摘要
针对传统AdaBoost算法中训练耗时及权值扭曲现象,提出了一种融合Haar和MB-LBP特征的车牌检测算法.首先,分别计算使得分类误差最小的Haar与MB-LBP特征,经比较2种特征的误差,选择最优的特征构成弱分类器,并利用AdaBoost算法将各弱分类器进行有效的级联;其次加入了权重阈值,调整了样本权值的更新规则,避免了训练过程中权值扭曲现象.该算法用于检测图像中的数字0～9,利用非最大抑制合并检测到的数字区域,通过车牌的灰度跳变特征过滤候选区域,更精确地定位车牌.实验表明,该方法有效地缩短了训练时间、减少了特征的个数、避免了训练中出现的权值扭曲现象、检测率较高且误检率较低.
Abstract： To address the training timeconsuming and the phenomenon of weights distortions of AdaBoost in license plate detection, a license plate detection algorithm based on the Haar and MBLBP （multiblock local binary patterns） features is presented. First, the Haar and the MBLBP features which make the minimum classification error are calculated, respectively. The best features are chosen after comparing the error of two features to constitute the weak classifiers, and the AdaBoost algorithm is used to obtain a cascade of weak classifiers. Secondly, the weight threshold is added and the updated rules of sample weights are adjusted to avoid the phenomenon of weights distortions in the training process. The proposed algorithm is applied to detect digitals from 0 to 9 in the image. The nonmaximum suppression is adopted to merge detection digital region, and the grayscale transitions of the license plate characteristics are used to filter the candidate region to help locating license plate precisely. The experimental results show that the proposed method can effec tively decrease the training time, reduce the number of features, avoid the phenomenon of weights distortions, and obtain a higher detection rate while reducing the false alarm rate.
出处
《东南大学学报（自然科学版）》
EI
CAS
CSCD
北大核心
2012年第A01期74-77,共4页
Journal of Southeast University：Natural Science Edition
基金
国家自然科学基金资助项目(61005008)
