透過梯度資訊偵測影像主體之對稱性

Abstract

本論文旨在偵測並擷取出影像中主體對稱，只要透過輸入的影像以及其像素所提供的梯度資訊就能擷取出影像的主體對稱軸和其對稱範圍。這個方法主要分成三大步驟: (1)確定待判定區域 (2)對稱性強弱偵測 (3)確認對稱軸和對稱範圍。

首先利用對稱影像輪廓相似的特性，擷取出可能存在影像對稱性的區域，再利用像素的梯度資訊(如梯度方向、梯度強度)進行判定，並依據這些判定結果，確認該影像是否存在對稱影像。

透過不斷的實驗與觀察，以調整各種參數所造成的影響，讓最後偵測出的對稱軸和其範圍能更加精確。

In this thesis, computational processes for capturing the axis of symmetry and identifying the areas where symmetry exists will be report.

The computation basically exploits such gradient-derived information as gradient strength and local homogeneity of orientations, and consists of three phases of operations. (1)Areas of interests are determined by image contrast distribution. (2)Measures of bilateral-symmetry with respect to axes in various orientations are applied to areas of potential interests, through which axes with top five scores above threshold are preliminarily determined. (3)Based on gradient-derived information associated with interested pixels on both sides of the axes, further refinement is carried out to make the axial length more exact and to demarcate the extent of within which the symmetry exists more precisely.

Through a series of experiments, the effects of the factors governed by adjusting the gradient-derived parameters are observed and the computation processes are then tuned accordingly, making the determination of axis and extent of symmetry visually satisfactory.