基於單一臉部深度影像之人臉辨識系統設計

Abstract

近年來，由於深度取像的技術越來越進步，以及深度取像的設備越來越普及，深度影像的應用逐漸成為研究的重要對象，使用深度影像進行臉部辨識的研究也越來越多，其中大部分的研究都是使用高價位的高解析度深度攝影器材。 本研究希望能透過經由演算法的改進，探討對於解析度較低的深度攝影機 所能造成的改進，並希望能夠有較快的辨識速度，使其能有實用上的價值。 本研究的貢獻有三點：(1) 提出一套結合深度解析度強化的方法，改善解析度較低的深度攝影機所獲得的深度影像品質，成功的改善了系統的影像品質；(2) 提出一套提取特徵點雲進行旋轉平移的方法，同時克服頭部姿態對於辨識率的影響並節省計算時間；(3) 提出一套將點雲重新映射回深度影像的方法，在進行轉正的處理之後重新回到深度影像的層面進行相似度比較，在辨識率相似的情況下同時能獲得較短的辨識時間。

Because of the progress of 3D range sensor technology, such equipment is now widely used in many applications and become a popular topic. There are more and more research discussing about the face recognition done by using depth image, most of which was captured by costly high-resolution range detecting sensor. In this research, we try to improve the performance of low-resolution ToF depth camera by our algorithm design so that short recognition time and high recognition rate could both be fulfilled. There are three contributions in this research: (1) A method combined with super-resolution of depth image is proposed and is successfully raise the quality of depth image acquired by camera; (2) A time-saving method to overcome the rotation and displacement during the processing is proposed; (3) A re-projection from rotated and shifted face point cloud to depth image method is proposed, so that the comparison could be made on depth image. It has been proved that it’s both efficient and accurate.