以特徵熵為基礎之ADABOOST演算法應用於人臉偵測之研究

人臉偵測是所有臉部相關分析演算法的關鍵步驟。雖然AdaBoost演算法結合Haar-like特徵應用於即時人臉偵測上已達到不錯的成效，但因為龐大的特徵數量致使AdaBoost學習過程相當的費時，並且Haar-like 特徵的對光照變化較無穩定性。儘管許多方法被提出用以加速AdaBoost學習過程或增強特徵的描述力與穩定性，仍然存在很大的改良空間。本論文基於鄰接區塊二元圖樣(NBBP)特徵與以特徵熵為基礎的AdaBoost(EBAdaBoost)學習演算法，提出一創新的人臉偵測方法。NBBP特徵是延伸自基本局部二元圖樣(LBP)，擁有光照不變性及描述能力較佳的優點。而所提出的EB-AdaBoost學習演算法則藉由計算NBBP 的特徵熵，於AdaBoost學習過程中，以非暴力且有系統的方式來決定每一回合最佳的弱學習子。因此，本文所提出之ED-Aaboost學習演算法可同時滿足快速學習與高正確率的需求。為凸顯EB-AdaBoost學習演算法的特性，本論文將以ROC曲線與訓練效率分析圖表呈現與其他相關方法比較的結果。由結果可知，NBBP特徵相較於Haar-like特徵更適合用於描述人臉，且基於熵的特徵選擇方法提供了一個較為準確的標準用以評估其所擁有之辨別力。因此，此方法可以做為AdaBoost演算法之仲裁者，藉此有效的篩選弱學習子，以達到加速學習之目的。

關鍵字

人臉偵測； AdaBoost ； EB-AdaBoost ；鄰接區塊二元圖樣； NBBP特徵熵

並列摘要

Face detection is the step stone to all facial analysis algorithms. AdaBoost has achieved good performance for real-time face detection with Haar-like features. However, the amount of Haar-like features is too huge to learn efficiently and they are not robust to lighting variations. A lot of approaches have been proposed either to accelerate the AdaBoost learning phase or enhance the feature discriminability. There are still existent rich space for improvement, both on the learning efficiency and the recognition efficacy. This thesis proposes a novel face detection scheme based on the Neighboring-Block Binary Pattern (NBBP) features and Entropy-Based AdaBoost (EB-AdaBoost) algorithm. The NBBP features are extended from the Local Binary Pattern (LBP) such that they are illumination invariant and more efficacious in discrimination power than Haar-like ones. By associating each NBBP feature with a entropy, the best weak-classifier on each learning iteration of AdaBoost is determined systematically in a non-brute-force manner. These manifest that the proposed EB-AdaBoost algorithm achieves both rapid learning and high recognition precision. To highlight the characteristics of EB-Adaboost learning algorithm, various comparisons with other approaches will be made in the thesis, including receiver operating characteristic (ROC) curves and training efficiencies analysis. Our experimental results show that the entropy can provide a useful criterion to measure the discriminability of available features and, hence, can serve as a “referee” for AdaBoost algorithm to screen weak classifiers effectively.

並列關鍵字

Face detection ； AdaBoost ； EB-AdaBoost ； NBBP ； NBBP entropy

參考文獻

[2] E. Hjelmas, “Face detection: A survey,” Computer Vision and Image Understanding, vol. 83, pp. 236–274, 2001. [Online]. Available: http://linkinghub.elsevier.com/retrieve/pii/S107731420190921X

[4] A. King, “A survey of methods for face detection,” IEEE Trans on Pattern Analysis and Machine Intelligence, vol. 24, pp. 34–58, 2003. [Online]. Available: http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.51.657

[7] C. Zhang and Z. Zhang, Boosting-Based Face Detection and Adaptation, 1st ed. Morgan and Claypool, 2010.

[9] S. Li, L. Zhu, Z. Zhang, A. Blake, H. Zhang, and H. Shum, “Statistical Learning of Multi-view Face Detection,” in In Proceedings of the 7th European Conference on Computer Vision, 2002, pp. 117–121. [Online]. Available: http://dx.doi.org/10.1007/3-540-47979-1_5

[14] J. Friedman, T. Hastie, and R. Tibshirani, “Additive Logistic Regression: a Statistical View of Boosting,” Annals of Statistics, vol. 28, 1998. [Online]. Available: http://citeseerx.ist.psu.edu/viewdoc/summary?doi=10.1.1.30.3515

國際替代計量

以特徵熵為基礎之ADABOOST演算法應用於人臉偵測之研究

未授權

主題瀏覽