A Deep Joint Learning Approach for Age Invariant Face Verification
Age-related research has become an attractive topic in recent years due to its wide range of application scenarios. In spite of the great advancement in face related works in recent years, face recognition across ages is still a challenging problem. In th
- PDF / 766,247 Bytes
- 10 Pages / 439.37 x 666.142 pts Page_size
- 7 Downloads / 226 Views
Guangzhou University, Guangzhou 510006, China Sun Yat-sen University, Guangzhou 510006, China [email protected], [email protected], [email protected], [email protected]
2
Abstract. Age-related research has become an attractive topic in recent years due to its wide range of application scenarios. In spite of the great advancement in face related works in recent years, face recognition across ages is still a challenging problem. In this paper, we propose a new deep Convolutional Neural Network (CNN) model for age-invariant face verification, which can learn features, distance metrics and threshold simultaneously. We also introduce two tricks to overcome insufficient memory capacity issue and to reduce computational cost. Experimental results show our method outperforms other state-of-the-art methods on MORPH-II database, which improves the rank-1 recognition rate from the current best performance 92.80% to 93.6%.
Keywords: Face verification CNN · Joint learning
1
· Age invariant · Face recognition · Deep
Introduction
Age-related research has become an attractive topic in recent years due to its wide range of application scenarios. Age information is useful in many applications, such as age-specific human-computer interaction, security surveillance monitoring, age-based face images retrieval, automatic face simulation and intelligent advertisement system etc.. In spite of the great advancement in face related works in recent years, face recognition across ages is still a challenging problem. In paper [1], face verification achieved near-human performance on Labeled Faces in the Wild (LFW) dataset using high-dimensional Local Binary Pattern feature (HD-LBP). The work in paper [2,3] even achieved exceed-human ability on face verification using deep learning method. To our best knowledge, there is no such good result on age invariant recognition. The challenges on age invariant recognition include large intra-subject variations and great inter-subject similarity. As well known, human face appearance will change greatly with the aging process. The changes are different in the different age period as shown in Fig.1(a). From birth to adulthood, the greatest change is the craniofacial growth, that is shape change; and from adulthood to c Springer-Verlag Berlin Heidelberg 2015 H. Zha et al. (Eds.): CCCV 2015, Part I, CCIS 546, pp. 296–305, 2015. DOI: 10.1007/978-3-662-48558-3 30
A Deep Joint Learning Approach for Age Invariant Face Verification
297
old age, the most perceptible change becomes skin aging, that is texture change [4]. These changes of same person are the intra-subject variations. Meanwhile, different persons on same age period maybe look like same, that is the intersubject similarity as shown in Fig.1(b). Therefore, enlarging the inter-subject
3
18
4
5
7
21
23
31
12
(a) Intra-subject variation
36
15
38
(b) Inter-subject similarity
Fig. 1. Example images showing intra-subject variations and the inter-subject similarity. (a) Shows face appearance changes with the aging process. Imag
Data Loading...
