Elsevier

Pattern Recognition

Volume 39, Issue 4, April 2006, Pages 717-720
Pattern Recognition

Rapid and brief communication
An algorithm for semi-supervised learning in image retrieval

https://doi.org/10.1016/j.patcog.2005.11.009Get rights and content

Abstract

We study the problem of image retrieval based on semi-supervised learning. Semi-supervised learning has attracted a lot of attention in recent years. Different from traditional supervised learning. Semi-supervised learning makes use of both labeled and unlabeled data. In image retrieval, collecting labeled examples costs human efforts, while vast amounts of unlabeled data are often readily available and offer some additional information. In this paper, based on support vector machine (SVM), we introduce a semi-supervised learning method for image retrieval. The basic consideration of the method is that, if two data points are close to each, they should share the same label. Therefore, it is reasonable to search a projection with maximal margin and locality preserving property. We compare our method to standard SVM and transductive SVM. Experimental results show efficiency and effectiveness of our method.

Introduction

Due to the rapid growth in the volume of digit images, there is an increasing demand for effective image management tools. Consequently, content-based image retrieval (CBIR) is receiving widespread research interest [1].

In recent years, much research has been done to semi-supervised learning. Different from traditional supervised learning which only makes use of labeled data; semi-supervised learning makes use of both labeled and unlabeled data. Generally, the unlabeled data can be used to better describe the intrinsic geometrical structure of the data space, and hence improve the classification performance. Also, most previous learning algorithms only consider the Euclidean structure of the data space. However, in many cases, the objects of interest might reside on a low-dimensional manifold, which is nonlinearly embedded in the ambient space (data space). In such cases, the nonlinear manifold structure is much more important than the Euclidean structure. Specifically, the similarity between objects should be described by the geodesic distance rather than the Euclidean distance.

In this paper, we introduce a new algorithm for image retrieval. Our algorithm is intrinsically based on support vector machines (SVM) and locality preserving projections (LPP). LPP is a recently proposed algorithm for linear dimensionality reduction. We first build a nearest-neighbor graph over all the data points (labeled and unlabeled) which model the local geometrical structure of the image space. By combining SVM and LPP, we can obtain a classifier which maximizes the margin and simultaneously preserves the local information. In many information retrieval tasks, the local information is much more reliable than the global structure. Also, our method explicitly considers the manifold structure of the data space. It would be important to note that Euclidean space is just a special manifold. As a result of these properties, our method can produce better results.

The rest of this paper is organized as follows. Section 2 gives a brief description of SVM and LPP. In Section 3, we present a new semi-supervised method for image retrieval. The experimental results are shown in Section 4. Finally, we give conclusions in Section 5.

Section snippets

A brief review of SVM and LPP

In this section, we give a brief review of SVM [2] and LPP [3].

Semi-supervised induction

Semi-supervised learning has received a lot of attentions in recent years. So far most of the efforts have been invested in a transductive setting that predicts only for observed inputs. Yet, in many applications there is a clear need for inductive learning, for example, in image retrieval or document classification. Unfortunately, most existing semi-supervised learners do not readily generalize to new test data. A brute force approach is to incorporate the new test points and re-estimate the

Experimental design

We performed several experiments to evaluate the effectiveness of the proposed approaches over a large image database. The database we use consists of 10,000 images of 79 semantic categories selected from the Corel Image Gallery. Three types of color features and three types of texture features are used in our system. The dimension of the feature space is 435. We designed an automatic feedback scheme to simulate the retrieval process conducted by real users. At each iteration, the system marks

Conclusions

In this paper, we introduce a new semi-supervised learning algorithm, which combines the support vector machines and locality preserving projections. The new algorithm maximizes the margin between two classes and simultaneously preserves the local information. The new method explicitly considers the manifold structure. We have applied the new method to image retrieval. Experimental results show that the unlabeled can be used to enhance the retrieval performance.

There are still several questions

About the Author—KE LU received his B.S. degree in thermal power engineering from Chongqing University, China in 1996, and his M.S. degree in Computer Engineering in 2003 from the University of Electronic Science and Technology of China where he is currently a lecturer. His research interests include pattern recognition and multimedia information retrieval.

References (4)

  • X. He et al.

    Learning a semantic space from user's relevance feedback for image retrieval

    IEEE Trans. Circuits Systems Video Technol.

    (2003)
  • V. Vapnik, The Nature of Statistical Learning Theory, Springer, New York, 1995, pp....
There are more references available in the full text version of this article.

Cited by (29)

  • Estimation of wind turbine power coefficient by adaptive neuro-fuzzy methodology

    2017, Neurocomputing
    Citation Excerpt :

    Wind speed increases with height, higher hub height leads to increased and smooth extraction of wind energy [2]. To optimize different nonlinear systems, various control techniques have been used such as fuzzy logic [20,21], artificial neural network [22,23], neuro-fuzzy [24,2] and support vector machines (SVMs) [25–27]. The relationship between power coefficient, TSR and pitch angle is totally nonlinear, so there is a need to apply some powerful soft computing techniques to realize the nonlinear relationships.

  • Orthogonal optimal reverse prediction for semi-supervised learning

    2016, Pattern Recognition
    Citation Excerpt :

    SVMs combining other algorithms such as locality preserving projection [8], called semi-supervised induction, have been proposed for images retrieval tasks. Experimental results have shown its efficiency and effectiveness in [9]. LapSVM and LapRLS [10] are two important regularizer based semi-supervised learning methods.

  • Analyzing of flexible gripper by computational intelligence approach

    2016, Mechatronics
    Citation Excerpt :

    The signals from embedded sensors were used for estimation of the grasping object shapes and therefore for optimal input displacement prediction. Aiming at ensuring optimal functioning of the gripper, many soft computing techniques are used today such as the fuzzy logic (FL) [47,48], ANN [49], neuro-fuzzy [50,51] and support vector machines (SVM) [52,53]. The basic idea behind the soft computing methodology is to collect input/output data pairs and to learn the proposed network from these data.

  • Dynamic label propagation for semi-supervised multi-class multi-label classification

    2016, Pattern Recognition
    Citation Excerpt :

    Label propagation (LP) [58] specifically assumes that nodes connected by edges of large similarity tend to have the same label through information propagated within the graph. A wide range of applications such as classification [33], clustering [18,7], dimension reduction [40] and retrieval [31,11,32] have adopted the label propagation strategy. Another type of semi-supervised learning, co-training [9], utilizes multi-modality features to help each other by pulling out unlabeled data to re-train and enhance the classifiers.

View all citing articles on Scopus

About the Author—KE LU received his B.S. degree in thermal power engineering from Chongqing University, China in 1996, and his M.S. degree in Computer Engineering in 2003 from the University of Electronic Science and Technology of China where he is currently a lecturer. His research interests include pattern recognition and multimedia information retrieval.

About the Author—JIDONG ZHAO received his B.S. degree in 1998, and his M.S. degree in Computer Engineering in 2003 from the University of Electronic Science and Technology of China where he is currently a lecturer. His research interests include computer network and multimedia information retrieval.

View full text