自動眼睛偵測及眼鏡反光消除

Abstract

眼睛偵測經常用在許多研究上，如：人臉辨識、瞌睡偵測、人眼注視追蹤等等；但是，當佩戴眼鏡時，眼睛偵測常會因為眼鏡本身顏色及鏡片上反光的干擾，造成誤判，因此需要對相關的干擾作處理。在本篇論文中，我們提出一套演算法，使其能夠在一張人臉影像中偵測眼睛的位置，並且在該人臉有戴眼鏡時，去除眼鏡及眼鏡內反光的干擾。此系統包含三個模組：臉部位置偵測，眼睛偵測，以及當受測者有戴眼鏡時，去除眼鏡反光的方法。首先，我們使用通用膚色圖偵測臉部區域，以確保在周遭光源條件變化時，仍有足夠的適應性。接著，我們採用邊緣及角的偵測再搭配非等向性擴散轉換，來偵測眼睛區域以及分離出眼鏡反光區域內關於眼睛的資訊。反光分離的原理是根據當反光影像可以成功的分離成兩張分別為前景層圖及反射層圖時，則該兩張成功分離影像的邊緣和角的總和，是所有可能分解的解中最小值。結果顯示，該反光原理用於鏡片上的反光分離處理是有效的且可以得到良好的反光分離。

Eye detection has been applied to many applications, for instance, human or faces recognition, eye gaze detection, drowsiness detection, and so on. However, eye detection often misdiagnoses for the interference caused by glasses when one wears spectacles. This thesis addresses an algorithm to automatically detect the eye location from a given face image and separate reflections within glasses while one has worn glasses. Our system consists of three modules: face segmentation, optic-area detection, and the separating of glasses reflections while one has worn glasses. First, we use the universal skin-color map to detect the face regions, which can ensure sufficient adaptability to ambient lighting conditions. Then, we proposed a novel method to detect the eye region and separate the reflection within glasses based on edge detection, corner detection, and anisotropic diffusion transform. The principle of separating reflection is based on that the correct decomposition of the reflection image whose summation of corners and edges is the smallest among all possible decompositions. The simulation and results demonstrate that the principle of separating reflection can be applied to the reflection within glasses effectively and result in good reflection separation.