生活照片之人物分類系統研究

Abstract

本論文採用膚色偵測法找出彩色生活照片可能的人臉位置，透過賈伯小波(Gabor wavelet)抽取特徵進行類神經網路訓練，以判斷是否為人臉。在人臉辨識階段，先使用主動式外觀模型(active appearance model)和可導引濾波器(steerable filter)進行人臉正規化，接下來採用稀疏編碼(sparse coding)演算法，在五個訓練樣本下，生活照人臉辨識率可達80%，使用其他正面人臉資料庫(AR資料庫)辨識率更可高達98%。並提出直方圖統計法來減少稀疏編碼的權重數目為原來的60%，除了降低系統運算量，同時特徵向量仍然具有代表性。整體而言，可適用於家庭數位相簿管理，或數位相框之分類系統。

This thesis adopts skin-color model to find the candidate face region, then Gabor wavelets transformation is adopted to extract the entire face features. Afterward, neural network is trained to determine whether the candidate region is a human face or not. Finally, this thesis adopts active appearance model and steerable filter to normalize all faces for face recognition. Then this thesis implements sparse coding algorithm with 5 training faces to increase the face recognition rate up to 80% for photographs, and for frontal face of AR database also increases by 98%. Furthermore, this thesis proposes using histogram method to reduce 60% of sparse coding needed which also reduces the amount of system computational cost, and then the features are still representative. As a whole, this system is suitable for digital media classification of family photograph albums or digital photograph frames.