以創新之高效率與高效能叢集分析技術應用於動態影像分析

A New Efficient and Effective Data Clustering Technique for Dynamic Image Analysis

王友哲(Yu-Che Wang)

指導教授 ： 蔡正發

國立屏東科技大學/管理學院/資訊管理系所/碩士(2013年)

https://doi.org/10.6346/NPUST.2013.00253

摘要

本論文提出一創新之使用以叢集為基礎的醫療視力檢查系統。本醫療視力檢查系統運用一創新的以正三角型為基礎的資料叢集分析技術為架構，並以新開發的距離公式應用於標記邊界點（Marked Boundary Object,簡稱MBO）之取得，可改善執行叢集分析時間之耗用。經由實驗結果顯示，本論文中所提出之方法能有效的降低MBO 擴散時資料重複判斷之數量及減少種子點產生之速度，進而可大幅的減少分群時間，其正確率與雜訊濾除率在不同資料集之分析均可超過百分之九十五以上。實驗結果證實本論文所提出之資料叢集分析技術其執行叢集分析之時間皆低於DBSCAN、IDBSCAN、QIDBSCAN、KIDBSCAN 等國際上已知之著名密度式叢集分析技術，它是筆者所知之目前全球執行密度式叢集分析最快速之技術。本論文首先呈現視力量測符號供受測者辨識，並利用隨機方式以提高量測的準確度，透過webcam的影像擷取及影像辨識技術測試受測者之視力。之後再依據視力檢測規則給予新的視力量測符號供受測者辨識，並在系統之螢幕上呈現視力檢查之結果。本論文以新開發的EMIDBSCAN叢集分析演算法應用於動態影像辨識，經由叢集分析演算法的執行，偵測獲取的視力檢測連續圖片得知受測者所示之視力檢測方向訊息，即時判斷量測是否正確並記錄之。最後，再將量測數據計算產出。操作過程中，檢驗人員只需於量測前執行系統設定，待受測者就定位後，即可開始進行量測。

關鍵字

影像辨識 ； 叢集分析演算法 ； 資料探勘 ； 視力量測

並列摘要

The contents of abstract in this thesis:
This thesis presents a novel image recognition technique with clustering based scheme for medical eyesight inspection. The proposed image recognition method employs a density-based clustering technique with equilateral triangle sampling and new clustering distance measure method to obtain three MBOs (Marked Boundary Objects), and thus it can lower the data clustering execution time. According to the experimental results, the proposed clustering method is faster than the existing well-known density-based clustering techniques because it eliminates the search for expansion seeds. The experimental results confirm that the proposed technique has very high clustering accuracy and noise filtering rate, and is faster than the well-known DBSCAN, IDBSCAN, QIDBSCAN and KIDBSCAN schemes. As the authors’ best knowledge, the proposed clustering method is the fastest density-based clustering technique in the world currently. Firstly, the work gives the symbols of the medical eyesight inspection randomly and automatically and then utilizes webcam to capture the hand gesture of the testee. Moreover, the medical eyesight inspection system records the inspection results and give the next appropriate symbol to testee according to medical inspection rules. Finally, the eyesight inspection system displays the exact eyesight of the testee.

並列關鍵字

image recognition ； density-based clustering ； data mining ； eyesight inspection

參考文獻

[5] 黃士育, "新的有效率之密度式分群技術之設計與應用," 碩士, 資訊管理系所, 屏東科技大學, 屏東縣, 2009.

[6] 宋俊毅, "一個有效率的密度式分群演算法與取樣技術之研究," 碩士, 資訊管理系所, 屏東科技大學, 屏東縣, 2009.

[4] 黃堂維, "QIDBSCAN：一個快速的改良密度式分群演算," 碩士, 資訊管理系, 國立屏東科技大學, 2011.

[7] wiki. (2013). YUV. Available: https://zh.wikipedia.org/wiki/YUV

[10] B. Borah and D. K. Bhattacharyya, "An improved sampling-based DBSCAN for large spatial databases," International Conference on Intelligent Sensing and Information Processing, pp. 92-96, 2004.