基於部分類別最小平方法之特徵提取與分類

Abstract

人臉辨識是圖形識別應用發展的重要指標，其相關衍伸方法更為圖形識別中重要里程碑，例如Eigenface與Fisherface。現今廣泛使用統計模型作人臉辨識中，主要有主成分分析法及最具代表性的線性判別分析法為基礎的方法，然而線性判別分析法卻很容易面臨小樣本的問題，因此必須使用不同的解決方法。此外，一般的特徵提取法在分類器的選擇上也易受到原本方法的特性，必須使用適當的分類器方可達到良好的辨識率。 本論文提出一個使用回歸模型的方法 －部分類別最小平方法(Classwise-Partial Least Squares, CW-PLS)，得以 1) 同時將分類隱含進特徵提取的程序內、2) 可直接使用訓練完成的回歸模型做預測，而不需其他的分類器、3) 具有高度可平行化之優點、4) 完全避免掉一般使用部分最小方法為基礎之方法所面臨收斂耗時的問題、以及5) 使用少數主特徵即可達到相當高之辨識率。 實驗的部分則採用ORL、Yale、PIE人臉資料庫做測試，並與Fisherface、Eigenface和傳統的部分最小平方法做比較，結果發現CW-PLS除了在運算效能上比傳統的部分最小平方法快上數倍之外，其辨識率也較其他三者高，並且在降維度上有大幅度的提升，僅需少數個特徵值即可表示大部分的資料。

Face recognition is an old but important issue in the field of pattern recognition, and many of its related research and methods have become milestones, such as Eigenface (PCA-based method) and Fisherface (LDA-based method). Until now, PCA and LDA are the most widely used statistical-based techniques, and the latter is the most popular and a still ongoing research topic; however, LDA is prone to suffer from the small sample size problem and needs to be coped with. On the other hand, general feature extraction methods are restrict to their own properties, thus, the selection of classifier could affect the recognition rate. This thesis proposed a regression-based technique ‒ Classwise Partial Least Squares (CW-PLS), which has the following properties: 1) implicitly classify data during feature extraction, 2) the trained regression model can be used to predict new testing data without other classifier, 3) computation is highly parallelizable, 4) completely avoid the convergence problem that occurs in traditional PLS, and thus reduces computational time, 5) high recognition rate with few principal components. In the experiment, ORL, Yale, and PIE databases are used for verification, where Eigenface, Fisherface, and traditional PLS are compared with CW-PLS. The results shows that CW-PLS works much faster than traditional PLS and the recognition rates are the highest among all three methods; moreover, the dimension reduced is significantly lower, thus, only a few principal components are sufficient to represent the original data.