中文摘要：

针对已有基于图嵌入的半监督算法的缺点,提出了一种半监督有局部差异的图嵌入算法。算法的思想是在保持数据的几何结构同时,最大化样本的差异信息,可有效地防止过学习问题。为了解决小样本问题,采用了差形式的目标函数,并通过参数来调整两部分样本所起作用的大小。最后在ORL和UMIST人脸库上进行了实验,实验结果明显优于已有2种经典算法的识别结果,最优时识别率提高了2.25%和2.23%。

英文摘要：

Aiming at the shortcomings of semi-supervised algorithm based on graph embedding,a novel method called semi-supervised local diversity graph embedding algorithm（ SLDGEA） is proposed.The idea of this algorithm preserves the local structure and simultaneously maximizes the diversity of data,SLDGEA can avoid the data over-learning problem.In order to solve small sample problem,SLDGEA adopts differential form of the criterion function,which can adjust samples effect of the two parts through parameters.Experimental results on ORL and UMIST face databases demonstrate that SLDGEA is better than the existing two kinds of classical algorithms and the optimal recognition rates are improved by 2.25 % and 2.23 %.