| Summary: | 碩士 === 國立臺灣科技大學 === 機械工程系 === 100 === An appearance-based coding scheme, called Cascade Local Deformation Code (CLDC), is proposed for expression recognition. CLDC has two component codes, Human Observable Code (HOC) and Haar-like Feature Code (HFC). The HOC encodes the local deformation regions caused by facial muscle contractions observable to humans, and the HFC encodes the Haar-like features selected by an AdaBoost algorithm. Given a training set, one first selects the observable local deformation regions, and trains a HOC detector which encodes the local deformation regions into HOC codewords according to seven predefined expressions. The training set is also used for the extraction of Haar-like features and encoding of the features into HFC codewords for the seven expressions. The combination of HOC and HFC gives the CLDC, which is proven to outperform either component in the decoding phase for the expression recognition on disjoint testing sets. Experiments on the CK+, JAFFE and the latest FERA databases show that the performance of the CLDC is competitive to the state-of-the-art approaches.
|
|---|
