3D Point Clouds Face Recognition by BP Neural Networks

In this paper, coordinates of 3D point clouds data are used directly for face recognition. 3D face point clouds database were built from neutral expressions and variant poses (left, frontal, and right). We are currently concentrating on a procedure of the coarse and fine transformation adjustment for face recognition. The transformation adjustment is based on the spatial fitted line to the point clouds data, centered at the nose tip, along longitudinal and cross-section profiles. Our 3D face point clouds data are combinations of 69 same-person pairs and 19,531 different-person pairs. We present an artificial neural networks with nose tip/root alignment geometry concept as network inputs for face recognition. The experiment presents that the success rates of same-person pairs is 100%. An appropriate amount for different-person pairs are 820 and their success rates can reach 99.3%. Compared to other related works, the framework has following highlights: (1) a coordinate transformation has to be done before face recognition alignment stage; (2) face recognition is directly completed by using point clouds without building 3D modality; (3) artificial intelligence neural networks for face recognition only with the feature points of nose tip/ root have illustrated that the proposed system is feasible.

關鍵字

3D face recognition ； 3D face point clouds database ； Artificial Intelligence neural networks

並列摘要

本研究利用3D點雲坐標資料直接進行人臉辨識，建立帶有正常表情及不同姿態（左側、正面，和右側）的三維人臉點雲資料庫，以基於通過鼻尖、鼻根縱橫剖面空間擬合直線線，經大略及精細調整轉換完全自動貼合。本研究將鼻尖／根排列的幾何概念作為人臉識別特徵以為人工神經網路輸入。三維人臉點雲數據包括相同人計69對，和不同人19,531對。驗證結果，同一人者的成功率是100%。不同人者，較合適的數量是820，成功率為99.3%。本研究相較其它相關研究框架具有以下特點：（1）做貼合比較之前必須進行坐標轉換；（2）不須三維塑模直接以點雲資料完成辨識；（3）僅依靠鼻尖和鼻根特徵點之神經網路進行人臉辨識，實驗顯示可行。

並列關鍵字

3D人臉辨識； 3D人臉點雲資料；人工智慧神經網路

主題瀏覽