Face Gender Classification using Boosting

[[abstract]]人臉性別辨識系統可以運用於許多現實生活的場合，例如：性別限制場所的門禁系統、賣場或展場的廣告看板、統計特定場所中性別比率做為市場調查分析的依據。但由於影像辨識易受到外界環境所干擾，例如：光源、角度、遮蔽物等...，因此如何排除這些外在因素又能以最短的時間達到辨識效果，即是人臉辨識的一大挑戰。 在論文中系統架構可分為三個部分：（1）前處?模組、（2）特徵萃取模組、（3）分?學習模組。使用Haar-Like特徵擷取人臉影像，再利用Haar-Like萃取出代表男女臉部特徵向量。在實驗中使用五折交叉測試(5-fold cross validation) 來分類，將所有資料亂數分成五組(每組中男女個數相同)，留下一組測試，其它組為訓練樣本。重複到每一組資料都要被測試過，最後將五次的測試結果取平均值。將人臉特徵區域分為三個部分，分別是外在特徵、一般特徵和內在特徵，搭配三種不同的Boosting做分類測試。最後再使用不同的Boosting分類器進行性別辨識。 實驗所採用的資?庫為FERET人臉資料庫，包含不同人種、年齡、表情、人臉旋轉角度以及亮度變化，在當中取出900張（450名男性和450名女性）正面人臉圖像實驗。結果顯示經由Haar-Like萃取出男女特徵搭配Logitboost的分類方法可達到較好的性能88%。[[abstract]]The face recognition system of gender can be applied to many real-life situations, such as: access control system of the place with gender restrictions, advertising billboards of store or exhibition, sex ratio as a market research analysis of the statistical site-specific basis. In the thesis, the system structure can be divided into three parts:（1）preprocessing module,（2）feature extraction module and（3）classification learning module. Firstly, we use Haar-Like feature algorithm to get the part of face. Secondly, we use Haar-Like to extract the facial features vector of male and female. In our experiment, we use 5-fold cross validation to classify the data. The data are dividedinto five groups(the numbers of male and female in each group is equal). One group is for testing, the others are training sample. Repeat the tests of each group. We take the average of test results in the five groups. After that, we divided the facial features into three parts: external features, General features and internal features. We use a different boosting classifier to identify gender after the three different Boosting tests for classification. In this thesis,the experimental data come from the FERET database, contains different race, age, expression, face rotation, and brightness changes. We pick out 900（450 males and 450 females）frontal face images implement experiments. The experimental results show that the best performance of 88% is achieved with the Haar-Like feature extraction and Logitboost classification methods.[[note]]碩