48 results about "Sparse methods" patented technology

The invention discloses a video motion identification method based on a sparse time slicing network. The method includes steps: extracting information from each training video, and performing first training and optimization on the time slicing network; adding a sparse to the network after the first optimization, and performing second training and optimization; performing cutting and dimension adjustment on the network after the second optimization; performing third training and optimization on the network after dimension adjustment until the identification precision or the sparsity reaches theexpectation; and extracting information from a to-be-identified video, inputting the extracted information to the network after the third optimization, and performing output and fusion on the time slicing network to obtain a motion identification result. According to the method, the information of the longer video can be obtained through the time slicing network, and a double-flow convolutional network structure can fully utilize the video information and greatly improve the motion identification precision; and a structured sparse method can enable the weight of a convolution layer to be sparse in groups, a model is further simplified by network cutting, and the storage space is reduced.

The invention provides a monocular infrared video three-dimensional reconstruction method based on a visual odometer. Firstly, a thermal infrared imager is calibrated; then, a direct method and sparse method visual odometer model is constructed; then, frames are managed, a key frame is generated from a continuous infrared image sequence to be processed, and the key frame is marginalized; point management is conducted on the pixel on the key frame; finally, sliding window optimization is conducted on the key frame, a photometric error of the direct method and sparse method visual odometer model is minimized, the reverse depths of thermal infrared imager pose, internal references and spatial points are solved in an iterative algorithm by a gauss-newton method, and finally, three-dimensional point cloud of a scene is obtained. According to the monocular infrared video three-dimensional reconstruction method based on the visual odometer, the direct method is applied to photometric error minimization, traditional feature point detection and matching are skipped, through direction operation on a grey value of the pixel, all variables which the photometric error to be optimized depend on is calculated directly, and therefore three-dimensional reconstruction of a night vision scene is achieved, the real-time capability of the three-dimensional reconstruction is ensured, and the spaciousness of the scene is enhanced.

The invention discloses an array sparse method for a broadband non-frequency-variable multi-beam imaging sonar. With the Bessel function, fitting of influences on array guiding vectors by different frequency points in the broadband signal bandwidth is performed and a broadband signal multi-beam forming model under the far-field situation is established; on the premise that the formed multiple beams approximate a reference beam, a minimum number of effective array elements are searched and multiple sets of weighting coefficients are calculated; a highly nonlinear sparse array optimization problem is transformed into a sparse signal reconstruction problem in the compressed sensing theory, a reconstruction weighting coefficient is calculated iteratively by an underdetermined system localizedsolution algorithm, and a sparse array structure is determined; a convex optimization theory is introduced so as to form a plurality of low-side-lobe beams and a multi-beam array sparse side-lobe suppression model for array element excitation is established. According to the invention, the main lobes of a plurality of formed beams are not extended with changes of signal operating frequencies; andpeak side-lobe levels of multiple beams formed by the sparse array are reduced effectively.

The invention discloses a knowledge assisted-based sparsity recovery STAP color loading method. The steps include: establishing a sample data model; calculating a priori space-time steering vector matrix; calculating a sparsity recovery vector; calculating sample clutter information; calculating a sample covariance matrix and a prior clutter covariance matrix; and calculating a STAP color loadingmatrix and performing filtering. The solution of the invention combines color loading and a sparsity method, that is, the sample covariance matrix of the color loading is a covariance matrix obtainedfrom a sparsity sample after sparsity processing is performed on a sample. In a process of sparsity recovery processing of the sample, an algorithm to suppress dense interference is performed. The sample sparsity processing is performed in a solving process of the sample covariance matrix of the color loading. A knowledge assisted-based method used to suppress the dense interference, and a minimumthreshold method for determining a clutter element solve a problem of moving target cancellation, and the algorithm has better robustness.

The invention discloses a neural network column sparse method based on weight saliency. When the neural network model carries out feedforward operation, the four-dimensional weight tensor can be expanded into a two-dimensional matrix in a tensor expansion mode, and therefore convolution operation is converted into matrix multiplication. Therefore, on the basis of a tensor expansion operation mode,the invention discloses a method for carrying out column sparsity on a weight matrix, namely deleting columns of the weight matrix in a neural network parameter layer to obtain a sparse model, and the dimension of the operation matrix is reduced at the moment, thereby reducing the model scale of the neural network and accelerating the operation of the neural network. Network re-training is performed on the remaining weight parameters to callback the accuracy, and re-training is stopped when the accuracy of the model is not increased any more to obtain a final model. According to the method, the deep learning model can be deployed on mobile and embedded equipment, a certain actual acceleration effect is obtained, and the application of an intelligent algorithm on a mobile terminal is promoted.

The invention discloses a method for reducing a frequency aliasing effect of undersampled magnetic resonance imaging. The method comprises the following steps of (1) obtaining space K through a nuclear magnetic resonance spectrometer and performing undersampling on the original data of the space K by using a random oscillometric method to obtain undersampling data; (2) forming an over-complete dictionary by selecting a sparse method; and (3) performing recovery reestablishment on the undersampled data in the step (1) by using a recovery algorithm according to the over-complete dictionary to obtain original data, performing Fourier transform to obtain images and performing real-time imaging. The random oscillometric method comprises a variable-density oscillating spiral track and a variable-density oscillating radiation track.

This invention discloses one dialogue control and sparse method based on flexible exchange in the telephone network, which comprises the following steps: making statistics for router link and aim number set, if the volume achieves the set length, then combing the calls according to the set router and pattern to change call directions and to comb set router to achieve dialogue control and combing without human interface.

The invention discloses a mechanical fault sparse representation method based on a wolf pack parameterization joint dictionary, and belongs to the technical field of mechanical fault diagnosis. The invention provides a grey wolf optimization sparse method based on a joint dictionary in combination with a sparse representation principle. An optimization algorithm is fused into orthogonal matching pursuit, a joint dictionary atom most matched with a fault is selected through inner product operation, the optimal atom is used for reconstructing a fault signal, and then envelope analysis is performed on the reconstructed signal, so that extraction of transient characteristic components in the signal is realized, fault information is effectively represented, and the fault type is judged more accurately. The method is suitable for rolling bearing fault diagnosis.

The invention discloses a main speed reducer multi-fault intelligent diagnosis method based on multi-channel data deep mining, and the method comprises the steps: obtaining a vibration signal of a main speed reducer through a multi-channel sensor, and carrying out the noise reduction of the vibration signal through the combination of a wavelet contraction noise reduction method and a structured sparse method; based on the vibration signal after noise reduction processing, adopting a deep belief network to perform multi-channel deep feature extraction, and fusing multi-channel deep features; and based on a deep feature fusion result, performing online state monitoring and multi-fault intelligent diagnosis on the main speed reducer by adopting a sparse Bayesian extreme learning machine. According to the invention, single-fault and multi-fault modes can be identified at the same time, and quantitative analysis of state monitoring and intelligent fault diagnosis of the main speed reducer is realized.

The invention provides a reverse operation sparse method of a neural network and a computing chip. The method is used for executing sparse processing in neural network operation and then performing reverse operation. The technical scheme provided by the invention has the advantages of small calculation amount and low power consumption.

The invention relates to an ultrahigh-voltage power transmission line point cloud data semantic segmentation method based on deep learning, which belongs to the field of ultrahigh-voltage power transmission line inspection, and specifically comprises the following steps: for actual ultrahigh-voltage power transmission line point cloud data, firstly establishing a deep learning training data set by using a scene division method based on a safety buffer distance; secondly, realizing feature acquisition in a deep learning network model by using a two-step sparse method based on a combination coefficient and a feature abstraction method based on importance of sparse data and local position related information; training the network model to obtain accurate network parameters; and finally, inputting the test sample set into the network to obtain a category label of each point, thereby realizing semantic segmentation of the point cloud data. According to the method, the problems of lack of data sets, low calculation efficiency and low calculation precision in deep learning semantic segmentation of the point cloud data of the power transmission line at present are effectively solved.

The invention relates to a zero attraction penalty and attraction compensation combined sparse LMS method, and belongs to the field of signal processing. According to the method, zero attraction penalty and attraction compensation are combined, coefficients of an estimation filter are divided into a near-zero coefficient, a small coefficient and a large coefficient, and then different attraction methods are adopted; in each iterative update, for estimating a near-zero coefficient of the filter, only a product term in an iterative update formula is used for calculation; then trace attraction compensation is carried out on the large coefficient of the estimation filter, so that the convergence speed of the coefficient of the estimation filter to approach the large coefficient of the channel is accelerated; for the small coefficient of the estimation filter, if the coefficient approaches the zero coefficient value of the channel or the large coefficient value of the channel in the iteration process, processing is carried out according to the method for estimating the near zero coefficient and the large coefficient of the filter, otherwise, simple zero attraction punishment is carried out on the coefficient. The method is high in convergence speed, low in complexity and wide in tuning parameter application range.

The invention provides a neural network sparse method and a computing chip. The method is used for executing sparse processing in neural network operation and then performing operation. The technicalscheme provided by the invention has the advantages of small calculation amount and low power consumption.

The invention discloses a mixed granularity-based combined sparse method for a neural network. The method comprises independent vector magnitude fine granularity sparsification and block level coarsegranularity sparsification. According to the method, bitwise logic AND operation is performed on a pruning mask which is independently generated through the two sparsification modes, so that a final pruning mask can be obtained, and a weight matrix of the sparsificated neural network is obtained. According to the method, the reasoning speed between the block sparsification mode and balance sparsification mode is always obtained through joint sparsity, and the vector line size of vector magnitude fine-grained sparsification and the vector block size of the block level coarse-grained sparsification do not need to be considered. The method is used for pruning a convolutional layer and a full connection layer of the neural network, and has the advantages of variable sparse granularity, accelerated general hardware reasoning and high model reasoning precision.

The invention discloses a sparse technology based on an improved adjacency matrix. Aiming at that an adjacency matrix of a storage network topological structure has the characteristic of high sparsity, pattern of manifestation of the adjacency matrix is improved, and the improved adjacency matrix is applied to node optimal numbering, in-advance determination of retrieving information and formation of an admittance matrix. In a LU decomposition process, in order to realize nonzero retrieval of line direction, storage information of the line direction is increased, and corresponding retrieval modes are customized. According to correlation property of newly added branches and factoring nonzero injection elements in the optimal numbering process, while executing optimal numbering, positions of the newly added elements are recorded and a storage framework is formed. The sparse technology based on the improved adjacency matrix is suitable for being used in calculation process of an electric power system, and can effectively increase efficiency of related calculation through improving the retrieval modes and the storage framework.

The invention relates to the technical field of industrial data acquisition visualization, in particular to a high-frequency time series data real-time sparse display method based on the number of pixels. The method comprises the following steps: when an oscillogram of a time series data point is displayed on a computer graphical interface, a point tracing display method is adopted, points in an actual coordinate system are converted into pixel coordinate system points for point tracing, the time sequence is horizontal coordinates, the amplitude is vertical coordinates, the time sequence is converted into horizontal coordinate pixels, the amplitude is converted into vertical coordinate pixels, and then the points are connected through folding lines. Under the condition of high-frequency data acquisition, data can be acquired; the data point density is very large, if a signal waveform curve is displayed in real time without distortion, a large amount of computer performance and networkperformance can be consumed, the data sparse method based on the number of pixel points is provided, and the signal waveform curve can be displayed in real time without distortion under the conditionthat the computer hardware performance is not improved.