54 results about "Low-rank approximation" patented technology

An apparatus and method of training a convolutional neural network (CNN) are provided. A method of training a CNN including a plurality of convolution layers stored in a memory involves approximating, using a processor, a convolution layer among the plurality of convolution layers using a low-rank approximation; reducing the number of output reconstruction filters of the approximated convolution layer; and modifying a structure of the CNN based on an approximation result and the reduced number of output reconstruction filters.

The invention provides a neural network training method, a neural network training device, a data processing method and a data processing device. The neural network training method comprises the following steps: S210, transforming a set of initial convolution kernels corresponding to each of at least one set of convolution layers of a convolutional neural network into a corresponding set of transformed convolution kernels by use of a low-rank approximation method; S220, training the convolutional neural network based on the transformed convolution kernels corresponding to the at least one set of convolution layers; S230, judging whether the trained convolutional neural network meets a predetermined standard, going to S240 if the trained convolutional neural network meets the predetermined standard, or going to S250; S240, decomposing the product of the set of trained convolution kernels corresponding to each of the at least one set of convolution layers into a corresponding set of compressed convolution kernels; and S250, taking the set of trained convolution kernels corresponding to each of the at least one set of convolution layers as a set of initial convolution kernels corresponding to the set of convolution layers, and returning to S210. Through the methods and the devices, the amount of computation can be saved.

A method and apparatus for spatio-temporal compressive sensing, which allows accurate reconstruction of missing values in any digital information represented in matrix or tensor form, is disclosed. The method of embodiments comprises three main components: (i) a method for finding sparse, low-rank approximations of the data of interest that account for spatial and temporal properties of the data, (ii) a method for finding a refined approximation that better satisfies the measurement constraints while staying close to the low-rank approximations obtained by SRMF, and (iii) a method for combining global and local interpolation. The approach of embodiments also provides methods to perform common data analysis tasks, such as tomography, prediction, and anomaly detection, in a unified fashion.

The invention discloses an MCMC sampling and threshold low-rank approximation-based image de-noising method. According to the method, firstly, during the denoising process, an image block is generated through the MONTE-CARLO sampling process. Secondly, based on a plurality of statistical features in a histogram, a similarity judging function that meets the condition of the Markov chain can be obtained. Thirdly, the singular value decomposition on all kinds of image block clusters is conducted and the self-adaptive threshold estimation for singular values is conducted according to the prior information corresponding to an image. Fourthly, on the basis of a decomposed low-rank structure, the image reconstruction is conducted according to the low-rank approximation algorithm, so that the de-noising purpose is realized. According to the invention, the characteristics of the similar non-local geometric structure information of images and the better treatment of high-dimensional data based on a low-rank matrix are fully utilized. Meanwhile, the defect in the prior art that the conventional non-local mean-value traversing search method is high in block-selecting complexity can be overcome. The block-selecting robustness is therefore improved. moreover, to a certain extent, the algorithm operating period is shortened.

The invention relates to a modal parameter recognizing method based on multi-input multi-output signal noise reduction. The method is characterized by comprising the steps of step 1, performing block-Hankel matrix conversion on noise-containing multi-point pulse responding signals to obtain a first block-Hankel matrix; step 2, processing the first matrix by utilizing structural low rank approximating according to the rank of the first matrix to obtain a first non-block-Hankel matrix after the low rank approximating; step 3, replacing the elements of each block in a second matrix with math average values of the block elements on a back-diagonal to obtain a second block-Hankel matrix; step 4, repeating the steps 2 and 3 until convergence, thus obtaining the noise-reduced multi-point pulse responding signal; and step 5, performing modal parameter recognizing. The modal parameter recognizing method based on multi-input multi-output signal noise reduction can save machine time and improve recognizing accuracy.

The present invention discloses an image de-noising processing method and device, and relates to the field of image processing technology. The objective of the invention is to solve the problem that a noise reduction method aiming at the magnetic resonance image design cannot effectively remove noise in the prior art. The method comprises: performing variance stabilization transformation of obtained nuclear magnetic resonance image signals, and obtaining transformation data with Gaussian distribution noise; performing estimation of the transformation data with Gaussian distribution noise through low rank approximation, and obtaining a reference image of reconstruction estimation and a nucleus Wiener filtering coefficient; performing convolution of the nucleus Wiener filtering coefficient and the transformation data with Gaussian distribution noise, and obtaining the noise reduction data with biased estimation; and performing unbiased inverse transformation of the noise reduction data with biased estimation, and obtaining a noise-free and unbiased estimation nuclear magnetic resonance image.

The invention discloses a multitask machine learning method and a multitask machine learning device both used for image classification. The method and the device are characterized in that low rank approximation of a residual structure and a covariance matrix of a regression matrix are utilized simultaneously, probability modeling is performed on the residual structure, the regression matrix, a low rank decomposition of the regression matrix and the covariance matrix of the regression matrix, learning of parameters of a probability model is performed through a variational deduction method or a sampling method, and a regression matrix high in accuracy is acquired finally and used for image classification. By the scheme, on one side, correlativity information among multitasks in the residual structure is utilized, so that parameter learning accuracy can be improved to improve classification accuracy; on the other side, by performing low rank approximation on the covariance matrix of the regression matrix, calculating complexity of an algorithm can be effectively lowered.

The invention provides an interference sample eliminating method based on a generalized inner-product arbitrary array, and solves the technical problem that the echo sample of an arbitrary array configurational antenna cannot be selected at present. The method comprises the following implementation steps: using a radar transmission signal, and selecting corresponding echo data as training samples; constituting the low-rank approximation matrix Ur of a clutter subspace in an off-line manner; calculating the inverse matrix of a clutter covariance matrix in the off-line manner; calculating the generalized inner-product value of each training sample; setting a detection threshold eta; carrying out elimination on disturbance samples, screening the training samples, and finally obtaining samples to be measured after the elimination of the disturbance samples so as to carry out space-time two-dimensional self-adaption processing in next step. The inverse matrix of the clutter covariance is constructed in the off-line manner, wherein the inverse matrix contains the position and beam pointing information of each array element, so that interference samples under the arbitrary array configurational antenna can be eliminated, the processing result is not influenced by the training samples, and the computational burden is low. The method is used for onboard and interspace radar space-time two-dimensional signal processing.

Systems and methods for generating a low-rank approximation of a matrix are provided. The various systems and methods may identify at least a first set of right singular vectors and a first set of singular values of a subset of the matrix, reduce the subset by an amount of energy of a selected data entry of the subset based on the first set of right singular vectors and the first set of singular values, incorporate a new data entry from the matrix into the subset, update the first set of right singular vectors and the first set of singular values of the subset based on the new data entry by a singular value decomposition (SVD) update, and generate the low-rank approximation of the matrix based on the updated first set of right singular vectors and the updated first set of singular values.

A system and method is provided for generating a map of a tissue property in a subject using magnetic resonance fingerprinting (MRF) and a compressed MRF dictionary, where the compressed MRF dictionary has a significantly reduced memory requirement relative to a standard MRF dictionary. The method includes performing a randomized singular value decomposition (rSVD) on a MRF dictionary to produce the compressed MRF dictionary. MRF data is then acquired and compared to the MRF dictionary to identify the tissue property from the region of interest in the subject. A tissue property map is then generated based on the tissue in the region of interest of the subject.

A high-resolution image reconstruction method based on a low-rank tensor and hierarchical dictionary learning is provided. The method comprises steps of: up-sampling and down-sampling an original image by using bilinear interpolation, and using a processed result and the original image as a dictionary learning training set; training the original image and a down-sampled image, extracting the down-sampled image gradient, arranging the original image and the down-sampled image gradient as the tensor, and subjecting the down-sampled image gradient tensor to low-rank approximation; subjecting the original tensor and the approximate down-sampled gradient tensor to sparse dictionary learning to obtain an image restoration dictionary; training a low-resolution image and the up-sampled image, extracting the low-resolution image gradient, arranging the low-resolution image gradient and up-sampled image as the tensor, learning to update the dictionary; transferring the original image to a YCbCr space, reconstructing the Y with the dictionary, reconstructing the Cb and the Cr by bilinear interpolation to obtain the original restored image; and subjecting the original restored image to global enhancement by using iteration back projection to obtain a final result. The method retains the structure information of the image by using the tensor and improves the precision of image reconstruction.

In a computerized data mining context, user input relating to positive and negative information is incorporated into node proximity measurements on a weighted, directed graph. Random walk results are updated without a full matrix inversion by using selective update to a low rank approximation and to inversion results.

The invention discloses a micro-video popularity prediction method based on low-rank constraint and multi-view characteristic fusion. The micro-video popularity prediction method includes the steps that four view modal characteristics are subjected to low-rank approximation processing respectively, and four kinds of noise-removing low-rank characteristic information are obtained; through multi-view-information typical correlation analysis, the four kinds of low-rank characteristic information are subjected to characteristic fusion; fused characteristic information is adopted, laplacian matrixes for expressing graph relationships between all micro videos are established; based on the laplacian matrixes, the popularity of the micro videos is predicted with the semi-supervised method. According to the micro-video popularity prediction method based on the low-rank constraint and multi-view characteristic fusion, the limitation of single-view characteristic on popularity prediction is avoided, the characteristics of all the views are processed through low-rank approximation processing, and the laplacian matrixes established between the characteristics have the higher stability.

The invention discloses a video denoising method based on a scale mixture model and a low-rank approximation, which mainly solves the problem that it is difficult to accurately remove Gaussian pulse mixture noise in the prior art. The scheme is: 1. Use the median filter method to obtain the initial estimate of the video, and find similar image block matrices in the front and rear frames of the test image; 2. Use the Laplacian scale mixture model to model the abnormal point set, and the abnormal points The problem of point estimation is transformed into a joint solution of outliers and hidden factors to remove outliers caused by mixed noise; 3. Perform low-rank approximation to similar image block matrices, and use non-local low-rank models to calculate denoised images; 4 .Use the Laplacian scale mixture model and the non-local low-rank model to iteratively calculate the restored single-frame image; 5. Repeat 1‐4 to obtain the denoised video. The invention can remove mixed noise, retain image detail information, have better visual effect, and can be used for denoising video multimedia, remote sensing images and medical images.

The exemplary embodiment disclosed herein comprises several aspects of defining and using optimal color test patches for calibrating a printing system for various purposes (e.g., multi-media, gamut extension, drifting correction, etc.). One aspect is the creation and use of color patches where the color patches are selected optimally to result in minimum mean square error when those color patches are used to determine weights of a limited number of basis vectors that model the system (e.g., model the print engine response). A joint optimization technique is given as a method to select the optimal color test patches for calibration.

The invention discloses a proton exchange membrane fuel cell non-linear state space model identification method. According to the method, firstly, a hydrogen flow and a load current are selected as input variables, a voltage is selected as an output variable, and plenty of data is acquired; secondly, a Hankel matrix is constructed through the acquired data, and a dual matrix is solved; thirdly, amatrix projection is constructed through utilizing the result of a matrix equation; and lastly, inclined projection is decomposed through utilizing a singular value to acquire state sequence estimation of a system. The method is utilized repeatedly to solve the system matrix, and the low-rank approximation technology is utilized to acquire non-linear system characteristic estimation; actually-measured input data is taken as an abscissa, a non-linear identification characteristic is taken as an ordinate, and a MATLAB curve fitting tool is utilized to carry out polynomial fitting of the system non-linear characteristic. The method is advantaged in that the non-linear characteristic of a proton exchange membrane fuel cell can be quite accurately described, not only can solution schemes be provided for proton exchange membrane fuel cell modeling of actual engineering personnel and subsequent control system design, and relatively good reference values are for modeling of similar systems.

Disclosed in the invention is an abnormal event detection method based on low-rank approximation structured sparse representation. The method comprises three steps of feature extraction, training and testing. To be specific, step (1), a multi-scale three-dimensional gradient feature of a video sequence is extracted; step two, dimensionality reduction is carried out on the multi-scale three-dimensional gradient feature, thereby forming a training feature set and testing feature set; step three, a residual training feature and a correlated parameter are initialized; step four, iterative learning group sparse dictionary is carried out on the residual training feature, thereby obtaining a normal-mode dictionary; step five, sparse reconstruction is carried out on a testing feature by using the group sparse dictionary obtained by the training process; and step six, according to a reconstruction error, whether the testing feature is an abnormal feature is determined. Therefore, defects that the low-rank characteristic of video data is not dug fully and the detection is carried out slowly according to the abnormality detection technology can be solved.

The invention provides a video image denoising processing method and an apparatus. The method comprises the following steps of using a preset denoising algorithm to carry out pre-denoising processing on a first video image to be processed so as to acquire a reference video image; determining a reference image block group possessing a similarity with an image block corresponding to an image block to be processed in the reference video image, mapping a coordinate position relation of the reference image block group into the first video image from the reference video image so as to acquire an image block group which is similar with the image block to be processed; and carrying out low rank approximation processing on the image block group in the first video image so as to acquire a second denoising video image. The reference video image after pre-denoising processing is taken as a basis of similarity measurement so that a similarity measurement result is accurate and quality of the final second denoising video image can be guaranteed. Based on a low rank approximation mode, the final second denoising video image is acquired so that a denoising effect of the video image is increased.

The invention provides a tire defect detection method based on singular value decomposition. A to-be-detected image I containing n pixels is divided to n m*m image blocks; each image block Pi is converted to a column vector ci; column vectors corresponding to all image blocks are used to construct an image block matrix M; through sequentially calculating the ratio of adjacent two singular values, the rank r of the image block matrix M is determined; the former r maximum singular values of the image block matrix M and the corresponding left singular vectors and right singular vectors are used to reconstruct a low-rank approximation matrix Mr of the image block matrix M; each column vector in the low-rank approximation matrix Mr is converted to an image block; all image blocks P<r> are used to reconstruct an approximate image Ir of the to-be-detected image I; hard threshold segmentation is carried out on a residual image I-Ir, a binary image Ib is obtained, and coordinates corresponding to a pixel with a gray value to be one in the image are the defect position. The detection method can position the specific defect position, missed detection and error detection caused by manual operation are avoided, and the defect detection accuracy is improved.