臉部辨識是最具人性化的一種生物驗證,其應用和重要性隨著社會安全要求的日益嚴謹而更加受到重視。臉部自動辨識系統主要包含臉部自動偵測和臉部辨識兩部份,而臉部偵測技術是臉部自動辨識重要的一部分,故擷取人臉的速度及準確度,將嚴重影響臉部自動辨識系統的效率和準確度,尤其是針對複雜彩色影像來說,更突顯出臉部偵測技術的重要性,對此本文提出一個簡易有效的臉部偵測方法來擷取具有多臉孔、臉部重疊或部分遮蔽和部分背景為膚色的複雜彩色影像中人臉。首先,將取得的影像經過色彩平衡與光線補償後,在YCbCr的色彩空間中以貝式分類器進行膚色分割,接著以邊緣偵測強化各個膚色分割區域,再利用影像形態學及影像拓撲學,標記出可能為臉部的膚色框架,並進行臉部框架的分割,之後再進行膚色框架的區域填充,隨之將影像正規化處理,最後以支援向量機(Support Vector Machine, SVM)進行臉部的驗證,實驗結果顯示文中的方法可以在複雜影像得到高效率和高成功率的臉部偵測。而在臉部辨識部份本文提出以餘弦相似性進行範本相片選擇,改進範本相片以人工挑選的困難,使用支援向量機SVM進行實驗,可獲得與人工挑選近似乎相同的辨識率。
The face detection technology has an important role in automatic face recognition. The speed and the accuracy of the human face detection will seriously affect the efficiency and the accuracy of the automatic face recognition system. Especially, the face detection technology is more important for the complex image. In the thesis, we propose a simple and useful face detection method for detecting faces in the complex images with multiple faces, face-overlapped or face with glasses. In the preprocessing stage of the method, the images are processed by color balance and lighting compensation and skin color pixels of the images are classified in the YCbCr color space by Moghaddam’s Bayesian classifier approach. The Sobel operator is used to enhance skin color region edge for accurately segmenting skin color regions in a processed image into a series of frames by using 4-connected operator and the morphological dilation and erosion operator are used to delete some small segmented frames which can be regarded as noisy to obtain candidate face frames. For the candidate face frames with face overlapped, a frame segmentation algorithm is developed to achieve optimum face frame division. A region filling algorithm is then performed to obtain complete candidate face frames. Finally, after normalizing the candidate face frames, the Support Vector Machine (SVM) is applied to verify them as human faces. The experimental results show that the proposed method can obtain a very high success ratio for complex images. In the face recognition, the Cosine transformation applied to select a subset of images formed templates from a series of individual person image is presented. Almost same recognition results obtained by SVM can be achieved by using the proposed selected template images against the template images they are artificially selected.
為了持續優化網站功能與使用者體驗,本網站將Cookies分析技術用於網站營運、分析和個人化服務之目的。
若您繼續瀏覽本網站,即表示您同意本網站使用Cookies。