基于子分类器融合的部分遮挡人耳识别

Ear recognition under partial occlusion based on sub-classifier fusion

遮挡是人耳识别中一个难以回避的问题,本文对人耳受到部分遮挡的识别问题进行了研究。在分析人耳不同位置的鉴别能力的基础上,提出了一种基于决策层的子分类器融合的识别方法:首先将图像分割为若干连续但不重叠的子窗口;对每个子窗口,利用邻域保留嵌入算法进行特征提取,然后利用最近邻分类器进行识别;根据这些子分类器识别率的高低,可以得到相应的子窗口的鉴别能力;接下来再利用具有较高鉴别能力的子分类器进行融合识别来解决部分遮挡问题。在USTB人耳图像库上的实验结果表明人耳图像中确实有部分区域具有更高的鉴别能力,利用这些区域即可进行身份识别,而且本文提出的基于局部信息融合的方法比基于原始图像直接进行识别的方法具有更高的识别率,尤其适合于解决人耳识别中的部分遮挡问题。

As a new biometrics authentication technology,ear recognition remains many unresolved problems;one of them is ear occlusion.This paper deals with ear recognition for partially occluded ear images on the basis of analyzing the discrimination ability on different locations.The ear image is firstly divided into continuous but non-lapped sub-windows.For each sub-window,Neighborhood Preserving Embedding（NPE） algorithm is used for feature extraction and nearest neighbor rule is used for classifier design.According to the recognition rates of these sub-classifiers,we can get the discrimination abilities of corresponding sub-windows.Final classification is made by sub-classifier fusion of the sub-windows with top ranked discrimination abilities.This fusion strategy is helpful for solving the partial occlusion problem.Experimental results on the USTB ear image database have illustrated that there indeed exist some sub-regions possessing higher discrimination ability.Using only a few sub-regions we can perform ear recognition.The proposed sub-classifier fusion approach can obtain higher recognition rate than the method using whole image for recognition,especially for ear recognition under partial occlusion.