A low rank model based improved eye detection under spectacles

Eye detection is a primary step in many applications such as face recognition, iris recognition, driver fatigue detection, gaze tracking etc. Occlusion by spectacles, glare and secondary image formations deteriorate its performance. In this paper, we formulate the glare/reflection removal as a classification problem and employ Low rank decomposition technique to overcome these challenges. We provide an in-depth analysis by comparing various low rank decomposition formulations and propose a simple preprocessing step to improve the detection accuracy. Experimentation on CASIA NIR-VIS 2.0 facial database validates the proposed preprocessing method.