Elsevier

Neurocomputing

Volume 140, 22 September 2014, Pages 77-83
Neurocomputing

A novel supervised feature extraction and classification fusion algorithm for land cover recognition of the off-land scenario

https://doi.org/10.1016/j.neucom.2014.03.034Get rights and content

Abstract

In this paper, a novel supervised feature extraction and classification fusion algorithm based on neighborhood preserving embedding (NPE) and sparse representation is proposed. Specifically, an optimal dictionary is adaptively learned to bate the trivial information of the original training data; then, in order to obtain the sparse representation coefficients, a sparse preserving embedding map is sought to reduce the dimensionality of high-dimensional data, and the test data is classified by the corresponding sparse representation coefficients. Finally, the novel supervised fusion algorithm is applied to the land cover recognition of the off-land scenario. Experimental results show that the proposed method leads to promising results in fusing feature extraction and classification.

Introduction

Techniques for dimensionality reduction in unsupervised and supervised learning tasks have attracted much attention in computer vision, machine learning and biometrics. Among them, subspace learning and manifold learning methods have been dominantly and successfully used in dimensionality reduction for high-dimensional data. Principal component analysis (PCA) [1] and linear discriminant analysis (LDA) [2] are two most popular linear subspace learning methods. In the past, many LDA extensions have been developed to deal with the small sample size problem, but they fail to realize the essential data structures nonlinearly embedded in high-dimensional space. In order to overcome this limitation, some known manifold learning methods are presented such as neighborhood preserving embedding (NPE) [3], locality preserving projection (LPP) [4], local linear embedding (LLE) [5], local Fisher discriminant analysis (LFDA) [6], Laplacianfaces [7] and unsupervised discriminant projection (UDP) [8], marginal Fisher analysis (MFA) [9], linear discriminant projection (LDP) [10], graph-optimized locality preserving projections (GoLPP) [11] graph-based Fisher analysis [12], hypergraph analysis approach [13], [14] and optimized multigraph-based semi-supervised learning (OMG-SSL) [15].

In recent years, sparse representation (or sparse coding) has been attracting a lot of attention due to its great success in image processing, and it has been used for face recognition and texture classification. Transform-invariant sparse representation [16] recovers the sparse representation of a target image and the image plane transformation between the target and the model images simultaneously. Wright et al. [17], [18] presented a sparse representation-based classification (SRC) method and successfully applied it to recognize human faces with varying lighting condition, occlusion and disguise. Yang et al. [19] applied the sparse representation for face recognition with occlusion based on the Gabor feature. Meanwhile, some known feature extraction algorithms based on sparse learning are proposed, sparse principal component analysis [20] uses Lasso (elastic net) to produce modified principal components with sparse learning; sparse projection [21] based on a graph embedding model learns a set of sparse basis function by applying regularized regression; sparse preserving projection [22], [23] aims to preserve the sparse reconstructive relationship of the data by minimizing a regularization-related objection function; Yang and Chu [24] used the decision rule of SRC to steer the design of a dimensionality reduction method, i.e. the sparse representation classifier steered discriminative projection (SRC-DP); Zhang et al. [25] proposed a novel linear subspace learning approach via sparse coding.

Although the above methods have shown success in classification and feature extraction, there are some limitations as follows:

  • 1.

    The feature extraction and classification are separate. In [1], [2], [3], [4], [5], [6], [7], [8], [9], [10], [11], [12], the feature extraction algorithms are firstly used to reduce the dimensionality of high-dimensional data, than classifiers are applied to measure the performance of the feature extraction criterions, similarly, the classifiers are applied to classify the dimension reduced data [13], [14], [15], [16], [17], [18], [19] independently. Therefore these methods are implemented through two stage: feature extraction stage and classification stage.

  • 2.

    SRC cannot be applied to high-dimensional data directly. For high-dimensional data, the sparse representation coefficients are hardly obtained since the dictionary is not over-complete. Therefore, in order to obtain sparse representation coefficients, the dimensionality of data is pre-reduced by random matrix [17], [18], [19] and PCA [24], [25].

  • 3.

    The feature extraction criterions proposed in [20], [21] and [24], [25] cannot extract features directly. In [20], [21], a new feature extraction criterion is proposed under the sparse constraint of the projection vectors. In [24], [25], the scatter matrices of the samples are redefined based on the results of SRC. In order to obtain the sparse projection vectors and the representation coefficients, PCA is used to preprocess data, rather than extract features directly, so these criterions are equivalent to two stage feature extraction.

  • 4.

    The entire training samples used as dictionary may affect the performance of sparse representation. In [22], [23], [24], [25], [26], the original image samples are used to represent the input data, actually the original training samples have much redundancy as well as noise and trivial information that can be negative to the recognition. In addition, if the training samples are huge, the computation of the sparse representation will be time consuming, it is needed a more compact and robust dictionary such that each sample in the test set can be represented as a sparse linear combination of its atoms.

In order to overcome the above limitations, in this paper, a novel fusion algorithm, namely feature extraction and classification fusion algorithm (FECFA), is developed to implement feature extraction and classification simultaneously. More specifically, an optimal over-complete dictionary is adaptively learned from the original training data to represent the test data, and the learned optimal dictionary may bate the redundancy as well as noise and trivial information of the original training data. Meanwhile, in order to obtain the sparse representation coefficients, a sparse preserving embedding map is sought to reduce the dimensionality of the test data. Lastly, the test data can be classified by the sparse representation coefficients. For FECFA, need of special note is that the sparse preserving embedding map is learned based on the classification criterions, and the sparse preserving embedding map and the sparse representation coefficients can be obtained by solving an optimization problem alternately. In contrast to the state-of-the-art feature extraction and classification methods, FECFA has the following advantages:

  • 1.

    FECFA can reduce the dimensionality of a query sample and classify it simultaneously. Therefore, the features of the high-dimensional data need not be pre-processed before classification, and the feature extraction method need not classifier to measure its performance.

  • 2.

    The sparse preserving embedding map is learned from test data and training data, rather than from training data only. So FECFA utilizes the prior knowledge of test data, which may be practical in the real application. Furthermore, the high-dimensional data pre-processed by the sparse preserving embedding map will improve positive recognition of classification because the sparse preserving embedding map is learned based on the classification criterions.

  • 3.

    Through sparse preserving, we need not decide how many k-nearest neighbors to be selected to reconstruct the samples such as in [5], [6], [7], [8], [9], [10], [11], so FECFA is more adaptive.

  • 4.

    The test data is represented by the learned optimal dictionary. In contrast to the original sample, the learned dictionary may bate the redundancy as well as noise and trivial information of the original training data.

The rest of the paper is organized as follows. We review the related work in Section 2. In Section 3, a novel supervised fusion algorithm for feature extraction and classification is proposed. Experiments are presented in Section 4. Conclusions are summarized in Section 5.

Section snippets

Brief review of the related work

In this section, we introduce the basic idea of the Neighborhood Preserving Embedding (NPE), sparse representation-based classifier and k-SVD dictionary learning algorithm.

The novel supervised feature extraction and classification fusion algorithm

In this section, we introduce the basic ideas of FECFA for feature extraction and classification. Specifically, an optimal dictionary is learned from the training data set to represent the test data; then a sparse preserving embedding map and sparse coefficients are optimized alternately. From the sparse preserving embedding map and sparse coefficients, we can reduce the dimensionality of the test data and classify it simultaneously. Furthermore, we generalize FECFA to the incremental test data

Experiments

In this section, we systematically apply FECFA on two land cover databases. The first land cover database created in 2012 by the Nanjing University of Science and Technology (NJUST), which is composed of 6 classes of land cover (such as dirt roads, sandy roads, tree, vegetation (green), water and vegetation (yellow)), contains 12,000 cropped images with 16×16 pixels derived from six different road condition video files. The second land cover database derived from the Outex Texture Database,

Conclusion

In this paper, we present a novel supervised feature extraction and classification fusion algorithm based on dictionary selection. According to Section 4.1, we know that the size of dictionary affects the positive recognition rate of the proposed fusion algorithm and the learned dictionary may bate the redundancy as well as noise and trivial information. According to Section 4.2, for the all cropped images databases, the dimensionality of the original data can be reduced by the proposed fusion

Acknowledgments

This work is supported by the National Science Foundation of China (Grant no. 61233011, 61373063, 613750071, 91220301, 61125305 and 61005005), the National Science Fund for Distinguished Young Scholars (Grant no. 61125305) and the Doctoral Candidate Creative Fund of Jiangsu province (Grant no. CXZZ12_0207).

Yan Cui was born in Shandong, China, in 1985. She received her B.S. and M.S. degrees from Liaocheng University, Liaocheng, China, in 2008 and 2011, respectively. Now she is studying for her Ph.D. degree in pattern recognition and intelligence system from School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, China. She visited the Department of Electrical and Computer Engineering, University of Miami, USA, from May 2013 to November 2013; she is

References (32)

  • X.F. He et al.

    Face recognition using Laplacianfaces

    IEEE Trans. Pattern Anal. Mach. Intell.

    (2005)
  • J. Yang et al.

    Globally maximizing, locally minimizingunsupervised discriminant projection with applications to face and palm biometrics

    IEEE Trans. Pattern Anal. Mach. Intell.

    (2007)
  • Sh.Ch. Yan et al.

    Graph embedding and extensionsa general framework for dimensionality reduction

    IEEE Trans. Pattern Anal. Mach. Intell.

    (2007)
  • H.P. Cai et al.

    Learning linear discriminant projections for dimensionality reduction of image descriptors

    IEEE Trans. Pattern Anal. Mach. Intell.

    (2011)
  • Y. Gao et al.

    3-D object retrieval and recognition with hypergraph analysis

    IEEE Trans. Image Process., Hypergraph Anal.

    (2012)
  • Y. Gao et al.

    Visual-textual joint relevance learning for tag-based social image search

    IEEE Trans. Image Process.

    (2013)
  • Cited by (4)

    • A KD curvature based corner detector

      2016, Neurocomputing
      Citation Excerpt :

      Image feature detection and matching are the basic problems in computer vision and image processing [1–3]. It has been applied to many vision tasks including scene analysis, 3D matching, robot navigation, panoramic image mosaic, feature extraction and target tracking [4–7]. Corners [8] are the most interesting features extracted from visual images, which are ubiquitous in the current big data era.

    • Robust corner detection using linear fitting error estimation

      2020, ACM International Conference Proceeding Series
    • A Multiple Feature-Based Image-Switching Strategy in Visual Sensor Networks

      2015, International Journal of Distributed Sensor Networks

    Yan Cui was born in Shandong, China, in 1985. She received her B.S. and M.S. degrees from Liaocheng University, Liaocheng, China, in 2008 and 2011, respectively. Now she is studying for her Ph.D. degree in pattern recognition and intelligence system from School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, China. She visited the Department of Electrical and Computer Engineering, University of Miami, USA, from May 2013 to November 2013; she is currently working as a research assistant at the Institute of Textiles and Clothing, Hong Kong Polytechnic University. Her current research interests include pattern recognition, machine learning and image retrieval.

    Zhong Jin received the B.S. degree in Mathematics, the M.S. degree in Applied Mathematics and the Ph.D. degree in pattern recognition and intelligence system from Nanjing university of Science and Technology (NUST), Nanjing, China, in 1982, 1984, and 1999, respectively. He is a professor in the School of Computer Science and Engineering, NUST. He had a stay of 15 months as a research assistant at the Department of Computer Science and engineering, the Chinese University of Hong Kong from 2000 to 2001. He visited the Laboratoire HEUDIASYC, Universite de Technologie de Compiegne, France, from October 2001 to July 2002. He visited the Centre de Visio per Computador, Universitat Autonoma de Barcelona, Spain, as the Ramon y Cajal program Research Fellow from September 2005 to October 2005. His current interests are in the areas of pattern recognition, computer vision, face recognition, facial expression analysis and content-based image retrieval.

    Jielin Jiang received the B.S. degree from Huainan Normal University, Huainan, China in 2007, the M.S. degree from Inner Mongolia University of Technology, Hohhot, China in 2010. He is currently pursuing the Ph.D. degree in pattern recognition and intelligence system from School of Computer Science and Engineering, Nanjing University of Science and Technology, Nanjing, China. From February 2013 to August 2013, he was an Exchange Student with the Department of Computing, the Hong Kong Polytechnic University, Hong Kong. His current research interests include image denoising and image classification.

    View full text