基于Fisher准则特征融合策略的PNN车牌汉字识别  

PNN Recognition of Chinese Characters on License Plate Based on Fisher Discriminant Criterion Feature Fusion Strategy

作　　者：高全华[1] 张士勇[1] 孙锋利[2] 

机构地区：[1]长安大学理学院,陕西西安710064 [2]西北工业大学电子信息学院,陕西西安710077

出　　处：《现代电子技术》2010年第8期106-110,共5页Modern Electronics Technique

基　　金：国家自然科学基金资助项目(10774076)

摘　　要：为克服图像二值化预处理的不利影响,提高汉字字符图像特征的表征能力,降低传统算法的训练复杂度,在此直接应用车牌字符灰度图像,基于Fisher准则对提取的Pseudo-Zernike矩特征和Gabor变换特征进行融合,在最佳鉴别矢量方向上求得表达能力更强的融合特征。训练结构简单、速度较快的概率神经网络作为识别器,实现车牌汉字的正确、快速识别。试验表明,此算法无需对车牌汉字图像二值化,与所采用的概率神经网络识别器相比,传统BP,SVM等的训练复杂度降低,速度提高,经表征能力更强的融合特征训练后,具有更高的识别准确率。A novel algorithm for recognizing Chinese characters on license plate is proposed so as to achieve three purpo- ses： overcoming disadvantages of two-value preprocessing, raising the characterization ability of image features and decreasing training complexity of traditional algorithms. According to Fisher discriminant criterion, two kinds of image features, Pseudo- zernike moments and Gabor transforming coefficients are fused along the direction of the optimal discriminant vectors. The new fused feature has better performance to characterize the image. Probabilistic neural network（PNN） is trained by new fu- sion features as the classifier due to its simple structure and quick learning rate. Numeral experiment shows this algorithm does not need two-value prelsrocessing and the obtained classifier has low computational complexity as well as high recognition rate compared with traditional classifiers such as BP and SVM.

关 键 词：FISHER准则 鉴别矢量 特征融合 概率神经网络 车牌汉字识别 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]