Knowledge-embedded symmetry transform for accurate eyeball location

It is common sense that the more we know, the better we do. By means of embedding more distinctive task-specific knowledge into a model, we achieved superior performance in our application of eyeball location. In this paper, a general formulation of a knowledge-embedded symmetry transform (KEST) is proposed. The key to this new design are the three newly introduced knowledge operators. The first is the symmetry-related weighting, the second is the mask operator serving as a localization factor, the last is the expected directional template. It is due to the embedding of the effective constraint knowledge that better screening efficiency, accuracy, simplified preprocessing and post-processing treatment are achieved.