2D facial landmark model design by combining key points and inserted points. Issue 21 (30th November 2015)

• Record Type: Journal Article
• Authors: Chen, Fangmei; Xu, Yong; Zhang, David; Chen, Kai
• Is Part Of: Expert systems with applications. Volume 42:Issue 21(2015)
• Page Start: 7858
• Page End: 7868
• Publication Date: 2015-11-30
• Subjects: Approximation error -- Geometric facial representation -- Inserted point -- Key point -- Landmark model; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2015.06.015
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: An optimized landmark model (LM) for face representation is developed. The landmarks are divided into key points and inserted points. Key points include anatomical and mathematical landmarks. Inserted points help describe contours and are optimized by iterative search. The LM shows superiority in delicate and efficient face shape representation. Abstract: Facial landmarks can be used as a compact and effective representation of facial geometry. Although different facial landmark models (LMs) have been used for various face image analysis tasks, there are few works investigating the design of the LMs, which takes care of how many landmarks to use to represent facial geometry and where to locate them. Existing LMs are either not sufficient to approximate face shapes precisely or including many redundant landmarks. This paper aims to develop an optimized LM which can delicately represent facial geometry while keeping as few landmarks as possible. We divide the landmarks into two categories, namely key points (KPs) and inserted points (IPs). KPs include anatomical and mathematical landmarks which are defined with consideration of an anthropometry face model; IPs are pseudo-landmarks that help describe the face shape accurately. An iterative searching scheme is proposed to find a series of optimal positions for IPs until some acceptable error is obtained. The obtained LM includes 54 KPs and 44 IPs distributed on different facial regions. We compare our LM and other LMs with similar number of points in terms of shape approximation error and landmark detection error. Experimental results show the superiority of our LM in both criteria. The proposed LM is useful for applications that require delicate shape representations and promotes standardization of facial LMs.