Learning low-rank and discriminative dictionary for image classification

Dictionary learning plays a crucial role in sparse representation based image classification. In this paper, we propose a novel approach to learn a discriminative dictionary with low-rank regularization on the dictionary. Specifically, we apply Fisher discriminant function to the coding coefficients to make the dictionary more discerning, that is, a small ratio of the within-class scatter to between-class scatter. In practice, noisy information in the training samples will undermine the discriminative ability of the dictionary. Inspired by the recent advances in low-rank matrix recovery theory, we apply low-rank regularization on the dictionary to tackle this problem. The iterative projection method (IPM) and inexact augmented Lagrange multiplier (ALM) algorithm are adopted to solve our objective function. The proposed discriminative dictionary learning with low-rank regularization (D²L²R²) approach is evaluated on four face and digit image datasets in comparison with existing representative dictionary learning and classification algorithms. The experimental results demonstrate the superiority of our approach.     

Learning group-based dictionaries for discriminative image representation

Dictionary learning is a critical issue for achieving discriminative image representation in many computer vision tasks such as object detection and image classification. In this paper, a new algorithm is developed for learning discriminative group-based dictionaries, where the inter-concept (category) visual correlations are leveraged to enhance both the reconstruction quality and the discrimination power of the group-based discriminative dictionaries. A visual concept network is first constructed for determining the groups of visually similar object classes and image concepts automatically. For each group of such visually similar object classes and image concepts, a group-based dictionary is learned for achieving discriminative image representation. A structural learning approach is developed to take advantage of our group-based discriminative dictionaries for classifier training and image classification. The effectiveness and the discrimination power of our group-based discriminative dictionaries have been evaluated on multiple popular visual benchmarks.     

Modeling continuous visual features for semantic image annotation and retrieval

Automatic image annotation has become an important and challenging problem due to the existence of semantic gap. In this paper, we firstly extend probabilistic latent semantic analysis (PLSA) to model continuous quantity. In addition, corresponding Expectation-Maximization (EM) algorithm is derived to determine the model parameters. Furthermore, in order to deal with the data of different modalities in terms of their characteristics, we present a semantic annotation model which employs continuous PLSA and standard PLSA to model visual features and textual words respectively. The model learns the correlation between these two modalities by an asymmetric learning approach and then it can predict semantic annotation precisely for unseen images. Finally, we compare our approach with several state-of-the-art approaches on the Corel5k and Corel30k datasets. The experiment results show that our approach performs more effectively and accurately.     

Effective image annotation for searches using multilevel semantics

There is an increasing need for automatic image annotation tools to enable effective image searching in digital libraries. In this paper, we present a novel probabilistic model for image annotation based on content-based image retrieval techniques and statistical analysis. One key difficulty in applying statistical methods to the annotation of images is that the number of manually labeled images used to train the methods is normally insufficient. Numerous keywords cannot be correctly assigned to appropriate images due to lacking or missing information in the labeled image databases. To deal with this challenging problem, we also propose an enhanced model in which the annotated keywords of a new image are defined in terms of their similarity at different semantic levels, including the image level, keyword level, and concept level. To avoid missing some relevant keywords, the model labels the keywords with the same concepts as the new image. Our experimental results show that the proposed models are effective for annotating images that have different qualities of training data.     

Visual word spatial arrangement for image retrieval and classification

We present word spatial arrangement (WSA), an approach to represent the spatial arrangement of visual words under the bag-of-visual-words model. It lies in a simple idea which encodes the relative position of visual words by splitting the image space into quadrants using each detected point as origin. WSA generates compact feature vectors and is flexible for being used for image retrieval and classification, for working with hard or soft assignment, requiring no pre/post processing for spatial verification. Experiments in the retrieval scenario show the superiority of WSA in relation to Spatial Pyramids. Experiments in the classification scenario show a reasonable compromise between those methods, with Spatial Pyramids generating larger feature vectors, while WSA provides adequate performance with much more compact features. As WSA encodes only the spatial information of visual words and not their frequency of occurrence, the results indicate the importance of such information for visual categorization.     

A probabilistic semantic model for image annotation and multi-modal image retrieval

This paper addresses automatic image annotation problem and its application to multi-modal image retrieval. The contribution of our work is three-fold. (1) We propose a probabilistic semantic model in which the visual features and the textual words are connected via a hidden layer which constitutes the semantic concepts to be discovered to explicitly exploit the synergy among the modalities. (2) The association of visual features and textual words is determined in a Bayesian framework such that the confidence of the association can be provided. (3) Extensive evaluation on a large-scale, visually and semantically diverse image collection crawled from Web is reported to evaluate the prototype system based on the model. In the proposed probabilistic model, a hidden concept layer which connects the visual feature and the word layer is discovered by fitting a generative model to the training image and annotation words through an Expectation-Maximization (EM) based iterative learning procedure. The evaluation of the prototype system on 17,000 images and 7736 automatically extracted annotation words from crawled Web pages for multi-modal image retrieval has indicated that the proposed semantic model and the developed Bayesian framework are superior to a state-of-the-art peer system in the literature.     

Recognition of attentive objects with a concept association network for image annotation

With the advancement of imaging techniques and IT technologies, image retrieval has become a bottle neck. The key for efficient and effective image retrieval is by a text-based approach in which automatic image annotation is a critical task. As an important issue, the metadata of the annotation, i.e., the basic unit of an image to be labeled, has not been fully studied. A habitual way is to label the segments which are produced by a segmentation algorithm. However, after a segmentation process an object has often been broken into pieces, which not only produces noise for annotation but also increases the complexity of the model. We adopt an attention-driven image interpretation method to extract attentive objects from an over-segmented image and use the attentive objects for annotation. By such doing, the basic unit of annotation has been upgraded from segments to attentive objects. Visual classifiers are trained and a concept association network (CAN) is constructed for object recognition. A CAN consists of a number of concept nodes in which each node is a trained neural network (visual classifier) to recognize a single object. The nodes are connected through their correlation links forming a network. Given that an image contains several unknown attentive objects, all the nodes in CAN generate their own responses which propagate to other nodes through the network simultaneously. For a combination of nodes under investigation, these loopy propagations can be characterized by a linear system. The response of a combination of nodes can be obtained by solving the linear system. Therefore, the annotation problem is converted into finding out the node combination with the maximum response. Annotation experiments show a better accuracy of attentive objects over segments and that the concept association network improves annotation performance.     

Learning a hierarchical image manifold for Web image classification

Image classification is an essential task in content-based image retrieval.However,due to the semantic gap between low-level visual features and high-level semantic concepts,and the diversification of Web images,the performance of traditional classification approaches is far from users’ expectations.In an attempt to reduce the semantic gap and satisfy the urgent requirements for dimensionality reduction,high-quality retrieval results,and batch-based processing,we propose a hierarchical image manifold with novel distance measures for calculation.Assuming that the images in an image set describe the same or similar object but have various scenes,we formulate two kinds of manifolds,object manifold and scene manifold,at different levels of semantic granularity.Object manifold is developed for object-level classification using an algorithm named extended locally linear embedding(ELLE) based on intra-and inter-object difference measures.Scene manifold is built for scene-level classification using an algorithm named locally linear submanifold extraction(LLSE) by combining linear perturbation and region growing.Experimental results show that our method is effective in improving the performance of classifying Web images.     

Incorporating multiple SVMs for automatic image annotation

In this paper, a novel automatic image annotation system is proposed, which integrates two sets of support vector machines (SVMs), namely the multiple instance learning (MIL)-based and global-feature-based SVMs, for annotation. The MIL-based bag features are obtained by applying MIL on the image blocks, where the enhanced diversity density (DD) algorithm and a faster searching algorithm are applied to improve the efficiency and accuracy. They are further input to a set of SVMs for finding the optimum hyperplanes to annotate training images. Similarly, global color and texture features, including color histogram and modified edge histogram, are fed into another set of SVMs for categorizing training images. Consequently, two sets of image features are constructed for each test image and are, respectively, sent to the two sets of SVMs, whose outputs are incorporated by an automatic weight estimation method to obtain the final annotation results. Our proposed annotation approach demonstrates a promising performance for an image database of 12 000 general-purpose images from COREL, as compared with some current peer systems in the literature.     

Photograph retrieval and classification by visual keywords and thesaurus

As we collect more digital images with the advent of digital cameras, we need effective content-based search and categorization functions. In this paper, we propose a novel notion of visual keywords to describe and compare digital visual contents. Visual keywords are visual prototypes extracted from a visual content domain with semantics labels. They can be further abstracted to form visual thesaurus. An image is indexed as a spatial distribution of visual keywords. Both retrieval and classification evaluation tasks on professional natural scene photographs have demonstrated the usefulness of this new methodology. Joo-Hwee Lim: He received his B.S. (Hons I) and M.S. degrees in Computer Science from the National University of Singapore in 1989 and 1991 respectively. He has joined Kent Ridge Digital Labs (KRDL), Singapore since Oct 1990 and is currently an associate research staff of KRDL. He has published widely in his areas of research interests which include content-based processing, pattern recognition, and neural networks.     

Support vector description of clusters for content-based image annotation

Continual progress in the fields of computer vision and machine learning has provided opportunities to develop automatic tools for tagging images; this facilitates searching and retrieving. However, due to the complexity of real-world image systems, effective and efficient image annotation is still a challenging problem. In this paper, we present an annotation technique based on the use of image content and word correlations. Clusters of images with manually tagged words are used as training instances. Images within each cluster are modeled using a kernel method, in which the image vectors are mapped to a higher-dimensional space and the vectors identified as support vectors are used to describe the cluster. To measure the extent of the association between an image and a model described by support vectors, the distance from the image to the model is computed. A closer distance indicates a stronger association. Moreover, word-to-word correlations are also considered in the annotation framework. To tag an image, the system predicts the annotation words by using the distances from the image to the models and the word-to-word correlations in a unified probabilistic framework. Simulated experiments were conducted on three benchmark image data sets. The results demonstrate the performance of the proposed technique, and compare it to the performance of other recently reported techniques.     

Evaluation for uncertain image classification and segmentation

Each year, numerous segmentation and classification algorithms are invented or reused to solve problems where machine vision is needed. Generally, the efficiency of these algorithms is compared against the results given by one or many human experts. However, in many situations, the location of the real boundaries of the objects as well as their classes are not known with certainty by the human experts. Furthermore, only one aspect of the segmentation and classification problem is generally evaluated. In this paper we present a new evaluation method for classification and segmentation of image, where we take into account both the classification and segmentation results as well as the level of certainty given by the experts. As a concrete example of our method, we evaluate an automatic seabed characterization algorithm based on sonar images.     

Hidden annotation for image retrieval with long-term relevance feedback learning

Hidden annotation (HA) is an important research issue in content-based image retrieval (CBIR). We propose to incorporate long-term relevance feedback (LRF) with HA to increase both efficiency and retrieval accuracy of CBIR systems. The work contains two parts. (1) Through LRF, a multi-layer semantic representation is built to automatically extract hidden semantic concepts underlying images. HA with these concepts alleviates the burden of manual annotation and avoids the ambiguity problem of keyword-based annotation. (2) For each learned concept, semi-supervised learning is incorporated to automatically select a small number of candidate images for annotators to annotate, which improves efficiency of HA.     

Combining global, regional and contextual features for automatic image annotation

This paper presents a novel approach to automatic image annotation which combines global, regional, and contextual features by an extended cross-media relevance model. Unlike typical image annotation methods which use either global or regional features exclusively, as well as neglect the textual context information among the annotated words, the proposed approach incorporates the three kinds of information which are helpful to describe image semantics to annotate images by estimating their joint probability. Specifically, we describe the global features as a distribution vector of visual topics and model the textual context as a multinomial distribution. The global features provide the global distribution of visual topics over an image, while the textual context relaxes the assumption of mutual independence among annotated words which is commonly adopted in most existing methods. Both the global features and textual context are learned by a probability latent semantic analysis approach from the training data. The experiments over 5k Corel images have shown that combining these three kinds of information is beneficial in image annotation.     

A general framework for image fusion based on multi-scale transform and sparse representation

In image fusion literature, multi-scale transform (MST) and sparse representation (SR) are two most widely used signal/image representation theories. This paper presents a general image fusion framework by combining MST and SR to simultaneously overcome the inherent defects of both the MST- and SR-based fusion methods. In our fusion framework, the MST is firstly performed on each of the pre-registered source images to obtain their low-pass and high-pass coefficients. Then, the low-pass bands are merged with a SR-based fusion approach while the high-pass bands are fused using the absolute values of coefficients as activity level measurement. The fused image is finally obtained by performing the inverse MST on the merged coefficients. The advantages of the proposed fusion framework over individual MST- or SR-based method are first exhibited in detail from a theoretical point of view, and then experimentally verified with multi-focus, visible-infrared and medical image fusion. In particular, six popular multi-scale transforms, which are Laplacian pyramid (LP), ratio of low-pass pyramid (RP), discrete wavelet transform (DWT), dual-tree complex wavelet transform (DTCWT), curvelet transform (CVT) and nonsubsampled contourlet transform (NSCT), with different decomposition levels ranging from one to four are tested in our experiments. By comparing the fused results subjectively and objectively, we give the best-performed fusion method under the proposed framework for each category of image fusion. The effect of the sliding window’s step length is also investigated. Furthermore, experimental results demonstrate that the proposed fusion framework can obtain state-of-the-art performance, especially for the fusion of multimodal images.     

Learning extremely shared middle-level image representation for scene classification

Learning middle-level image representations is very important for the computer vision community, especially for scene classification tasks. Middle-level image representations currently available are not sparse enough to make training and testing times compatible with the increasing number of classes that users want to recognize. In this work, we propose a middle-level image representation based on the pattern that extremely shared among different classes to reduce both training and test time. The proposed learning algorithm first finds some class-specified patterns and then utilizes the lasso regularization to select the most discriminative patterns shared among different classes. The experimental results on some widely used scene classification benchmarks (15 Scenes, MIT-indoor 67, SUN 397) show that the fewest patterns are necessary to achieve very remarkable performance with reduced computation time.     

Joint learning and weighting of visual vocabulary for bag-of-feature based tissue classification

Automated classification of tissue types of Region of Interest (ROI) in medical images has been an important application in Computer-Aided Diagnosis (CAD). Recently, bag-of-feature methods which treat each ROI as a set of local features have shown their power in this field. Two important issues of bag-of-feature strategy for tissue classification are investigated in this paper: the visual vocabulary learning and weighting, which are always considered independently in traditional methods by neglecting the inner relationship between the visual words and their weights. To overcome this problem, we develop a novel algorithm, Joint-ViVo, which learns the vocabulary and visual word weights jointly. A unified objective function based on large margin is defined for learning of both visual vocabulary and visual word weights, and optimized alternately in the iterative algorithm. We test our algorithm on three tissue classification tasks: classifying breast tissue density in mammograms, classifying lung tissue in High-Resolution Computed Tomography (HRCT) images, and identifying brain tissue type in Magnetic Resonance Imaging (MRI). The results show that Joint-ViVo outperforms the state-of-art methods on tissue classification problems.     

Maximum Margin Multisurface Support Tensor Machines with application to image classification and segmentation

Virtually all previous classifier models take vectors as inputs, performing directly based on the vector patterns. But it is highly necessary to consider images as matrices in real applications. In this paper, we represent images as second order tensors or matrices. We then propose two novel tensor algorithms, which are referred to as Maximum Margin Multisurface Proximal Support Tensor Machine (M³PSTM) and Maximum Margin Multi-weight Vector Projection Support Tensor Machine (M³VSTM), for classifying and segmenting the images. M³PSTM and M³VSTM operate in tensor space and aim at computing two proximal tensor planes for multisurface learning. To avoid the singularity problem, maximum margin criterion is used for formulating the optimization problems. Thus the proposed tensor classifiers have an analytic form of projection axes and can achieve the maximum margin representations for classification. With tensor representation, the number of estimated parameters is significantly reduced, which makes M³PSTM and M³VSTM more computationally efficient when handing the high-dimensional datasets than applying the vector representations based methods. Thorough image classification and segmentation simulations on the benchmark UCI and real datasets verify the efficiency and validity of our approaches. The visual and numerical results show M³PSTM and M³VSTM deliver comparable or even better performance than some state-of-the-art classification algorithms.