利用最佳化參數與主成分分析預測人臉老化之研究

Abstract

人臉預測技術經常被廣泛運用在各個領域上，例如：醫藥科學、法醫檢測、臉部模擬及人臉辨識。本篇論文提出一個基於擷取人臉的特徵進行數學統計分析的人臉老化模擬方法。 目前現有的年齡估測方法有WAS (Weighted Appearance Specific), AAS (the Appearance and Age Specific Classifiers), 和AGES (AGing pattErn Subspace)等。 我們採用AGES的年齡估測技術，在加上父母臉部資訊以增強個人老化的預測準確度。由於使用 FG-NET 人臉資料庫進行訓練，大多數的年齡的資料有所欠缺，用 PCA with missing data 來填補預測的結果欠缺不足。此外由於數張資料庫的人臉影像過於模糊或解析度甚低，這種情形使得能夠取得臉部細節紋理的資料更少。因此我們考慮增加數張父母成長影像，藉由父母臉部的細節加強以合成適當的臉部紋理，實作且比較其差異。

Human face prediction is an interesting task in many applications, such as medical science, forensic science, face synthesis, and identification. This thesis proposes a statistic method based on human face features, which is used for face aging simulation. The existing age estimation methods are WAS (Weighted Appearance Specific), AAS (the Appearance and Age Specific Classifiers), and AGES (AGing pattErn Subspace) etc presently. We adopt a method: parent-enhanced aging prediction for repairing the aging prediction result from AGES method. Since we use the FG-NET face image database and train them by PCA with missing data to predict aging human face, the results are not appropriate for those images which are not from the training samples. In addition, several face images of FG-NET database are blurred and lack details for aging texture. So we consider annexing several images from one’s parents to enhance his/her image detail display in our experiment. Our experiment shows that the proposed method achieves more faithful and detailed aging simulation.