藉由降維演算法來提升局部迴歸並應用在人臉年齡辨識上

Using Dimensionality Reduction to Improve Local Regression on Facial Age Estimation

簡嘉宏(Jia-Hong Jian)

指導教授 ： 張智星

國立臺灣大學/電機資訊學院/資訊工程學研究所/碩士(2015年)

https://doi.org/10.6342/NTU.2015.01669

摘要

在本論文裡，我們將專注在使用過去文獻提及可以提升近鄰(k-Nearest Neighbors, kNN) 的結果之演算法，其可於提升局部迴歸的效能並用於人臉的年齡識別上。而在此篇中的辨識架構為基於Chao的方法做改進。這裡我們使用人臉年齡估計上常用的主動外觀模型 (Active Appearance Model, AAM) 來做為特徵的擷取，其可以準確的描述人臉的形狀與紋理不同。考慮數據流形局部的特性，我們改用最大邊界投影 (Maximum Margin Projection, MMP)，其可以使高維空間中近鄰同類別的資料點在低維空間中仍然在附近，不同類別的資料點在低維空間中使其較遠。而為了增加或減少降維後資料其每個維度間的影響力，並且同時克服原作者用RCA其並沒有利用到資料間不同類別的特性之缺點，我們使用統計學的方法判別組件分析 (Discriminant Component Analysis, DCA)，根據不同類別間共變異數的分析來做距離度量的調整。因為在年齡估計我們打算使用局部性的迴歸分析而非總體的迴歸分析，因此在降維上選用了上述保留局部鄰近資訊的方法。另外加入了一些方法來平衡資料庫在收集上可能會造成的不均與偏差。而在效果的評估上使用近年來廣泛被使用的MORPH人臉資料庫並儘量降低整體的平均絕對誤差 (Mean Absolute Error, MAE)。最後我們做出來的平均絕對誤差約5.6242，在相同的資料上使用我們取出的特徵，比起Chao的方法降低了0.2的誤差。

關鍵字

主動外觀模型 ； 流形學習 ； 距離度量學習 ； 局部迴歸 ； 平均絕對誤差

並列摘要

In this thesis, we focus on using dimensionality reduction algorithms to let k-nearest neighbor have better results so that we could get better performance on local regression which we used on human face age determination. Our framework was based on Chao’s method and we do some optimization. The common feature extraction technique is applied in this field, active appearance model (AAM), which can jointly represent the shape and texture variations of human face. In order to discover both geometrical and discriminant structures of the data manifold, maximum margin projection (MMP) is used for dimensionality reduction and causes the margin between relevant and irrelevant classes is maximized. For enhancing the influence of some specific dimensionality and overcome the disadvantage of RCA from Chao’s method, we leverage discriminant component analysis (DCA) instead, which is according to covariance analysis to do distance metric adjustment. Because of the advantage of neighbor preserving as mentioned, we use local regression instead of global regression for age determination. Furthermore, the proposed method can lighten the problem of dataset imbalance and ordinal relationship. Finally, we get 5.6242 MAE which is improved by 0.2 errors, comparing to Chao’s method.

並列關鍵字

active appearance model ； manifold learning ； distance metric learning ； local regression ； mean absolute error

參考文獻

[8] Cai, Deng, et al. "Locality Sensitive Discriminant Analysis." IJCAI. 2007.

[1] Wei-Lun Chao, Jun-Zuo Liu, and Jian-Jiun Ding. "Facial age estimation based on label-sensitive learning and age-oriented regression." Pattern Recognition46.3 (2013): 628-641.

[3] Lee, Tai Sing. "Image representation using 2D Gabor wavelets." Pattern Analysis and Machine Intelligence, IEEE Transactions on 18.10 (1996): 959-971.

[4] Lanitis, Andreas, Christopher J. Taylor, and Timothy F. Cootes. "Toward automatic simulation of aging effects on face images." Pattern Analysis and Machine Intelligence, IEEE Transactions on 24.4 (2002): 442-455.

[5] Guo, Guodong, et al. "Human age estimation using bio-inspired features."Computer Vision and Pattern Recognition, 2009. CVPR 2009. IEEE Conference on. IEEE, 2009.