具表情混合比例及強度之人臉表情辨識系統

Abstract

本論文發展出一套可用來辨識出人臉表情中各基本表情所占比例以及人臉表情強度之人臉表情辨識系統。首先以主動外觀模型(Active Appearance Model, AAM)訓練出人臉影像的形狀及紋理模型;並採用改良式Lucas-Kanade 影像校正進行輸入影像校正;然後利用之前訓練出的AAM紋理模型取得紋理特徵參數。本論文所使用的紋理特徵參數分為全臉、上半臉、下半臉三種，分別作為強度辨識、上半臉Action Units (AU)組合辨識、下半臉AU組合辨識各自的輸入。強度辨識、上半臉AU組合辨識、下半臉AU組合辨識都使用倒傳遞類神經網路(Back propagation Neural Network)給出評價分數;接著我們提出一套整合設計可用這些評價分數得到表情強度及表情比例。

In this thesis, a facial expression recognition system which can recognize facial expressions as well as the expression intensity and mixture ratio of basic expressions is developed. Active Appearance Model (AAM) is used to train shape model and texture model. The improved Lucas-Kanade image alignment algorithm is then applied to align the input images to obtain texture features. A novel method is proposed to recognize ratio of basic expressions and intensity of facial expression. Three kinds of texture features are used in this method: 1. texture features of whole face, which are used as inputs of facial expression intensity recognition, 2. texture features of upside face, which are used as inputs of upper face action units recognition, 3. texture features of downside face, which are used as the inputs of lower face action units recognition. Back propagation neural networks are used to obtain the recognition scores, which are then exploited to classify the facial expression results, including basic facial expression ratio and the facial expression intensity.