Search

With the rapid development of social websites, recent years have witnessed an explosive growth of social images with user-provided tags which continuously arrive in a streaming fashion. Due to the fast query speed and low storage cost, hashing-based methods for image search have attracted increasing attention. However, existing hashing methods for social image retrieval are based on batch mode which violates the nature of social images, i.e., social images are usually generated periodically or collected in a stream fashion. Although there exist many online image hashing methods, they either adopt unsupervised learning which ignore the relevant tags, or are designed in the supervised manner which needs high-quality labels. In this paper, to overcome the above limitations, we propose a new method named Weakly-supervised Online Hashing (WOH). In order to learn high-quality hash codes, WOH exploits the weak supervision by considering the semantics of tags and removing the noise. Besides, We develop a discrete online optimization algorithm for WOH, which is efficient and scalable. Extensive experiments conducted on two real-world datasets demonstrate the superiority of WOH compared with several state-of-the-art hashing baselines., Comment: Accepted by ICME 2021

Magnetoencephalography(MEG) is a functional, noninvasive technology for recording neuronal activity of human brain in real-time. MEG has been approved as an efficient and powerful functional detection method which has been used in clinical practices relate with epilepsy, depression, and other diseases that cannot be well-diagnosed by traditional structural methods, e.g., CT and MRI. By using MEG, very weak magnetic signals typically in the range of 10 femtotesla (fT)-1 picotesla (pT) that originate from electric currents flowing in neurons, can be recorded. Compared with scalp EEG, MEG provides much better spatial resolution. Meanwhile, MEG is a noninvasive technology for preoperative workup for epilepsy surgery, which means it may replace traditional invasive methods. As the development of high-sampling rate MEG systems as well as the discovery high-frequency oscillations (HFOs) in the brain, MEG is now typically digitally analyzed in very high-frequency range. Although MEG has been clinically used for two decades, it is not so widely-used as EEG. The lack of a powerful software platform is an important reason which hinders the development of MEG. A MEG-specific software, MEG processor, was released here. With the friendly graphical user interface (GUI) of MEG Processor, researcher can easily identify/mark HFOs and project the signal to MRI receiving a precise 3D reconstruction. We hope such a platform will help and encourage more clinicians to uncover the magnificent features of MEG.

Neuronal activities of two monkeys are analyzed to explore whether moving tasks can trigger specific motor-controlled signal as response. A multi-array contains 96-channel electrodes was used to record the long-term neuronal activity. In the time domain analysis, we found the cues of tasks can stably stimulate some spindle waves which do not exist in the task-free stages. Most spindles sparkled after the participant issued real and correct actions as responses to the cues of tasks, indicating the spindle as a motor-controlled wave-marker. In a further time-frequency analysis, the motor-controlled spindles indicate their high intensity in the spectrogram of β band. A spindle detection algorithm based on SVM was constructed to annotate the motor-controlled spindles automatically. Considering the individual and the day-to-day differences, the subject-specific model under leave-one-day-out cross-validation was framed to guarantee the model robustness. The algorithm performed well with F1 scores at 83.73 ± 13.08% and 79.26 ± 16.95% on the two monkeys, respectively. This finding might be used in the future as a new marker in certain research areas, e.g., mental detection, autism spectrum disorder (ASD).

Dense views are needed urgently for autostereoscopic three-dimensional (3D) displays. A dense view synthetization system with three input views which are captured by a three-camera rig is presented. In the system, an image preprocessing step is presented before rendering dense views, in which both color and depth images are optimized by the proposed adaptive joint view optimization algorithm. Then visual artifacts in generated views are avoided fundamentally. With the preprocessing instead of post-processing on each virtual image, the computing time is reduced apparently, especially when the number of output views is great. Experimental results show that the generated dense views are well presented on the autostereoscopic display. Both the quality of virtual images and the continuity among viewpoints are improved by the system.

Nowadays, with the increasing usage of the internet, the movie ratings on the SNS website related to movies influence our choice of movies remarkably. However, a newly released film has insufficient rating counts to reflect the quality of the movie, and it can not avoid the influence of malicious rating by some people. Therefore, this paper proposes a method of rating prediction based on Gaussian Mixture Model (GMM), enabled by imitating rating behavior of audience. Meanwhile, this model can avoid the influence of malicious rating because GMM is not sensitive to exception. In GMM, 4 features of the movies are taken into consideration. In order to verify the validity of our model, data from Douban website is used in the implementation. Experimental results exhibit the effectiveness of the method and an improved performance of rating prediction is achieved compared with the benchmark of linear regression.

In this paper, a high-speed permanent magnet motor with 10kw and 100krm which used in an air compressor was designed at first, then, the loss as one of the key researches of high speed permanent magnet motor was analyzed from the angle of the magnetic field: the stator core loss was analyzed under rotor magnetic field and armature current magnetic field, meanwhile, the stator losses of different materials were compared; there was a comparison between rotor eddy current loss generated by space harmonics and time harmonics and the influence of the number of turns on the eddy current loss of the rotor was analyzed. Besides, at last, Influence of stator fission on loss distribution of the motor was analyzed, it was adopted that the rotor temperature calculation is important for high-speed permanent magnet motor design.

The mass matrix resulting from a finite-element-method (FEM) based electromagnetic analysis possesses favorable properties that can be exploited to reduce the computational time and memory cost for large-scale circuit modeling. In this paper, a mass-matrix based frequency-domain FEM method is developed for the analysis of integrated circuits. Its convergence is theoretically proved, and its efficiency is accelerated by a reduced eigenspace method.

State-of-the-art electromagnetics-based circuit simulators perform circuit simulation on the physical layout of an integrated circuit, the size of which can be large even for an optimal-complexity simulator to analyze it efficiently. We find a frequency- and time-independent minimal-order model of the integrated circuit from DC to high frequencies for any prescribed accuracy based on electromagnetics. The model size is only two for RC circuits and minimal such as tens for RLC and fullwave circuits, regardless of the original circuit size and the number of input/output ports. Numerical experiments demonstrate the superior performance of the proposed minimalorder model based fast electromagnetic simulation.

The prevailing circuit simulation paradigm heavily relies on extraction to perform circuit simulation. In addition, no linear complexity has been reported by existing circuit simulation approaches. In this work, we develop a circuit simulator that bypasses the step of extraction. It produces an RLC (resistor-inductor-capacitor) representation of the linear network without any numerical computation. Such a representation is inherently passive and stable. In addition, it allows for the simulation of a circuit including both nonlinear devices and the linear network in linear complexity. The proposed circuit simulator rigorously captures the coupling between nonlinear circuits and the linear network. As an added bonus, it possesses electromagnetic-physics based accuracy. Application to very large-scale on-chip circuits involving close to 1 million CMOS transistors and interconnects having hundreds of millions of unknowns has demonstrated the superior performance of the proposed extraction-free circuit simulator.

An explicit and unconditionally stable time-domain finite-element method of linear complexity is developed for the simulation of 3-D global interconnect network. Numerical experiments have demonstrated that the proposed unconditionally stable explicit method outperforms both the conditionally stable explicit method and the unconditionally stable implicit method in computational efficiency.

This paper presents a design of a 100 krpm and 10 kw-class permanent magnet synchronous motor (PMSM), which is developed for a spindle motor of a 5-axis blade grinding machine. In addition, a comparison between PMSM and induction motor is given.

In this work, a direct domain-decomposition of linear complexity is developed to overcome the challenge of simulating a wide range of geometrical scales spanning at least 10000∶1 present in an integrated circuit system. This method is developed in the framework of a time-domain finite-element method. It allows for the use of different meshes in different domains. Numerical experiments have validated the proposed method.

From an angle of information science, this paper researches on image sequence which express the working condition information. By means of time sequence analysis the feature sequence extraction is done, further time sequence analysis of image sequence is achieved, whereby the forecasting of quality characteristic can be realized to meet quality prediction of complex industrial production process and production optimization. The research results show that this technical idea is a new way to control and optimize the complex industrial production process.

Aiming at modeling of the complex and large industrial process, this paper proposes the multi-clustering based on the fusion of hierarchical clustering algorithm of field potential topology and fuzzy c-means clustering, on which the structure of fuzzy neural network model is determined. According to the membership degree calculated, the fuzzy neural network model has been configured based on the multi criteria information fusion and fuzzy technology. Simulation results verify the validity of the model, which has some reference value and significance on the coal mining.

This paper presents a wavelet-based image fusion method, in which the self-organizing particle swarm optimization algorithm is applied to image fusion based on region energy. In order to find image fusion by the optimal threshold particles are firsthand formed by thresholds. With comprehensive consideration of the parameters that affect the performance the algorithm, acceleration factors and inertia weights are adaptively adjusted in the self-organizing particle swarm optimization algorithm. Also, to improve the shortcoming of premature convergence of particle swarm optimization, the particles are organized to track the local and global optimal region. The experimental results show that the proposed method can achieve very good fusion performance.

The co-simulation of electromagnetic structures and nonlinear circuits has been explored in the framework of the finite-difference time-domain method, the time-domain finite-element method [1], and the time-domain integral equation method. However, they often have been found not amenable for on-chip VLSI design because of unique on-chip modeling challenges and large number of nonlinear devices. Recently, a time-domain layered finite-element reduction-recovery (LAFE-RR) based nonlinear-linear co-simulation algorithm was developed for solving large-scale integrated circuits and package problems [2–3]. This method can reduce the system matrix of O(N) rigorously to that of O(M) for any multilayered structure, with N being the number of unknowns in the entire 3-D structure, and M the number of unknowns on a single surface. Furthermore, the reduction from O(N) to O(M) was achieved analytically. The reduced linear-nonlinear coupled system is solved and the rest of the unknowns are then recovered in linear complexity. However, the reduced single-surface linear-nonlinear system is a strongly coupled system which either introduces a dense submatrix in the Jacobian matrix of the Newton's method or introduces extra iterations in the solution procedure. When the number of devices is large, the large dense submatrix or the large number of iterations will limit the capacity of this algorithm. In this work, we develop a fast electromagnetics-based nonlinear-linear co-simulation algorithm based on a recently developed time-domain orthogonal finite-element reduction-recovery method (OrFE-RR) [4]. In this method, the linear part and the nonlinear part are naturally decoupled due to the diagonal matrix in the single surface system. Meanwhile, as the devices are usually not connected to the same mesh element, the devices are also decoupled. Thus, the linear part and the nonlinear devices can be solved separately with the coupling accounted in the right hand side. This method neither generates large dense matrix nor requires a staggered marching iteration. Hence, it is capable of simulating large-scale integrated circuits that consist of both linear network and nonlinear devices.

Driven by the continued advance of integrated circuit (IC) technology, large-scale electromagnetic (EM) modeling and simulation problems have been encountered across the design of on-chip circuits, package, and die-package interfaces. To give a few examples, full-chip post-layout performance verification that includes transmission line and full-wave effects, system-level signal and noise integrity analysis, and the characterization of broadband die-package interaction for clean power delivery. Two major difficulties exist for the analysis of the aforementioned large-scale problems. One is the large problem size. The modeling of a die, a package, and a combined die-package system results in numerical problems of ultra-large scale, requiring billions of parameters to describe them accurately. The other is the multiscale nature of the problem. An electromagnetic simulator is required to span scale ranges of at least 10000:1 to analyze a combined die-package system. Recently, a time domain Orthogonal Finite-Element Reduction-Recovery method (OrFE-RR) was developed to handle the large problem size associated with the design of very large-scale integrated circuits [1]. In this method, a set of orthogonal prism vector bases are developed. With this set of bases, the original ultralarge scale 3-D system of order N is rigorously reduced to a 2-D single-layered system of order M with negligible computational cost, where M is much less than N. The reduced 2D system is diagonal, and hence can be solved readily. After obtaining the solution of the reduced system, the rest of the solutions can be recovered in linear complexity. This method entails no theoretical approximation. It applies to any arbitrarily-shaped multilayer structure involving inhomogeneous materials. Numerical experiments have demonstrated its accuracy and capacity in simulating very large-scale on-chip, package, and die-package interaction problems. The multiscaled nature of the integrated circuit system has not been addressed in the OrFE-RR method. The method projects the meshes in all the layers to the same layer. Namely, all the layers share the same mesh. This results in redundant mesh elements, and hence redundant number of unknowns. In this work, we extend the OrFE-RR method to model the orders-of-magnitude difference in scale of the die and the package by developing a multiscale modeling scheme. With such a scheme, we allow for the use of different mesh sizes in different layers and regions. As a result, we do not need to project the meshes in different layers to the same layer. In addition, we permit the modeling of the orders-of-magnitude difference in scale of a physical structure by using the minimum number of mesh elements.

Focusing on the problem in production practice of sintering process, a novel classifier based on BP learning algorithm is proposed for on-line quality inference of sintered ore. In order to speed up the convergence rate of BP learning algorithm, the learning algorithm with adaptive variable step-size is adopted. On the basis of the above work a quality prediction model is proposed in this paper. Experimental results indicate the higher prediction accuracy rate and better generalization ability of the model.

This paper presents an efficient finite difference time domain (FDTD) algorithm to calculate the transient electromagnetic fields above conductive ground near lightning channel. Significant computational time and memory savings are achieved by using surface impedance boundary condition (SIBC) to avoid simulating the underground region. Numerical results are shown to validate the proposed method.

The co-simulation of a die and a package results in numerical problems of ultra-large scale, requiring billions of parameters to describe them accurately. In this paper, an orthogonal finite-element reduction-recovery method is proposed to overcome the large problem sizes encountered in the simulation of an integrated die-package system. In this method, a set of orthogonal prism vector basis functions are developed to render the surface-unknown-based matrix in each layer diagonal. An arbitrary 3D multilayered system such as a combined package and die is then reduced to a single-layer system with negligible computational cost. The reduced single-layer system is diagonal due to the orthogonal bases, and hence can be solved readily. From the solution of the reduced single-layer system, the solution in other layers is recovered in linear complexity. The method entails no approximation. It applies to arbitrarily-shaped multilayer structures and arbitrary inhomogeneity. In addition, it permits nonlinear device modeling and broadband simulation within one run. Numerical and experimental results have demonstrated its accuracy and high capacity in simulating on-chip, package, and die-package interface problems.

A fast nearfield method for calculating the pressure generated by a polynomial apodized rectangular piston is obtained. The fast method is based on the instantaneous impulse response given by Scarano et al. A trigonometric transform of the integrand is performed and the order of integration is exchanged to obtain the 1D integral for the apodized FNM. Comparisons between the 1D polynomial apodized FNM and the Rayleigh-Sommerfeld integral show that the 1D polynomial apodized FNM is 18.58, 81.17, and 205.11 times faster than the Rayleigh-Sommerfeld integral for 10%, 1%, and 0.1% peak normalized errors, respectively.

A nesting combination of the non-overlapping domain decomposition and the algebraic multigrid method is developed in this paper for solving eddy current problems. The iterations between sub-domains in the domain decomposition method can be decreased greatly due to the application of the algebraic multigrid method. The finite element method with edge element is used for the subdomain computation. The harmonic fields in railgun are solved using this method, And the results show its efficiency and accuracy compared with ANSYS

Previous simulation studies have shown that the fast nearfield method (FNM) is faster and more accurate than the impulse response and rectangular radiator methods for nearfield pressure calculations. The FNM achieves this advantage by subtracting a singularity from a one-dimensional integral expression, which improves the numerical convergence relative to these other techniques. A related approach subtracts a singularity from the two-dimensional (2D) Rayleigh integral, thereby separating the Rayleigh integral into two parts. In this approach, one integral is evaluated analytically and the remaining terms are evaluated numerically. Numerical calculations are performed with a seven point Gauss quadrature rule. This approach is evaluated for a triangular source and compared to results obtained from FNM calculations. For larger triangular sources, the 2D Rayleigh integral is subdivided into smaller triangular patches, and then the results are superposed. Results show that for an equilateral triangle with sides equal to one wavelength, this approach achieves a 10% peak error without subdividing the source, and a 1% peak error is obtained when the source is subdivided into four smaller triangles. Computation times with this approach are 0.140s and 0.2190s for 10% and 1% peak errors, respectively, whereas the FNM results for the same peak errors are 0.031s and 0.047s for a reduction in the computation time by a factor of 4.52 and 4.66, respectively. This shows that FNM calculations converge more rapidly than the method that subtracts the singularity from the 2D Rayleigh integral. Future efforts will replace this 2D Rayleigh integral calculation with the FNM in large-scale simulations of ultrasound devices designed for thermal therapy applications

Planar ultrasound (US) phased arrays with square element are easier to fabricate and model than other concave shaped phased arrays. Combined with phases derived for single spot focusing, the sector-vortex phase scheme can be applied to the square planar array to generate a vortex shaped pressure field that can be controlled by the mode number M. Increasing the mode number M also increases the diameter of the annular focus. The location of the annular focus can be moved along the axis or off axis by steering the single spot focus. The weight of the phase for single spot focusing can control the axial length of the sector-vortex pattern. Superposing the SAR of multiple modes can heat large tumors with diameters up to 5 cm by broadening the temperature distribution.

From the view of pattern recognition, a novel fault location method based on impedance angle analysis at different frequencies is proposed in this paper. It firstly deduces the sign change rule of the impedance angle (Impedance Angle Reverse) at different frequency is related with the length from the sending end to the fault location. Further, from the theoretic analysis, the conclusion that the influence of transient resistance can be negligible in this method is shown. It samples the transient voltage at the sending end and the transient currents flow through the transmission line during the fault and analyzes the phase relation between voltage and current signals at different frequency with Fast Fourier transforms (FFT) algorithm. As a result, the impedance angle reverse at different frequency can be used to locate the fault. The accordance of the theoretic analysis and the simulating results indicates the feasibility of the method.

This paper is devoted to subjective/objective evaluation on medical image quality. A model based on neural network is proposed to mimic radiology doctors' perception. Medical images were compressed by different algorithms such as Dct and wavelet with different compression rate. After decompression, subjective and objective methods are used to evaluate image quality. Subjective rating is given by common person, clinic and radiology doctors. Objective properties are obtained from three categories: image difference from traditional parameters such as MSE, SNR and so on; image feature based on HVS model and image structure distortion from a universal image quality index. Then all data are used to train a BP neural network, whose inputs are objective scores and outputs are yielded to subjective rating by professionals.

In the optimization process of some ordering problems, we compare them to the traveling salesman problem (TSP) and put forward the conception of a virtual map. Based on this conception, the problems are all put into the framework of the dynamic traveling salesman problem (DTSP). The paper makes a lot of comparisons between TSP and DTSP about their characteristics, and finds out the commonality and differences between them. At the same time, we consider that the presentation of DTSP conception is attributed to the rise of modern optimization methods founded on probability, and evolutionary computation is one of the best algorithms to solve DTSP nowadays.

A mirror-symmetric feature extraction approach is proposed. When a point on the assumed symmetry axis is given, an arbitrary 1D signal is decomposed into a pair of odd and even components which maintain the information of absolute symmetry and dissymmetry, respectively. The average difference between the energy of the odd and even components reflects the degree of symmetry with regard to the given symmetry point. An objective function is obtained to search the most symmetric point in a 1D signal as an optimization process. The most likely mirror-symmetry axis is obtained from the set of symmetric points in all the rows in an image.

The effect of population isolation is discussed by means of an analysis of the domains of attraction of local optima. Separation among populations and adaptive gathering of the initial population are achieved by local evolution, so as to transform the multi-modal function optimization into a uni-modal function optimization. Combining the space-division-based (/spl mu/+1) selection approach, which has a rapid convergence speed in uni-modal function optimization, a new evolutionary algorithm is presented to automatically separate and gather the initial population according to its domains of attraction. Numerical simulation results show the global searching ability of the new algorithm.