基於賈伯小波及區域保留投影之人臉偵測與人臉辨識

Abstract

人臉辨識在長久以來一直都是很熱門的研究主題，因為可以廣泛的應用在各個不同的領域之中，例如身份識別、基於內容的影像檢索、電腦視覺與人類電腦介面等。而做為必要前處理的人臉偵測也同樣重要，因為要先能把人臉偵測出來，才能夠做接下來的人臉辨識。因此本論文結合了賈伯小波與區域保留投影的優點而提出一個可在數位相片中執行人臉偵測與人臉辨識的方法。先使用膚色偵測來找出可能是人臉的區域，然後使用賈伯小波轉換取得人臉整體特徵，並利用區域保留投影將人臉特徵降至很低的維度，再訓練類神經網路來判別是否為人臉，根據人臉偵測的結果用人工做分類並建立資料庫，透過資料庫中的人臉來訓練類神經網路做辨識，最後在測試相片輸入時，就能夠對於所選定的人臉做辨識，藉此辨識結果就可以將資料庫中含有相同人物的相片挑選出來，達到基於人物身份的影像檢索的目的。本論文的主要貢獻為利用賈伯小波轉換能夠同時在時域和頻域中兼顧足夠的分辨率的特性來取得人臉的整體特徵，再使用區域保留投影能保留局部非線性資訊做降維的優點，將賈伯小波轉換後過於龐大的特徵向量降低至很小的維度，實驗結果顯示本論文的方法在人臉偵測與人臉辨識都有不錯的效能。

Face recognition has been a popular research topic for a long time because it can be widely used in many different fields, such as identity identification, content-based image retrieval, computer vision and human computer interaction. However, face detection, which serves as the preprocessing procedure, is equally important since it has to be done first before face recognition is taken. This thesis, therefore, proposes a method which takes the advantages of Gabor transformation and locality preserving projection to implement face recognition and face detection on a digital picture. The first step adopts face detection to find candidate face region. Second, Gabor wavelets transformation is adopted to extract face features of human face, and locality preserving projection is applied to project features of human face into lower dimension space. Afterwards, neural network is trained to decide whether candidate region is human face or not. Then, database is constructed manually according to the result of face detection. Finally, a neural network is trained by the faces is stored in the database. When a test picture is input the proposed method is able to identify the faces of the chosen people. According to the result of identification, pictures of the same person can be chosen from database and implement the identify-based image retrieval. The main contribution of this thesis is to employ the specialty of Gabor wavelets transformation, which is to maintain sufficient recognition rate in both time domain and frequency domain, to obtain face feature; moreover apply the strength of locality preserving projection, which preserves the local structure of the multidimensional structure, the immense feature vectors of Gabor wavelets transformation is lowered to minimum. The experiment results show the proposed method has good performance in both face detection and face recognition.