在人臉偵測及人臉辨識的相關研究中，所需加強的不外乎準確率的提升，而輸入影像的前處理步驟及輸入影像的品質也將明顯影響後續處理的準確性。在本論文中，利用人臉特徵資訊來判斷人臉旋轉方向，以藉此找出連續影像中最接近正面的人臉影像。其中特徵抽取方法包括了膚色區域切割及邊緣資訊的結合，根據所抽取之人臉特徵影像，提出了三種方法分別解決了人臉三方向的旋轉問題。 第一種方法利用特徵垂直集中性，判斷人臉特徵左右集中位置，以得到人臉左右轉動角度及方向。第二種方法使用放射型樣板統計人臉特徵影像中每個區域的特徵點數，將所得到的放射型直方圖加以分析及根據人臉分佈特性，可以判斷出人臉正面旋轉角度。第三種方法藉由頸部位置與人臉區域及特徵區域的相對位置關係，判斷其俯仰的可能性。 利用上述三種方法的旋轉偵測，配合人眼位置計算及對稱度的比較，以計算出最接近正面的人臉影像，經由自行拍攝的影像測試結果顯示，我們所提出的方法在人臉正面判定上，的確具有可行性及其正確。 It is an urgent desire for researchers to uplift the accuracy in human face detection and recognition. The main bottleneck lies mainly on the quality of input images which definitely drastically affect the accuracy of the system. In this thesis, the information of facial feature is adopted to determine the direction of face rotation and then find the best shot of faces in a surveillance video stream so that the quality of input images can be improved. The features to be extracted include skin color information and edge information. Skin color information can be obtained by analyzing the skin color distribution in YCbCr color space and edge information can be detected by applying Sobel edge detector. Moreover, three strategies are proposed to determine the three directions of face rotation based on the feature image. The first strategy uses the collection of vertical feature projection. The direction and the angle of face turning can be determined by analyzing the distribution of vertical projection histogram. The second strategy uses a novel model called radial template to detect the presence of face rotation. This template is designed to find the angle of center-rotated objects. According to the characteristics of skin detection and edge extraction, the extracted feature will be stable under this kind of template. The third strategy is to determine the presence of bending or lifting of faces based on the relationship of feature area and neck position. By integrating the three strategies and the geometry of eye position and judgment of symmetry, the best shot of frontal face image can be identified which can be employed in later face recognition task. Experiments were conducted on various video images containing faces. Experimental results verify the feasibility and validity of our proposed approach in determining the most frontal faces in video sequences.
