Multimed Tools Appl
of recognition performance. One solution is to use the class specific dictionary learning
(CSDL) algorithm, which has greatly improved the classification accuracy. However, the
CSDL algorithm fails to consider the constraints to sparse codes. In particular, it cannot
guarantee that the sparse codes in the same class will be concentrated based on the learned
dictionary for each class. Such concentration is actually beneficial to classification. To
address these limitations, in this paper, we propose a class specific centralized dictionary
learning (CSCDL) algorithm to simultaneously consider the desired characteristics for both
dictionary and sparse codes. The blockwise coordinate descent algorithm and Lagrange
multipliers are used to optimize the corresponding objective function. Extensive experimen-
tal results on face recognition benchmark datasets demonstrate the superior performance of
our CSCDL algorithm compared with conventional approaches.
Keywords Centralized dictionary learning · Face recognition · Class specific
1 Introduction
After decades of effort, the power of dictionary learning for sparse representation has
been gradually revealed in visual computation areas, such as image annotation [25], image
inpainting [30], image classification [14, 16], face recognition [32], object detection [28],
transfer learning [29], and image denoising [6], due to its impressive performance. Differ-
ent from traditional decomposition frameworks like principal component analysis (PCA),
non-negative matrix factorization [31], and low-rank factorization [27], sparse representa-
tion allows coding under over-complete bases (that is, the number of bases is greater than
the dimension of input data), and thus generates sparse codes capable of representing the
data more adaptively.
Face recognition, one of the successful applications of sparse representation, is a classi-
cal yet challenging research topic in computer vision and pattern recognition [38]. Effective
face recognition usually involves two stages: 1) feature extraction, 2) classifier construc-
tion and label prediction. For the first stage, Eigenfaces were proposed by performing
PCA [24]. Laplacianfaces were proposed to preserve local information [9]. Fisherfaces
were suggested to maximize the ratio of between-class scatter to within-class scatter [1].
Yan et al. [33] proposed a multi-subregion based correlation filter bank algorithm to extract
both the global-based and local-based face features. For the latter stage, a nearest neighbor
method was proposed to predict the label of a test image using its nearest neighbors in the
training samples [5]. Nearest subspace methods were proposed to assign the label of a test
image by comparing its reconstruction error for each category [10, 22].
Under the nearest subspace framework, a sparse representation based classification
(SRC) system was proposed and achieved impressive performance [32]. Given a test sam-
ple, the sparse representation technique represents it as a sparse linear combination of the
training samples. The predicted label is determined by the residual error from each class. To
analyze SRC, collaborative representation based classification (CRC) was proposed as an
alternative approach [36]. CRC represents a test sample as the linear combination of almost
all the training samples. An interesting observation in [36] is that it is the collaborative rep-
resentation rather than the sparse representation that makes the nearest subspace method
powerful for classification.
Despite their promise, both SRC and CRC algorithms directly use the training sam-
ples as the dictionary for each class. By contrast, a well learned dictionary, especially by
