特徵選擇與擷取對辨識娃娃臉之研究

Abstract

在社交場合中，娃娃臉這種臉部特徵在外表上會具有吸引力而且給人友善的感覺。人們可以很簡單的去判斷一個人是否有娃娃臉，然而，構成娃娃臉的特質十分模糊。在我們的論文中，將去分析人臉上的特徵，並挑選出哪些特徵對於判斷一個人是否具有娃娃臉是有幫助的。我們使用特徵選擇(Feature selection)方法去挑選出最佳的特徵組合以及使用卷積神經網路(Convolutional Neural Network)去自動的學習出特徵來判斷是否為娃娃臉。在實驗當中，我們比較使用心理學的特徵、特徵選擇以及卷積神經網路三種方法的差別，在使用卷積神經網路方法的結果會比其他兩種方法來得更好。

Babyface is a type of face that is usually attractive and friendly in appearance. People can recognize this special face easily. However, the components that compose a babyface remain unclear. In this paper, we analyze the features in a human face and determine which features are useful for determining a babyface. In particular, we use feature selection methods to choose the best combination in discriminative capability and the convolutional neural networks to automatically learn the features. We compare our result with the psychological studies and showed that the features obtained by using the convolutional neural networks technique outperform the other methods under testing.