基於時間強化設計之情緒辨識方法

Abstract

本論文之主旨在研究基於連續影像之情緒辨識方法，文中提出一套基於連續時間關聯資訊之情緒辨識與描述之方法。本方法首先透過主動外觀模型(Active appearance model, AAM)產生人臉影像樣本之形狀模型與紋理模型，擷取人臉特徵點及幾何特徵值，再由相關向量機(Relevance vector machine, RVM) 辨識情緒狀態。在辨識設計方面，本研究透過時序分析，辨識情緒類別之可能性(Likelihood)，並將辨識結果轉換至二維Arousal與Valence平面(Arousal-Valence Plane, A-V Plane)，以利系統之反應設計。所發展之方法能針對情緒程度、類別比例等資訊做更細微之辨識，且能夠分析情緒之轉變過程。經由實驗驗證，所發展之方法確能有效提升情緒辨識之效能，對基本表情之辨識率可達95%以上，對複雜情緒亦能做有效之辨識。為驗證本方法線上(on-line)辨識之效果，本論文設計一套基於人臉情緒辨識之智慧音樂選曲系統，此系統可藉由即時人臉情緒辨識，選取適當之音樂進行播放，透過音樂將使用者情緒逐漸導向至目標情緒。

In this thesis, a temporal-reinforced approach to enhancing emotion recognition from facial images has been developed. Shape and texture models of facial images are computed by using active appearance model (AAM), from which facial feature points and geometry feature values are extracted. The extracted features are used by relevance vector machine (RVM) to recognize emotional states. In this work, we propose a temporal analysis approach to recognizing likelihood of emotional categories, such that more subtle emotion, such as degree and ratio can be obtained. Furthermore, a method is developed to map the recognition result to the Arousal-Valence Plane (A-V Plane). Experimental results verify that the performance of emotion recognition is enhanced by the proposed method. Furthermore, the A-V values are applied to an intelligent music selection system. With emotion recognition of current A-V values, appropriate songs are selected and played by this system to change a person emotion towards a target emotion.