【作者】 李昱；

【导师】 任获荣；

【作者基本信息】 西安电子科技大学 ， 测试计量技术及仪器， 2010， 硕士

【摘要】 作为一种非线性降维技术,流形学习算法能更好地发现复杂数据集的内在结构,为数据的进一步处理提供基础。目前已出现一些成熟的流形学习算法,并在模式识别,机器视觉等领域取得了成功应用。目前的流形学习算法多是无监督的算法,没有利用到样本的先验信息。如能获得部分样本的先验信息,可以在训练阶段利用这些信息来提高分类器的分类性能,对普通学习算法进行推广得到其半监督算法。本文主要研究了流形学习算法的半监督推广。在研究和分析了目前的一些方法后,基于传统的流形学习方法拉普拉斯特征映射(LE)算法,提出了半监督的拉普拉斯特征映射(SS-LE)算法。该算法利用少量样本的已知信息,可以大幅提高所求解的低维嵌入坐标的精度。另外从计算复杂度和准确度方面比较了半监督拉普拉斯和半监督局部线性嵌入算法(SS-LLE)的性能。随着邻域数k的增加,SS-LE的计算复杂度远低于SS-LLE,而精度只是相对略有下降,且当k取较小的值时,SS-LE算法就已经可以取得和SS-LLE算法最高精度接近的结果。最后使用人造数据和真实数据验证了SS-LE算法在数据降维,人脸识别,可视化和视频目标跟踪中的应用,取得了预期的效果。更多还原

【Abstract】 As a non-linear dimensionality reduction method, manifold learning can discover the intrinsic construction of complex data for further processing. Several manifold learning algorithms have been developed and were widely used in pattern recognition and machine vision etc. area.Most of the existing manifold learning algorithms are unsupervised method without using prior-information. If we can obtain some prior-information of training samples, we can use this information to help the study step to increase the classification ability, so improving the original method to semi-supervised method.The objective of this paper is to research about the semi-supervised manifold learning. After analyzing existing methods, we propose a semi-supervised manifold learning algorithm based on Laplacian Eigenmaps, named semi-supervised Laplacian Eigenmaps (SS-LE). SS-LE uses prior information of very few samples to calculate more accurate low dimensional embedding coordinates. We also analyze its superior in accuracy and computation complexity compared with other methods. Compared with SS-LLE, the computation time of SS-LE is dramatically reduced and the drop of the accuracy is within an acceptable range. Besides, the accuracy of our method can achieve nearly best result of SS-LLE only need to set K (the number of neighbors) to a relative small value. We demonstrated the usefulness of our algorithm by synthetic and real world problems, especially its efficiency in object tracking problem.更多还原