Local Orientation Patterns for 3D Surface Texture Analysis of Normal Maps: Application to Facial Skin Condition Classification

In this paper we investigate methods for analysing 3D surface texture for automated facial skin health assessment. We propose a Texture Spectrum inspired method for analysing surface texture from normal maps. A number of approaches for extracting invariant region descriptors from 3D volumetric data have been proposed, yet 3D surface texture analysis has been somewhat neglected. The method we introduce characterizes a normal map with a descriptor based on an extension of Texture Spectrum. We propose two methods for assessing the variation of orientation between two normals. The first applies a threshold on their dot product, while the second variant compares their polar and elevation angles directly. We tested both variants by classifying some facial skin conditions from high resolution normal maps. The results show a clear improvement using the second proposed pattern function over the first on classifying high frequency skin conditions such as visible pores and wrinkles.