整合式機器學習演算法與人臉年齡估測

基於可接觸資料的增加與計算技術的快速發展，過去十年，機器學習已因為人類生活中大量的自動化需求而吸引了許多注意力。現今在物件識別、機器人學、人工智慧、電腦視覺、甚至是經濟學等科學領域當中，機器學習已成為從資料當中抽取與探索重要資訊不可或缺的角色。另一方面，在過去幾十年，人臉偵測與辨識等人臉相關的主題已漸漸成為物件識別與電腦視覺中的重要研究領域。其原因來自於自動化識別與監視系統的需求、對於人類視覺系統在人臉感知上的興趣、與人機互動介面的設計開發等。在本篇論文當中，我們專注在使用機器學習技術來估測人類的年齡。我們提出了一個人類年齡估測的架構，其包含了特徵抽取、距離度量調整、降低維度、與年齡測定等四個步驟。我們使用了相當受歡迎的主動外觀模型(active appearance model, AAM)來抽取特徵。此特徵可以共同地表述人臉形狀與質地的變化。為了增強監督式降維演算法的效果，我們使用相關組件分析(relevant component analysis, RCA)技術來達成距離度量的調整。因為年齡標籤本身具有順序性的關係，我們提出了一個標籤敏感(label-sensitive)的概念，以更有效的在距離度量與降維學習的過程當中運用標籤的資訊。此外，我們還提出了一些修改來減輕資料庫當中可能存在的不平衡問題。根據降維過程中保存局部或鄰近資訊的特性，我們使用局部性回歸而非總體性回歸來測定年齡。由實驗結果中與其它方法的比較，我們所提出的架構與修改在最普遍使用的FG-Net資料庫中可以達到最低的平均絕對誤差。

關鍵字

機器學習；降維；流形學習；距離度量學習；人類年齡估測；回歸

並列摘要

Based on the increasing of accessible data and the fast development of the computational technology, machine learning attracted lots of attention in the last ten years because of the great demand of automation in human life. Now in the disciplines of pattern recognition, robotics, artificial intelligences, computer vision, and even economics, machine learning has been an indispensible part to extract and discover the valuable information from data. On the other hand, human face related topics such as face detection and recognition became important research fields in pattern recognition and computer vision during the last few decades. This is due to the needs of automatic recognition and the surveillance system, the interest in the human visual system on human face perception, and the design of human-computer interface, etc. In this thesis, we focus on using machine learning techniques for human age estimation. A human age estimation framework is proposed, which includes four steps: feature extraction, distance metric adjustment, dimensionality reduction, and age determination. The popular active appearance model (AAM) is exploited in our work for feature extraction, which can jointly represent the shape and texture variations of human faces. To enhance the performance of supervised dimensionality reduction, the relevant component analysis (RCA) is used for distance metric adjustment. Because human ages are with ordinal relationship, we proposed the “label-sensitive” concept to better exploit aging labels during distance metric learning and dimensionality reduction learning. In addition, several modifications are proposed to alleviate the possible unbalance problem in the database. According to the neighbor / locality preserving characteristic during dimensionality reduction, the proposed framework utilizes local regression instead of global regression for age determination. From the experimental results, the proposed framework and its modification versions achieve the lowest mean absolute error (MAE) over other existing techniques in the most widely-used FG-Net database, both in the LOPO and cross-validation settings.

並列關鍵字

machine learning ； dimensionality reduction ； manifold learning ； distance metric learning ； human age estimation ； regression

參考文獻

[1] M. Turk and A. Pentland, “Eigenfaces for recognition,” Journal of Cognitive Neuroscience, vol. 3, no.1, pp. 72-86, 1991.

[2] P. N. Belhumeur, J. P. Hespanha, and D. J. Kriegman, “Eigenfaces vs. Fisherfaces: Recognition using class specific linear projection,” IEEE Trans. Pattern Analysis and Machine Intelligence, vol. 19, no. 7, 711-720, 1997.

[3] P. Viola and M. Jones, “Rapid object detection using a boosted cascade of simple features,” Proc. IEEE Conf. Computer Vision and Pattern Recognition, pp. 511 -518, 2001.

[4] P. Viola and M. Jones, “Robust real-time face detection,” Int’l Journal of Computer Vision, vol. 57, no. 2, pp. 137-154, 2004.

[5] Flicker: http://www.flickr.com/

國際替代計量

整合式機器學習演算法與人臉年齡估測

全文下載

主題瀏覽