徵的自適應性增強(Adaptive Boosting)的機器學習算法，其準確率都接近100%。牌照定位完成後，以投影的方式對偵測到的牌照進行文字切割，最後文字辨識的準確率也都有95%以上。然而幾乎所有的牌照辨識系統都是針對單行車牌做處理，遇上雙行牌照案例的話，便會導致較差的辨識率。為了解決這類雙行牌照的問題，本論文提出基於影像處理的方法，進行歪斜牌照的校正與多行牌照的處理。首先，我們會進行牌照的歪斜校正並判斷是否出現雙行的情況，在經過分行處理與單行文字切割後，便可以套用文字辨識系統。經過實驗，本論文提出的方法確實能夠有效地處理多行牌照的情況。

As automatic license plate (LP) localization and recognition getting popular, the requirement of accuracy is rising, too. Most of LP localization systems can be divided into two categories, one is based on technique of image processing and computer vision using edge detection methods such as Sobel operator for gradient computation and search for areas with high texture density as reasonable LP positions; the other is based on machine learning using adaptive boosting with Haar-like features. The accuracy of both methods is close to 100%. Based on the results of localization, character segmentation can be performed by projection and final license plate recognition (LPR) rate up to 95% can be achieved. However, almost all LPR methods are aimed to handle single-line LPs, and may lead to poor recognition rate for the case of double-line LP. To handle such a problem, we propose an image processing method to handle skewed and double-line LP. The proposed approach first performs skew correction of LP and determines whether it has double lines. After line separation, followed by character segmentation performed for each line, single-line LPR can be applied. Experimental results show that the proposed approach is indeed effective in dealing with the situation of skewed LP which may have single or double lines.