人們可以透過對方臉上的表情變化來了解對方的喜好與意圖,因此,自動化表情辨識在電腦視覺與人工智慧領域都提供了重要的學術和商業潛力。在本研究中,我們針對靜態臉部影像,提出了基於深度學習之可分類四種不同表情類別的表情辨識系統,我們利用深度卷積神經網路 (Deep CNN) 來訓練及辨識表情,並且使用 Haar 檢測器來提取眼睛和嘴巴的區域,以簡化輸入圖像,此外,我們取出區域二值化紋理 (LBP) 的特徵作為 JEFFE 資料庫的深度學習模型輸入。根據實驗結果顯示,本研究所提出的方法與其他的結果相比較,能有效地提升表情辨識的準確度。
Emotion recognition provides significant academic and commercial potential in the fields of computer vision and artificial intelligence. In this thesis, we propose an emotion recognition system for classifying four different emotion classes from static facial images. We leverage the deep convolutional neural network (Deep CNN) to recognize the emotions. The Haar-cascade detectors are used to extract regions of eyes and mouth for simplifying the problem domain from the input images. Then, we extract the local binary patterns (LBP) features as input to CNN models for JEFFE database. The experimental results are compared with the other results to demonstrate the classification accuracy of emotions.