Elsevier

Pattern Recognition

Volume 63, March 2017, Pages 15-29
Pattern Recognition

Sequential dimensionality reduction for extracting localized features

https://doi.org/10.1016/j.patcog.2016.09.006Get rights and content

Highlights

  • We propose a new nonnegative matrix underapproximation model for images.

  • We design an algorithm that runs in linear time in the dimensions of the input matrix.

  • This allows us to extract sequentially localized and spatially coherent features.

  • We illustrate the effectiveness of our approach on a synthetic data set, facial and hyperspectral images.

  • We show that it competes favorably with comparable state-of-the-art techniques.

Abstract

Linear dimensionality reduction techniques are powerful tools for image analysis as they allow the identification of important features in a data set. In particular, nonnegative matrix factorization (NMF) has become very popular as it is able to extract sparse, localized and easily interpretable features by imposing an additive combination of nonnegative basis elements. Nonnegative matrix underapproximation (NMU) is a closely related technique that has the advantage to identify features sequentially. In this paper, we propose a variant of NMU that is particularly well suited for image analysis as it incorporates the spatial information, that is, it takes into account the fact that neighboring pixels are more likely to be contained in the same features, and favors the extraction of localized features by looking for sparse basis elements. We show that our new approach competes favorably with comparable state-of-the-art techniques on synthetic, facial and hyperspectral image data sets.

Introduction

Linear dimensionality reduction (LDR) techniques are powerful tools for the representation and analysis of high dimensional data. The most well-known and widely used LDR is principal component analysis (PCA) [14]. When dealing with nonnegative data, it is sometimes crucial to take into account the nonnegativity in the decomposition to be able to interpret the LDR meaningfully. For this reason, nonnegative matrix factorization (NMF) was introduced and has been shown to be very useful in several applications such as document classification, air emission control and microarray data analysis; see, e.g., [7] and the references therein. Given a nonnegative input data matrix MR+n×m and a factorization rank r, NMF looks for two matrices UR+n×r and VR+r×m such that MUV. Hence each row M(i,:) of the input matrix M is approximated via a linear combination of the rows of V: for 1in, M(i,:)k=1rUikV(k,:).In other words, the rows of V form an approximate basis for the rows of M, and the weights needed to reconstruct each row of M are given by the entries of the corresponding row of U. The advantage of NMF over PCA (that does not impose nonnegativity constraints on the factors U and V) is that the basis elements V can be interpreted in the same way as the data (e.g., as vectors of pixel intensities; see Section 3 for some illustrations) while the nonnegativity of the weights in U make them easily interpretable as activation coefficients. In this paper, we focus on imaging applications and, in particular, on blind hyperspectral unmixing which we describe in the next section.

A hyperspectral image (HSI) is a three dimensional data cube providing the electromagnetic reflectance of a scene at varying wavelengths measured by hyperspectral remote sensors. Reflectance varies with wavelength for most materials because energy at certain wavelengths is scattered or absorbed to different degrees, this is referred to as the spectral signature of a material; see, e.g., [21]. Some materials will reflect the light at certain wavelengths, others will absorb it at the same wavelengths. This property of hyperspectral images is used to uniquely identify the constitutive materials in a scene, referred to as endmembers, and classify pixels according to the endmembers they contain. A hyperspectral data cube can be represented by a two dimensional pixel-by-wavelength matrix MR+n×m. The columns M(:,j)R+n of M (1jm) are original images that have been converted into n-dimensional column vectors (stacking the columns of the image matrix into a single vector), while the rows M(i,:)R+m of M (1in) are the spectral signatures of the pixels (see Fig. 1). Each entry Mij represents the reflectance of the i-th pixel at the j-th wavelength. Under the linear mixing model, the spectral signature of each pixel results from the additive linear combination of the nonnegative spectral signatures of the endmembers it contains. In that case, NMF allows us to model hyperspectral images because of the nonnegativity of the spectral signatures and the abundances: Given a hyperspectral data cube represented by a two dimensional matrix MR+n×m, NMF approximates it with the product of two factor matrices UR+n×r and VR+r×m such that the spectral signature of each pixel (a row of matrix M) is approximated by the additive linear combination of the spectral signatures of the endmembers (rows of matrix V), weighted by coefficients Uik representing the abundance of the k-th endmember in the i-th pixel. For all i, we have: M(i,:)k=1rUikV(k,:),where r is the number of endmembers in the image. The matrix U is called the abundance matrix while the matrix V is the endmember matrix. Fig. 1 illustrates this decomposition on the urban hyperspectral data cube.

Unfortunately, as opposed to PCA, NMF is a difficult problem (NP-hard) [22]. Moreover, the decomposition is in general non-unique and has to be recomputed from scratch when the factorization rank r is modified. For these reasons, a variant of NMF, referred to as nonnegative matrix underapproximation, was recently proposed that allows us to compute factors sequentially; it is presented in the next section.

Nonnegative matrix underapproximation (NMU) [8] was introduced in order to solve NMF sequentially, that is, to compute one rank-one factor U(:,k)V(k,:) at a time: first compute U(:,1)V(1,:), then U(:,2)V(2,:), etc. In other words, NMU tries to identify sparse and localized features sequentially. In order to keep the nonnegativity in the sequential decomposition, it is natural to use the following upper bound constraint for each rank-one factor of the decomposition: for all 1kr, U(:,k)V(k,:)Mp<kU(:,p)V(p,:).Hence, given a data matrix MR+n×m, NMU solves, at the first step, the following optimization problem:minuR+n,vR+mMuvTF2suchthatuvTM,referred to as rank-one NMU. Then, the nonnegative residual matrix R=MuvT0 is computed, and the same procedure can be applied on the residual matrix R. After r steps, NMU provides a rank-r NMF of the data matrix M. Compared to NMF, NMU has the following advantages:

  • 1.

    As PCA, the solution is unique (under some mild assumptions) [10].

  • 2.

    As PCA, the solution can be computed sequentially, and hence the factorization rank does not need to be chosen a priori.

  • 3.

    As NMF, it leads to a separation by parts. Moreover the additional underapproximation constraints enhance this property leading to better decompositions into parts [8] (see also Section 3 for some numerical experiments).

NMU was used successfully for example in hyperspectral [10] and medical imaging [16], [17], and for document classification [3].

Modifications of the original NMU algorithm were made by adding prior information into the model, as it is also often done with NMF; see, e.g., [4] and the references therein. More precisely, two variants of NMU have been proposed:

  • 1.

    One adding sparsity constraints on the abundance matrix, dubbed sparse NMU [11].

  • 2.

    One adding spatial information about pixels, dubbed spatial NMU [12].

In this paper, we include both sparsity constraints and spatial information in NMU. This allows us to extract localized features in images more effectively. We present our algorithm in Section 2 and show in Section 3 that it competes favorably with comparable state-of-the-art techniques (NMF, sparse NMF and several variants of NMU) on synthetic, hyperspectral and facial image data sets.

Section snippets

Algorithm for NMU with priors

In this section, we describe our proposed technique that will incorporate both spatial and sparsity priors into the NMU model. This will allow us to extract more localized and more coherent features in images.

Experimental results

In this section, we conduct several experiments to show the effectiveness of PNMU.

In the first part, we use synthetic data sets for which we know the ground truth hence allowing us to quantify very precisely the quality of the solutions obtained by the different algorithms.

In the second part, we validate the good performance of PNMU on real data sets, and compare the algorithms on two widely used data sets, namely, the Cuprite hyperspectral image and the CBCL face data set.

The following

Conclusions

In this paper a variant of NMU was proposed, namely PNMU, taking into account both sparsity and spatial coherence of the abundance elements. Numerical experiments have shown the effectiveness of PNMU in correctly generating sparse and localized features in images (in particular, synthetic, hyperspectral and facial images), with a better trade-off between sparsity, spatial coherence and reconstruction error.

Acknowledgment

We would like to thank the reviewers for their insightful feedback that helped us improve the paper significantly. NG acknowledges the support by the F.R.S.-FNRS (incentive grant for scientific research no. F.4501.16) and by the ERC (starting grant no. 679515)

Gabriella Casalino received her Ph.D. degree in Computer Science in 2015, from University of Bari (Italy), the topic of her thesis being “Non-negative factorization methods for extracting semantically relevant features in Intelligent Data Analysis”. From University of Bari she also received the M.Sc. degree and the B.Sc. degree in Computer Science in 2013 and 2008, respectively. She is currently a postdoctoral research fellow at the Department of Pharmaceutical Sciences, University of Bari,

References (23)

  • N. Gillis et al.

    Accelerated multiplicative updates and hierarchical ALS algorithms for nonnegative matrix factorization

    Neural Comput.

    (2012)
  • Cited by (24)

    View all citing articles on Scopus

    Gabriella Casalino received her Ph.D. degree in Computer Science in 2015, from University of Bari (Italy), the topic of her thesis being “Non-negative factorization methods for extracting semantically relevant features in Intelligent Data Analysis”. From University of Bari she also received the M.Sc. degree and the B.Sc. degree in Computer Science in 2013 and 2008, respectively. She is currently a postdoctoral research fellow at the Department of Pharmaceutical Sciences, University of Bari, working on microarray data analysis. Her research interests include computational intelligence, intelligent data analysis, and non-negative matrix factorization.

    Nicolas Gillis received a master degree and a Ph.D. degree in Applied Mathematics from Université Catholique de Louvain (Belgium) in 2007 and 2011, respectively. He is currently an associate professor at the Department of Mathematics and Operational Research, Faculté Polytechnique, Université de Mons, Belgium. His research interests lie in optimization, numerical linear algebra, machine learning, data mining and hyperspectral imaging.

    View full text