Improved locally linear embedding and its application on multi-pose ear recognition

As a supplement of biometrics recognition, human ear recognition has been researched widely and made greatly progress. However, human ear recognition technology still has many problems to be resolved in depth such as multi-pose ear recognition. Principal component analysis and linear discriminant analysis have their limitations when data set represents highly nonlinear especially such as the changes of ear pose. Locally linear embedding, an unsupervised learning algorithm, is proposed in recent years. This method can better solve the problems of non-linear dimensionality reduction. But lacking of the label information of data set, it is not suitable for ear classification and recognition if directly applied to multi-pose ear recognition. In this paper, an improved locally linear embedding algorithm is presented. Experiments show that the rate of multi-pose ear recognition can be improved in some extent, compared to that of ear recognition using locally linear embedding algorithm only.