Face Recognition using Parallel Associative Memory 2009/12/31 學生：羅國育
Outline • Introduction • Auto-associative memory for face recognition • Parallel associative memory for face recognition • Face similarity measure for recognition • Experimental Results • Conclusion
Introduction • This paper proposes parallelization of the auto-associative memory in order to apply it for recognition of high resolution face images. • The Olivetti Research Laboratory (ORL) face database.
Auto-associative memory for face recognition (1/3) • Associative memories mimic the capacity of human brain to recall information in a robust and associative access mode. • Neural network.
Auto-associative memory for face recognition (2/3) • The input-output relationship in a linear associative memory is described by M × N gray scale images. W is an (MN)×(MN) matrix. (1) (2)
Auto-associative memory for face recognition (3/3) • Now substituting W from (2) we have (3)
Parallel associative memory for face recognition (1/4) • The system mainly performs three tasks: (i) Information storage, (ii) information retrieval based on some input pattern and (iii) matching of the information.
Face similarity measure for recognition (2/2) • Run length: 兩個1之間，0出現的次數分別是 3、9、3、2 的話，這些值就稱為0的 run length。 Fig. 2. A grid structure to explain the concept of run length count
Experimental Results (1/4) • The ORL database contains total 400 images of 40 individuals from various ethnicity and sex under various pose, light, scale and expression. • One set is considered as training set and the remaining 39 sets are used as testing set.
Experimental Results (2/4) • In this experiment values of N, n and Tr are set to 128, 16 and 6 respectively. • It is found that the best suitable value of Th1 and Th2 are 40 (i.e, 65 % portion of input image) and 25 (i.e, 40 % portion of input image) respectively.
Experimental Results (4/4) • False Rejection Ratio (FRR) is 1% and 5.2% respectively. • For false accept experiments the system was tested with 40randomly chosen face images which are not present in the database.The observed FAR is 3.1%.
Conclusion • In this paper a parallel associative memory based efficient face recognition system has been proposed. • The goal is to scale the associative memories to high resolution images. • A novel run length count based measure of face similarity suited for parallel associative memory is also proposed.
Thanks for your attention! 2014/11/21