PCA-whitening CSS shape descriptor for affine invariant image retrieval

In this paper, we introduce the Principal Component Analysis (PCA)-whitening Curvature Scale Space (CSS) affine-invariant shape descriptor, which extends the CSS shape descriptor to an affine invariant shape descriptor. PCA-whitening is applied to transform the shape contours into their canonical forms. The CSS image maxima are then extracted from those canonical contours as shape descriptors, which can be compared through maxima matching. We tested the descriptors by using them as features in shape based silhouette image retrieval. Experiments on a 1890 silhouette image database show that the proposed descriptor has a promising overall retrieval rate of 91.02%, which is 10.03% more than that of its countparter using traditional affine parameterization.