Ear Recognition Using LLE and IDLLE Algorithm - Semantic Scholar
Ear Recognition Using LLE and IDLLE Algorithm Zhaoxia Xie and Zhichun Mu School of Information Engineering, University of Science and Technology Beijing, 100083, China [email protected] Abstract Although ear recognition has been researched widely, there still exist some problems to be resolved in depth such as multi-pose ear recognition which is rarely focused on. In this paper, different from previous methods used for ear recognition ， one nonlinear algorithm, called locally linear embedding (LLE) belonging to manifold learning technique is introduced, and an improved locally linear embedding algorithm (IDLLE) is proposed considering the disadvantage of the standard LLE algorithm. Comparison to PCA and KPCA, experimental results demonstrate that applying LLE for multi-pose ear recognition can obtain better recognition results; and using the IDLLE can further improve the recognition performance as for multi-pose ear recognition, which greatly shows the validity of this improved algorithm.
1. Introduction Biometrics is being more and more widely used in recently year owing to the irreproducible characteristics of the human body. As one kind of biometrics, the ear has its own characters [7]: the structure of the ear is rich and stable, and does not change radically over time; the ear is less variability with expressions, and has a more uniform distribution of color than faces. These unique characters of the ear make it possible to make up the drawbacks of other biometrics and to enrich the biometrics identification technology. At present ear recognition technology has been developed from the initial feasible research to the stage of how to enhance ear recognition performance further, for instance, 3D ear recognition [1], [2], ear recognition with occlusion [3], and multi-pose ear recognition etc. Multi-pose ear recognition is referred to when the angle between the ear and the camera changes, the shape of the ear will be distorted, resulting in the decrease of the recognition performance. Therefore it is necessary to discuss this problem deeply for many researchers. This
978-1-4244-2175-6/08/$25.00 ©2008 IEEE
paper mainly addresses the issue of the multi-pose ear recognition using 2D ear images without occlusion under the invariant illumination condition. Methods using ear geometrical features which are extracted for ear recognition were easily influenced by pose variations, and evidently are not feasible for human ear recognition with varying poses [4], [5], [6], [7], [8]. Principal component analysis (PCA) was used for ear recognition [9]. However, when data points are distributed in a nonlinear way such as pose variations, PCA fails to discover the nonlinearity hidden in data points. Kernel principal component analysis (KPCA) [10] was also used for ear recognition, but projection results aren’t visual using KPCA, and the performance of this method is greatly influenced by kernel parameters. To deal with the problem of pose changes for ear recognition, manifold learning technique such as locally linear embedding (LLE) [11] is introduced for multi-pose ear recognition. Different from previous methods for ear recognition, this method not only can find the intrinsic structure of data points by constructing embedded space which preserves their topology relationship, but also can transform the nonlinear problem into the linear problem by local linear fitting and enable us to visualize data structure clearly, hence this method may better deal with the nonlinear problem, especially for multi-pose ear recognition. The remainder of this paper is organized as follows: Section 2 presents the standard LLE algorithm. The improved LLE algorithm (IDLLE) is given in section 3. Experiment results conduced on USTB ear database are shown in section 4. Finally, the conclusion and future work are discussed.
2. Locally linear embedding Input vectors {X1, X 2 ,", X N }, i = 1,", N , each of dimensionality d ( X i ∈ R d ) are given. Output
vectors {Y1,Y2 ,", YN }, i = 1,", N , will be obtained, each of dimensionality m ( Yi ∈ R ), where m
1. Introduction Biometrics is being more and more widely used in recently year owing to the irreproducible characteristics of the human body. As one kind of biometrics, the ear has its own characters [7]: the structure of the ear is rich and stable, and does not change radically over time; the ear is less variability with expressions, and has a more uniform distribution of color than faces. These unique characters of the ear make it possible to make up the drawbacks of other biometrics and to enrich the biometrics identification technology. At present ear recognition technology has been developed from the initial feasible research to the stage of how to enhance ear recognition performance further, for instance, 3D ear recognition [1], [2], ear recognition with occlusion [3], and multi-pose ear recognition etc. Multi-pose ear recognition is referred to when the angle between the ear and the camera changes, the shape of the ear will be distorted, resulting in the decrease of the recognition performance. Therefore it is necessary to discuss this problem deeply for many researchers. This
978-1-4244-2175-6/08/$25.00 ©2008 IEEE
paper mainly addresses the issue of the multi-pose ear recognition using 2D ear images without occlusion under the invariant illumination condition. Methods using ear geometrical features which are extracted for ear recognition were easily influenced by pose variations, and evidently are not feasible for human ear recognition with varying poses [4], [5], [6], [7], [8]. Principal component analysis (PCA) was used for ear recognition [9]. However, when data points are distributed in a nonlinear way such as pose variations, PCA fails to discover the nonlinearity hidden in data points. Kernel principal component analysis (KPCA) [10] was also used for ear recognition, but projection results aren’t visual using KPCA, and the performance of this method is greatly influenced by kernel parameters. To deal with the problem of pose changes for ear recognition, manifold learning technique such as locally linear embedding (LLE) [11] is introduced for multi-pose ear recognition. Different from previous methods for ear recognition, this method not only can find the intrinsic structure of data points by constructing embedded space which preserves their topology relationship, but also can transform the nonlinear problem into the linear problem by local linear fitting and enable us to visualize data structure clearly, hence this method may better deal with the nonlinear problem, especially for multi-pose ear recognition. The remainder of this paper is organized as follows: Section 2 presents the standard LLE algorithm. The improved LLE algorithm (IDLLE) is given in section 3. Experiment results conduced on USTB ear database are shown in section 4. Finally, the conclusion and future work are discussed.
2. Locally linear embedding Input vectors {X1, X 2 ,", X N }, i = 1,", N , each of dimensionality d ( X i ∈ R d ) are given. Output
vectors {Y1,Y2 ,", YN }, i = 1,", N , will be obtained, each of dimensionality m ( Yi ∈ R ), where m
