关键词: 核主元分析(KPCA), 字符识别, 图像, 降维, 车牌

Abstract: Analyzing the shortages of KPCA in feature extraction of license plate characters/figures, a new algorithm is proposed using multiple mean-vectors instead of original image information vectors to compute kernel matrix. Thus, the kernel matrix's dimensions can be reduced substantially with the image information of characters/figures kept up efficiently. Experimental results showed that the algorithm does not reduce the accuracy of recognition, during reducing greatly the number of dimensions of input data and the computing time can be shortened greatly, thus meeting satisfactorily the technological requirement for real-time recognition of vehicle license plates. The comparative test results revealed that this algorithm's performance indices are higher than both the conventional PCA and FLD algorithms.