An image statistics approach towards efficient and robust refinement for landmarks on facial boundary

In the real-world unconstrained face recognition scenarios, automatic facial landmarking scheme using the active shape model (ASM) usually gives non-ideal results, especially at the facial boundary. This is because the local subspace methods such as the principal component analysis (PCA) used in the ASM does not properly discern skin texture and background with very similar photometric and textual properties, thus fails to accurately locate the facial boundary. In this work, we have novelly developed a robust image statistics approach to efficiently refine the landmarks on facial boundary. Moreover, with the aid of banana wavelets to highlight the facial boundary, our proposed approach can deal with even more difficult task. This algorithm can dramatically increase the accuracy of landmarks on facial boundary for unconstrained facial images with minimum computational expense since this method is purely based on image statistics with no training stages involved at all. We have shown the effectiveness of our proposed methods on the GBU database where the refined landmarks yield much lower MSE from the ground truth.