Image Classification by Non-Negative Sparse Coding, Low-Rank and Sparse
Decomposition
Chunjie Zhang
1
, Jing Liu
1
, Qi Tian
2
, Changsheng Xu
1
,Hanqing Lu
1
, Songde Ma
1
1
National Laboratory of Pattern Recognition, Institute of Automation, Chinese Academy of Sciences,
P.O.Box 2728, Beijing, China
2
University of Texas at San Antonio, One UTSA Circle, San Antonio Texas, 78249-USA
{cjzhang, jliu, csxu, luhq}@nlpr.ia.ac.cn, qitian@cs.utsa.edu, masd@most.cn
Abstract
We propose an image classification framework by lever-
aging the non-negative sparse coding, low-rank and sparse
matrix decomposition techniques (LR-Sc
+
SPM). First, we
propose a new non-negative sparse coding along with max
pooling and spatial pyramid matching method (Sc
+
SPM) to
extract local features’ information in order to represent im-
ages, where non-negative sparse coding is used to encode
local features. Max pooling along with spatial pyramid
matching (SPM) is then utilized to get the feature vectors
to represent images. Second, motivated by the observation
that images of the same class often contain correlated (or
common) items and specific (or noisy) items, we propose to
leverage the low-rank and sparse matrix recovery technique
to decompose the feature vectors of images per class into a
low-rank matrix and a sparse error matrix. To incorporate
the common and specific attributes into the image represen-
tation, we still adopt the idea of sparse coding to recode
the Sc
+
SPM representation of each image. In particular,
we collect the columns of the both matrixes as the bases
and use the coding parameters as the updated image repre-
sentation by learning them through the locality-constrained
linear coding (LLC). Finally, linear SVM classifier is lever-
aged for the final classification. Experimental results show
that the proposed method achieves or outperforms the state-
of-the-art results on several benchmarks.
1. Introduction
As a fundamental problem in computer vision, image
classification has attracted a lot of attention in recent years.
Among many image representation models, the bag of vi-
sual words (BoW) model [1] has been widely used by many
researchers [2-4] and shown very good performance. The
BoW model contains mainly two modules: (i) codebook
generation and quantization of features extracted from local
image patches; (ii) histogram based image representation
and prediction. Recently, it has been shown that combining
the two modules with sparse representation is very effective
and can achieve the state-of-the-art performance.
As to the first module of the BoW model, k-means is usu-
ally used to generate codebook and quantize visual descrip-
tors extracted from local image patches by nearest-neighbor
search. A histogram is then computed to represent each im-
age by counting the occurrence number of each visual word
within this image. Recently, Yang et al. [4] developed an
extension by generalizing vector quantization to sparse cod-
ing. By using sparse coding instead of k-means, they are
able to learn the optimal codebook and coding parameters
for local features simultaneously, hence are able to reduce
the quantization loss. Multi-scale max pooling is then used
to get the feature representation of images. However, sparse
coding has no constraints on the sign of coding coefficients.
To satisfy the objective of sparse coding, negative coeffi-
cients are sometimes needed, while large numbers of zero
coefficients are inevitable. Since non-zero components typ-
ically provide useful information, the encoding process with
max pooling will bring the loss in terms of those negative
components, and further degrade the classification perfor-
mance.
Instead of learning sparse representations for local fea-
tures [4], the use of sparse representation for the final classi-
fication has also been widely applied to many visual appli-
cations and can achieve the state-of-the-art performances,
e.g., image restoration [5] and classification tasks [6-11].
These holistically sparse representations on the whole im-
age ensure robustness to occlusions and image corruptions.
Training images are often chosen as the bases for sparse rep-
resentation and test images are then classified by assigning
the class with the lowest reconstruction error. Ideally, a test
image can be reconstructed by the training samples and only
the coefficients of the samples within the same class may be
nonzero. This means a test image can be sufficiently recon-
1673
剩余7页未读，继续阅读
资源评论
Ccery1232014-04-06很有参考价值。
星空liang2014-09-24还可以吧，没有期待的那么好- 魅雪紫梦2013-07-29不错的资源，着重介绍稀疏编码在图像分类的应用
