Search

Highlights • We exploit the distribution information of principal orientations of dataset by learning the projection matrix with trajectories on both spatial and temporal domains for extracting features informatively. • We exploit the residual information of projected features in the projection subspace by maximizing the residual value of features from principal orientations. • We consider the correlation between RGB channel and depth channel for RGB-D based action recognition and jointly learn the projection matrices on corresponding channels. Abstract In this paper, we propose an unsupervised representation method to learn principal orientations and residual descriptor (PORD) for action recognition. Our PORD aims to learn the statistic principal orientations and to represent the local features of action videos with residual values. The existing hand-crafted feature based methods require high prior knowledge and lack of the ability to represent the distribution of features of the dataset. Most of the deep learned feature based methods are data adaptive, but they do not consider the projection orientations of features nor the loss of locally aggregated descriptors of the quantization. We propose a method of principal orientations and residual descriptor considering that the principal orientations reflect the distribution of local features in the dataset and the residual of projection contains discriminative information of local features. Moreover, we propose a multi-modality PORD method by reducing the modality gap of the RGB channels and the depth channel at the feature level to make our method applicable to RGB-D action recognition. To evaluate the performance, we conduct experiments on five challenging action datasets: Hollywood2, UCF101, HMDB51, MSRDaily, and MSR-Pair. The results show that our method is competitive with the state-of-the-art methods. [ABSTRACT FROM AUTHOR]

The characteristic signals of the machined surface are a mixture of actual signals and noise. It is feasible to make the features distinct through wavelet denoising. However, some of the deterministic signals may be lost with noise removed resulting in the loss of energy which make it difficult to judge the real components of the surface. An improved signal determination method —— wavelet denoising with compensation of the loss (WDCL) is proposed in this paper. The compensation method uses ensemble empirical mode decomposition (EEMD) and transfer function in which instantaneous frequency is calculated by Hilbert transform (HT). The coefficients of the transfer function are adjusted by improving the passing rate of the deterministic signals and lowering the passing rate of noise. The result shows that the WDCL can enhance the resolution of the real signals and reduce noise further. [ABSTRACT FROM AUTHOR]

Previous study shows that if the total integration time of intensity measurements is fixed, the variance of DOLP estimator depends on the distribution of the degree of linear polarization for two intensity measurements. However, the optimal time dependents on the quantity being measured. In this paper, the expected variance is used to define a cost function for optimization of the estimator, which overcomes the above limitation. Actually, minimizing the expected variance is equivalent to minimizing the noise power of estimator, so as to improve the precision of the estimator. We also deduce the closed-form solution of the optimal distribution of the integration time for additive Gaussian noise by employing Lagrange multiplier method. According to the theoretical analyses, it is shown that the variance of DOLP estimator can be decreased for most values of DOLP without prior knowledge, and the proposed method statistically improve the measurement accuracy of the polarimetry system. [ABSTRACT FROM AUTHOR]

Convolutional neural networks (CNNs) consist of alternating convolutional layers and pooling layers. The pooling layer is obtained by applying pooling operator to aggregate information within each small region of the input feature channels and then down sampling the results. Typically, hand-crafted pooling operations are used to aggregate information within a region, but they are not guaranteed to minimize the training error. To overcome this drawback, we propose a learned pooling operation obtained by end-to-end training which is called LEAP (LEArning Pooling). Specifically, in our method, one shared linear combination of the neurons in the region is learned for each feature channel (map). In fact, average pooling can be seen as one special case of our method where all the weights are equal. In addition, inspired by the LEAP operation, we propose one simplified convolution operation to replace the traditional convolution which consumes many extra parameters. The simplified convolution greatly reduces the number of parameters while maintaining comparable performance. By combining the proposed LEAP method and the simplified convolution, we demonstrate the state-of-the-art classification performance with moderate parameters on three public object recognition benchmarks: CIFAR10 dataset, CIFAR100 dataset, and ImageNet2012 dataset. [ABSTRACT FROM AUTHOR]

Profile error of free-form surface is evaluated in this paper based on sequential quadratic programming (SQP) algorithm. The optimal localization model is established with the minimum zone criterion firstly. Subsequently, the surface subdivision method or STL (STeror Lithography) model is used to compute the point-to-surface distance and the approximate linear differential movement model of signed distance is deduced to simplify the updating process of alignment parameters. Finally, the optimization model on profile error evaluation of free-form surface is solved with SQP algorithm. Simulation examples indicate that the results acquired by SQP method are closer to the ideal results than the other algorithms in the problem of solving transformation parameters. In addition, real part experiments show that the maximum distance between the measurement points and their corresponding closest points on the design model is shorter by using SQP-based algorithm. Lastly, the results obtained in the experiment of the workpiece with S form illustrate that the SQP-based profile error evaluation algorithm can dramatically reduce the iterations and keep the precision of result simultaneously. Furthermore, a simulation is conducted to test the robustness of the proposed method. In a word, this study purposes a new algorithm which is of high accuracy and less time-consuming. [ABSTRACT FROM AUTHOR]

Important task for acoustic emission (AE) monitoring involves detecting frequency shift phenomenon and intense periodic components. In the present research, we investigate time dynamics embedded in AE signal acquired in the laboratory rock burst experiment on limestone sample. By applying the Singular Spectrum Analysis (SSA)-based algorithm developed in this research, we reconstruct the decomposed components and then select the main component with a decision-making process based on the criterion that it should be significant both in the eigenvector space and spectral domain, termed eigenfrequency. The frequency shift phenomenon is represented by the eigenfrequencies of the first main component consistently. Precursory waves of the first main component represents time dynamics of the rock burst process by elastic wave over the low-level loading phase, high-frequency wave with self-oscillating envelopes at unloading, low-frequency quasi-shock waves during the rheological delay phase and low-frequency shock wave at complete rock burst failure. [ABSTRACT FROM AUTHOR]

To describe the random movement of ocean wave exactly, the random field theory for local average sampling is introduced. In this work, the rotated empirical orthogonal function analysis (REOF) is proposed to estimate the significant wave height (SWH) from the data captured with marine X-band radar. After obtaining the rotated principal components (PC) of the radar image sequences, the standard deviation of the rotated PC is a new means of estimating SWH. The results show the advanced algorithm has a good correlation between the SWH retrieved by radar and that of the measured with the buoy. The root mean squared error (RMSE) of the SWH between the advanced model and the buoy is only 0.1440 m, which decreased about 22% compared with the EOF analysis. [ABSTRACT FROM AUTHOR]

Non-convex l 2 / l q (0 < q < 1) minimization method can efficiently recover the block-sparse signals whose non-zero coefficients occur in a few blocks. However, in many applications such as face recognition and fetal ECG monitoring, real-world signals also exhibit intra-block correlations aside from standard block-sparsity. In order to recover such signals exactly and robustly, the block sparse Bayesian learning framework is studied in this paper. In contrast to l 2 / l q norm minimization the proposed method involves a quadratic Mahalanobis distance measure on the block and a covariance matrix on the intra-block correlation. The improved iteratively reweighted least-squares algorithm for the induced framework is proposed than the recent known for mixed l 2 / l q optimization. The proposed algorithm is tested and compared with the mixed l 2 / l q algorithm on a series of signals modeled by autoregressive processes. Numerical results demonstrate the outperformance of the proposed algorithm and meanfulness of the novel strategy, especially in low sample ratio and large unknown noise level. [ABSTRACT FROM AUTHOR]

Abstract: Sparse representation is being proved to be effective for many tasks in the field of face recognition. In this paper, we will propose an efficient face recognition algorithm via sparse representation in 2D Fisherface space. We firstly transformed the 2D image into 2D Fisherface in preprocessing, and classify the testing image via sparse representation in the 2D Fisherface space. Then we extend the proposed method using some supplementary matrices to deal with random pixels corruption. For face image with contiguous occlusion, we partition each image into some blocks, and define a new rule combining sparsity and reconstruction residual to discard the occluded blocks, the final result is aggregated by voting the classification result of the valid individual block. The experimental results have shown that the proposed algorithm achieves a satisfying performance in both accuracy and robustness. [Copyright &y& Elsevier]

Abstract: Sparse representation is being proved to be effective for many tasks in the field of pattern recognition. In this paper, an efficient classification algorithm based on concentrative sparse representation will be proposed to address the problem caused by insufficient training samples in each class. We firstly compute representation coefficient of the testing sample with training samples matrix using subspace pursuit recovery algorithm. Then we define concentration measurement function in order to determine whether the sparse representation coefficient is concentrative. Subspace pursuit is repeatedly used to revise the sparse representation until concentration is met. Such a concentrative sparse representation can contribute to discriminative residuals that are critical to accurate classification. The experimental results have showed that the proposed algorithm achieves a satisfying performance in both accuracy and efficiency. [Copyright &y& Elsevier]

Abstract: In this paper, we propose a novel content-based blind watermarking algorithm which is robust to arbitrary affine transformation attacks. Although watermarking technology has improved remarkably nowadays, algorithm that resistant to geometric transformation attacks remains to be a challenge to the researchers. A small distortion of the watermarked image may lead the failure of many existing watermark detecting and extracting algorithms because of the destruction of synchronization. We present a synchronization method based on the barycentric coordinate representation of the host image. We first segment the host image according to the pixel values, then calculate the weighted center (of mass) of each part, and then select three of these centers to establish a barycentric coordinate system. The host image is represented using this barycentric coordinate system. A squared area is chosen and is decomposed using 3-level discrete biorthogonal wavelet transform. The watermark is embedded in the middle bands of the wavelet decomposition of the represented host image using a CDMA based algorithm. The watermarked image is obtained by inverse wavelet transform and inverse barycentric coordinate transform. In the watermark extraction phase, the same segmentation should be done on the received image, and the received image should be represented using the barycentric coordinate system, and a corresponding squared area is chosen and decomposed using the same biorthogonal wavelet transform as used in the watermark embedding phase, and then the watermark is extracted from the middle bands using the CDMA based algorithm. The proposed algorithm is absolutely blind. Multi-bit information can be embedded in the host image. Experimental results show that the proposed algorithm is robust to arbitrary affine transformation attacks and many other attacks. [Copyright &y& Elsevier]

Abstract: This work is concerned with approximation of a signal from local averages. It improves a result of Butzer and Lei [P.L. Butzer, J. Lei, Approximation of signals using measured sampled values and error analysis, Commun. Appl. Anal. 4 (2000) 245–255]. [Copyright &y& Elsevier]

Atrial fibrillation is a most common arrhythmia. An early and accurate detection for the cure and even spread of this disease is considerably critical. The visual examination of electrocardiogram signals is the most extensively used diagnosis approach, but this method is cumbersome and low-efficient. In this work, we propose an intelligent network model based on the modified Elman neural network for signals discrimination. Motivated from the exponential moving average strategy, the proposed model is capable of fully modeling the information feedback and also effectively and efficiently striking a balance between current information representation and historical information representation in original Elman neural network. To evaluate its practicability, the model is also plugged into a convolutional neural network framework and two control subjects are established for a fair comparison. Experiments on the MIT-BIH atrial fibrillation and arrhythmia databases show that the proposed model can enjoy a consistent improvement in classification performance with the accuracy of 98.2% and 97.2% respectively and exhibit lower convergence rate than existing Elman network. Thanks to its high model performance, we are planning to develop the model into a computer-aided diagnosis system to assist physicians. • We design modified Elman network (MENN) for atrial fibrillation (AF) detection. • Patient-independent validation ensures the model robustness. • The feature extraction and classification are not required. • To our knowledge, this is the first time to redesign ENN for AF detection. [ABSTRACT FROM AUTHOR]

Thresholding based iterative algorithms have the trade-off between effectiveness and optimality. Some are effective but involving sub-matrix inversions in every step of iterations. For systems of large sizes, such algorithms can be computationally expensive and/or prohibitive. The null space tuning algorithm with hard thresholding and feedbacks (NST+HT+FB) has a mean to expedite its procedure by a suboptimal feedback, in which sub-matrix inversion is replaced by an eigenvalue-based approximation. The resulting suboptimal feedback scheme becomes exceedingly effective for large system recovery problems. An adaptive algorithm based on thresholding, suboptimal feedback and null space tuning (AdptNST+HT+subOptFB) without a prior knowledge of the sparsity level is also proposed and analyzed. Convergence analysis is the focus of this article. [ABSTRACT FROM AUTHOR]

As free-form surface is widely used in engineering, it is urgently needed to develop advanced methodology of detecting and evaluating the profile error. To this end, the semantic of profile tolerance in ASME Y14.5.1M is reviewed and the mathematical definition of profile tolerance is discussed. Subsequently, a mathematical model for error evaluation is built. This mathematical model is augmented based on distance function by considering the second-order terms in the computation of the distance from point to surface. Then, a profile error evaluation algorithm, which combines Differential Evolution (DE) algorithm and Nelder–Mead (NM) algorithm, is developed to solve this model. The proposed model and optimization algorithm are validated with simulation results from a case study. Additionally, the model is superior to Least-squares (LS) model in simplicity, efficiency and robustness. [ABSTRACT FROM AUTHOR]

• An accelerated class of adaptive scheme of iterative thresholding algorithms based on the feedback mechanism of the null space tuning techniques is studied. • Accelerated convergence rate and improved convergence conditions are obtained by selecting an appropriate size of the index support per iteration. • The theoretical findings are sufficiently demonstrated and confirmed by extensive numerical experiments. An accelerated class of adaptive scheme of iterative thresholding algorithms is studied analytically and empirically. They are based on the feedback mechanism of the null space tuning techniques. The main contribution of this article is the accelerated convergence analysis and proofs with a variable/adaptive index selection and different feedback principles at each iteration. The convergence analysis requires no longer a priori sparsity information s of a signal. It is shown that uniform recovery of all s -sparse signals from given linear measurements can be achieved under reasonable (preconditioned) restricted isometry conditions. Accelerated convergence rate and improved convergence conditions are obtained by selecting an appropriate size of the index support per iteration. The theoretical findings are sufficiently demonstrated and confirmed by extensive numerical experiments. It is also observed that the proposed algorithms have a clearly advantageous balance of efficiency, adaptivity and accuracy compared with all other state-of-the-art greedy iterative algorithms. [ABSTRACT FROM AUTHOR]

In this paper, we propose a novel semi-supervised method to predict clothing attributes with the assistance of unlabeled data like fashion shows. To this end, a two-stage framework is built, i.e., the unsupervised triplet network pre-training stage that ensures frames in the same video having coherent representations while frames from different videos having larger feature distances, and a supervised clothing attribute prediction stage to estimate the value of attributes. Specifically, we first detect the clothes of frames in the collected 18,737 female fashion shows and 21,224 male fashion shows which contain no extra labels. Then a triplet neural network is constructed via embedding the temporal appearance consistency between frames in the same video and the representation gap in different videos. Finally, we transfer the triplet model parameters to multi-task clothing attribute prediction model, and fine-tune it with clothing images holding attribute labels. Extensive experiments demonstrate the advantages of the proposed method on two clothing datasets. [ABSTRACT FROM AUTHOR]