Constrained Local Model With Independent Component Analysis And Kernel Density Estimation: Application To Down Syndrome Detection

Statistical shape models generally characterize shape variations linearly by principal component analysis (PCA), which assumes that the non-rigid shape parameters are drawn from a Gaussian distribution. This practical assumption is often not valid. Instead, we propose a constrained local model based on independent component analysis (ICA) and use kernel density estimation (KDE) for non-parametrically modeling the distribution of the shape parameters. The model fitting is achieved by maximum a posteriori via the expectation-maximization algorithm and results in a mean shift-like update optimizer. The proposed approach is capable of modeling non-Gaussian shape priors and significantly outperformed the PCA-based model (p=0.03) and ICA-based model with Gaussian shape prior (p=0.01) in experiments to detect facial landmarks. Moreover, we applied the model to Down syndrome detection from frontal facial photographs and obtained higher accuracy than the best results reported in literature.