以區域二元圖樣為基礎之階層式人臉表情辨識

Abstract

自動化的人臉表情辨識迄今一直是電腦視覺領域的一個大難題，這項技術與許多領域息息相關，例如人機互動或以資料驅動的動畫製作。在這篇論文中，我將使用人臉辨識中常用的「區域二元圖樣」方法來辨識表情，並進一步提出一個新的方法來改善結果。這個方法是「階層式的表情辨識」。在傳統的表情辨識中，研究員使用訓練資料來做出一個表情分類器，並將測試資料直接歸類為可能的幾種表情之一。而階層式表情辨識則將辨識的過程分成很多個階段，在每個階段使用特定的方法來處理特定的表情，一步一步地辨識出目標照片為何種表情。使用基本的區域二元圖樣方法在我採用的兩個表情資料庫上分別可以達到約80%及86%的辨識率，若加上階層式方法可提升至約87%及89%的辨識率。這個結果顯示區域二元圖樣方法用在表情辨識上有不錯的結果，而階層式表情辨識的方法可以更進一步改善辨識率。

So far, automatic human facial expression recognition has been a challenging problem in computer vision area, and brings strong impacts in important applications in many areas such as human-computer interaction and data-driven animation. In this thesis, I use the Local Binary Patterns method, which was used commonly in face recognition, to recognize the facial expressions. Furthermore, I propose a novel idea to improve the result. This new method is called “Hierarchical Facial Expression Recognition”. In traditional facial expression recognition, the researchers use training data to make a classifier, and classify the test expression data into one of the several basic expressions directly. Hierarchical method separates the recognition process into few stages, each stage using specific methods to deal with the specific expression to step by step recognize the target image. The overall recognition rates of my system on two expression databases are about 80% and 86% with pure Local Binary Patterns algorithm, and reach 87% and 89% after using the hierarchical method. This work shows that the Local Binary Patterns method is practical in facial expression recognition, and the idea of hierarchical facial expression recognition can further improve the result.