人臉偵測與辨識的問題是電腦視覺與人機互動中的重要議題,也是相當困難的任務。近年來隨著深度學習的發展,已有許多針對人臉偵測以及表情辨識的相關技術,尤其是卷積神經網路之技術在此領域中的表現更為突出。 本論文先將多任務串級卷積神經網路之技術應用於人臉偵測,再設計基於密集連接卷積神經網路(DenseNet)之即時表情辨識系統。此系統首先將輸入影像進行縮放形成圖像金字塔,再利用階層網路判斷可能為人臉的候選窗口,若為人臉區域,則將其輸入至表情辨識系統,經由DenseNet來進行人臉表情辨識。DenseNet擁有特徵再利用的特性,能夠有效減少參數量以及計算量,因此有助於即時辨識系統的開發。為了捕捉不同表情的細微面部肌肉差異,在架構設計上採用步幅為1的卷積運算,也嘗試不同數量的密集塊,透過實驗證實本論文所設計的系統確實能夠達到30FPS的即時辨識以及優於人眼辨識的準確率。
Face detection and recognition is an important issue and a difficult task in computer vision and human-computer interaction. Recently, with the development of deep learning, several related technologies have been proposed for face detection and facial expression recognition, and the outstanding convolutional neural networks are the most common used in this field. This thesis applies the multi-task cascade convolutional neural network to face detection, and then designs the real-time facial expression recognition system based on densely connected convolution network (DenseNet). The system first scales the input image to an image pyramid, and then uses the hierarchical network to determine whether a candidate window includes a human face. If a face exists, then send the candidate window to the facial expression recognition system. Since DenseNet possesses the property of feature reuse, it can effectively reduce the amount of parameters and computation efforts, beneficial to develop the real-time system. In order to capture the variation of facial muscle in different expressions, this architecture adopts convolution operations with a stride 1 and tries different numbers of dense blocks. Through experiments, the proposed system can achieve real-time recognition in 30FPS and with recognition accuracy better than human eyes.