針對受過人工視網膜手術之病患之視覺輔助系統

Abstract

受過人工視網膜手術的病患的視覺品質往往被低解析度的影像畫面所影響，由於相關技術上面的瓶頸，使得如何有效地利用現有的低解析度畫面成為一個重要的議題。本篇論文提出一個眼控視覺輔助系統，期望可以運用眼控的方式針對使用者感興趣的物件做縮小放大：當使用者凝視時放大畫面，閉眼時縮小畫面。以達到有效率地使用低解析度畫面的目的。為了更真切地模擬患者實際使用的情況，我們特別設計了一個頭戴式眼控顯示器方便研究人員穿戴時可以獲得近似病患低解析度的視覺感受。本論文的核心演算法主要是圍繞在眼控這個主題上，在以往的相關研究中，得到的成果大多對側眼以及受到眼皮遮擋的眼睛無法做出準確判讀，而我們提出一種以影像梯度向量射線為基礎的演算法來克服前述的問題。首先，我們利用一特殊設計過的動態閥值來去除掉眼球反光部分，然後再將每一個像素對其負梯度向量作一射線，之後在此射線圖中尋找較大的連通群體並進而分析其關係然後選取出瞳孔中心。實驗顯示在正眼、側眼以及被遮擋的眼睛，此演算法都有著優異的成果，除此之外對於光影變化劇烈的環境也能獲得相當良好的表現。因此在眼控縮放目標物的實驗中也獲得相當令人滿意的結果。

The vision quality of the visually impaired people treated by retinal prosthesis is largely affected by the vision resolution. Due to some related technical bottlenecks, how to make effective use of existing low-resolution screen becomes an important issue. This paper presents an eye-controlled visual aid system to achieve an efficient use of low-resolution vision by zooming in the user's interested objects (when staring, zoom-in; when closing eye, zoom-out). In order to better simulate the actual low-resolution vision of the patients, we have also designed a head-mounted eye-controlled display for the research purpose. In this thesis, the proposed algorithm is mainly about eye-tracking. The existing eye-tracking methods often have difficulties in correctly detecting pupil center from side eyes or occluded eyes. While we have proposed an emission of gradient orientation based algorithm to overcome the aforementioned problems. First, we use a specially designed dynamic threshold to remove the reflection of eye, then we emit a negative gradient orientation ray from each pixel. After obtaining this ray diagram, we extract the maximum cluster and analyze it to find the pupil center. The proposed algorithm exhibits many desired properties of front eyes, side eyes and occluded eyes. Furthermore, the method is invariant to severe lighting changes. Thus we have a desirable result in the zoom-in/out experiment.