51results about How to "Improve Noise Robustness" patented technology

A noise reduction device is configured by use of: means for calculating a predetermined constant, and a predetermined reference signal Rω(T) in the frequency domain, respectively by use of adaptive coefficients Wω(m), and for thereby obtaining estimated values Nω and Qω(T) respectively of stationary noise components, and non-stationary noise components corresponding to the reference signal, which are included in a predetermined observed signal Xω(T) in the frequency domain; means and for applying a noise reduction process to the observed signal on the basis of each of the estimated values, and for updating each of the adaptive coefficients on the basis of a result of the process; and an adaptive learning means and for repeating the obtaining of the estimated values and the updating of the adaptive coefficients, and for thereby learning each of the adaptive coefficients.

The present invention provides a method and apparatus for detecting speech spectral peaks and a speech recognition method and system. The method for detecting speech spectral peaks comprises detecting speech spectral peak candidates from power spectrum of the speech, and removing noise peaks from the speech spectral peak candidates according to peak duration and / or peak positions of adjacent frames, to detect speech spectral peaks. In the present invention, reliable speech spectral peaks can be obtained by removing noise peaks using the limitations of peak duration and adjacent frames in the detection of the speech spectral peaks. Further the energy values of the speech spectral peaks are used to extract the MFCC feature of speech instead of a sample sequence of the whole power spectrum in the conventional technique, the noise robustness of speech recognition can be enhanced while not increasing the speech feature dimensions.

The present invention relates to implementing a system and method to improve speech recognition and speech enhancement of noisy speech. The present invention discloses a way to improve the noise robustness of a speech recognition system by providing additional input to a Neural Network speech classifier. The additional information characterizes the noise environment of the speech. The present invention further discloses a speech separation system that uses the output of the neural network. The speech separation system employs models for the speech and for the distractor or noise. The neural network is used to identify the most likely combinations of speech and noise. Furthermore, a system for efficiently finding the most likely clean speech log-spectrum value is disclosed.

The invention discloses a speaker recognition method based on a Gaussian super vector and a deep neural network. The method comprises a speaker feature extraction stage, a deep neural network design stage, and a speaker identification and decision-making stage. According to the invention, the deep neural network is fused with a speaker recognition system model, and the obvious effect of a multilayer structure combining the Gaussian super vector and the deep neural network in the aspect of improving the characterization capability of an evaluation model is achieved. The speaker recognition method provided by the invention can effectively improve the recognition performance of a system in the environment of background noise, reduces the influence of the noise on the system performance, improves the robustness of the system noise, optimizes the system structure, and improves the competitiveness of a corresponding speaker recognition product.

The invention provides a satellite navigation interference signal type intelligent identification method and system, and the method comprises the steps of carrying out the medium-frequency digital signal collection of a satellite interference signal, and carrying out the preprocessing; carrying out shallow feature extraction on the preprocessed digital intermediate frequency sampling signal by using a PWVD transformation algorithm; taking a high-dimensional time-frequency domain feature vector extracted from an unknown interference signal as an input sample; extracting a high-dimensional time-frequency domain feature vector and a corresponding category from the known interference signal to construct a training set; training an improved GoogLeNet convolutional neural network model by usingthe training set; performing deep feature extraction and type judgment on an input sample by using the trained improved GoogLeNet convolutional neural network model to obtain a category to which an unknown type belongs. According to the invention, a traditional identification algorithm in the field of satellite navigation interference identification at present is greatly improved, and reference isfurther provided for realizing real-time interference detection and identification by combining software and hardware.

The present invention discloses a part identification and positioning method based on a mixed model, belonging to the field of image processing technology. The method comprises the steps of: S1, collecting part images, identifying parts in the part images by employing a template matching algorithm, and obtaining one pose corresponding to each part; S2, extracting one edge point set of each identified part, and extracting an edge point set of a template image and a direction vector thereof; S3, based on a Gaussian mixture model, constructing one likelihood function of geometrical errors betweeneach part edge point set and the edge point set of the template image; S4, employing an EM algorithm to perform optimization of the constructed likelihood functions, and obtaining optimal values of the geometrical errors; and S5, employing the optimal values of the geometrical errors to perform correction of the poses of the corresponding parts, obtaining one corrected pose of each part, and identifying each part in the part images. The part identification and positioning method based on the mixed model has accurate identification and positioning effects and has good robustness.

The invention relates to a speaker identification method based on the deep stack autoencoder network. The method comprises steps of S1, speaker feature extraction; S2, stack autoencoder network design; and S3: speaker identification and decision making. The method is advantaged in that compared with traditional speaker identification, the deep stack autoencoder network is fused with a speaker identification system model, in combination with the multi-layer structure of a stack autoencoder to improve the characterization ability of an evaluation model, system identification performance in the presence of background noise can be finitely improved, influence of the noise on the system performance is reduced, system noise robustness is improved, the system structure is optimized, and identification timeliness is effectively enhanced.

The invention relates to a speaker understandability detection method of an artificial cochlea signal under the noise environment, and belongs to the field of voice signal processing. Firstly the artificial cochlea processing algorithm is performed on a pure reference voice signal and waveform reconstruction is performed so that the pure voice after artificial cochlea processing can be obtained. Then the voice model of the specific speaker is established after feature extraction; and as for the identification phase, the voice and the noise are purely identified and then the identified voice with noise is formed, and matching with the speaking understandability model is performed after feature extraction so that the final detection result can be obtained. The advantages are that the important theoretical basis can be provided for enhancing the voice perception capacity of the artificial cochlea user, the influence of the noise in the matching process can be reduced, the detection accuracy can be enhanced, and the noise robustness of the detection method can be further enhanced by using the combined feature parameters based on the dynamic Gammachirp filter bank.

The invention discloses a noise-robustness acoustic modeling method based on posterior knowledge supervision and belongs to the voice man-machine interaction technology field. The method comprises thefollowing steps of acquiring the posterior probability distribution of a clean voice through the training of a teacher model; and taking the posterior probability distribution of the clean voice as astandard so as to supervise the training of a student model so that the student model infinitely approaches the posterior probability distribution of the teacher model, wherein the teacher model is aclean voice training model and the student model is a voice training model with a noise. By using the modeling method, an established acoustic model has high environment robustness and shows excellent anti-noise performance.

The invention provides a method for quantitative digital microscopic phase contrast imaging, which is realized on a normal digital optical microscope platform and a general personal computer and comprises the following four steps of: step one: establishing a light intensity transmission equation and carrying out discretization treatment on the equation; step two: designing a discretized light intensity transmission equation numerical algorithm; step three: collecting the light intensity data; and step four: carrying out quantitative phase recovery and phase contrast imaging. The method is on the normal digital optical microscope, needs not to increase complex optical parts, realizes the phase contrast imaging on the basis of phase recovery numerical algorithm, improves the contrast of the transparent or semitransparent sample imaging, and quickly realizes quantitative microscopic phase contrast imaging and quantitative phase analysis. The method is also applicable to a microscopic imaging system of other modes and has wide practical values and application prospects in the technical field of microscopic optical imaging.

An utterance detector for speech recognition is described. The detector consists of two components. The first part makes a speech / non-speech decision for each incoming speech frame. The decision is based on a frequency-selective autocorrelation function obtained by speech power spectrum estimation, frequency filter, and inverse Fourier transform. The second component makes utterance detection decision, using a state machine that describes the detection process in terms of the speech / non-speech decision made by the first component.

The invention discloses an MIMO underwater acoustic channel estimation method based on an improved adaptive hybrid optimization smoothing L0 norm. An improved algorithm (MAReSL0) of an objective function with a better approaching L0 norm is adopted to adaptively generate a reliable regularization factor, and an equilibrium iteration process is based on the sparseness of a solution in the objectivefunction and the error of residual; and in order to ensure the accurate convergence of the iteration to an optimal point, an initial value of the Nesterov gradient acceleration method (NAG) subjectedto inner loop iteration is used as the initial value of the newton method to perform hybrid optimized channel estimation. Numerical simulation indicated that the method is has better robustness thanother classical algorithms and effectively improves the channel estimation performance.

The invention discloses a variable working condition bearing fault diagnosis method based on a multi-scale dispersion entropy partial mean value and nonlinear mode decomposition, which belongs to thetechnical field of fault diagnosis. The method comprises the following steps of: acquiring a fault original signal of a to-be-diagnosed bearing under a variable working condition, decomposing the acquired original fault signal by adopting nonlinear mode decomposition, calculating the obtained component by adopting a multi-scale dispersion entropy index to obtain a multi-scale dispersion entropy partial mean value, selecting a component corresponding to the maximum partial mean value, and identifying the fault type of the bearing by adopting order analysis. According to the method, the variableworking condition bearing fault signals are decomposed through nonlinear mode decomposition, the multi-scale walk entropy index-partial mean value is adopted to select the component containing the most fault information, and the fault type can be accurately judged.

The invention discloses a multi-antenna broadband spectrum detection method based on compressed sensing and entropy, and the method comprises the specific steps: carrying out the sparse representationof a signal through employing a discrete fractional Fourier transform (DFRFT) algorithm; secondly, reconstructing a sparse coefficient based on a multi-task compressed sensing technology of a hierarchical Bayesian probability model by sharing priori information of sparse coefficients among array element receiving signals, so as to obtain an original signal; and finally, carrying out frequency spectrum judgment on the frequency domain statistics based on the information entropy. According to the method, on one hand, the single-node detection capability can be improved, the measurement uncertainty is solved, the signal reconstruction precision is improved, on the other hand, the problem of noise uncertainty is solved, and the noise robustness of the system is enhanced.

The invention relates to a fingerprint texture image denoising enhancing method. According to the method, a non-local diffusion tensor diffusion-fluctuation partial differential equation model is provided. The method comprises the following steps that an image is preprocessed; the non-local structure tensor is calculated; the non-local diffusion tensor is calculated through the non-local structure tensor; iteration updating is carried out, and a result image is obtained. The texture characteristic can be enhanced while the noise is eliminated, a certain repairing function of the fracture texture is achieved, and the fingerprint recognition can be carried out more accurately in the later process.

The invention relates to a video anomaly detection method and system based on generation of a collaborative discrimination network. The method comprises the steps: collecting normal video monitoring data, converting the normal video monitoring data into an original frame, selecting an original future frame, and adding noise into the original future frame to obtain a noise future frame; inputting the original frame into a generator to obtain a predicted future frame; calculating optical flow information between the predicted future frame and a previous frame of the original future frame; calculating the optical flow information between the original future frame and the previous frame of the original future frame and the difference between the two frames, inputting the predicted future frame and the original future frame into a discriminator, and constructing a target function of the discriminator; inputting the noise future frame and the predicted future frame into a coordinator, and constructing a target function of the coordinator; constructing a target function of the generator; updating the generator, and determining a predicted future frame during testing; and calculating an abnormal score, and determining whether a to-be-detected frame is abnormal or not according to the abnormal score. According to the method, the detection precision of the anomaly in videos can be improved.

The invention discloses an image focusing measure implementation method based on block PCA, and belongs to the technical field of information processing. The method comprises the following steps: firstly, preprocessing an image, then partitioning the processed image, and carrying out PCA calculation on each sub-image to obtain a characteristic value corresponding to a principal component vector ofthe corresponding sub-image; and constructing the definition of the sub-images through the feature values, and finally, solving the variance of a set formed by the definitions of all the sub-images,and taking the variance value as the focusing measure of the whole image. According to the method, the image detail information is extracted by adopting the image blocks and the PCA, the method has the advantages of simple principle and low calculation complexity, the interference of noise on the image detail information is reduced through the implementation of the steps, and particularly, the noise robustness is high under a low-contrast imaging condition. The method is suitable for a passive imaging system of a camera and has high popularization and use value.

The invention belongs to the field of radar signal processing, and discloses a multi-baseline phase unwrapping method based on a plane approximation model. The method comprises steps: an interferencephase diagram is acquired; the wrapping phase gradient of the interference phase diagram is calculated; fuzzy number gradient estimation based on plane approximation is then carried out; a fuzzy number is solved through an MCF model; the absolute phase is calculated according to the fuzzy number and the wrapping phase; and a phase unwrapping result is obtained. Thus, the problem of poor noise robustness for multi-baseline phase unwrapping under noise interference can be solved.

The invention is suitable for the technical field of deep learning, and provides an online learning method and device for a deep learning model, equipment and a medium. The online learning method comprises the following steps: performing offline training on the deep learning model into which the suppression signal is introduced; after training, publishing the deep learning model on a line; enabling the received online training image to be subjected to image identification; cutting an unrecognizable online training image through a sliding window; obtaining a corresponding basic feature set; carrying out similarity matching on the obtained basic feature set and a training image set; setting the basic features corresponding to the similarity lower than a similarity threshold value in all theobtained similarities as singular features; and according to the singular feature set formed by the singular features and a preset model training algorithm, training the deep learning model again to complete online learning of the deep learning model, thereby improving the noise robustness of the deep learning model by introducing a suppression signal, and improving the model recognition precisionthrough personalized training.

The invention discloses a structural damage identification method based on ALO-INM and a weighted trace norm. The method comprises the steps of: building a structural finite element model comprising Nel units according to a model correction theory and a finite element principle, and calculating the first Nm-order inherent frequency and vibration mode of the model; respectively establishing an original objective function O([alpha]), a first conjugate objective function and a second conjugate objective function, namely O*(alpha) and O**([alpha]), of the structural damage identification constraint optimization problem according to the frequency relative change rate and the modal confidence criterion; and solving O**([alpha]) by using an ALO-INM algorithm to obtain a structural damage identification result. According to the invention, an INM local search strategy is introduced on the basis of a meta-heuristic algorithm, the global optimization capability of the algorithm is enhanced to a certain extent, a weighting strategy and trace sparse regularization are introduced into a target function, so that the recognition precision and the noise robustness are improved, the influence of damage sensitivity and noise of different structures on the recognition precision can be reduced, and the method has relatively strong global optimization capability, relatively high recognition precision and relatively good noise robustness.

The present invention relates to implementing a system and method to improve speech recognition and speech enhancement of noisy speech. The present invention discloses a way to improve the noise robustness of a speech recognition system by providing additional input to a Neural Network speech classifier. The additional information characterizes the noise environment of the speech. The present invention further discloses a speech separation system that uses the output of the neural network. The speech separation system employs models for the speech and for the distractor or noise. The neural network is used to identify the most likely combinations of speech and noise. Furthermore, a system for efficiently finding the most likely clean speech log-spectrum value is disclosed.

The invention discloses a compressed sensing signal reconstruction method and system based on dictionary double learning. The method comprises the following steps: obtaining a reconstructed signal sample according to a received signal and a finite equidistant principle-based observation matrix optimized via feature decomposition; performing signal processing on the reconstructed signal sample to obtain a compressed sample; carrying out signal reconstruction on the compressed sample according to a sparse dictionary through a preset robustness dictionary learning model to obtain a reconstructedsignal. The invention provides an unknown narrowband signal compressed sensing signal receiving method based on observation sparse double learning, and the method has objective robustness, adaptivityand cognitive ability at the same time.

The invention discloses an image focusing measure implementation method based on local edge detection, and belongs to the technical field of information processing. The method comprises the followingsteps: firstly, carrying out Gaussian filtering on an image, then partitioning the filtered image, and carrying out edge detection based on a moment function on each sub-image to obtain an edge imageof the corresponding sub-image; constructing the definition of the sub-images through the edge images, and finally, solving the variance of a set formed by the definitions of all the sub-images, and taking the variance value as the focusing measure of the whole image. According to the method, on the basis of Gaussian filtering and image blocking, edge detection based on a moment function is adopted to extract image detail information, the method has the advantages of being simple in principle and low in calculation complexity, through implementation of the steps, interference of noise on the image detail information is reduced, and particularly, the noise robustness under the low-contrast imaging condition is high. The method is suitable for a passive imaging system of a camera and has high popularization and use value.

An apparatus may include a driving information input unit for receiving driving information generated while a vehicle travels, a steering angle location control unit for receiving a command steering angle for autonomous driving and a current motor steering angle of a driving motor and outputting an autonomous driving command through location control, and a motor-driven power steering control unit for driving the driving motor based on the autonomous driving command in an autonomous driving mode, determining whether a driver intervenes in steering, based on the driving information during the autonomous driving, computing a driver command according to the driver' steering based on a result of the determination, computing a compensation output between the autonomous driving command and the driver command by applying a weighting according to a steering angular speed, and making a mode transition from the autonomous driving mode to a driver mode while driving the driving motor.

The invention discloses an electromagnetic vector sensor array angle estimation method based on double quaternions. The method comprises steps that the double quaternions are introduced on the basis of a quaternion model, and a time-domain tri-orthogonal electric dipole sensor array model is built; the orthogonal information between array element electric dipoles is effectively reserved, orthogonality between quaternions has higher constraint than vector orthogonality, and better angle measurement performance can be achieved. According to the method, arrival angle estimation is carried out through a double quaternion multi-signal classification method BQMUSIC, under the condition of polarization parameter priori, better angle measurement performance than that of LVMUSIC is achieved through the BQMUSIC, the operand is reduced, and meanwhile robustness of related noise is improved. In addition, the double quaternion method not only expands a component on the basis of a two-component sensor quaternion model, but also has the potential of solving other problems of the quaternion model.

The invention discloses a co-prime array beam forming method and system based on low tube rank tensor decomposition and a medium, and the method comprises the steps: 1, building a far-field narrowband incoherent signal model under a co-prime array, carrying out the initialization interpolation of a virtual array element signal, and rearranging the virtual array element signal into a multi-snapshot virtual array element signal matrix; 2, based on the low-tube-rank characteristic, the initialization tensor is complemented; 3, decomposing the complemented tensor to obtain three factor matrixes; 4, solving and matching the target angle and power by using the corresponding relation between each factor matrix and each forward slice of the complemented tensor; and 5, calculating a beam forming weight vector by using the parameter estimation value in the step 4. According to the technical scheme, the noise robustness of parameter estimation can be improved by using the multi-dimensional structure of the tensor; the calculation amount can be obviously reduced by using the alternating direction multiplier method and the rapid convergence of tensor decomposition; and the target direction is obtained through parameter inversion and is less influenced by a prior error.

The invention discloses a fingerprint positioning method based on a long and short time memory network model and an access point selection strategy, and the method at least comprises the steps: S1, collecting indoor RSS signal data, and S3, completing the screening of access point data based on a preset judgment strategy. S4, carrying out localized extraction on the features of a fingerprint database by using a sliding window; and S5, training the updated fingerprint database by adopting a long-term and short-term memory network model to obtain a trained network model so as to realize indoor positioning. The access point selection strategy is adopted to reconstruct the database so as to simplify the scale of the database. Data selected and output by the AP is processed by utilizing a feature point extraction method so as to extract features with relatively large information amount; and the extracted features are input into the LSTM network model, so that the effect of enhancing the noise robustness while reducing the calculated amount is achieved, the positioning problem in an indoor complex environment is solved, and the influence on the positioning precision in a non-line-of-sight transmission background is reduced.