局部區塊相減樣式使用於性別辨識

Local Block Difference Pattern for Gender Classification

王彥棋(Yen-chi Wang)

指導教授 ： 范國清

國立中央大學/資訊電機學院/資訊工程學系/博士(2014年)

摘要

近年來，圖形識別已經廣泛的應用在日常生活上。像是國道上的自動收費系統，除了利用RFID感應之外，還必須結合車牌的偵測和辨識，才能確保系統的穩定度。或是以掌紋，虹膜和人臉這些生物特徵為基礎的門禁系統，透過這些生物特徵來做認證進而達到門禁控管的目的。這些應用中最關鍵的技術為如何擷取出具有辨識能力的特徵，在這些特徵中局部紋理特徵的方式，更是廣泛且頻繁的應用在各種研究中，其中最為著名的是LBP(local Binary Pattern)，許多的研究學者也發展出基於LBP特點的各種局部紋理特徵，在本研究中，我們提出以雙攝影機系統為平台的性別辨識系統，全景攝影機擷取姿態影影像，PTZ攝影機則擷取人臉影像，並以我們提出的區域紋理方法LBDP(Local Block Difference Pattern) 擷取姿態與人臉的特徵進行性別辨識，如此結合姿態與人臉資訊的方式，可以提高性別辨識的正確率。
我們提出的區域紋理特徵除了繼承LBP計算簡單的優點，而且為了加強其描述物體的能力，我們採用區塊衡量相關程度的概念取代LBP單純只使用中心點像素來計算描述物體的缺點，由於此計算機制的改變，可以更進一步有效的減少像素變化的影響，我們編碼的方式有別於LBP。而在實際描述物體上則是類似LBP，將影像區分成相同大小但是不重疊區塊，將每個區塊在經過LBDP擷取特徵之後，利用統計方式將其轉為histogram，最後將這些histogram連接成特徵向量以描述此物體。
為了提升演算法和系統的穩定性，我們先以姿態影像為基礎做性別辨識來輔助之後階段的人臉性別辨識，其中擷取姿態和人臉特徵擷取部分都是以LBDP區域紋理特徵來達成，而最後的實驗結果，辨識率都能顯示出我們系統有顯著的效果。

關鍵字

局部紋理特徵 ； 性別辨識 ； 姿態影像

並列摘要

Recently, pattern recognition has been widely applied in many areas including ETC (electronic toll collection) system and vision-based access control system. ETC system not only uses RFID sensor but also combines license plate detection and recognition to ensure system’s stability. As to access control system, it needs to extract biometric features such as face, palm print and iris information for authentication. The most critical issue inherent in these applications is the extraction of discriminative features. One of the well-known feature extraction algorithms is Local Binary Pattern (LBP). In the past, many researchers developed local texture descriptors aim at inheriting the advantages of LBP.
In this dissertation, a novel gender classification scheme operating on dual cameras is proposed. The fix camera is used to extract gait image sequences for gait based gender classification and the PTZ camera is used to extract face images for face based gender classification. The classification features of both gait and face are described by the proposed novel texture descriptor called Local Block Difference Pattern (LBDP).
In conventional LBP, the problem of sensitivity to intensity change usually constrains its practicality due to its pixel-based comparison in the encoding mechanism. Different from LBP, the proposed LBDP describes the local texture information by extending the encoding mechanism from pixel-based comparison to block-based comparison so as to extracting more detailed information. The discrimination capability of LBDP is thus enhanced because the difference of local structures and the correlation degree of neighboring blocks are both considered in the proposed encoding mechanism. Moreover, the proposed LBDP can decrease the influence resulting from intensity change because of the expanding of encoding range. To obtain the image descriptor, we first convert the LBDP of each pixel in an image into a decimal value and divide the image into non-overlapping sub-regions. Form a histogram of each sub-region by accumulating the decimal values of all pixels in the sub-region. An image descriptor can then be obtained by concatenating the histograms of all sub-regions to be used as the feature for gender classification.
To improve the performance and stability of the system, we adopt the result generated in gait-based gender classification to assist facial gender classification. The proposed LBDP is utilized in extracting both the gait and facial features. Experimental results verify the validity and excel performance of the proposed system in gender classification.

參考文獻

[1] J.Bekios-Calfa, J.M.Buenaposada and L.Baumela, “Revisiting Linear Discriminant Techniques in Gender Recognition,” IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol.33, No.4, pp. 858-864, 2011.

[2] B. Moghaddam and M.-H. Yang, “Learning Gender with Support Faces,” IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 24, No. 5, pp. 707-711, May 2002.

[3] S. Baluja and H.A. Rowley, “Boosting Sex Identification Performance,” International Journal of Computer Vision, Vol. 71, No. 1, pp. 111-119, 2007.

[4] E. Ma ¨kinen and R. Raisamo, “Evaluation of Gender Classification Methods with Automatically Detected and Aligned Faces,” IEEE Transactions on Pattern Analysis and Machine Intelligence, Vol. 30, No. 3, pp. 541-547, Mar. 2008.

[5] S.A. Khan, M. Nazir,S. Akram and N. Riaz, “Gender classification using Image Processing Techniques: A Survey,” in Proceedings of International Conference on Mul-titopic, pp 25-30,2011.