36results about How to "Reduce the number of multiplications" patented technology

The embodiments of the invention disclose a convolution acceleration method, a convolution calculation processing method, a convolution acceleration device, a convolution calculation processing device, an electronic apparatus and a storage medium. The acceleration method includes the following steps that: the to-be-processed data of a preset size in a to-be-processed task are sequentially read from the off-chip memory of an accelerator through an FIFO (first-input-first-output) port, and are stored in the input cache regions of a first cache region in the on-chip memory of the accelerator; to-be-processed input window data are sequentially read from the input cache regions in response to a condition that the to-be-processed data are stored into the input cache regions, and convolution calculation is performed on the to-be-processed input window data through convolution kernel data, so that output window data are obtained and are stored in a third cache region; and output window data inthe third cache region are sequentially stored to the off-chip memory through the FIFO port. With the methods, devices, electronic apparatus and storage medium provided by the embodiments of the invention adopted, a condition that the on-chip memory and bandwidth of the accelerator are insufficient and a condition that processor resources are limited can be avoided, and the efficiency of convolution calculation processing is improved. The methods and devices can be applied to hardware platforms such as an FPGA and an ASIC.

The invention discloses an interpolating method of self-adapting medical sequence image among fault space based on interesting area, which utilizes tissue relativity and body relativity to judge whether the interpolated point belongs to the interesting area to select fitful interpolating method. The invention considers the change of grey and structural outline, which solves the insufficiency of traditional interpolating method to make new interpolated image approach actual condition. The interpolating method modifies the visual effect and image quality greatly, which reduces the noise and shortens the calculating time to be applied in three-dimensional reconstruction and display system.

The invention discloses a pedestrian real-time detection method based on a neural network, and the method comprises the steps: 1) collecting data, and constructing an original training data set; 2) calibrating pedestrian positions corresponding to the acquired images for the images in the original training data set, and constructing real position output of the training data set; 3) constructing aneural network model; 4) in order to increase the network training data volume and enhance the applicability of the method, performing data enhancement on the original input image by using a data enhancement method; 5) setting training parameters for the designed neural network model for training, and storing the trained neural network model parameters; and 6) acquiring image data of pedestrian detection to be performed by using an image acquisition device, and then inputting the image of the pedestrian to be detected into the stored neural network model to obtain the pedestrian position of the image of pedestrian detection to be performed. The method can reduce a large amount of detection time on the premise of meeting the detection accuracy of pedestrian detection.

The invention discloses a spherical multi-probe antenna test data processing method in order to solve the technical problem that the existing method is of poor practicability. According to the technical scheme, a field established in a space by a to-be-tested antenna is expanded into the sum of spherical wave functions, a weighted function in the expansion contains the information of a far-field directional pattern, the weighted function can be worked out according to near-field sampling data and the far-field directional pattern is thus worked out, multi-probe spherical near-field scanning is adopted in near-field sampling, the to-be-tested antenna azimuthally rotates around the antenna center, probes carry out sampling at a distance of a few wavelengths from the antenna aperture, and therefore, the far-field coverage angle provided by spherical near-field measurement is within the range of +/-180 degrees for azimuth angle and +/-90 degrees for pitch angle, and all radiation information of the to-be-tested antenna can be obtained. Thus, the method is applicable to measurement of all types of beam antennas. The number of times of program looping is reduced, the speed of program running is improved, and the time for antenna test data processing is shortened.

The invention discloses a method for predicting self-interaction effect of protein. The method comprises the steps of selection and establishment of a data set, generation of a PSSM matrix, extractionof a feature value by a Fourier descriptor, construction of a training set and a testing set and construction of a classifier model. According to the method, the feature value of a sample set is extracted by the Fourier descriptor, the number of times of multiplication required for discrete Fourier transform of a computing data set of a computer is greatly reduced, and the computing quantity is reduced. A model can be constructed by a random projection method, the prediction precision is greatly improved, and a good prediction effect can be achieved. The method is low in computing cost and small in power consumption; the self-interaction effect of the protein can be predicted effectively; and the prediction effect can reach 93% or above.

The invention discloses a self-adaptive Cartesian grid generation method for a three-dimensional streaming problem of any shape, and the method comprises the steps: generating an isotropic self-adaptive Cartesian grid suitable for an immersed boundary method based on geometric information in the three-dimensional streaming problem, carrying out the calculation of a flow field, and encrypting an area containing key flow features according to the calculation result of the flow field. In order to solve the problem of computational fluid mechanics numerical simulation with complex three-dimensional streaming, a surface set composed of triangles is adopted as input, a grid intersection judgment method based on a separation axis theory and a grid inside and outside judgment method based on an improved ray algorithm are adopted for grid classification, a grid subdivision method based on a unit is adopted for encrypting and coarsening grid units, and a self-adaptive Cartesian grid meeting the requirements of an immersed boundary method and flow field calculation resolution can be efficiently and robustly generated; and a region containing a feature structure is selectively encrypted according to flow field parameters obtained subsequently, and the flow field feature structure in the current flow field area is displayed in real time.

The invention provides a RUMSWF based low-complexity reduced rank balancing method in an MIMO system, which is improved from an MSWF based self-adapting reduced rank linear balancing method and is a reduced rank self-adapting MIMO linear balancing method which realizes a multilevel Weiner filter based on a rectangular block matrix and by a related subtraction structure. By improving the block matrix of the unitary multilevel Weiner filter, a rectangular matrix block of the square block matrix is selected as the block matrix, and the number of dimension of a received signal is reduced step by step in the forward recursive decomposition of the unitary multilevel Weiner filter, thereby reducing the iteration complexity of the self-adapting balancing and simultaneously increasing the convergence rate. The theoretical analysis and simulation result show that the low-complexity quick reduced rank self-adapting balancing method has the advantages of low complexity and quick convergence rate. In a V-BLAST system provided with 4 transmitting antennas and 8 receiving antennas and adopting the BPSK modulation, by only utilizing one half of complexity of the multilevel Weiner filter based balancing method, the error code performance which is only 0.78 dB lower than that of the multilevel Weiner filter based balancing method can be achieved.

The invention discloses a data processing method, device and equipment, and a storage medium; the method is applied to a first electronic device, and the first electronic device and at least one second electronic device jointly participate in secure multi-party computing, and the method comprises the steps of obtaining to-be-processed service data; generating a first random number matrix and a second random number matrix, and generating a third random number matrix according to a preset encoding mode, the first random number matrix and the second random number matrix, the number of elements ofthe random number matrix encoded by the preset encoding mode being smaller than the number of elements of the random number matrix before encoding; and encrypting the service data in a secure multi-party computing process according to the first random number matrix, the second random number matrix and the third random number matrix. According to the invention, the generation efficiency of the matrix multiplication triad can be improved, and the encryption efficiency of service data in secure multi-party calculation is further improved.

The invention discloses a fast Fourier analysis algorithm, and relates to the technical field. The algorithm comprises the steps of obtaining a sequential sorting sequence x (n) of input data of N sampling points; calculating inverted sequence y(n) of input data by y(n)=x (G(m)+2Lmax-1*l); taking 8-point, 10-point or 12-point data from the inverted sequence y (n) for the first time for Fourier transform; amplifying the number of analysis points by two times each time to perform Fourier transform; and carrying out Fourier transform with mutually independent N points. According to the algorithm,the inverted sequence y (n) of input sequences by y(n)=x (G(m)+2Lmax-1*l) is calculated according to the difference between the total number of points participating in the sequence and the number ofpoints participating in the Fourier transform for the first time, and then the first-time 8-point, 10-point or 12-point data Fourier transform is rapidly completed through a set of formulas; the number of analysis points is amplified by two times each time to perform Fourier transform; according to the invention, the time complexity during calculation is reduced, the calculation rate is increased, and electric signal spectrum analysis is facilitated.

The embodiment of the invention provides an interpolation data processing method, device and equipment and a computer readable storage medium. The method comprises the following steps: acquiring information of two adjacent key frames of a spline curve, wherein the information of the two adjacent key frames comprises first time and second time of two-dimensional coordinate points respectively corresponding to the two adjacent key frames on a time axis; determining interpolation data calculation parameters according to the information of the two adjacent key frames; and when the third time is between the first time and the second time, determining target interpolation data corresponding to the third time according to the interpolation data calculation parameters. According to the method, interpolation data calculation is carried out through the spline curve, so that the calculation amount of the interpolation data is reduced, and the processing efficiency of the interpolation data is improved.

The invention discloses a portable multi-frequency electrical impedance tomography front-end data acquisition and processing method which comprises the following steps: converting a sinusoidal signal generated by a DAC into a current signal by using a voltage-controlled constant current source, and performing current excitation under the action of a first analog switch; obtaining a reference input signal through a sampling resistor, gating different groups of electrodes through a second analog switch, and performing program control gain by using an inverted T-shaped program control gain amplifier to obtain a measurement input signal; synchronously acquiring the reference input signal and the measurement input signal, performing multiple single-point Fourier demodulation according to the periodicity and symmetry of the digital signal and the symmetry of the trigonometric function, and obtaining the phase information of the signal. The existing demodulation method is improved, the required multiplication frequency is greatly reduced, and the demodulation speed is improved.

The invention provides a method, a system and equipment for accelerating image processing and a storage medium. The method comprises the following steps: dividing obtained picture data into data blocks, and executing the following steps during convolution operation of each data block: carrying out picture data conversion and convolution kernel data conversion on the data blocks to respectively obtain a picture data conversion item and a convolution kernel data conversion item; and accelerating the convolution operation of the image data conversion item and the convolution kernel data conversion item in the data block through a GPU acceleration unit, and performing inverse transformation on the result of the convolution operation to obtain a calculation result. The invention has the advantages that the image convolution operation speed can be greatly increased, and the performance requirement during reasoning is met. In addition, the invention also provides an optimal data partitioning structure for executing the method disclosed by the invention, so that further implementation of the method disclosed by the invention is facilitated.

The invention discloses a tensor and neural network-based dynamic electromagnetic spectrum posture method and system. The method comprises the steps of obtaining spectrum data of an area surrounded by current unmanned aerial vehicle flight; performing analog-to-digital conversion on the frequency spectrum information, then reading the frequency spectrum information after performing fast Fourier transform, and checking whether the frequency spectrum data displayed for one time is full or not; then an original matrix is constructed, and the constructed original matrix is visualized; performing data completion on the visualized original matrix to obtain an optimized final completion matrix; and inputting the optimized final completion matrix result value into a BP neural network for fuzzification processing to obtain a final completion matrix value, drawing a completed spectrum posture two-dimensional graph by using the final completion matrix value, drawing a three-dimensional posture spatial resolution graph, and displaying a spectrum posture signal of the dynamic signal according to time and frequency changes. According to the invention, by monitoring the radio signal and the electromagnetic spectrum posture, the purposes of eliminating electromagnetic interference and improving the spectrum utilization rate are achieved.

The invention discloses an audio frequency equalizing method of a decoder based on sub-band filter frame, which realizes relatively low complexity and relatively small consumed resource. The technical scheme is as follows: the method comprises the following steps: (1) obtaining time sampling points by an inverse quantization step in a decoder based on the sub-band filter frame; and limiting the bandwidth of the signal constituted by the time sampling points in the frequency range corresponding to each decoder sub-band; and (2) performing plus adjustment on the time sampling point in the frequency band which is obtained by the inverse quantization step and is adjusted by an audio frequency equalizer before the sub-band of the decoder composites filtering.