Ruizhen Zhao

Hessian Semi-Supervised Sparse Feature Selection Based on

Semi-supervised sparse feature selection, which can exploit the small number labeled data and large number unlabeled data simultaneously, has become an important technique in many applications on large-scale web image owing to its high efficiency and effectiveness. Recently, graph Laplacian-based semi-supervised sparse feature selection has obtained considerable attention, but it suffers with only few labeled data because Laplacian regularization is short of extrapolating power. In this paper we propose a novel semi-supervised sparse feature selection framework based on Hessian regularization and l2,1/2- matrix norm, namely Hessian sparse feature selection based on L2,1/2- matrix norm (HFSL). Hessian regularization favors functions whose values vary linearly with respect to geodesic distance and preserves the local manifold structure well, leading to good extrapolating power to boost semi-supervised learning, and then to enhance HFSL performance. The l2,1/2-matrix norm model makes HFSL select the most discriminative sparse features with good robustness. An efficient iterative algorithm is designed to optimize the objective function. We apply our algorithm into the image annotation task and conduct extensive experiments on two web image datasets. The results demonstrate that our algorithm outperforms state-of-the-art sparse feature selection methods and is promising for large-scale web image applications.

{L_{2,1/2}}

Abstract Surface defect inspection of TFT-LCD panels is a critical task in LCD manufacturing. In this paper, an automatic defect inspection method based on the low-rank matrix reconstruction is proposed. The textured background of the LCD image is a low-rank matrix and the foreground image with defects can be treated as a sparse matrix. By utilizing the Inexact Augmented Lagrange Multipliers (IALM) algorithm, the segmentation of a LCD image can be converted into the reconstruction of a low-rank matrix with a fraction of its entries arbitrarily corrupted. This low-rank matrix reconstruction problem can be exactly solved via convex optimization that minimizes a combination of the nuclear norm and the l1-norm. Also, adaptive parameter selection strategy is proposed by conducting deep analysis on the IALM algorithm, which improves the generality of the IALM algorithm for different defect types. Experiment results show that our inspection algorithm is robust for the defect shapes and types under different illumination conditions. The shapes and edges of defect areas in the LCD images can be well preserved and segmented from textured background by our detection algorithm.

-Matrix Norm

Abstract Recently, a greedy algorithm called Atomic Decomposition for Minimum Rank Approximation (ADMiRA) was proposed. It has solved the low-rank matrix approximation problem when the rank of the matrix is known. However, the rank of the matrix is usually unknown in practical application. In this paper, a Rank Adaptive Atomic Decomposition for Low-Rank Matrix Completion (RAADLRMC) algorithm is proposed based on the Atomic Decomposition for Minimum Rank Approximation. The advantage of RAADLRMC is that it works when the parameter rank-r of matrix is not given. Furthermore, the step size of iteration is decreased adaptively in order to improve the efficiency and accuracy. As illustrated by our experiments, our algorithm is robust, and the rank of matrix can be predicted accurately.

Defect inspection for TFT-LCD images based on the low-rank matrix reconstruction

A novel system consisting of a camera and a light emitting diode (LED) is presented for measuring food portion size. The LED is positioned at a fixed distance besides the camera with its optical axis parallel to the optical axis of the camera. The distance to and oblique angle of the object plane are calculated according to the deformation of the projected spotlight pattern. Experimental results show that satisfactory measurements of food portion size can be obtained with this simple system.

Rank adaptive atomic decomposition for low-rank matrix completion and its application on image recovery

Wavelet threshold denoising is a powerful method for suppressing noise in signals and images. However, this method uses a coordinate-wise processing scheme, which ignores the structural properties in the wavelet coefficients. We propose a new denoising method using sparse representation which is a powerful mathematical tool developed only recently. Instead of thresholding wavelet coefficients individually, we minimize the number of coefficients in the sparse representation frame work under certain conditions. The denoised signal is reconstructed by solving an optimization problem. We show that, by using an iterative algorithm, the solution to the optimization problem can be obtained uniquely and the estimates are unbiased, i.e., the statistical means of the estimates are equal to the ideal wavelet coefficients. Our experiments on test signals show that this new denoising method is effective and efficient for a wide variety of signals including those with a low signal-to-noise ratio.

A food portion size measurement system for image-based dietary assessment

In this paper, a new video restoration approach is proposed. By using a modified version of random PatchMatch algorithm, nearest-neighbor patches among the video frames can be grouped quickly and accurately. Then the video restoration problem can be boiled down to a low-rank matrix recovery problem, which is able to separate sparse errors from matrices that possess potential low-rank structures. Furthermore, the reweighted low-rank matrix model is used to improve the performance of video restoration by enhancing the sparsity of the sparse matrix and the low-rank property of the low-rank matrix. Experimental results show that our system achieves good performance in denosing of joint multi-frames and inpainting in the presence of small damaged areas.

A new denoising method based on wavelet transform and sparse representation

Electroencephalography (EEG) is a major tool for clinical diagnosis of neurological diseases and brain research. EEGs are often collected over numerous channels and trials, providing large data sets that require efficient collection and accurate compression. Compressive sensing (CS) emphasizing signal sparseness enables the reconstruction of signals from a small set of measurements, at the expense of computationally complex reconstruction algorithms. In this paper we show that using Discrete Prolate Spheroidal Sequences, rather than sine functions, it is possible to derive a sampling and reconstruction method which is similar to CS. Assuming non-uniform sampling our procedure can be connected with compressive sensing without complex reconstruction methods.

Video restoration based on PatchMatch and reweighted low-rank matrix recovery

Visual vocabulary is the core of the Bag-of-visual-words (BOW) model in image retrieval. In order to ensure the retrieval accuracy, a large vocabulary is always used in traditional methods. However, a large vocabulary will lead to a low recall. In order to improve recall, vocabularies with medium sizes are proposed, but they will lead to a low accuracy. To address these two problems, we propose a new method for image retrieval based on feature fusion and sparse representation over separable vocabulary. Firstly, a large vocabulary is generated on the training dataset. Secondly, the vocabulary is separated into a number of vocabularies with medium sizes. Thirdly, for a given query image, we adopt sparse representation to select a vocabulary for retrieval. In the proposed method, the large vocabulary can guarantee a relatively high accuracy, while the vocabularies with medium sizes are responsible for high recall. Also, in order to reduce quantization error and improve recall, sparse representation scheme is used for visual words quantization. Moreover, both the local features and the global features are fused to improve the recall. Our proposed method is evaluated on two benchmark datasets, i.e., Coil20 and Holidays. Experiments show that our proposed method achieves good performance.

Discrete Prolate Spheroidal Sequences for compressive sensing of EEG signals

Sparse representation based on dictionary has gained increasing interest due to its extensive applications. Because of the disadvantages of computational complexity of traditional dictionary learning, we propose an algorithm of analytic separable dictionary learning. Considering the differences of sparse coefficient matrix and dictionary, we divide our algorithm into two phases: 2D sparse coding and dictionary optimization. Then an alternative iteration method is used between these two phases. The algorithm of 2D-OMP (2-dimensional Orthogonal Matching Pursuit) is used in the first phase because of its low complexity. In the second phase, we create a continuous function of the optimization problem, and solve it by the conjugate gradient method on oblique manifold. By employing the separable structure of the optimized dictionary, a competitive result is achieved in our experiments for image de-noising.

Separable vocabulary and feature fusion for image retrieval based on sparse representation

The sparse property of the signal to be processed is very important and directly affects the efficiency of compressive sensing. A signal pre-processing method suitable for compressive sensing is given, which is helpful to effective sensing and accurate reconstruction. Under the condition that the sparse characteristic of the signal is unknown, a frequency modulation pattern is introduced to pre-process the signal to increase the sparse proportion of the signal. Then we choose a difference matrix as the reconstruction matrix, the signal could be reconstructed accurately in the process of sparse reconstruction. Theoretical analysis and experimental results show that the proposed pre-processing method for compressive sensing is very effective and efficient.