146results about How to "Reduce computing power consumption" patented technology

The invention belongs to the application technical field of AR intelligent glasses, and discloses a local image recognition method and a system based on the AR intelligent glasses. The method comprises steps calibrating the consistency of an intelligent glasses imaging screen, a front camera picture of the intelligent glasses and a visual field picture of a real environment around the intelligentglasses; recognizing the human eye image and calculating the eye motion vector to obtain the eye motion coordinates; identifying whether the focal point of the extended line of human binocular visionis an imaging screen or a real three-dimensional space; obtaining the mapping relationship of human eyes in the real world around them; according to the embedded mapping algorithm, the coordinates ofthe human eye fixation point on the glasses imaging screen and the front camera image being obtained respectively. The invention obtains the eye image information of the human eye through the eye movement instrument, synchronously calibrates the eye image and the scene camera, obtains the fixation point area of the human eye in the real scene, and performs image recognition only on the fixation point area during processing, thereby greatly reducing the image processing pressure of the GPU and improving the image processing efficiency.

The invention provides a compressed convolutional neural network-oriented parallel convolution operation method and apparatus. The method comprises the steps of determining an adopted operation mode according to input control signal convolution data shift chain length selection, accumulated offset enabling and convolution calculation enabling; and by adopting two serial shift register chains, inputting convolution data, convolution parameters and channel offset, and performing 3X3 and 1X1 convolution operation at the same time for a same input convolution data stream. According to the method, a multiplier, an accumulator, a parameter register and an offset register are added only based on original serial shift register chain-based 3X3 convolution operation; the realization method is simple; the executive efficiency is high; and the convolution operation in a compressed neural network algorithm can be effectively accelerated. According to the apparatus, a plurality of characteristic graphs can be output at the same time through simple hardware unit expansion and copying; and the apparatus has the advantages of low power consumption, high function unit utilization rate and high processing speed.

The invention discloses a block outputting method. The method comprises the following steps: a node selection step: selecting at least two nodes from all nodes of a blockchain network, respectively serving as an accounting node and a consensus node; a verification step: verifying whether transactions in a transaction pool are legal or not by the accounting node, and signing and broadcasting the legal transactions; a synchronization step: receiving the transactions signed by the accounting node by the consensus node, verifying the signature, and updating the transaction pool with the transactions that pass the verification; a block outputting step: packaging the transactions signed by the accounting node in the transaction pool by the consensus node into alternative blocks, and broadcastingthe alternative blocks; and a block saving step: carrying out consensus on the alternative blocks by all consensus nodes, saving the alternative blocks by all consensus nodes after the consensus is successful, and executing the transactions in the alternative blocks by the accounting node. The invention further discloses an implementation system of the block outputting method. The block outputting method provided by the invention greatly reduces the repetitive work of the consensus nodes, and reduces the hash rate consumption in the blockchain network.

The invention discloses an embedded processor of a storage and calculation integrated chip, an instruction set and a data processing method. The method comprises the following steps that: 1) an embedded processor directly reads data from an input data cache through a data input interface; 2) when a storage and calculation integrated array performs artificial neural network calculation, the storageand calculation integrated array sends data to be processed into storage and calculation integrated array/logic, controls a DAC to convert the data into analog signals and then perform matrix multiplication and addition operation, and then controls an ADC to convert operation results into digital signals; 3) the embedded processor reads back the operation results and performs operation or activation operation on the operation results, caches intermediate data, and then judges whether current operation is the last layer of neural network operation or not; (4) if the current operation is not the last layer of neural network operation, the step (2) and the step (3) are carried out circularly, and if the current operation is the last layer of neural network operation, the step (5) is carriedout; and (5) the embedded processor carries out final result processing, judges an artificial neural network recognition result and determines whether to output a result or not according to the recognition result.

The invention, which belongs to the technical field of communication, provides an cloud service comprehensive scheduling optimization system and method fusing edge facilities. The system comprises a terminal device, an edge server connected with the terminal device, and an edge resource device and a cloud server connected with the edge server. According to the invention, various features are extracted from tasks uploaded by terminal facilities through pattern recognition and are classified according to the extracted features; reinforcement learning mechanism correction classification is performed through the running state data features to implement scheduling; when the tasks run, the edge server keeps controlling and monitoring the edge resources, resource allocation and scheduling are carried out, so that system resources are fully utilized, the system efficiency is optimized, the bandwidth load of a core network is effectively reduced, the overall resource utilization rate of the system is improved, the task execution efficiency is improved, and tasks which cannot meet time delay requirements and are caused by centralization of a cloud computing system are made up through edge computing equipment and technologies.

The invention is suitable for the technical field of flash memory storage, and provides a method, a system, and a flash memory of an ECC dynamic adjustment. The method comprises configuring a plurality of ECC algorithms with different error correction capability for the flash memory; grading error bits with different data in the flash memory into different error grades, and grading the error correction capability of the ECC algorithms and the ECC check bits, with the ECC algorithms and the ECC check bits of different error correction grades being corresponding to different error grades; and updating to corresponding ECC algorithms and the ECC check bits when maximum error bit of the data in the flash memory reaches different error grades. Therefore, the method dynamically adjusts ECC bits at different stages in the life cycle for an NAND Flash, thereby realizing balance of an access rate and a service life of NAND Flash.

This invention provides an impulsive noise suppression method in orthogonal frequency division multiplexing. The method comprises an equalization and de-mapping step for estimating a preliminary estimation of signal and a total noise estimation by utilizing ideal channel estimation, de-mapping, and pilot insertion technique on received signal; and a SNR comparison step for determining a SNR by dividing said preliminary estimation of signal and said total noise estimation and comparing said SNR with a threshold value.

The invention discloses a synaptic transistor based on a two-dimensional semiconductor material and a preparation method of the synaptic transistor. The synaptic transistor comprises an insulating substrate, and a channel, a source electrode, a drain electrode and a gate electrode which are arranged on the substrate, wherein the channel is a two-dimensional semiconductor material; the source electrode and the drain electrode are arranged at the two ends of the channel respectively and form an ohmic contact with the channel material; the gate electrode and an electrical interconnection system formed by the channel, the source electrode and the drain electrode are kept in electronic insulation; an organic electrolyte covers a channel region and most of the gate electrode and comprises an organic carrier capable of being electrically insulated and ions capable of being migrated, and effective ion control of the gate to the channel material is formed. According to the synaptic transistor based on the two-dimensional semiconductor material and the preparation method of the synaptic transistor, an ion attachment-intercalation mechanism is utilized, and the characteristics of large surface area and adjustable resistance value of the two-dimensional material are combined, so that the device shows long-term and short-term synaptic plasticity, and the two characteristics can change witheach other along with the change of a gate signal. Meanwhile, the device has good linearity and ultralow operational power consumption, and the implementation and large-scale integration application of a high-precision neuromorphic device are facilitated.

The invention discloses a photonic convolutional neural network architecture based on optical delay line caching. The architecture can achieve the large-scale optical caching through employing the wavelength division multiplexing technology and the optical delay line technology, thereby completing the data shift operation which must be included in convolution calculation. Moreover, the architecture simultaneously realizes signal input and weight loading by using the array input electro-optical modulator and the array weight unit, can realize convolution calculation of the non-von Noemann architecture when the scale is expanded, gives full play to photon broadband and rate advantages, and reduces convolution calculation power consumption.

The invention discloses a configurable convolutional array accelerator structure based on Winograd. The structure comprises an activation value caching module, a weight caching module, an output caching module, a controller, a weight preprocessing module, an activation value preprocessing module, a weight conversion module, an activation value matrix conversion module, a dot multiplication module, a result matrix conversion module, an accumulation module, a pooling module and an activation module. According to the configurable convolutional array accelerator structure based on the Winograd, the convolutional array accelerator with the configurable bit width is designed according to the operation characteristics of a Winograd convolution algorithm of a fixed normal form, and the requirements of different neural networks and different convolutional layers for the bit widths are flexibly met. In addition, a special multiplier unit with configurable data bit width is also designed, so that the calculation efficiency of the neural network convolution operation is improved, and the calculation power consumption is reduced.

The invention discloses a method and device for realizing a Bayesian neural network by using memristor intrinsic noise. The method comprises the steps: obtaining a Bayesian network, and carrying out the training of the Bayesian network according to a selected data set, and obtaining the weight distribution of the Bayesian network; and processing the weight distribution of the Bayesian network, calculating according to the processed weight distribution and the conductivities of the plurality of memristors to obtain a target conductivity value, and mapping the target conductivity value into thememristors. According to the method, the Bayesian neural network is realized by utilizing the memristor cross array, and the power consumption is low, and the calculation speed is high, and the calculation energy efficiency is high.

The invention provides a processor applied to a sparse neural network. The processor includes a storage unit used for storing data and instructions, a control unit used for getting the instructions stored in the storage unit and sending out control signals, and a calculation unit used for getting the node values in a layer of a neural network and corresponding weight value data from the storage unit in order to get the node values in a next layer. When any of elements to be calculated is equal to zero, the calculation unit does not implement the multiplication of the calculation element, wherein the elements to be calculated include node values and weight values. By using the processor of the invention, the computing speed of the neural network can be improved, and power can be saved.

The invention discloses a neural network model real-time automatic quantification method, which is based on an embedded AI accelerator, and comprises the following steps: carrying out embedded AI neural network training at a PC end, establishing a PC end deep learning neural network, and training an input floating point network model of an embedded AI model; quantizing the floating point network model into an embedded end fixed point network model; preprocessing data needing to be quantized, and realizing all acceleration operators of each layer of the model network through a hardware mode; deploying embedded AI hardware of the embedded end and transplanting the neural network model of the embedded end, and transplanting the neural network model of the built AI hardware platform. The invention further discloses a neural network model real-time automatic quantification system. According to the invention, algorithm acceleration is realized based on an embedded AI accelerator hardware mode, the storage occupied space of a neural network model can be reduced, the operation of the neural network model can be accelerated, the computing power of embedded equipment can be improved, the operation power consumption can be reduced, and the effective deployment of the embedded AI technology can be realized.

The invention provides a dynamic differential positioning method based on mobile edge computing (MEC). The method comprises the following steps: dynamic differential positioning calculation modules are deployed on an MEC server; the MEC server receives regional differential correction values that are broadcast by a cloud computing platform in a real-time manner, wherein the cloud computing platform is based on a satellite navigation ground-based augmentation system; the MEC server receives observation data, which is based on a satellite navigation system, from a mobile terminal; the MEC servercarries out differential calculation according to the regional differential correction values and the observation data of the mobile terminal and through the dynamic differential positioning calculation modules so as to obtain a high-precision positioning result of the mobile terminal, and sends the high-precision positioning result to the mobile terminal; and the mobile terminal receives the high-precision positioning result from the MEC server. The method provided by the invention has the characteristics of millisecond-scale delays, large system capacity, high disaster tolerance, and high simplicity of a communication protocol.

The invention discloses an image data enhancement method and device, a medium and equipment, and the method comprises the steps: carrying out the training of a deep learning network according to an image training set, and obtaining a soft label generator, wherein the image training set comprises K types of image sample sets; obtaining a soft label and a category of the first image sample, whereinthe soft label represents the category association degree of the first image sample and the K categories; obtaining category representativeness of the image samples in the K-1 categories of image sample sets according to a soft label generator, wherein the K-1 categories do not contain the category of the first image sample; selecting a second image sample from the K-1 types of image sample sets according to the soft label of the first image sample and the type representativeness of the image sample; and fusing the first image sample and the second image sample to obtain a target image sample.The invention relates to the field of computer vision, and by generating a new image sample with controllable difficulty, the training difficulty degree of a classification network model can be adjusted to improve the generalization ability of the model.

The invention discloses a power transmission line laser radar point cloud acquisition and external damage prevention real-time monitoring method. The method comprises the following steps: remotely aligning a to-be-monitored power transmission line channel; the laser radar scans for multiple times to obtain point cloud data of a power transmission line channel, and high-density point cloud data is obtained; power transmission line point cloud data and non-power transmission line point cloud data are separated and stored; the laser radar scans the power transmission line channel and acquires point cloud data of the power transmission line channel in real time; and according to the point cloud data of the power transmission line, taking the point cloud data of a single power transmission line, searching whether other point cloud data exist in a safe range at an interval distance set by a user along the erecting direction of the power transmission line, and judging whether the power transmission line has an external damage risk. According to the invention, the remote power transmission line point cloud data can be collected, the problem that the remote power transmission line is difficult to generate the point cloud data in real time and monitor in the laser radar is solved, the point cloud data in a specified range is searched through traversal along the power transmission line, and real-time and rapid external damage prevention monitoring is realized.

The present invention provides a tumble pre-judgment method based on human body key point behavior recognition and LSTM. According to the tumble pre-judgment method based on human body key point behavior recognition and LSTM, the human body is further divided into a head area, a trunk area and a leg area for behavior recognition, the calculation amount is greatly reduced, and therefore the detection efficiency is improved; and on the basis, the memory function of the acquired video is realized by adopting an LSTM (Long Short Term Memory), namely a long-term and short-term memory neural networkmechanism, so that the functions of analyzing and identifying the behavior change of the human body are realized, and finally, the identification results are classified into three types, namely tumble, non-tumble and others. According to the method, the calculation power consumption is reduced, and the fall detection time is saved, so that the functions of real-time detection and fall detection pre-judgment are realized.

The invention provides a less complex successive cancellation decoding method for a Polar code, aiming to lower the calculation complexity of a Successive Cancellation (SC) decoding solution. The invention solves a problem that the original SC decoding solution is highly complex in calculation. A log likelihood ratio is calculated in the receiving terminal by comparing a constant error linear piecewise function to a hyperbolic function in an SC decoding solution rather than by using recursive calculation of a hyperbolic tangent function directly, that is, multiplication and summation operations replace log, exponential and division operations in the SC method. Compared with the original SC decoding solution, the improved method greatly lowers the calculation complexity while maintaining the property.

The present invention discloses a panoramic image storage and display methods and devices. The storage method comprises: dividing a to-be-stored target panoramic image into a preset quantity of image slices according to a pre-determined coordinate system, and acquiring coordinates of each image slice; and storing each image slice and index information corresponding to the image slice, so as to complete storage of the target panoramic image. According to the method, when the panoramic image is displayed, the consumption of system bandwidth and computing power can be reduced; waste of system resources can be avoided; and display efficiency of the panoramic image can be improved.

The invention provides a device and method for treating multiplication and addition calculations. The device comprises a first adder, which is used for conducting an addition calculation on first data and second data which are input and obtaining first middle data, a second adder, which is used for conducting an addition calculation on third data and fourth data which are input and obtaining second middle data, and a logarithm adder, wherein the values of the first data and the second data are logaA and logaB respectively, the values of the third data and the fourth data are logaC and logaD respectively, and the value of the second middle data is n; the logarithm adder is used for obtaining a<n-m> according to m and n input by the first adder and the second adder, and approximately determining the sum of m and a<n-m> as the value of (loge)*loga (A*B+C*D), and the first adder, the second adder and the logarithm adder are achieved by a hardware circuit. According to the device and method for treating multiplication and addition calculations, the power consumption of the calculations can be reduced in the calculation process.

The invention discloses a video snow noise detection method, and belongs to the technical field of video image processing. The method comprises the following steps: 1), image preprocessing: enabling a high-resolution video image to be compressed and cached at an equal proportion after a current video image is obtained, calculating the difference of a current compressed image frame and a former compressed image frame, and obtaining a difference graph; 2), edge point extraction: carrying out the Sobel operator processing of the difference graph, obtaining an edge point result, and carrying out the quantitative analysis of the edge point result, wherein a snow noise result can be obtained under the extreme condition; 3), uniformity analysis: dividing an edge point result, which is not under the extreme condition, into N longitudinal and lateral detection zones for uniformity analysis; 4), result output: obtaining a snow noise detection result at stages of P2 or P3. The method is low in consumption of calculation power, can achieve the intervention at any moment, and is good in detection capability.

The invention discloses a low-power-consumption human body state detection system based on NB-IoT. The low-power-consumption human body state detection system comprises a shoe detection device and a cloud server which are disposed on a shoe, and the shoe detection device comprises an MCU main control module, and a sensor and an NB-IoT module which are connected with the MCU main control module; the sensor is used for monitoring human bodies and generating human body monitoring data; the MCU main control module is used for reading human body monitoring data generated by the sensor and outputting the human body monitoring data to the NB-IoT module; the NB-IoT module is used for forwarding the human body monitoring data to the cloud server through the NB-IoT network; and the cloud server is used for calculating and obtaining human body state information according to the human body monitoring data. According to the low-power-consumption human body state detection system based on NB-IoT, the power consumption of the system can be reduced, the hardware configuration cost can be reduced, and the accuracy of detection result can be improved.