Statistical LLE for Multi-view TSL Hand Shape Recognition
碩士 === 國立成功大學 === 資訊工程學系碩博士班 === 94 === The image-based object recognition problem becomes complicated when the objects of interest are not posed at a fixed view. In recognition of sign language, the variation of a hand shape due to multiple views and the large number of hand shapes (classes) yield...
| Main Authors: | , |
|---|---|
| Other Authors: | |
| Format: | Others |
| Language: | en_US |
| Published: |
2006
|
| Online Access: | http://ndltd.ncl.edu.tw/handle/45152572968684954377 |
| Summary: | 碩士 === 國立成功大學 === 資訊工程學系碩博士班 === 94 === The image-based object recognition problem becomes complicated when the objects of interest are not posed at a fixed view. In recognition of sign language, the variation of a hand shape due to multiple views and the large number of hand shapes (classes) yield a data distribution with a complicated nonlinear structure in the feature space. This makes it difficult to preserve the class separability under a linear transformation of dimensionality reduction. The locally linear embedding (LLE) is an unsupervised nonlinear dimensionality reduction approach that utilizes the local linearity to discover the low dimensional manifold embedded in the high dimensional space. This suggests that LLE may preserve the neighborhood configuration for the nonlinear structure of the multi-view hand shape data distribution. Although LLE has capability to recover the global nonlinear structure from locally linear fits, the class label information is not taken into account when mapping samples from the high dimensional space to a low dimensional feature space. The statistical LLE is thus proposed herein to improve the capability of LLE associated with classification by incorporating the class label information statistically. In experiments, the statistical LLE was applied to a multi-view TSL hand shape dataset to reduce dimensionality prior to classification. Several UCI datasets were also utilized to validate the proposed approach. Experimental results show that the statistical LLE gave a superior classification performance compared to the original LLE and the linear dimensionality reduction methods such as LDA and PCA.
|
|---|