Towards a framework for recognising Sign language alphabets captured under arbitrary illumination

Our work addresses the problem of automatically recognising a Sign Language alphabet from a given still image obtained under arbitrary illumination. To solve this problem, we designed a computational framework that is founded on the notion that shape features are robust to illumination changes. The statistical classifier part of the framework uses a set of weighted, self-learned features, i.e., binary relationship between pairs of pixels. There are two possible pairings: an edge pixel with another edge pixel, and an edge pixel with a non-edge pixel. This two- pairing arrangement allows a consistent 2D image representation for all letters of the Sign Language alphabets, even if they were to be captured under varying illumination settings. Our framework, which is modular and extensible, paves the way for a system to perform robust (to illumination changes) recognition of the Sign Language alphabets. We also provide arguments to justify our framework design in term of its fitness for real world application.