Lin Feng

Global Correlation Descriptor

Global Correlation Descriptor (GCD) is proposed to represent image information.Global Correlation Vector (GCV) characterizes the color feature.Directional Global Correlation Vector (DGCV) characterizes the texture feature.GCD obtains superior performance in CBIR. The image descriptors based on multi-features fusion have better performance than that based on simple feature in content-based image retrieval (CBIR). However, these methods still have some limitations: (1) the methods that define directly texture in color space put more emphasis on color than texture feature; (2) traditional descriptors based on histogram statistics disregard the spatial correlation between structure elements; (3) the descriptors based on structure element correlation (SEC) disregard the occurring probability of structure elements. To solve these problems, we propose a novel image descriptor, called Global Correlation Descriptor (GCD), to extract color and texture feature respectively so that these features have the same effect in CBIR. In addition, we propose Global Correlation Vector (GCV) and Directional Global Correlation Vector (DGCV) which can integrate the advantages of histogram statistics and SEC to characterize color and texture features respectively. Experimental results demonstrate that GCD is more robust and discriminative than other image descriptors in CBIR.

Scatter Balance: An Angle-Based Supervised Dimensionality Reduction

Subspace selection is widely applied in data classification, clustering, and visualization. The samples projected into subspace can be processed efficiently. In this paper, we research the linear discriminant analysis (LDA) and maximum margin criterion (MMC) algorithms intensively and analyze the effects of scatters to subspace selection. Meanwhile, we point out the boundaries of scatters in LDA and MMC algorithms to illustrate the differences and similarities of subspace selection in different circumstances. Besides, the effects of outlier classes on subspace selection are also analyzed. According to the above analysis, we propose a new subspace selection method called angle linear discriminant embedding (ALDE) on the basis of angle measurement. ALDE utilizes the cosine of the angle to get new within-class and between-class scatter matrices and avoids the small sample size problem simultaneously. To deal with high-dimensional data, we extend ALDE to a two-stage ALDE (TS-ALDE). The synthetic data experiments indicate that ALDE can balance the within-class and between-class scatters and be robust to outlier classes. The experimental results based on UCI machine-learning repository and image databases show that TS-ALDE has a lower time complexity than ALDE while processing high-dimensional data.

Robust activation function and its application: Semi-supervised kernel extreme learning method

Semi-supervised learning is a hot topic in the field of pattern recognition, this paper analyzes an effective classification algorithm - Extreme Learning Machine (ELM). ELM has been widely used in the applications of pattern recognition and data mining for its extremely fast training speed and highly recognition rate. But in most of real-world applications, there are irregular distributions and outlier problems which lower the classification rate of ELM (kernel ELM). This is mainly because: (1) Overfitting caused by outliers and unreasonable selections of activation function and kernel function and (2) the labeled sample size is small and we do not making full use of the information of unlabeled data either. Against problem one, this paper proposes a robust activation function (RAF) based on analyzing several different activation functions in-depth. RAF keeps the output of activation function away from zero as much as possible and minimizes the impacts of outliers to the algorithm. Thus, it improves the performance of ELM (kernel ELM); simultaneously, RAF can be applied to other kernel methods and a neural network learning algorithm. Against problem two, we propose a semi-supervised kernel ELM (SK-ELM). Experimental results on synthetic and real-world datasets demonstrate that RAF and SK-ELM outperform the ELM which use other activation functions and semi-supervised (kernel) ELM methods.

Semantic Discriminative Metric Learning for Image Similarity Measurement

Along with the arrival of multimedia time, multimedia data has replaced textual data to transfer information in various fields. As an important form of multimedia data, images have been widely utilized by many applications, such as face recognition and image classification. Therefore, how to accurately annotate each image from a large set of images is of vital importance but challenging. To perform these tasks well, it is crucial to extract suitable features to character the visual contents of images and learn an appropriate distance metric to measure similarities between all images. Unfortunately, existing feature operators, such as histogram of gradient, local binary pattern, and color histogram, care more about the visual character of images and lack the ability to distinguish semantic information. Similarities between those features cannot reflect the real category correlations due to the well-known semantic gap. In order to solve this problem, this paper proposes a regularized distance metric framework called semantic discriminative metric learning (SDML). SDML combines geometric mean with normalized divergences and separates images from different classes simultaneously. The learned distance metric can treat all images from different classes equally. And distinctions between similar classes with entirely different semantic contents are emphasized by SDML. This procedure ensures the consistency between dissimilarities and semantic distinctions and avoids inaccuracy similarities incurred by unbalanced locations of samples. Various experiments on benchmark image datasets show the excellent performance of the novel method.

Maximal Similarity Embedding

In recent times the dimensionality reduction technique has been widely exploited in pattern recognition and data mining. The global linear algorithms characterize the local sampling information, thereby making it superior to Principal Component Analysis (PCA). However, these algorithms are all inefficient for extracting the local data feature, which leads to incomplete learning. A new global linear algorithm is proposed in this paper, which is named Maximal Similarity Embedding (MSE). The preserving local feature of this new algorithm makes it distinct from most other methods. The MSE algorithm utilizes the Cosine Metric to describe the geometric characteristics of neighborhood and thus seeks to maximize the global similarity for dimensionality reduction. This new proposal method is robust for sparse dataset and naturally helps in avoiding the problem of small sample size cases. Extensive experiments have been performed on both synthetic and real-world images to prove the efficiency of the MSE algorithm.

Metric learning with geometric mean for similarities measurement

Distance metric learning aims to find an appropriate method to measure similarities between samples. An excellent distance metric can greatly improve the performance of many machine learning algorithms. Most previous methods in this area have focused on finding metrics which utilize large-margin criterion to optimize compactness and separability simultaneously. One major shortcoming of these methods is their failure to balance all between-class scatters when the distributions of samples are extremely unbalanced. Large-margin criterion tends to maintain bigger scatters while abandoning those smaller ones to make the total scatters maximized. In this paper, we introduce a regularized metric learning framework, metric learning with geometric mean which obtains a distance metric using geometric mean. The novel method balances all between-class scatters and separates samples from different classes simultaneously. Various experiments on benchmark datasets show the good performance of the novel method.

Multi-view Sparsity Preserving Projection for dimension reduction

Abstract In the past decade, we have witnessed a surge of interests of learning a low-dimensional subspace for dimension reduction (DR). However, facing with features from multiple views, most DR methods fail to integrate compatible and complementary information from multi-view features to construct low-dimensional subspace. Meanwhile, multi-view features always locate in different dimensional spaces which challenges multi-view subspace learning. Therefore, how to learn one common subspace which can exploit information from multi-view features is of vital importance but challenging. To address this issue, we propose a multi-view sparse subspace learning method called Multi-view Sparsity Preserving Projection (MvSPP) in this paper. MvSPP seeks to find a set of linear transforms to project multi-view features into one common low-dimensional subspace where multi-view sparse reconstructive weights are preserved as much as possible. Therefore, MvSPP can avoid incorrect sparse correlations which are caused by the global property of sparse representation from one single view. A co-regularization scheme is designed to integrate multi-view features to seek one common subspace which is consistent across multiple views. An iterative alternating strategy is presented to obtain the optimal solution of MvSPP. Various experiments on multi-view datasets show the excellent performance of this novel method.

A Novel Multi-Feature Representation of Images for Heterogeneous IoTs

With the applications heterogeneous of Internet of Things (IoT) technology, the heterogeneous IoT systems generate a large number of heterogeneous datas, including videos and images. How to efficiently represent these images is an important and challenging task. As a local descriptor, the texton analysis has attracted wide attentions in the field of image processing. A variety of texton-based methods have been proposed in the past few years, which have achieved excellent performance. But, there still exists some problems to be solved, especially, it is difficult to describe the images with complex scenes from IoT. To address this problem, this paper proposes a multi-feature representation method called diagonal structure descriptor. It is more suitable for intermediate feature extraction and conducive to multi-feature fusion. Based on visual attention mechanism, five kinds of diagonal structure textons are defined by the color differences of diagonal pixels. Then, four types of visual features are extracted from the mapping sub-graphs and integrated into 1-D vector. Various experiments on three Corel-datasets demonstrate that the proposed method performs better than several state-of-the-art methods.

Robust discriminative extreme learning machine for relevance feedback in image retrieval

Relevance feedback (RF) has long been an important approach for multi-media retrieval because of the semantic gap in image content, where SVM based methods are widely applied to RF of content-based image retrieval. However, RF based on SVM still has some limitations: (1) the high dimension of image features always make the RF time-consuming; (2) the model of SVM is not discriminative, because labels of image features are not sufficiently exploited. To solve above problems, we proposed robust discriminative extreme learning machine (RDELM) in this paper. RDELM involved both robust within-class and between-class scatter matrices to enhance the discrimination capacity of ELM for RF. Furthermore, an angle criterion dimensionality reduction method is utilized to extract the discriminative information for RDELM. Experimental results on four benchmark datasets (Corel-1K, Corel-5K, Corel-10K and MSRC) illustrate that our proposed RF method in this paper achieves better performance than several state-of-the-art methods.

Locality Structured Sparsity Preserving Embedding

In recent years, the theory of sparse representation (SR) has been widely exploited in sparse subspace learning (SSL). Among all these methods, SR is a parameter-free global algorithm in nature which is mostly utilized to construct the correlations between samples to avoid some negative effects incurred by k-nearest neighbor (KNN) or some other methods. However, these SSL algorithms always lack obvious discrimination because of the ignorance of samples distribution. Meanwhile, some incorrect correlations are taken into consideration owing to the global feature of SR. To solve these two problems, a new SSL algorithm called locality structured sparsity preserving embedding (LSPE) is proposed in this paper. We add the local structured information to SR and construct correlations between samples. However, LSPE is an unsupervised method which wastes all label information. Therefore, LSPE is extended to semi-supervised LSPE (SLSPE) in this paper. SLSPE not only makes good use of the label information but also enhances the discriminative power of LSPE. Extensive experiments have been performed on three image datasets (CMU, COIL20, ORL) and two UCI datasets (Glass, Segment) to prove the efficiency of the LSPE and SLSPE.