機器人之表情辨識快速學習法則

Abstract

應用在機器人的表情辨識系統，會因為使用者呈現表情的方式有所不同，而產生系統無法辨識的新人臉表情。為了使機器人能夠適應新人臉之表情，本篇論文提出了一套能夠學習新臉孔之表情辨識系統。主要的想法是利用調整支持向量機(Support vector machine, SVM)的切割平面(Hyperplane)係數，來達到辨識新表情資料的目的。支持向量追蹤學習法(Support vector pursuit learning, SVPL)的概念被引入在高斯核空間(Gaussian kernel space)中來調整切割平面。為了加快訓練學習的速度，只有錯誤的表情資料和一定數量的關鍵舊集合被拿來重新訓練，藉以產生新的SVM分類器。經過調整切割平面後，不僅可以辨識之前無法辨識之新臉孔表情，並且還可以保持對舊有資料的辨識率。另外，我們使用蓋伯小波(Gabor wavelet)特徵擷取的方法來強化擷取之效能，以確保欲學習表情特徵值之正確性。所提出的表情學習演算法已成功的應用在實驗室之娛樂機器人的平台上，線上測試結果顯示即使是新的表情資料，也可以透過學習系統把辨識率由58%提升到81.3%，並且還可以對舊有表情資料保持78.7%之辨識率。

A robotic facial expression recognition system very often misclassifies data from a new face because different people may show their expressions in different ways. This thesis aims to study a facial expression recognition system that can learn new facial data and facilitate a robot to accommodate itself to various persons. The main idea of the proposed method is to adjust parameters of the hyperplane of support vector machine (SVM) for classifying new facial data. The concept of support vector pursuit learning (SVPL) is adopted to retrain the hyperplane in the Gaussian kernel space. To expedite the training procedure, we propose to retrain the new SVM classifier by using only samples classified incorrectly and the critical sets (CSs) from previous samples. After adjusting hyperplane parameters, the new classifier not only recognizes new facial data but also keeps acceptable performance of classifying previous data. Further, to obtain reliable facial features, we adopted Gabor wavelet to develop a feature extraction method in the system. The proposed algorithms have been successfully implemented on an entertainment robot platform. On-line experimental results show that the proposed system learns new facial data with a recognition rate of 81.3% increased from an original recognition rate of 58%. The proposed method also keeps satisfactory recognition rate of old facial samples with a recognition rate of 78.7%.