利用深層卷積神經網絡之多層次特征偵測臉部關鍵點

Abstract

因為日常生活中人臉照片中常常出現的大幅度頭部移動、物體遮擋和光照變化，臉部關鍵點的偵測一直是一個富有挑戰而又急需解決的問題。在此論文中，我們嘗試利用深層卷積神經網絡中多層次分佈的特征去協助我們更好地解決這個問題。通過實驗研究，兩層額外的卷積層幫助我們有效地提取出在不同層中的特征圖並將它們有效結合在一起令偵測結果得到提升。除此之外，我們還進一步研究了不同種類激活函數對於模型快速收斂的幫助。通過在不同網絡層中使用不同激活函數，我們不僅加快了模型收斂，還得到更好地實驗結果。在文章最後，我們比較我們的模型與過去知名的方法在不同數據上的表現，並發現我們的模型有足夠的能力超越它們。

Facial landmarks detection has received a significant amount of attention due to its application to face alignment, recognition, verification, and identification. Yet, various factors including the issues of head pose variations, occlusion, and illumination uncertainly made this task challenging. In this study, we exploit hierarchically distributed features which can be learned through deep convolutional neural network (DCNNs). We modify the standard DCNNs via extending two convolutional layers for extracting semantic features in each convolutional layer. We further explore various activation functions with model their convergence efficiency. Experimental results demonstrate that the combination of PReLU and Linear activation performs well in facial landmark detection task. Our approach has also been evaluated on a benchmark dataset and compared with several state-of-the-art methods. The superior results demonstrate its effectiveness and general applicability.