37results about How to "Consumes less memory" patented technology

Electronic television program guide information is downloaded, stored and searched by a local device. A plurality of linked lists are used to remove redundancy, increase search speed, and allow dynamic categories. When a new category is encountered a new head node for a link list data structure is created. When a first program guide object containing information related to that category is downloaded, a link is made (i.e., a pointer is stored) from the head node to the guide object. When subsequent program guide objects containing information related to that category are downloaded, links are added to the end of the list or inserted in alphabetical order. As guide objects expire the links are removed. When no links remain in a particular category the head node for that category is deleted.

The invention relates to a map linear symbol drawing method and a map linear symbol drawing system based on a graphics processing unit (GPU). According to the method, under a three-dimensional rendering programmable assembly line system, a shader language is used for sending a function relation between map linear symbol coloring and linear parameters to the GPU for calculation, so that map linear symbols can be drawn with high efficiency. The method comprises the following steps of establishing a vortex array caching object by taking half-linear width as a buffering region according to a linear center line, recording xyz of the position of a vortex and transverse and longitudinal coordinate relation parameters uv of the vortex on the whole buffering zone; writing a vortex shader, sending the uv parameters into a fragment shader, and performing rasterization interpolation by using the GPU; and building a function of a fragment color and the uv in the fragment shader, and calculating pixel coloring. In an actual drawing process, the vortex caching objects are bound to be property parameters for sending; global parameters required in a shader file are set; and therefore, the linear symbols can be drawn.

The invention discloses a multi-modal dynamic gesture recognition method based on a lightweight 3D residual network and a TCN. Firstly, original videos in a data set are sampled and sorted and storedaccording to a time sequence; then, the lightweight 3D residual network is pre-trained by using a large public gesture recognition data set, and thus storing a weight file of the model; then, the RGB-D image sequence is used as input, the lightweight 3D residual network and the time convolution network are used as basic models to extract long-term and short-term spatial and temporal features, andan attention mechanism is used for weighted fusion of multi-modal information, wherein the RGB sequence and the Depth sequence are respectively input into the same network structure; finally, a full connection layer is used for classification, a cross entropy loss function is used for calculating a loss value, and accuracy and F1Score are used as evaluation indexes of the network model. The invention not only can achieve high classification accuracy, but also has the advantage of low parameter quantity.

The invention discloses a model identification method based on a local characteristic aggregation descriptor. First of all, SIFT characteristics of vehicle images in a model database are extracted; then Kmeans clustering is performed on the SIFT characteristics of all the vehicle images so as to form K cluster centers and obtain a dictionary with K visual words; then, each SIFT characteristic is endowed with a closest visual word for each model image; a residual error cumulant between SIFT characteristic vectors around each visual word and current visual words is counted, and the local characteristic agglomeration descriptor of a current vehicle picture is obtained; finally, an indexable coding image database of n model types is obtained through quantification coding from local characteristic agglomeration descriptors of n model images of a training module; and for a test vehicle image, the local characteristic agglomeration descriptor of the test vehicle image is extracted in the same way as a query vector for introduction into an image database for indexing, and the model of a test vehicle is identified through matching by use of an approximation nearest neighbor search method.

The present invention provides a storage method of executable code under an ARMv7m architecture. The method involves five logic units including a compilation and burning program, a startup program, an FPB association program, an execution program, and a bus exception handling program. According to the method, the FPB association program associates a RAM space to an address that is not associated with a physical flash, and uses the RAM space as a storage space where the executable code is stored, and a CPU kernel uses the storage space to acquire and execute a machine instruction. In addition, the RAM space may be dynamically updated to a new value, and may be re-reset and re-associated by the FPB to a new flash address. With the method according to the present invention, a small on-chip flash is used to store FPB management code. As such, a small RAM space and a cheap off-chip flash are used to expand the storage space of the executable code to 512 MB.

The invention provides a mononucleotide variation detecting method based on blood circulation tumor DNA, a device and a storage medium. The method comprises the steps of acquiring mutation data of each site of the blood circulation tumor DNA of a testing sample, wherein the mutation data comprises a site mutation frequency; acquiring a confidence range of each site background mutation frequency ofthe training sample, wherein the confidence range is obtained through performing learning modeling on all three-base mutation frequency and site mutation frequency in each training sample and training the model by means of an in-place updated list; comparing the site mutation frequency of each site of the testing sample with the confidence range of the background mutation frequency of each site in the model, and outputting the mononucleotide variation wherein the site mutation frequency of the testing sample is not in the confidence range. The mononucleotide variation detecting method, the device and the storage medium have advantages of greatly optimizing a calculation resource requirement and a detecting speed, improving sensitivity and accuracy in detecting ctDNA mononucleotide mutation, and satisfying a requirement for ctDNA mononucleotide mutation detecting reliability in clinical tumor detection.

The invention discloses a double-layer license plate character recognition method based on deep learning, and relates to the technical field of license plate recognition. The double-layer license plate character recognition method comprises the following steps of constructing a deep neural network model, wherein the deep neural network model comprises a character positioning network, a character synthesis network, a character recognition network, a character positioning network, a character synthesis network and a character recognition network which are sequentially connected with each other,obtaining a double-layer license plate training sample image set, training a deep neural network model, and recognizing a to-be-recognized double-layer license plate image through the trained deep neural network model. According to the present invention, the deep neural network model is smaller in memory consumption, smaller in calculation amount, more accurate in license plate character recognition result and higher in robustness, and the defects that a traditional license plate character recognition effect is poor and the robustness is low are overcome.

The invention discloses a semantic segmentation-based surface feature recognition and classification method and device. The method comprises the following steps: 1) acquiring multi-source remote sensing data of each region, and combining the data of the same region into a sample to obtain a sample set; 2) establishing a semantic segmentation model; training the model by using the sample set, wherein the semantic segmentation model is formed by connecting an encoder, a central module and a decoder in series; the encoder is formed by connecting N encoding modules in series, the decoder is formedby connecting N decoding modules and a point convolution module in series, and the center module is formed by connecting a convolution module C1 and a convolution module C2 in series, and each encoding module El is formed by connecting a convolution module El1, a convolution module El2 and a down-sampling module DSl in series, and each decoding module D1 is formed by connecting an up-sampling module USl, a convolution module Dl2, a convolution module Dl1 and a convolution module Dl0 in series; and 3) processing the remote sensing data to be identified by using the trained model to obtain an identification result of the ground object type.

The invention discloses a mixed analysis method for electromagnetic scattering of a cavity containing a medium target. The method comprises the following steps that a cavity body part containing the cavity target is filled to be solid metal, a target discrete model after filling is conducted is established, and a matrix equation is constructed; recursive solving is conducted on node electric field values on tangent planes in sequence; correction is conducted on the electric field values on the last tangent plane, electric field values near discrete nodes are obtained, and phase correction is conducted on the electric field values; solving is conducted on the cavity body part in the target independently by using a body area integral equation, metal surfaces and dielectric bodies in the cavity body are dispersed to obtain sub scatterer grouping, induced currents on the metal surface and polarization currents in inner bodies of the medium are solved by using a fast multilevel method, the electric field values of the discrete points needed by a parabolic equation on an opening surface of the cavity body are solved, and switch solving of a radar discrete section area is conducted on near-field electric field values. The body area integral equation is combined with a meshless parabolic equation, and the defect that cavity body scattering cannot be calculated by the parabolic equation is avoided.

A method and system for finding a longest prefix match for a key in a computer network is disclosed. The method and system include providing a main engine and providing an auxiliary engine. The main engine is for storing a first plurality of addresses and for searching the first plurality of addresses for the longest prefix match for the key. None of the first plurality of addresses is a prefix for another address of the first plurality of addresses. The auxiliary engine is for storing and searching a second plurality of addresses. A first address of the second plurality of addresses is capable of including the prefix for a second address of the first plurality of addresses or for a third address for the second plurality of addresses. None of the first plurality of addresses is the prefix for any of the second plurality of addresses. Each of the second plurality of addresses is distinct from each of the first plurality of addresses.

The invention provides a memory optimization method for video animation picture storage. The memory optimization method comprises the steps of extracting video animation pictures, and acquiring a PNG picture of each frame in the video animation pictures based on the video animation pictures; based on the PNG picture of each frame, converting the pictures into JPG pictures; and based on the JPG pictures, converting the pictures into video animation pictures for reproduction. The memory optimization method for the video animation picture storage according to the invention can reduce the memory consumption of the video animation picture processing by extracting the video animation pictures, converting the PNG pictures into the JPG pictures through alpha channel data and then deleting the alpha channel data, so as to achieve the purpose of memory optimization and reduce the memory of run-time pictures.

The invention discloses a robot navigation method oriented to virtual images. The method comprises steps as follows: a first bidirectional flow field between a first image and a second image is calculated; a first virtual view angle image and a second virtual view angle image are generated by adopting the first bidirectional flow field; a second bidirectional flow field between the first virtual view angle image and the second virtual view angle image is calculated, and field depth information of the first virtual view angle image and the second virtual view angle image is calculated; the first virtual view angle image and the second virtual view angle image are subjected to image close-up according to the field depth information, and a virtual image is obtained through fusion of an obtained first virtual close-up image and an obtained second virtual close-up image; a panoramic view image of an ambient environment of a robot is acquired and matched with the virtual image for azimuth positioning; a distance calculation function of the robot is fitted out according to a corresponding relation between an image feature point ratio and a distance, and a motion distance of the robot is output. The robot navigation method has the advantages of smaller storage data, low memory consumption and high calculation speed.

The invention relates to an LTE-A (LTE-Advanced) control channel resource demapping method and belongs to the technical field of communication. According to the method, data extraction of integral multiples is carried out on system bandwidths through utilization of FIR (Finite Impulse Response) filters and the bandwidths are normalized into 1.4M, so resource consumption for resource demapping is reduced, and the problem that the computing quantity is complex, the operation speed is slow and the source consumption is high due to high data quantity under high bandwidths is solved. According to the method, through decoding of PBCHs (Physical Broadcast Channels), the system bandwidths are acquired, all system bandwidths are normalized, the computing quantity of a terminal is reduced, the memory consumption is reduced, and the operation speed of the terminal is improved.

The invention relates to a bridge monitoring method and system based on binocular vision, and the method comprises: calibrating a first camera and a second camera, obtaining the relative pose relationbetween the two cameras, and determining the real-time three-dimensional coordinates of a central point on a calibration plate under a first camera coordinate system; calculating a coordinate conversion relationship between the first camera coordinate system and the calibration plate coordinate system according to the calibration information, and converting a three-dimensional coordinate of the central point under the first camera coordinate system into a real-time three-dimensional coordinate under the calibration plate coordinate system; and determining the displacement information of the bridge by combining the three-dimensional coordinates of the central point at the initial moment under the calibration plate coordinate system. By calibrating the relative pose relationship between thetwo cameras, the three-dimensional coordinate of the center point of the calibration plate under the first camera coordinate system is converted to be under the coordinate system of the calibration plate, and the displacement information of the bridge is determined by combining the coordinates of the center point of the initial moment under the coordinate system of the calibration plate; and manual measurement is replaced, and the method is high in precision, high in efficiency and low in price.

The invention relates to a method and a system for detecting fall of an elderly person based on depth learning, which comprises the following steps: acquiring scene image data through a camera to determine a target detection area; according to the scene image data, the instantaneous motion velocity field of all pixels in the current frame image being calculated by preset logic, and the pixel motion velocity field image being obtained; gathering similar pixels according to the pixel motion velocity field image and preset similarity conditions to form a candidate moving target region, and acquiring a candidate moving target in the candidate moving target region; screening out the new moving objects from the candidate moving objects; Identifying pedestrian targets from new moving targets according to depth learning method; updating target tracking list information of pedestrian targets; based on the change state of the target tracking list information and the preset judgment condition, whether the pedestrian target falls or not is judged and an alarm is given.

The invention discloses an image semantic segmentation method based on multi-receptive-field context semantic information. The method comprises the steps of 1, converting an input image into a pixel matrix through convolution operation; 2, converting the same pixel matrix into a plurality of feature maps with multi-receptive-field context semantic information by adopting expansion convolution with different expansion rates; 3, performing feature extraction and down-sampling processing on the feature map with the multi-receptive-field context semantic information through converter encoders in different subnets to obtain a plurality of down-sampling feature maps with different receptive fields; 4, carrying out step-by-step up-sampling processing on the down-sampling feature map through a decoder to obtain feature maps with the same size and dimension, and generating a final feature fusion map; and 5, the feature fusion image completes image segmentation prediction through a convolutional neural network. The method can be effectively applied to image semantic segmentation, does not lose deep low-resolution features and fine-grained features, consumes a small memory, is remarkable in effect, and is convenient to popularize.

The invention discloses an audio feature extraction method and device based on a re-parameterized decoupling mode. The method comprises the following steps: acquiring a to-be-detected voice sample of a target speaker; preprocessing the to-be-detected voice sample; extracting acoustic features of the preprocessed to-be-detected voice sample; and inputting the acoustic features into a network reasoning module to obtain voiceprint feature vectors, wherein the network reasoning module is a network model of a single-path structure converted by a trained multi-layer network training module through re-parameterization. According to the method, the multi-branch structure is used in the training stage to achieve a better convergence effect, the multi-branch structure is re-parameterized into the single-path structure in the reasoning stage to obtain a better effect than the multi-branch structure with equivalent parameter quantity, the speed is higher, and the memory consumption is lower.

The invention discloses a navigation method for correcting an inertia angle of a robot based on a visual angle. Pure inertia navigation is completed just through the data of a gyroscope in a short-time walking process of a mobile robot, low-pass filtering processing is carried out on a difference value between the angle of the gyroscope and the angle of a machine body acquired by a camera so as toobtain a filtering angle correction value which is kept in a stable state for a long time, and after the mobile robot takes a relatively long time to traverse a preset planning path in a corresponding room area, a filtering angle correction value obtained at a corresponding latest sampling moment is fed back to a visual initial navigation angle acquired and calculated on a preset planned path ina non-traversed room area so as to realize compensation of accumulated errors, so that the problem of offset errors accumulated by long-time inertial navigation can be solved; therefore, the navigation angle of the robot is kept high in precision for a long time, and the mobile robot is prevented from deviating in a short-time walking path and a long-time walking path. And the real-time performance of robot navigation is improved.

The invention belongs to the traveling wave tube simulation technology, and discloses an Euler method for simulating the linear injection wave interaction of a traveling wave tube. On the basis of the linear injection-wave interaction theory, the Euler linear injection-wave interaction model of the present invention is obtained by linearizing it. The Euler method of the present invention has the characteristics of fast calculation speed and low memory consumption, and can accurately simulate the linear injection wave interaction process in the helical traveling wave tube.

The invention discloses a simulation method and device of a mass spectrum isotope fine structure and an ultrafine structure. The simulation method comprises the following steps: calculating an isotopefine and ultrafine structures of a molecular formula formed by any one element through an isotope vector operation, wherein an isotope vector is a combination of a quality vector and an abundance vector; arranging the quality vector according to an ascending order or a descending order, wherein abundance vector element corresponds to quality vector elements; and building an isotope element cluster vector. A high-efficient calculation method of the isotope fine structure is provided, a high-efficient algorithm of a designated mass-to-charge ratio isotope ultrafine structure is given, and high-efficient calculation of the isotope fine structure and the ultrafine structure is realized by using a unified algorithm; the problems that less internal memories of the vector operation and a change-keeping method are consumed, the operation speed is high and the calculation precision is high are found, and the device can be applied to distribution and calculation of the isotope of any elements,and the self-defined isotope distribution is supported.

The invention discloses a construction method of a super-resolution convolutional neural network model based on a portable system. The method comprises the steps of 1, constructing a convolutional neural network model; 2, training the convolutional neural network model in the step 1; and 3, deploying the convolutional neural network model trained in the step 2 to realize a super-resolution function of the image. According to the method, an up-sampling algorithm of the sub-pixel convolution layer of the ESPCN is adopted, a network structure before the sub-pixel convolution layer is modified, and a residual block is introduced to construct a network.

The present invention provides an efficient image classification method based on structured pruning, which incorporates a spatial pruning method based on variation regularization, including steps such as image data preprocessing, inputting images to neural network, image model pruning and retraining, and new image class predication and classification. The present invention adopts a structured pruning method that removes unimportant weight parameters of the original network model and reduces unnecessary computational and memory consumptions caused by the network model in image classification to simplify the image classifier, and then uses the sparsified network model to predict and classify new images. The simplified method according to the present invention improves the original network model in image classification efficiency by nearly two times, costs about 30% less memory consumption and produces a better classification result.