PCA in On-Line Handwriting Recognition of Whiteboard Notes: A Novel VQ Design for Use with Discrete HMMs

In this work we further evaluate a recently published, novel vector quantization (VQ) design for discrete HMM-based on-line handwriting recognition of whiteboard notes. To decorrelate the features, a principal component analysis (PCA) is applied. The novel VQ design ensures a lossless representation of the pressure information while modeling the statistical dependencies between the pressure and the remaining features. This is necessary because standard k-Means VQ systems cannot quantize this binary feature adequately although they have been decorrelated, which is shown in this paper. Our experiments show that the new system provides a relative improvement of r = 2.8 % in character level accuracy and a relative improvement of r = 3.3 % in word level accuracy compared to a standard k-means VQ system. Additionally our system is compared and proven to be competitive to a stateof-the-art continuous HMM system yielding a relative improvement of r = 1.4 %. A relative improvement of up to r = 0.8 % in word level accuracy can be reported when using decorrelated features compared to a system omitting the decorrelation.