An efficient face verification method in a transformed domain

https://doi.org/10.1016/j.patrec.2006.12.005Get rights and content

Abstract

In this paper we propose a low-complexity face verification system based on the Walsh–Hadamard transform. This system can be easily implemented on a fixed point processor and offers a good compromise between computational burden and verification rates. We have evaluated that with 36 integer coefficients per face we achieve better Detection Cost Function (6.05%) than the classical eigenfaces approach (minimum value 6.99% with 126 coefficients), with a smaller number of coefficients.

Introduction

In the last decade significant advances have been achieved on biometrics, especially on face recognition (Jain et al., 1999). This has been possible due to the increase of computational power of the state-of-the-art computers. However, there are several application scenarios where a low-complexity algorithm, which can be implemented on a low-cost processor is desirable. Some examples of this situation are mobile telephone, PDA or standalone control access systems. Probably in these situations the processor will be a fixed point one, and the number of operations per second smaller than the state-of-the-art processors used to develop the best algorithms available nowadays.

Section snippets

Face recognition

Usually, a pattern recognition system consists of two main blocks: feature extraction and classifier. Fig. 1 summarizes this scheme. On the other hand, there are two main approaches for face recognition:

  • (a)

    Statistical approaches consider the image as a high-dimension vector, where each pixel is mapped to a component of a vector. Due to the high-dimensionality of vectors some vector-dimension reduction algorithm must be used. Typically the Karhunen–Loeve transform (KLT) is applied with a simplified

Walsh–Hadamard transform

The Walsh–Hadamard transform basis functions can be expressed in terms of Hadamard matrices. A Hadamard matrix Hn is a N × N matrix of ±1 values, where N = 2n.

In contrast to error-control coding applications, in signal processing it is better to write the basis functions as rows of the matrix with increasing number of zero crossings.

The ordered Hadamard matrix can be obtained with the following equations (Gonzalez and Woods, 1993):H(x,u)=1N(-1)i=0n-1bi(x)pi(u)where bk(x) is the kth bit in the

Results

This section evaluates the results achieved using the WHT and compares them with the classical KLT, eigenface, and DCT methods.

Conclusions

We have proposed a new approach to face recognition based on the Walsh–Hadamard transform, which can be easily implemented on a fixed point processor (Faundez-Zanuy et al., 2005). The experimental results reveal that it is competitive with the state-of-the-art statistical approaches to face recognition. Taking advantage of the minor differences of using different transforms (see Jain, 1989, p. 517), emphasis is focused on this items:

  • (a)

    We check that WHT performs reasonably good using small and

Acknowledgements

This work has been supported by FEDER and MEC, TIC-2003-08382-C05-02, TEC-2006-13141-C03-02.

References (11)

  • M. Faundez-Zanuy

    On the vulnerability of biometric security systems

    IEEE Aerospace Electron. Systems Mag.

    (2004)
  • M. Faundez-Zanuy

    Data fusion in biometrics

    IEEE Aerospace Electron. Systems Mag.

    (2005)
  • M. Faundez-Zanuy et al.

    A low-cost webcam and personal computer opens doors

    IEEE Aerospace Electron. Systems Mag.

    (2005)
  • J.D. Gibson

    Digital Compression for Multimedia, Principles and Standards

    (1998)
  • R.C. Gonzalez et al.

    Digital Image Processing

    (1993)
There are more references available in the full text version of this article.

Cited by (28)

  • Fusion of local normalization and Gabor entropy weighted features for face identification

    2014, Pattern Recognition
    Citation Excerpt :

    Among the most widely used methods for face recognition based on feature extraction are eigenfaces [10], a holistic method that uses Principal Component Analysis (PCA) to project the image data vector into a reduced space, maximizing the variance of the data; and Fisherfaces [11], which is based on Linear Discriminant Analysis (LDA), maximizes the distance between classes and minimizes the distance between prototypes within each class; and methods based on Independent Component Analysis (ICA) [12,13]. Other methods for face recognition are based on frequency space as a discrete cosine transform (DCT) [14,15], that compares DCT–based feature vectors from different images, and the Walsh–Hadamard transform (WHT) [16], a low complexity algorithm that compares WHT-based feature vectors. There are also methods that use a linear combination of training images to reconstruct the testing image, such as Sparse representation-based methods (SRC) [17].

  • Constrained large Margin Local Projection algorithms and extensions for multimodal dimensionality reduction

    2012, Pattern Recognition
    Citation Excerpt :

    We observe from Table 9 that our techniques can deliver comparable or even better results than other methods, including the class labels or sparsity representation guided LPP extensions, i.e., DLPP, DLPP/MMC, SPP and SOLPP, in most cases. DCT-1 and DCT-2 perform better than PCA in most cases, which is consistent with [52]. We also notice that these LPP extensions outperform the original LPP algorithm in most cases.

  • Methodological improvement on local Gabor face recognition based on feature selection and enhanced Borda count

    2011, Pattern Recognition
    Citation Excerpt :

    Our results, LMG-verif, are better than those published in [18]. In Table 10 our result is compared to eigenfaces, KLT, DCT and WHT for the same database [18]. Jet selection has two important results, the first one, is to improve recognition rate and the second, to reduce computational time.

View all citing articles on Scopus
View full text