Abstract:Due to its good property to provide an approximation to any distribution, GMM has been widely applied in the field of pattern recognition. Usually, the iterative EM algorithm is applied to estimate GMM parameters .The computational complexity at model training procedure will become very high when large amounts of training data and large mixture number are engaged. The CUDA technology provided by NVIDIA Corporation can perform fast parallel computation by running thousands of threads simultaneously on GPU. In this paper, a fast GMM model training implementation using CUDA is presented, which is especially applicable to large amounts of training data. The fast training implementation contains two parts, the K-means algorithm for model initialization and the EM algorithm for parameter estimation. Furthermore, this fast training method has been applied in language GMMs training. The experimental results show that language model training using GPU is about 26 times faster on NVIDIA GTS250 when compared to traditional implementation on one of the single core of Intel DualCore Pentium Ⅳ 3.0GHz CPU.

Abstract:As the precondition of correcting VCO nonlinearity precisely, the detection of VCO nonlinearity is worth to study theoretically and practically. For the noise immunity and the real-time performance of the method based on element analytic and FrFT still can not satisfy the desire, a VCO nonlinearity detection method based on overlapped subsection and FrFT is proposed. The noise immunity is improved by overlapped subsection. The real-time performance is enhanced by adopting the fast converged golden section method to search the peak value in fractional Fourier transform domain. Simulation shows the better noise immunity of the proposed method. The RMSE of the proposed method is about 1/2 of the present method when the overlapped ratio is 0.25 and the real-time performance of the proposed method will exceed the present one’s while the overlapped ratio is less than 2/3.

Abstract:A new weight-fusion algorithm was proposed for frequency estimation of multi-sinusoids with the known frequency-ratio and the same length (Multi-Sinusoids-KFRSL), due to its significant research value and disadvantages of its present frequency estimation methods. According to the known frequency-ratio, the frequency-ratio amend matrix is given, in order to make spectrums of the Multi-Sinusoids-KFRSL as the same as those of multi-sinusoids with the same frequency and length. For the phase-incoherent problem of the Multi-sinusoids-KFRSL, the phase compensating matrix is created to make phase coherent and noise eliminate. Next, spectrums of the Multi-Sinusoids-KFRSL are weight-fused by the phase compensating matrix to turn almost the same as the spectrum of the phase-coherent sinusoid signal, which is as the same length as the Multi-Sinusoids-KFRSL. Consequently, precise frequency estimation can be obtained through spectral peak searching of the weight-fusion spectrum. Algorithm analyses and simulation results show that, compared with the present methods, the proposed algorithm works better in term of precision, calculation, noise immunity, significant theoretical and practical value.

Abstract:Subspace technique is an efficient method for automatic facial attractiveness analysis. To enhance the intrinsic description for facial attractiveness, a dual subspace method on the subspaces of PCA and Generalized Low Rank Approximation Matrix (GLRAM) is proposed. Thus, their individual characteristics in characterizing the global and local intrinsic description of facial attractiveness can be collaboratively boosted. In addition, the Gaussian Field model (GF) is applied to reflect the geometry structure in sample space. The experiment is carried on a challenging database, which takes on significant variations in the aspects of illumination, background, facial expression, age, race, and so on. The experimental results show the advantages of the proposed dual subspace method for facial attractiveness analysis over individual subspace.

Abstract:To solve the localization precision descending about ME algorithm when the cut angles between partial sensor pairs are too small or big in a multi-sensor system, a multi-sensor management passive localization algorithm based on virtual measurement transform (VMT) is presented. The effect of cut angle on the GDOP is analyzed in global coordinate system firstly, and the cut angle restriction relationship is obtained from which the better localization precision of dual-sensor system can be achieved. Secondly, the VMT localization algorithm is presented when certain sensor pairs do not satisfy above restriction relationship, which can deploy sensor in an optimal way by means of sensor space management. The principle and characteristics of the algorithm are analyzed, especially the precision of intersection which after and before transform are compared from the perspective of GDOP. Simulation results indicate that the performance of VMT algorithm excels that of ME algorithm remarkably, which verifies it and the important role that sensor management plays in multi-sensor passive localization.

Abstract:When a radar target contains rotating structures, such as a rotor on a helicopter, the target’s body image will be contaminated due to the micro-Doppler effect generated by additional frequency modulations on the received signals. In this paper, an imaging algorithm is proposed for targets with rotating parts based on the Compressed Sensing theory. Because the target’s body almost keeps relative changeless in the imaging region and the rotating parts migrate in the imaging region during the imaging time, the image of main body can be reconstructed by the OMP algorithm, and also the interference generated by the rotating parts is eliminated. Because of compressed sampling, the datum of the returned signal is reduced. Finally, the computer simulations are given to prove the effectiveness of the proposed method, and the performance of the algorithm with different SNR is analyzed.

Abstract:Fingerprint segmentation has been considered as one of the critical processes of the automatic fingerprint identification system. Following the analysis of the relationship between the second order statistical characteristics and the grey-scale level, the offset value and the relative direction, an innovative fingerprint segmentation algorithm based on the gray level co-occurrence matrix (GLCM) is thus presented. Firstly, the fingerprint is split into a number of rectangular blocks to get the contrasts of GLCM for each in different directions. And then, to judge for whether a rectangular block is the prospect region or not, the proposed algorithm compares its variance of the contrast with the predefined threshold. The theoretical analysis and experiment results on the FVC2004 show that the proposed algorithm performs well and is robust in handling the varied qualities of fingerprint images collected in any circumstance.

Abstract:A Voice Activity Detection algorithm based on ensemble empirical mode decomposition domain statistical model is presented in this paper. The noisy speech is decomposed into Intrinsic Mode Function (IMF) components by using EEMD method. Two IMF components with the higher correlation with original speech are added to calculate statistical model characteristic parameter. The decision of the speech/noise is made by comparing characteristic parameter with threshold. The proposed VAD algorithm is tested on speech signals under various noise conditions with several SNRs. The results of experiments show that the proposed VAD algorithm outperforms some standard VAD algorithms, especially under low SNR noisy condition.

Abstract:The line search corner detection approach shows ordinary adaptability to uneven illumination for its global thresholds of parameters. In this paper, an adaptive parameter improvement based on the line search corner detector is put forward. The main contribution is the design of the adaptive thresholds of Univalue Segment Assimilating Nucleus (USAN) in every steps of the line search corner detector based on the local image contrast. The Gaussian smoothing is introduced for noise suppression. The dynamic values of other parameters are suggested. Experiments show that the parameter adaptation for line search corner detector proposed in this paper has good adaptability to uneven illumination and can keep high detection rate among various real images, while false response is acceptable.

Abstract:In order to decrease the length of the adaptive filter, this paper proposes a statistic model for adaptive algorithm in its early iterations as well as a novel algorithm for echo cancellation. The statistic model analyzes the expectation and variance of each coefficient of filter in the early iterations of adaptive algorithm. The modified algorithm based on this model identifies the location of the peak of the echo path and makes an estimation of the bulk delay. After the estimation a shorter adaptive filter centered about the peak coefficient is used to approach only the active coefficients instead of the whole echo path. Simulations with real echo path and theory both show that the peak coefficient is discriminated and the estimation of delay can be made in the early 75~100 iteration. A short filter is used to identify the echo path, which results in faster convergence speed and lower computational complexity.

Abstract:This paper presents a novel frequency offset estimation algorithm based on the linear regression. Frequency offset estimation is converted to the estimation of argument increment which satisfy the linear relationship. The algorithm correlates the signal received with local PN sequences, and then calculates the frequency offset based on the linear regression of arguments of the correlation results. Simulations indicate that the algorithm has advantages of accurate estimation over multi-path channels and excellent performance of channel noise suppression. The accuracy of the algorithm is not sensitive to frequency offset in the effective range. In addition, the algorithm reduces system resources by sharing the same PN sequence and correlation results with PN synchronization and can be easily implemented.

Abstract:The adaptive phase shift compensator is introduced to estimate the frequency difference of arrival (FDOA), by aligning the time-varying phase shift adaptively. According to the different computation methods, two estimation algorithms，the algorithm based on time average and the algorithm based on linear fitting，are proposed by using adaptive compensation factors. The estimator based on time average is asymptotically unbiased, with controllable error, small mean square error and low computational complexity. Simulation results also show that the performance of the algorithm based on linear fitting is close to CRLB when signals’ SNR are more than -3 dB.

Abstract:Unburnt carbon detection in fly ash is the key target in fly ash using. Comparative studied three algorithms, FFT frequency domain analysis, autocorrelation, virtual digital lock-in amplification based on cross-correlation. The results show that the data of using virtual digital lock-in amplification measurement detect standard fly ash carbon content was more accurate and reliable, the deviation and variance were relatively smaller, less than 3%. Measuring the carbon content fly ash sample of 0.5%-4%, photoacoustic signal voltage value vs. carbon content was a more intuitive and clear linear relationship. Virtual digital lock-in amplifier technology can detecting the unburnt carbon content in fly ash accurately, simply and fastly . It’s great significant for energy conservation and environmental protection.

Abstract:Health diagnosis of aero-generator is an issue of great importance to flight safety. One of the requirements is availability of extracting useful information from raw data. This paper presents a health diagnosis method based on neighborhood rough sets theory and support vector machine(SVM). Raw data were obtained from specific aero-generator test platform. An approach of attribute reduction using neighborhood rough sets theory is outlined and a diagnosis classifier is designed based on SVM to further carry out health diagnosis of aero-generator. The effectiveness of the proposed method is demonstrated through an experimental research. Result shows a better performance of the classifier that uses attributes reduction subsets as inputs, and also indicates the method may have wide popularization and application potential.

Abstract:Vehicle detection and recognition is of great importance to the development of Intelligent Transportation System(ITS)，but Target Recognition is a challenging problem for low_resolution Radar. This paper proposes a Vehicle Recognition approach using Doppler Radar, and the spectrum variation of one vehicle reflects its outline. Then, the dimension of effective spectrum feature can be reduced by the methods of Principal Component Analysis(PCA) and Linear Discriminant Analysis(LDA), then vehicles can be classified into three types by classifier algorithms such as Support Vector Machine(SVM), k-Nearest Neighbor(KNN). At last, the paper compares the experiment results of different algorithms by cross validation, and shows the algorithm based on PCA-LDA-SVM can achieve ideal result.

Abstract:Unaided acquisition of weak GPS signals requires both long coherent integration time and a great amount of noncoherent integration operations. The unknown navigation data bits and bit edges should also be considered. From the point of noncoherent integration of unaided weak GPS signal acquisition method, a new alternative is proposed. Besides the new one, four different noncoherent integration alternatives are introduced. They use different approaches to compute the noncoherent integration at each step in order to reduce calculations and save memory space. Under the same coherent integration time, noncoherent integration operation number and navigation data bit edge number, but different carrier-to-noise ratio, Monte Carlo method is applied to verify the probability of detection and do comparison analysis of the 4 alternatives.

Abstract:To analyze local condensation speciality of signal energy meticulously, the concept of energy condensation point on level of signal was presented. The scale factor and level factor in the definition of energy condensation point were used to regulate the precision and sensitivity of signal analysis, respectively. A method for analyzing signal feature was presented based on the energy condensation point, and it was used in signal detection of buried target. The result shows that the method helps a lot for improving signal detection properties.

Abstract:Firstly, a mathematical model of asymmetrical bi-directional multi-relay system is presented. Using opportunistic relay and network coding, an outage probability expression of the mode is given,which is proved accurate by simulation. Secondly, we analysis the outage performance of the model at different relay node location, different number of relay node, and different power allocation coefficient, indicating the inherent relationship between power allocation coefficient and relay node location, the number of relay node, and the total power of the system. Simulation results indicate that optimal system outage performance of asymmetrical bi-directional multi-relay system is attained when power allocation coefficient is chosen at 0.6 or about 0.6 while using Network Coding and opportunistic relay, and indicating that the location of relays is important while investigating the problem of outage probability.

Abstract:Compressed sensing is a new paradigm in signal processing that trades sampling frequency for computing power and allows accurate reconstruction of signals sampled at rates many times less than the conventional Nyquist frequency. Today, modern radar systems operate with high bandwidths and high resolution. Compared to complex radar system and mass data, often only a small amount of target parameters is the final output. Compressed sensing is one of good means to effectively reduce data size. This paper reviews the latest developments of compressed sensing in radar target detection and recognition and introduces the key technical problems of design of measurement matrix and reconstruction algorithm for sparse signal. Several possible applications are considered: PD radar, through wall radar, MIMO radar, radar target parameter estimation, radar imaging and radar target detection and recognition system. Then this paper also discusses the existing difficult problems in the study and looks into the future research directions on compressive sensing applied to radar.

Abstract:Considering fault feature extraction difficulty to the non-linear vibration signals,a feature extraction method is proposed based on the general dimension mean(MeanDq)and the parameters of singular spectrum(^a and ^f).Firstly，characteristic of multi-fractal for the vibration signals were analyzed,then MeanDq 、^a and ^f was calculated respectively used as a fault characteristic value, Finally，fault feature extraction method applies to fault detection for rolling bearing.Study shows that state of the vibration signals for rolling bearing are identify effectively with MeanDq 、^a and ^f used together,besides,MeanDq and ^a the have a stronger sensibility than ^f.The example proves that this integrated method is feasible.