一個即時人臉特徵偵測及表情辨識系統

本論文研製一個即時人臉特徵偵測及表情辨識系統，主要是改善表情辨識的方法。以往的方法是靠嘴巴及嘴型的變化來進行表情辨識，造成生氣和悲傷臉型的辨識率偏低，而本論文將眉毛眉型變化和嘴巴嘴型變化的因素考慮進去，故而提高了生氣和悲傷臉型辨識的正確率。在輸入影像時，我們以預先訓練的背景將非人臉部分的像素去除，然後用YCbCr色彩空間判斷膚色，將不是人臉的資訊再度剔除，接著用型態學中的侵蝕運算來去除零星物件雜訊，再以水平投影及垂直投影法從多個物件中找出人臉，最後再用膨脹運算回補侵蝕運算所造成的失真，進而完成了人臉偵測的部分。接著，本論文再提出自行設計的五官定位法找出眼睛、瞳孔、眉毛及嘴巴的區域，再針對各個區域進行濾波，並找出具代表性的特徵點，最後以這些特徵點來進行表情辨識。本文利用預先訂定的各種情緒所伴隨的特徵點間的變化，作為判斷表情的度量依據。由於將眉毛眉型變化考慮進去，因而提升了臉型辨識的正確率。

關鍵字

人臉偵測；人臉辨識；型態學；膨脹；侵蝕；表情辨識

並列摘要

This thesis develops a real-time facial feature detection and facial expression recognition system, the core contribution is to improve the facial expression recognition rate. The former methods of facial expression recognition mainly depend on the change of mouth shape. In some angry or sad facial expression cases, the recognition rate is somewhat low. However, our thesis used eyebrow feature in addition to mouth shape factor considered that therefore enhanced the recognition rate. At the beginning of our process, the image background was trained and used to delete unrelated pixels, then the method of skin color adjustment in YCbCr color space was used to remove the information irrelative to face. After that, we used the erosion operator of morphology to eliminate the small noise objects, and use the horizontal/vertical projection from multiple objects to find out the face object. And finally in the face detection phase, the dilation operator of morphology was used to expand the face object shrunken and broken by the previous erosion operator. The thesis again proposed an independently design, the facial features methods, locating the regions of eyes, the pupil of the eye, eyebrow and mouth, aimed at each region to carry on the filter again, and discovered the facial feature points. Our facial expression recognition method is to rely on these feature points. In this thesis, we used new features to distinguish different facial expression. Due to the use of feature points of eyebrow, the facial expression recognition rate was improved.