基於人臉影像之年齡與表情強度估計

Abstract

如何讓電腦藉由觀察臉上的特徵來了解使用者的情緒與資訊是個基礎且重要的議題。在本論文中，我們針對人臉的年齡估計以及表情分析等題目進行研究與探討。在臉部年齡估計方面，與過去常見預測人臉影像年齡的方式不同，我們將年齡估計轉換為排序學習的問題，藉由成對式的比較將原本的年齡估計問題轉換為一連串的子問題，並且統整一連串子問題的結果來推論出人臉影像的真實年齡。我們提出了一個基於相對順序來估計人臉影像年齡的方法，藉由實驗結果的證明，顯示我們所提出的排序方法比一般常見的方法更準確。在臉部表情分析的研究上，過去的研究大多著重在表情類別辨識，在本論文中我們同時考慮了表情類別以及表情的強度，並且提出一個藉由單張人臉影像同時辨識表情類別以及估計表情強度的方法。過去的方法將不同表情之間的類別視為是互相獨立，我們的方法利用成本敏感學習同時考慮了不同表情類別之間的相對關係，以及相同表情類別下不同強度之間的關係。利用廣泛的實驗，我們證明所提出的方法與過去的方法相比不只可以達到最低的成本，同時有效的降低分類錯誤率。

Face image understanding is important in computer vision and pattern recognition. In this dissertation, two important topics in face image understanding are studied, age estimation and facial expression analysis. To estimate human ages based on face images, we convey age estimation into a ranking problem. Instead of predicting a person's age directly, we infer the exact age of a face image based on a series of comparisons. Our approach is designed based on relative order, and experimental results show that our method performs better than conventional approaches. In facial expression analysis, we consider further the intensity of facial expression, and focus on both expression category identification and intensity level estimation. We present a facial expression recognition approach that can also infer the expression intensity rank based on a single image. Instead of regarding the labels of expressions as independent to each other, our approach takes the inter-label relationship into consideration, which can estimate both the facial expression and intensity under a cost-sensitive setting. Experimental results show that our method can provide minimal-cost results and low misclassification rates compared to existing approaches.