陳美勇Chen, Mei-Yung林高遠Lin, Kao-Yuan2019-09-032016-08-232019-09-032016http://etds.lib.ntnu.edu.tw/cgi-bin/gs32/gsweb.cgi?o=dstdcdr&s=id=%22G060132057A%22.&%22.id.&http://rportal.lib.ntnu.edu.tw:80/handle/20.500.12235/96988本論文提出一種以人類臉部影像為輸入資料，用類神經網路即時測得受測者疼痛水準的方法。情感運算在最近幾年來已經逐漸受到重視，而臉部影像疼痛水準自動估測，有助於醫療照顧、健康促進等領域自動化，有效減少第一線照顧者的負擔。但相關研究的數量與關注程度遠落後於其他表情辨識技術，使得相關應用發展受限。 本論文嘗試兩種輸入資料：一種是屬於低階外觀特徵的人類臉部眼、嘴區域Uniform LBP直方圖，取得118-D向量；另一種屬於使類神經網路自動尋找的高階抽象特徵，將臉部影像做最大池化(Max Pooling)處裡後，從32x32灰階影像取得1024-D向量。將兩者正規化，再輸入類神經網路做迴歸訓練與測試。 實驗結果方面，將The UNBC-McMaster Shoulder Pain Expression Archive Database隨機分割為兩份，分別做為訓練與測試樣本。將本論文提出的方法與Sebastian Kaltwang等人與Xiaopeng Hong等人的研究比較，可以達到較小的均方誤差(MSE=0.17)與較接近1的皮爾森相關係數(r=0.94)。速率表現方面，本論文以C#實作出的程式在i5雙核心的電腦上平均可以達24FPS。This thesis presents a method to estimate pain intensity which is revealed on human face image rapidly. Two types of data are extracted from the human face image: one of which is the Uniform LBP, the belong low-level appearance features, which is extracted from the eyes and mouth area; the other is the 32x32 face image data which is extracted using Max Pooling. Both will be computed by the regression neural network, and the neural network is trained and the training result will be verified. The data from the UNBC-McMaster Shoulder Pain Expression Archive Database is randomly assigned into two groups—one for training, another for testing. The result of this study achieves a MSE of 0.17 and a Pearson’s correlation coefficient close to 1 (r = 0.94), and the average computing speed achieves 24FPS on i5 dual-core computer.監督式機器學習特徵學習痛苦偵測電腦視覺Supervised Machine LearningFeature LearningPain detectionComputer Vision