影像上的轉折點具有幾何轉換(如平移、旋轉、縮放…等)之不變性，在電腦視覺的研究領域，一直都是重要的辨識特徵。近年來，偵測邊界轉折點已被廣泛應用在多邊形逼近、曲線密合、自動光學檢測、影像切割、影像校正與形變、物體辨識、運動速寫等各方面。偵測邊界轉折點時，需先將影像主體自背景分離出來，接著在物體邊界上找出曲率變化明顯的轉折點位置。然而影像在數位化過程，因量化處理與雜訊干擾，往往造成邊界上的鋸齒現象，影響偵測邊界轉折點的成效。 本文提出一種能有效抵抗量化處理與雜訊干擾的邊界轉折點偵測方法。此演算法包含三個要件：運用共變數矩陣特徵值衡量邊界像素之曲率，藉由折線模型估計任意角度之曲率門檻值，以及依據鑑別力指數高低決定支援區間長度。實驗顯示，不論是處理乾淨的或者帶雜訊的影像，我們提出的演算法在偵測邊界轉折點的表現上均優於其他對照方法。此結果植基於我們同時改進了傳統作法在衡量曲率屬性與決定支援區間長度兩方面潛藏之問題。 Corners have been one of the most important features in computer vision since they are invariant to geometric transformations, such as translation, rotation and scaling. Boundary-based corner detectors, segmenting objects from an image first and then locating the discontinuities on the object boundaries, have been widely applied to polygonal approximation, spline curve fitting, automated visual inspection, image segmentation, image registration, shape morphing, handwriting/environment/object recognition, motion sketch, etc. The accuracy of corner detection on boundaries is primarily influenced by quantization and noises. In this thesis, we propose a robust boundary-based corner detection algorithm for diverse images. The algorithm is composed of three components: a new measure of significance based on the eigenvalues of covariance matrices, threshold estimation of the measure of significance of any angle, and an optimization procedure based on a discriminant criterion for determining the length of region of support. The experimental results show that our algorithm outperforms other methods, even in the noisy samples. These robust results are due to not only the reliable measure of significance but also the discriminating optimization procedure of our algorithm.
