Shijie Hao

Scalable Semi-Supervised Learning by Efficient Anchor Graph Regularization

Many graph-based semi-supervised learning methods for large datasets have been proposed to cope with the rapidly increasing size of data, such as Anchor Graph Regularization (AGR). This model builds a regularization framework by exploring the underlying structure of the whole dataset with both datapoints and anchors. Nevertheless, AGR still has limitations in its two components: (1) in anchor graph construction, the estimation of the local weights between each datapoint and its neighboring anchors could be biased and relatively slow; and (2) in anchor graph regularization, the adjacency matrix that estimates the relationship between datapoints, is not sufficiently effective. In this paper, we develop an Efficient Anchor Graph Regularization (EAGR) by tackling these issues. First, we propose a fast local anchor embedding method, which reformulates the optimization of local weights and obtains an analytical solution. We show that this method better reconstructs datapoints with anchors and speeds up the optimizing process. Second, we propose a new adjacency matrix among anchors by considering the commonly linked datapoints, which leads to a more effective normalized graph Laplacian over anchors. We show that, with the novel local weight estimation and normalized graph Laplacian, EAGR is able to achieve better classification accuracy with much less computational costs. Experimental results on several publicly available datasets demonstrate the effectiveness of our approach.

Learning on Big Graph: Label Inference and Regularization with Anchor Hierarchy

Several models have been proposed to cope with the rapidly increasing size of data, such as Anchor Graph Regularization (AGR). The AGR approach significantly accelerates graph-based learning by exploring a set of anchors. However, when a dataset becomes much larger, AGR still faces a big graph which brings dramatically increasing computational costs. To overcome this issue, we propose a novel Hierarchical Anchor Graph Regularization (HAGR) approach by exploring multiple-layer anchors with a pyramid-style structure. In HAGR, the labels of datapoints are inferred from the coarsest anchors layer by layer in a coarse-to-fine manner. The label smoothness regularization is performed on all datapoints, and we demonstrate that the optimization process only involves a small-size reduced Laplacian matrix. We also introduce a fast approach to construct our hierarchical anchor graph based on an approximate nearest neighbor search technique. Experiments on million-scale datasets demonstrate the effectiveness and efficiency of the proposed HAGR approach over existing methods. Results show that the HAGR approach is even able to achieve a good performance within 3 minutes in an 8-million-example classification task.

Image quality assessment based on matching pursuit

Abstract The objective assessment of image quality is an essential part of many visual processing systems. The challenge lies in evaluating the image quality consistently under subjective perceptions. In this paper, we propose a novel image quality metric based on the matching pursuit algorithm. Under the principle of structural information distortion, we assume that various structure data contributes differently to the single image quality score. Specifically, we decompose the reference image using matching pursuit with a separable 2D Gabor dictionary, thus obtaining structural information, and develop a characterization for this information and its importance. We then discuss the relationship between the structural distortion intensity and the subjective quality measurement on the energy scale. In experiments, we compare the performance of our algorithm with both subjective ratings and state-of-the-art objective methods on image datasets with multi-type distortions. The experimental results validate our proposed method.

Automatic image annotation by semi-supervised manifold kernel density estimation

The insufficiency of labeled training data is a major obstacle in automatic image annotation. To tackle this problem, we propose a semi-supervised manifold kernel density estimation (SSMKDE) approach based on a recently proposed manifold KDE method. Our contributions are twofold. First, SSMKDE leverages both labeled and unlabeled samples and formulates all data in a manifold structure, which enables a more accurate label prediction. Second, the relationship between KDE-based methods and graph-based semi-supervised learning (SSL) methods is analyzed, which helps to better understand graph-based SSL methods. Extensive experiments demonstrate the superiority of SSMKDE over existing KDE-based and graph-based SSL methods.

Image detail enhancement with spatially guided filters

In recent years, enhancing image details without introducing artifacts has been attracting much attention in image processing community. Various image filtering methods have been proposed to achieve this goal. However, existing methods usually treat all pixels equally during the filtering process without considering the relationship between filtering strengths and image contents. In this paper, we address this issue by spatially distributing the filtering strengths with simple low-level features. First, to determine the pixel-wise filtering strength, we construct a spatially guided map, which exploits the spatial influence of image details based on the edge response of an image. Then, we further integrate this guided map into two state-of-the-art image filters and apply the improved filters to the task of image detail enhancement. Experiments demonstrate that our results generate better content-specific visual effects and introduce much less artifacts. We construct a spatially guided map exploiting spatial influences of edge features.We improve two state-of-the-art filters by integrating our spatially guided map.Better filtering results are obtained with more naturalness and less artifacts.

Person Re-identification based on nonlinear ranking with difference vectors

Abstract Matching people across non-overlapping camera views, a.k.a. the person re-identification problem, is important for video surveillance and gaining increasing attention. In this paper, we propose a re-identification method that uses Nonlinear Ranking with Difference Vectors (NRDV). Instead of trying to eliminate the differences between cameras or seek more reliable features, our strategy is to make full use of the targets’ differences to build a binary classifier. We then achieve re-identification through a ranking approach by employing a support vector machine with a nonlinear kernel based on radial basis function. We also propose to pre-cluster the training images using the affinity propagation clustering algorithm, and select representative images to form negative training instances. In this strategy, the classifier maintains its performance with fewer training samples, and has lower memory requirements. Extensive experiments are conducted on three public benchmark datasets, and the results demonstrate the state-of-the-art performance of the proposed method.

Image classification based on low-rank matrix recovery and Naive Bayes collaborative representation

Abstract Most image classification methods require an expensive learning/training phase to gain high performances. But they frequently encounter problems such as overfitting of parameters and scarcity of training data. In this paper, we present a novel learning-free image classification algorithm under the framework of Naive-Bayes Nearest-Neighbor (NBNN) and collaborative representation, where non-negative sparse coding, low-rank matrix recovery and collaborative representation are jointly employed to obtain more robust and discriminative representation. First, instead of using general sparse coding, non-negative sparse coding combined with max pooling is introduced to further reduce information loss. Second, we use the low-rank matrix recovery technique to decompose the training data of the same class into a discriminative low-rank matrix, in which more structurally correlated information is preserved. As for testing images, a low-rank projection matrix is also learned to remove possible image corruptions. Finally, the classification process is implemented by simply comparing the responses over the different bases. Experimental results on several image datasets demonstrate the effectiveness of our method.

Dimensionality reduction on Anchorgraph with an efficient Locality Preserving Projection

Manifold learning based dimensionality reduction methods have been successfully applied in many pattern recognition tasks, due to their ability to well capture the underlying relationship between data points. These methods, however, meet some challenges in terms of the storage cost and the computation complexity with the rapidly increasing data size. We propose an improved dimensionality reduction algorithm called Anchorgraph-based Locality Preserving Projection (AgLPP), trying to cope with the limitations via a novel estimation of the relationship between data points. We extend AgLPP into a kernel version, and reformulate it into a novel sparse representation. The experiments on several real-world datasets have demonstrated the effectiveness and efficiency of our methods.

Scale-Aware Spatially Guided Mapping

The scale information in images is important for guiding image-filtering configuration. The authors propose a scale-aware spatially guided mapping (SaSGM) model, which formulates and combines multiple spatial influences of simple edge responses under different levels of detail. The SaSGM model is thus more sensitive to image patterns at a large scale. The authors further incorporate the SaSGM into several image processing models, such as detail enhancement and image stylization models. Experiments show that by inheriting the characteristics of the SaSGM, the extended models are able to differentiate image contents in terms of their scales and thus generate more natural or diversified visual effects. This article is part of a special issue on quality modeling.

Spatially guided local Laplacian filter for nature image detail enhancement

Nature images make up a significant proportion of the ever growing volume of social media. In this context, automatic and rapid image enhancement is always among the favorable techniques for photographers. Among the image representation models, the Gaussian and Laplacian image pyramids based on isotropic Gaussian kernels were once considered to be inappropriate for image enhancement tasks. The recently proposed Local Laplacian Filter (LLF) updates this view by designing a point-wise intensity remapping process. However, this model filters an image with a consistent strength instead of a dynamical way which takes image contents into account. In this paper, we propose a spatially guided LLF by extending the single-value key parameter into a multi-value matrix that dynamically assigns filtering strengths according to image contents. Since it is still very challenging to recognize arbitrary image contents with machine learning methods, we propose a simple but effective technique, which only approximates the richness of image details instead of specific contents. This trade-off between concrete semantics and algorithm efficiency enables filtering strengths to be spatially guided in the LLF process with little extra computational cost. Experimental results validate our method in terms of visual effects and a conditionally faster LLF implementation.