Multi-pose Ear Recognition Based on Improved Locally Linear Embedding

In this paper, an improved Locally Linear Embedding method is proposed and used in the multi-pose ear recognition. In the traditional LLE method, the neighbors of a data point are selected by using k-nearest neighbor algorithm or e-nearest neighbor algorithm. Both of these methods neglect the different neighborhood of each data point and use a uniform way to select neighbors which leads to the sensitiveness of LLE method to the sparseness of data structure. Some sophisticated methods are used to avoid this problem, but they may be time-consuming. The improved idea solves this problem by selecting a basic value Kb first and partially calculating the reconstruction error within a ball with a radius fixed by multiplying the distance of the Kb-th nearest point with an h, selecting the value of K with minimal reconstruction error. Experimental results on multi-pose ear recognition show the improvement and effectiveness of this method.