Ocular Feature Detection Under Spectacles

Abstract

Ocular image processing is a primary step in many applications such as iris recognition, driver fatigue detection, gaze tracking etc. Occlusion, glare, and secondary reflections formed due to and on the spectacles—results in poor detection and localization of eye features. In this thesis, the challenges that arise from the usage of spectacles are termed as “The spectacle problem”. In literature most of the existing algorithms are targeted to remove occlusion due to spectacle frame and have not considered the influence of specularities. To alleviate the issues of specularities and to improve ocular feature detection under spectacles two new approaches viz. (i) Low rank model-based approach (multiple image-based approach), (ii) Dichromatic model-based approach (single image-based approach)—are proposed. In the proposed Low rank model-based approach, glare/reflection removal is formulated as a classification problem and Low-rank decomposition technique is employed to overcome these challenges. Experimental analysis on the proposed approach revealed that due to the inability to capture eye dynamics—this proposed approach suits for detection and localization of coarser eye features only. So, by employing the Dichromatic reflection model (DRM) and extending it to Hue-Saturation-Value colour space, a single image-based approach is proposed to overcome the limitations of multiple image-based approaches. By solving the least-squares problem of the DRM, reflection separation is implemented on a single-pixel level to decompose a given image into diffuse (spectacle problem free) and specular components. Human alertness-level detection from eye blinks (under spectacles) is considered as a case study to validate the proposed algorithm. Literature review on the available databases lead to an observation that a complete database which enlists the challenges of spectacle problem along with the benchmark data is unavailable. Therefore, an experiment is designed to yield a video database of 58 human subjects wearing spectacles and are at different levels of alertness. Experiments on this database demonstrate that the proposed Dichromatic model-based approach achieves desirable spectacle problem removal results within minimum execution time compared to the state-of-the-art. Additionally, an alertness-level index that aids in the cross-verification of any visual cue based alertness level detection approach is proposed.