以個人化表情辨識系統來實現虛擬頭像控制

Abstract

傳統的人臉表情辨識已經被研究多時，也有許多應用，如：數位相機這類的科技產品上、網路相片管理軟體…等。其中偵測臉部的方法不一，辨識表情的方式也各有所異，雖然準確度已能達到水準之上，仍無法突破至九成以上，故若應用表情辨識來做為動畫的工具仍顯不足，必須對辨識機制有所改進。因此，在這篇論文中，我們提出一套針對個人化的系統來辨識專屬於使用者各自的表情，以自動化的人臉偵測校正人臉，建出專屬於使用者個人化的資料庫，並搭配區域二元圖樣(Local Binary Pattern)的方法，即時對於網絡攝像機前的使用者進行表情辨識。 在動畫產業上，有許多擬真人的動畫人物需要豐富的表情，若是在演員直接做表情之下能準確判別其表情；或者在演員無法做出較誇張或不容易表現的表情時，能有效的判斷出並對應到動畫人物，應用其對應的表情來完成動畫，那麼動畫人物的操作將會方便許多，且更為生動。

In this thesis, we propose a personalized system for facial expression recognition. Since our target is for avatar control or puppet control, we have adopted a very different approach. First, the user number is limited and personalized, and therefore our approach deviates away from traditional facial expression recognition approaches because the latter are targeted for any kind of subjects. Second, our approach is for personal use and there is a mapping between one’s facial expressions and the actual expressions of the avatar. Therefore, very exquisite facial expressions of the avatar such as awkwardness and evil smile etc. can be done while traditional facial expression system is targeted toward recognizing the regular six expressions. Furthermore, since we were using personalized expressions, and therefore we can optimize the weighting function of the recognition kernel such as Local Binary Patterns to increase the inter-class distance while reducing the intra-class distance. As a result, our recognition rate for personalized data can be close to 100% rate for the standard six expressions, while the reported data for general purpose systems were around 85%. In our system, more than 6 regular expressions, actually up to 12, can be used, and the recognition rate is still at 92% based on data from performers.