相關濾波器適用於人臉辨識之設計與應用

Abstract

最小平均相關能量濾波器很常應用於影像辨識系統。當用來訓練濾波器的影像經過適當的選擇，我們可以得到滿意的辨識率。然而我們知道最小平均相關能量濾波器的效能對於失真非常敏感，因此我們利用滿足某些特定的條件來最佳化濾波器的效能、而不使用嚴格的限制。為了避免多於不必要的影像來訓練濾波器，而浪費記憶體和運算量，在這篇論文我們提供一個演算法可以自動從資料庫裡找尋最適合拿來訓練濾波器的影像。另外基於安全和隱私因素，我們必須把加密功能加入我們的濾波器設計中，避免資料被盜用的可能，然後我們會證明使用的加密功能並不會影響到辨識能力。最後為了處理資料量龐大的情形跟人臉的非線性變形，我們使用種類相關的特徵分析跟使用更高維的資料訊息來做相關性。整合這些技術，我們將濾波器改良的更可以克服現實生活當中可能產生的問題。

The minimum average correlation energy (MACE) filter is a well known correlation filter for pattern recognition. The recognition rates will be attractive while choosing the training images properly. But the MACE filter is sensitive to the distortion, thus we optimize the filter by removing the hard constraint and satisfying certain criterion, which is so-called unconstrained correlation filter. In order to avoid redundant training images, here we provide an algorithm to automatically choose the proper training images from a dataset. For security issue, we also use encryption method in the filter design and we will show that the encryption process do not affect the recognition performance. Finally for the large scale database and the nonlinear distortion in human face, we improve the recognition performance by using class-dependent feature analysis and correntropy function, where correntropy is a positive definite function that generalizes the concept of correlation by utilizing higher order moment information of signal structure. Using these technologies, we can make the filter more practical in real application.