基於影像小區塊方式之性別辨識

Patch-Based Gender Classification

林聖儒(Shen-Ju Lin)

指導教授 ： 黃仲陵

國立清華大學/電機資訊學院/電機工程學系/碩士(2010年)

摘要

在這篇論文中，我們使用一貝氏估計的架構去針對人臉區塊相似度及由訓練資料所累積的分布來達到對人臉照片做性別辨識的目的。我們的人臉性別辨識主要包含兩個程序。第一個程序是先使用Haar-like特徵從測試照片中尋找人臉候選者，接著使用尺寸濾波器及膚色偵測去除掉非人臉候選者。另外，我們利用圖像濾波及權重遮罩來得到標準化的照片。第二個程序主要利用預訂圖像資料庫及訓練照片來辨別出人臉測試照片的性別，這些人臉照片包含大量的變異例如光線、表情、人臉角度、背景等變化。我們認為圖像資料庫的組成對於由訓練資料所建立的男性及女性累積分布有密切的關係。我們提出一個以特徵臉(eigenface)與K均值分群法的結合方式，從我們自行收集並分類的男性及女性照片資料庫中選取出具有代表性的男性及女性照片，接著由剩餘的照片視為訓練照片並將其特性根據不同圖像資料庫的數量及組成建立出對應的男性及女性累積分布。
我們去對兩種圖像資料庫選擇方法(特徵臉與K均值分群法的結合以及隨機挑選)來做評估並去觀察性別辨識準確度的差異，這裡我們使用的測試照片包括從網路上收集1000張大量變異程度的男女照片、從FERET人臉資料庫得到的1364張均勻光線正面男女照片及Bao face database的團體男女照片。從實驗結果得知我們所提出來的圖像資料庫選擇方法相較於隨機挑選方法，根據不同張數的圖像資料庫及解析度的實驗條件下，在大多數的情況下擁有較好的性別辨識準確度。

關鍵字

性別辨識

並列摘要

In the thesis, we employ a Bayesian estimation framework that exploit image patch similarity and accumulated distribution to reach the goal of predicting the gender from the facial images. Our facial gender classification consists of two processes. The first process searches the face candidates using Haar-like feature from the images, and uses the size filter and skin color detection to remove the non-face candidates. In addition, we apply the image filtering and weighting mask to obtain the normalized facial images. The second process uses the predefined library and training images to identify human gender of the test images. The facial images require a large range of variation including lighting, expression, pose, background…etc. We consider that the formation of library images has a close relationship with the accumulated distribution about gender (male, female). We propose a library images selection scheme to choose the discriminative male and female images, which is based on the method of eigenface with K means clustering, then build the male and female accumulated distribution according to the characteristic of the rest part of training male and female images. We evaluate the two library selection method, eigenface with clustering method and random method, and observe the difference of gender prediction accuracy using 1000 web facial images of male and female, the color FERET face database which contain 1364 regular frontal facial images of male and female, and the Bao face database group images of male and female. The experimental results demonstrate that our proposed method is better than the random selection in most cases.

並列關鍵字

Gender Classification

參考文獻

[21] 陳曉瑩, “即時多角度人臉偵測 Real-time Multi-pose Face Detection” 國立清華大學電機工程系所.

[1] J. Aghajanian, J. Warrell, S.J.D. Prince, P. Li and J.L. Rohn and B.Baum “Patch-based Within-Object Classification,” IEEE 12th International Conference on Computer Vision, 2009.

Expression Using Active Appearance Models”. In: Proc. 7th International. Conf. on

[3] E Makinen, R Raisamo, “Evaluation of Gender Classification Methods with Automatically Detected and Aligned Faces”. IEEE Trans. On Pattern Analysis and Machine Intelligence, vol. 30, no. 3, pp. 541-547, March 2008.

[4] B. Moghaddam, M. H. Yang., “Learning Gender with Support Faces”. IEEE Trans. On Pattern Analysis and Machine Intelligence, vol. 24, no. 5, pp. 707-711, May 2002.