- 學位論文
以加伯濾波器之材質特徵為基礎的人臉表示與辨識技術
Face Representation and Recognition based on Texture Scale and Orientation through Gabor Filter
摘要
在電腦視覺和電腦圖學的領域中，人臉辨識是一項頗受學術界和產業界重視的技術，雖然目前已有為數不少的學術性論文和商業性產品，但辨識生活照中的人臉仍然是一大艱鉅的挑戰。在本論文中，我們將重點放在對生活照進行非監督式辨識(Unsupervised Recognition)，而非監督式辨識指的是在整個辨識過程中不需要訓練任何模型。生活照具有各種不利於辨識的因素，如表情、光影、遮蔽和模糊等等，大大增加了辨識的難度。由於加伯濾波器(Gabor Filter)能夠有效地擷取出不同尺度和方向的材質特徵，在本論文中，我們以加伯濾波器為基礎實做了兩種演算法，分別為區域性加伯二元圖樣統計圖串(LGBPHS)和加伯相位統計圖(HGPP)，這兩種演算法分別使用加伯量(Gabor Magnitude)和加伯相位(Gabor Phase)來描述臉部特徵。進一步地，我們使用了三種方法來合併區域性加伯二元圖樣統計圖串和加伯相位統計圖，並使用了多核處理器和圖形處理器技術來加速程式執行速度，而在實驗部分。除了生活照外，我們也使用FERET人臉照片資料庫來驗證我們的演算法在標準臉部相片中的辨識準確度。我們的演算法可將309張生活照分為109個群組，並有96.7%的準確率，另一組測試資料為838張生活照，我們的演算法將其分為252個群組，準確率為99.2%，而在FERET人臉照片資料庫中，準確率則可達到95.97%。同時，加伯濾波器可使用圖形處理器來加速，在我們的實做中，執行速度可加快140倍以上。
並列摘要
In the past ten years, face recognition has become a popular area in computer vision. This technique can be used in several applications, such as security system or photo categorization system. Although many technical papers and commercial systems have emerged, recognition of photos under uncontrolled environment is still a challenge. Here we will focus on recognizing different people in home photo datasets without any training procedure. Since Gabor filter has the multi-resolution and multi-orientation characteristics, we implement two algorithms, called Local Gabor Binary Pattern Histogram Sequence (LGBPHS) and Histogram of Gabor Phase Patterns (HGPP), which use Gabor magnitude and Gabor phase as the face descriptor respectively. How to combine LGBPHS and HGPP is also addressed here. Moreover, we use multi-thread and GPU programming to reduce the computation time, and evaluate our approach on general face images from the FERET Database. Our approach can result in 96.71% precision in dividing into 109 clusters from 309 home photos, and 99.22% precision in dividing into 252 clusters from 838 home photos. On FERET Database, precision of our approach is 95.97%, which is higher than the previous research. In our implementation, the Gabor filter using GPU programming is more than 140 times faster than the single core version.