68 results about "Regularization algorithm" patented technology

A method of acquiring scatter data and image projection data in computed tomography is provided that includes attenuating a radiation source using a pattern of blockers arranged to provide blocked and unblocked regions of the radiation source, and acquiring image data and scatter data of a target using an imaging device. A scatter map in the projection image can be estimated by interpolation and / or extrapolation of the projection image using an appropriately programmed computer, subtracting the estimated scatter map from the projection image to obtain scatter-corrected projections, reconstructing a CBCT volume using a total variation regularization algorithm, and applying an iterative regularization process to suppress the noise level on the reconstructed CBCT volume. Reconstructing a CBCT volume can include using a total variation regularization algorithm and applying an iterative regularization process to suppress the noise level on the reconstructed CBCT volume, where scatter-induced artifacts are corrected in the projection image.

The invention discloses a high-precision near-field acoustic holography algorithm adopting a weighted iteration equivalent source method, which is characterized in that a holographic plane H is arranged in a sound source near-field radiation area, and sound pressure PH on the holographic plane H is measured; an equivalent source plane Se is arranged at the side, which is away from the holographic plane H, of an object reconstruction plane T, and equivalent sources are arranged on the equivalent source plane Se; a relation between the sound pressure PH and each equivalent source is established by using a sound pressure transfer matrix between the equivalent source and the holographic plane H; and the source intensity Q of each equivalent source is solved by adopting a new iterative regularization algorithm with a posteriori weighted norm constraint penalty term, and then sound field data on the object reconstruction plane T is calculated by using the solved source intensity Q and the transfer matrix between the equivalent source and the object reconstruction plane T. According to the invention, the source intensity of each equivalent source is precisely solved by using the new iterative regularization algorithm with the posteriori weighted norm constraint penalty term, thereby avoiding source intensity energy leakage caused by a 2-norm penalty term in the Tikhonov regularization process. Compared with conventional equivalent source based near-field acoustic holography, a calculation result acquired by the method disclosed by the invention is more accurate.

The invention discloses an automatic building contour extraction algorithm based on a convolutional neural network and polygon regularization. The automatic building contour extraction algorithm comprises the following steps: constructing a sample library according to an existing image and a building coverage vector file; Constructing a multi-scale fusion full convolutional neural network, training the multi-scale fusion full convolutional neural network through a sample library, and predicting the remote sensing image by using the trained network model to obtain a segmentation result coveredby the surface building of the remote sensing image; Performing building edge initialization based on the building semantic segmentation result, and obtaining an initial vector polygon; Removing wrongpolygons and wrong edges and nodes of the polygons by using a coarse adjustment algorithm; conducting regularization on the vector polygons through a regularization algorithm, and obtaining regular building vector edges. According to the method, the multi-scale fusion full convolutional neural network is high in scale robustness, the regularization algorithm can adapt to vector edges under various conditions, and the workload of manually drawing building edges is greatly reduced.

A method of acquiring scatter data and image projection data in computed tomography is provided that includes attenuating a radiation source using a pattern of blockers arranged to provide blocked and unblocked regions of the radiation source, and acquiring image data and scatter data of a target using an imaging device. A scatter map in the projection image can be estimated by interpolation and / or extrapolation of the projection image using an appropriately programmed computer, subtracting the estimated scatter map from the projection image to obtain scatter-corrected projections, reconstructing a CBCT volume using a total variation regularization algorithm, and applying an iterative regularization process to suppress the noise level on the reconstructed CBCT volume. Reconstructing a CBCT volume can include using a total variation regularization algorithm and applying an iterative regularization process to suppress the noise level on the reconstructed CBCT volume, where scatter-induced artifacts are corrected in the projection image.

The invention relates to a voice evaluation method and system based on voice similarity and the method comprises the following steps: providing a training data set; using the training data set to train the dynamic time regularization algorithm and the support vector sequential regression algorithm to obtain a similarity scoring model; providing reference voice information; recording, mocking and reading the mocked voice information for the reference voice information; extracting a set of reference voice feature sequences in the reference voice information and a set of mocked voice feature sequences in the mocked voice information; and inputting the set of reference voice feature sequences in the reference voice information and the set of mocked voice feature sequences in the mocked voice information into the similarity scoring model to obtain and output the similarity scoring value between the mocked voice information and the reference voice information. On the basis of evaluating the correctness of sound making, the similarity evaluation method of the invention is added by the evaluation of the mocking of speech-making so as to assist a user to accomplish targeted mocking practice and improve their speech-making level.

The invention provides a train control onboard device failure diagnosis method with a LSTM (Long Short Term Memory Network) and a neural network combined. The method comprises steps: a log file of theonboard device is used to build an onboard device operation information corpus through text data mining processing, and original sample data are built; a multilayer network system with the LSTM and the BP (back propagation) network cascaded is built, and a Bayesian regularization algorithm is adopted to optimize the multilayer network system; training sample data are used to train the optimized multilayer network system, the well-trained multilayer network system is used to build a failure diagnosis model for the train control onboard device, the failure diagnosis model is used to diagnose anunknown failure sample of the train control onboard device, and a diagnosis result of the unknown failure sample is obtained. According to the train control onboard device failure diagnosis method with the LSTM and the BP network cascaded, intelligent train operation information classification is realized, demands on manual experience in the field are reduced, and failure diagnosis on the train control onboard device is carried out effectively.

A computationally efficient dictionary learning-based term is employed in an iterative reconstruction framework to keep more spatial information than two-dimensional dictionary learning and require less computational cost than three-dimensional dictionary learning. In one such implementation, a non-local regularization algorithm is employed in an MBIR context (such as in a low dose CT image reconstruction context) based on dictionary learning in which dictionaries from different directions (e.g., x,y-plane, y,z-plane, x,z-plane) are employed and the sparse coefficients calculated accordingly. In this manner, spatial information from all three directions is retained and computational cost is constrained.

The invention discloses a lithium battery SOC prediction method of a bayes regularization LM-BP neural network. The lithium battery SOC prediction method comprises the following steps: a, establishinga BP neural network model; b, establishing a bayes regularization LM-BP neural network algorithm; c, acquiring sample data and calculating sample SOC; and d, performing the normalization processing of data. The neural network has good nonlinear fitting capacity and does not need to consider a complicated chemical structure inside the battery, dynamic characteristics of the lithium battery can bewell fit, by combining the bayes regularization algorithm, the generalization capacity of the network can be improved, by combining the LM algorithm, the convergence rate of the network can be increased, and the approximation accuracy can be improved; and therefore, the lithium battery SOC prediction method of the bayes regularization LM-BP neural network has the characteristics of high predictionprecision, high convergence speed, and high generalization capacity and is suitable for various power batteries.

The invention discloses a face image super-resolution method based on local and sparse non-local regularities. The face image super-resolution method comprises the following steps of 1, obtaining image blocks of all pixel positions of a test image and a training sample image; 2, using a local PCA dictionary learning method, using a K-means clustering algorithm to divide and cluster the image blocks of the training sample image blocks, and learning a local PCA dictionary by each cluster; 3, for a low-quality image block, solving an optimal representation coefficient vector by applying a local constraint and sparse non-local dual-core norm regularization algorithm; 4, synthesizing a high-resolution image block on the high-resolution dictionary by using the optimal representation coefficientvector, updating a non-local coding coefficient, and putting the updated coefficient and the high-resolution image block into the step 3 for next iteration; A high-resolution image block is obtained through multiple times of iterative updating; And step 5, outputting a high-resolution image. The method has the advantage of improving the image quality.

The invention relates to a road abnormality detection model based on window partition and dynamic time regularization, comprising the following steps: 1) carrying out threshold detection and sliding window processingon z-axis acceleration data, and screening fragments to be determined; 2) comparing the acceleration data of the segment to be determined with the segments of several known abnormal types and normal road segments through a dynamic time regularization algorithm to obtain a difference degree vector; 3) determining the abnormal type of the segment to be determined. The road abnormality detection model has the beneficial effects that the abnormal section of the road can be intercepted completely, and the experiment result also proves that the abnormal condition of the road can be detected more accurately on different data sets, and the effect of the method is better than that of the existing method in terms of the two-classification or the multi-classification.

The invention discloses a social network event timing relation analysis method, which comprises the following steps: obtaining event detection result data, extracting event short text cluster time series, dynamically adjusting time series, and constructing a quantile-Quantile diagram to analyze the temporal relationship of events. At first, the invention extract the event short text cluster time series from the event short text cluster set, matches the time series of the event with a dynamic time regularization algorithm, then quantitatively calculates the time sequence distance of the time sequence correspondence relationship between the event short text clusters according to the matched result, and quantiles the time sequence distance of the event short text clusters according to the time sequence distance of the event short text clusters. Quantile graph visualization method qualitatively analyzes the temporal relationships among event short text clusters, which can significantly improve the recognition accuracy of event time series relationships in social networks.

The invention discloses a motion noise detecting method based on photoplethysmography signals. The method lays a foundation for follow-up work of heart rate measurement. In the method, multiple photoplethysmography signals and acceleration signals within the same period are acquired by a reflecting type photoelectric sensor and a motion sensor; the acceleration signals are processed with a principal component analysis method, a reference signal related with motion noise is produced, and part of motion noise is eliminated in combination with a least mean square adaptive filter; the multiple processed photoplethysmography signals and acceleration signals form a frequency spectrum matrix, sparse structure features of rows of the frequency spectrum matrix are extracted, and a sparse signal reconstruction model is constructed; finally, the sparse signal reconstruction model is optimized with a regularization algorithm, and the spectrum peak position of the motion noise in the frequency spectra of multiple photoplethysmography signals is acquired. The motion noise in the photoplethysmography signals can be detected accurately, and high-accuracy measurement of the heart rate is realized.

The invention provides a method for acquiring signal-to-noise ratio and sampling ratio of a sparse microwave imaging radar, comprising the following steps: S1. establishing a sparse microwave imaging radar echo wave model; S2. constructing a sparse microwave imaging radar sampling ratio model according to the sparsity of sparse target signals in the echo wave model; S3. constructing the needed sparse microwave imaging radar signal-to-noise ratio model according to the sparsity and sampling ratio of the scene; S4. according to the signal-to-noise ratio model, acquiring a sparse reconstructed regularization parameter; and S5. according to the sparse microwave imaging radar echo wave model and a set parameter range, carrying out sparse reconstruction on sparse target signals by utilizing an L1 regularization algorithm, thus obtaining a sparsely reconstructed sparse target signal.

The invention provides a seismic source wavelet optimal setting method and device. The seismic source wavelet optimal setting method comprises the following steps that seismic wavelets are extracted from seismic data to serve as a target wavelet and an initial seismic source wavelet; the initial seismic source wavelet is updated by means of the target wavelet and the initial seismic source wavelet through a regularization inverse algorithm; when a preset convergence condition is met, optimized seismic source wavelets are output. According to the seismic source wavelet optimal setting method and device, the optimized seismic source wavelets are obtained through regularization algorithm inversion, the initial seismic wavelets extracted from the seismic data can directly serve as the seismic source wavelets to be output to a two-dimensional migration and inversion process, and therefore the reliable seismic source wavelets can be provided for numerical simulation, migration imaging and full waveform inversion of acoustic waves and elastic waves.

The invention discloses a fused convolutional neural network-based wear surface three-dimensional topography measurement method, which comprises the following steps of: generating a random rough surface through a two-dimensional digital filtering technology, and obtaining a luminosity image sequence of the random wear surface by utilizing Blender rendering software so as to generate a data set forneural network training; designing a feature extraction module, a fusion module and a normal vector estimation and refinement module to obtain a fused convolutional neural network applied to wear surface normal vector estimation; defining a training loss function of the neural network, and training and adjusting a network model based on the data set; and in combination with priori knowledge of the abraded surface, solving the depth information of the abraded surface based on a regularization algorithm. According to the method, the neural network method and the photometric stereo technology are effectively combined, the problem that the reflection characteristics of the abraded surface are not matched with the Lambert model is solved, and accurate reconstruction of the abraded surface is achieved in combination with priori knowledge of the abraded surface.

The invention discloses a neural network model processing method and device and a data processing method and device. The neural network model processing method comprises the following steps: training a neural network model by using a singular value decomposition algorithm and a norm regularization algorithm to obtain a pre-trained model; obtaining a target mask matrix corresponding to the pre-training model, wherein the dimension of the target mask matrix is the same as the dimension of a weight matrix of the pre-training model; and processing the weight matrix of the pre-training model by using the target mask matrix to obtain a target model. The technical problems that the calculation amount of a neural network model is large, and hardware cost and power consumption are increased in the prior art are solved.

The invention provides a method for identifying a lateral driving condition of a vehicle. The method includes the following steps: identifying a vehicle steering process; calculating condition identification data in the identified steering process, wherein the condition identification data includes the sum of weighted distance among a heading angle change angle, a heading angel history curve and apreset typical lane change template curve, maximum yaw angular velocity and lateral displacement, and the sum of weighted distance is obtained based on a dynamic regularization algorithm; and determining a driving condition to which the steering process belongs based on the calculated condition identification data, wherein the driving condition includes a turning driving condition, a U-turn driving condition and a lane changing driving condition. The invention also provides a system for identifying the lateral driving condition of the vehicle. Since the lane changing driving condition in thevehicle lateral driving condition is identified by combining the dynamic time regularization algorithm, the identification method has good adaptability and high identification accuracy, and the accuracy rate of lane changing conditions is more than 90%, thereby more accurately determining the driver's control condition of the vehicle.

The invention discloses a sharp edge preserving resistance tomography image reconstruction method. The method comprises the steps of obtaining a relative boundary measurement value vector and a sensitivity matrix required by reconstruction according to a measured field domain; Setting initialization parameters; Determining an optimal regularization parameter by using an improved GCV method; Calculating a gradient and a Hessian matrix of the objective function; Updating the value of the solution by using a Gauss-Newton iteration method; Judging whether iteration is finished or not; And performing imaging according to the finally solved imaging gray value. According to the method, the defects that the edge of a reconstructed image is fuzzy and the image resolution is low in a traditional Tikhonov regularization algorithm and a total variation regularization algorithm are overcome, and the method has a good effect on improving the quality of an electrical tomography reconstructed image and keeping the keeping capability of a sharp edge.

The invention discloses a neural network based prediction method for bird damage state of transmission lines, which comprises the following steps: S1, detecting bird activity near transmission line pole and tower by Doppler radar, obtaining bird flight trajectory information and extracting flight trajectory characteristic information; S2, constructing a flight trajectory prediction mathematical model based on neural network; S3, train a mathematical model network of a flight path prediction neural network by adopting a Bayesian regularization algorithm; the invention can effectively reduce thestartup times of the bird prevention device in the aspect of bird damage prevention of the transmission line, thereby greatly reducing the electric power consumption of the existing bird prevention device, realizing that the existing bird prevention device achieves intelligent bird driving, providing great convenience for the subsequent patrol and inspection personnel, and having strong practicalvalue.

The invention provides an air pollution prediction method. The method comprises the following steps: (1) acquiring air pollution meteorological data; (2) converting the air pollution meteorological data into a pixel matrix and performing data filling; (3) carrying out time-space domain-oriented spatial-temporal feature unified modeling through a three-dimensional convolutional neural network model; (4) taking the output of the three-dimensional convolutional neural network model as the input part of a convolutional long-short term memory network, and carrying out long-time and short-time dependent modeling; (5) generating an air pollution prediction model based on a space-time dynamic advection method; and (6) optimizing the air pollution prediction model through an orthogonal regularization algorithm, and then carrying out environment prediction. The decoupled three-dimensional convolution is used to carry out spatial-temporal feature unified modeling, the representation capability of the spatial-temporal feature extraction is enhanced, the fusion of the spatial-temporal features is truly realized, and the training cost is reduced and the training speed is improved while the spatial-temporal convolution performance is improved.

The invention provides a method for obtaining a sensitive lithology identification factor based on inversion of pre-stack seismic data, which comprises the following steps of: establishing sandstone,limestone and mudstone models according to an actual working area for carrying out lithology sensitivity analysis to obtain a sensitive lithology identification factor, in order to reduce the error accumulation of indirect calculation of the factor, deriving an AVO approximate formula taking sensitive lithologic identification factor as a variable, deriving an elastic wave impedance formula expressed by the sensitive lithologic identification factor, a shear modulus and density by reference of an elastic impedance derivation principle, using a regularization algorithm to directly extract the sensitive lithologic identification factor, shear modulus and density data bodies from an actual elastic impedance data body through inversion for lithology prediction and fluid identification of a reservoir, the sensitive lithologic identification factor has better identification capability compared with a prior elastic parameter, and the calculation result is fast and the stability is strong.

The invention discloses a method for restoring a spatially varying blurred image based on TV and wavelet regularization, which comprises the following steps: (1) grayscaling the blurred image; (2) constructing a fuzzy kernel decomposition model according to the grayscale fuzzy image, and decomposing the fuzzy kernel into a basic filter matrix and a coefficient matrix by using a singular value decomposition method in the fuzzy kernel decomposition model; (3) applying the fuzzy kernel decomposition model, and combining the TV regularity term and the wavelet regularity term to construct a defuzzification model; 4) transforming defuzzification model into the augment Lagrangian form, the augmented Lagrangian form defuzzification model is improved to obtain a new defuzzification model; (5) solving The new deblurring model by ADMM algorithm, and the restored image is obtained. This method solves the problem of loss of detail information in the restoration process of TV regularization algorithm.

The embodiments of the invention disclose a method and a device for matching a subtitle file for a local video file. The method comprises: acquiring a first video file name of a local first video file to be matched with a subtitle file and a subtitle file name set including local subtitle file names; regularizing the first video file name and the subtitle file names in the subtitle file name set according to a preset regularization algorithm, to obtain a standardized first video file name and a standardized subtitle file name set including all standardized subtitle file names; performing matching degree calculation on the standardized first video file name and each standardized subtitle file name in the standardized subtitle file name set; and determining a subtitle file corresponding to the standardized subtitle file name of which the calculation result satisfies a preset condition in the standardized subtitle file name set as a subtitle file matched with the first video file. By adopting the embodiments, subtitles can be automatically matched for the local video file better and more conveniently, and the video watching experience of a user is improved.

The invention discloses a progressive building extraction method based on an original laser point cloud, and the method comprises the following steps: 1, carrying out the progressive mathematical morphological filtering of the original laser point cloud, and separating non-ground points; 2, adopting an improved 3D-Hough conversion algorithm to detect a plane in non-ground points, extracting building point cloud through the plane, and projecting the building point cloud to a two-dimensional plane; 3, determining key points of the building point cloud by adopting an adjacent point azimuth anglethreshold method; 4, based on the key points of the building point cloud, utilizing RANSAC fitting to obtain an initial contour line of the building; and 5, regularizing the initial contour line of the building by using a regularization algorithm to obtain a regularized building contour line. According to the method, a complete set of extraction process from the original point cloud to the building contour line is provided, processing can be directly started from the obtained laser radar point cloud data, and the type, range and precision of the extracted point cloud can be adjusted accordingto requirements.

A load identification method based on numerical operation and an improved regularization algorithm comprises the following steps: step 1, establishing a discrete finite element model of a system, andutilizing an explicit Wilson-theta based load identification method to obtain a load identification model of the structure system through a load identification algorithm; 2, applying a dynamic load tothe structure, and measuring the response of the dynamic load; Step 3, constructing a load identification regularization model; Step 4, determining regularization parameters of the load identification model by using an L curve method, and substituting the regularization parameters into the load identification model to carry out calculation of load identification; 5, calculating is finished, and outputting a load identification result. The method is simple and convenient to operate in practical application, only the parameter data of the structure needs to be known, a corresponding load identification regularization model is established, and an unknown dynamic load can be identified by utilizing a response signal obtained through measurement; In addition, a traditional Tikhonov regularization method is improved, and the dynamic load identification precision of the improved regularization method is higher.

The invention discloses a method for adaptively measuring film thickness through white light reflectance based on a Bayesian regularization algorithm. The method is characterized by carrying out learning fitting on signals polluted by noise in dynamic measurement through a BP nerve network using the Bayesian regularization algorithm to obtain a fit WLRS curve; and obtaining characteristic values of the fit WLRS curve, and carrying out fast and accurate measurement of the thickness of a film to be measured. Through actual measurement, the method is proved to have better anti-interference capability and a certain adaptability.

The invention relates to a prior knowledge fault diagnosis method based on the Tennessee Eastman process. The method comprises the steps that the offline historical data of the Tennessee Eastman process are acquired; an adjustment parameter matrix that U belongs to R<nxn> and k of a KNN algorithm are selected; an adjacent matrix W is constructed on an existing weighted undirected graph, a matrix D is accordingly calculated, a Laplacian matrix L=D-W is defined, and the Laplace regular term L<~> is calculated according to a Laplace regularization algorithm; the local regular term (I-A)<T>(I-A) is calculated according to a local regularization algorithm; a tag matrix is calculated according to F<*>=(UD<~>+L<~>+(I-A)<T>(I-A))<-1>UD<~>Y; and the unmarked samples are marked according to f=arg maxF<*><ij>, 1<=j<=c, and fault classification information of the industrial process is obtained after normalization. Characteristic information of the marked samples and the unmarked samples is fully mined and utilized to establish a fault diagnosis model and verification is performed by using the Tennessee Eastman process data, and the classifier is improved in the final classification phase so that the classification accuracy can be enhanced, and the classification error rate of the samples and the sample separation degree and other verification standards can be improved.

In order to solve the technical problems that a traditional lens-free camera image reconstruction method is relatively sensitive to noise and relatively low in system depth of field, the invention provides a lens-free camera image reconstruction method based on a coding mask and a Learned-TSVD algorithm. The method comprises the following steps: encoding a propagation process of light by using an encoding mask, converting an original large-scale system measurement matrix into a left system measurement matrix and a right system measurement matrix which are small in scale by utilizing the separable characteristic of the coding mask and a TSVD algorithm; thirdly, constructing neural network training to circularly train the left and right system measurement matrixes, and reducing an error of an approximate operation on a final result; and finally reconstructing an image through the TSVD algorithm and a regularization algorithm. According to the method, the learned system measurement matrixes are used for subsequent calculation, so that the noise influence resistance of the whole reconstruction process is higher; scene images at other distances can be well reconstructed by using the learned system measurement matrixes, and the problem of low depth of field of other reconstruction algorithms is solved.

The invention discloses an implicit identity authentication method for a mobile intelligent terminal, which comprises the following steps that: a mobile intelligent terminal authentication service system sets a password with a certain length and legal WiFi, and collects touch behavior data when the password is set, and the touch behavior data is stored in a database after being processed by a weight-based dynamic regularization algorithm; the legality of the WiFi environment is monitored in real time; a legal WiFi environment user can automatically enter the mobile intelligent terminal to check information without inputting a password; an illegal WiFi environment user needs to input a password, touch behavior data generated when the password is input are compared with touch behavior data stored in a database according to a weight-based dynamic regularization algorithm, the legality of the identity of the user is determined, and the user enters a mobile intelligent terminal legally to check information; according to the invention, the existing authentication mechanism is enhanced in convenience and security, and the user can be prevented from frequently inputting a password.

The invention discloses an air micro-station concentration prediction method based on an LSTM neural network. An isolated forest algorithm is adopted to preprocess pollutant concentration data obtained by an air micro-station, and a batch gradient descent algorithm is fused in deep learning to improve the stability of the whole system. Meanwhile, a Dropout algorithm and an L2 regularization algorithm are added into an input layer and a hidden layer to avoid the over-fitting phenomenon, and the whole algorithm is used for processing the complex space-time relation of particulate matter concentration, gas concentration input and multiple air quality outputs through a low-cost sensor.