128results about How to "Speed ​​up the convergence process" patented technology

The invention discloses a target tracking method and system of a convolution twin network based on multi-layer feature fusion. The method comprises the following steps of: according to the target position and size of the images, cutting out the target template images and the search area images of all images in the image sequence training set, and forming a training data set by image pairs composedof the target template images and the search area images; constructing a convolution twin network based on multi-layer feature fusion; training the convolution twin network based on the multi-layer feature fusion based on the training data set to obtain a trained convolution twin network based on the multi-layer feature fusion; using the trained convolution twin network based on multi-layer feature fusion for target tracking. In the process of tracking targets, the invention integrates scores of different layers, combines high-level semantic features and bottom-level detail features, better distinguishes the interference of similar or similar targets, and prevents the problems of target drift and target loss in the tracking process.

The present invention relates to lithographic apparatuses and processes, and more particularly to tools for optimizing illumination sources and masks for use in lithographic apparatuses and processes. According to certain aspects, the present invention significantly speeds up the convergence of the optimization by allowing direct computation of gradient of the cost function. According to other aspects, the present invention allows for simultaneous optimization of both source and mask, thereby significantly speeding the overall convergence. According to still further aspects, the present invention allows for free-form optimization, without the constraints required by conventional optimization techniques.

The present invention relates to lithographic apparatuses and processes, and more particularly to tools for optimizing illumination sources and masks for use in lithographic apparatuses and processes. According to certain aspects, the present invention significantly speeds up the convergence of the optimization by allowing direct computation of gradient of the cost function. According to other aspects, the present invention allows for simultaneous optimization of both source and mask, thereby significantly speeding the overall convergence. According to still further aspects, the present invention allows for free-form optimization, without the constraints required by conventional optimization techniques.

An in-line compensation method for the match error of multiple sensors integrated tracking system features that the maximal probability estimation method is used to determine the initial error, extracting the system error by use of the information measured by sensors, and the filtered tracking information, and performing in-line compensation.

A method of replicating a mental state of a first subject in a second subject comprising: capturing a mental state of the first subject represented by brain activity patterns; and replicating the mental state of the first subject in the second subject by inducing the brain activity patterns in the second subject.

The multi-parking-lot and multi-vehicle-type vehicle path scheduling control method comprises the steps of 1, establishing an objective function by taking the lowest total cost of all delivery vehicles as an objective; Step 2, performing a coding step; 3, performing population initialization; 4, evaluating all the individuals by adopting the objective function as a fitness function; Step 5, performing selection and crossover operation; step 6, performing mutation operation; 7, performing neighborhood search on each individual in the population by using an improved extreme value optimization algorithm; Step 8, calculating fitness of all individuals in the population; Step 9, performing selecting; Step 10, performing elite retention; Step 11, completing iteration in sequence; Step 12, judging whether a termination condition is met or not, the termination condition being that the number of iterations g reaches the maximum number of iterations MaxGen or the number of iterations Nu of whichthe Gb fitness value remains unchanged reaches the specified number of iterations Kbest, if yes, continuing to execute the step 13, and if not, returning to execute the step 5; Step 13, outputting the individual Gb and the fitness value fGb thereof; And 14, interpreting the optimal individual Gb and the fitness value fGb thereof. The invention aims to improve the search efficiency and convergencespeed of the algorithm.

Low density parity check (LDPC) decoders are described utilizing a sequential schedule called Zigzag LBP (Z-LBP), for a layered belief propagation (LBP) architecture. Z-LBP has a lower computational complexity per iteration than variable-node-centric LBP (V-LBP), while being simpler than flooding and check-node-centric LBP (C-LBP). For QC-LDPC codes where the sub-matrices can have at most one “1” per column and one “1” per row, Z-LBP can perform partially-parallel decoding with the same performance as C-LBP. The decoder comprises a control circuit and memory coupled to a parity check matrix. Message passage is performed within Z-LBP in a first direction on odd iterations, and in a second direction on even iterations. As a result, a smaller parity check matrix can be utilized, while convergence can be more readily attained. The inventive method and apparatus can also be implemented for partially-parallel architectures.

The present invention provides an automated design method and platform for intelligent hardware system development. The method comprises: creating various types of module resource libraries; performing demand analysis on a newly-developed product; forming a system parameter configuration table according to various requirements of the newly-developed product; importing the parameter configuration table into an automated design platform; enabling the platform to call a module resource library to perform scheme optimization; and according to the parameter configuration table, generating, by configuration, various types of documents, a circuit principle diagram, a circuit PCB layout, a hardware drive program, a mobile terminal application and a cloud server program, so as to perform debug detection. The platform is an operating platform of the method, and the platform comprises a plurality of application functional modules and module resource libraries. Powerful module resource libraries are built in the platform, and non-professional personnel can implement type selection of various types of hardware components by means of product functions and performance parameters; various types of software sub-modules and hardware sub-modules have a strong called property and high reusability, and the hardware is short in development cycle and high in development success rate; and excessive professional personnel are not required, and the research and development costs are low.

The invention discloses a generation method of convolutional neural networks for conducting expression recognition on the human face in an image, an expression recognition method, calculation equipment and a mobile terminal. The generation method of the convolutional neural network comprises the steps that the first convolutional neural network is established, wherein the first convolutional neural network comprises a first number of processing modules, a first overall average pooling layer and a first classifier which are connected in sequence; according to a pre-acquired facial image data set, the first convolutional neural network is trained, and the first classifier outputs and indicates an expression corresponding to the human face conveniently, wherein the facial image data set comprises multiple pieces of facial image information; the second convolutional neural network is established, wherein the second convolutional neural network comprises a second number of processing modules, a second overall average pooling layer and a second classifier which are connected in sequence; according to the facial image data set, the trained first convolutional neural network and second convolutional neural network are subjected to joint training, and the second classifier outputs and indicates the expression corresponding to the human face conveniently.

The invention relates to a semantic segmentation method based on a reverse attention model. The method mainly comprises the following steps: acquiring an image data set, and constructing a training set and a test set; constructing a deep semantic segmentation network model, wherein the deep semantic segmentation network model comprises a basic network model and a reverse attention model; and inputting the features output by the basic network into a reverse attention model to calculate attention views, gradually counteracting the attention views on the low-level output features of the basic semantic segmentation network, and fusing the attention views with the output features of the basic network and the up-sampling features thereof to obtain a final segmentation result. According to the model, only basic semantic segmentation network output features are used for calculating an attention view, and low-level features are guided to be fused into the basic semantic segmentation network output features, so that noise in the low-level features of the model is suppressed, and the robustness and segmentation precision of the semantic segmentation model are improved; meanwhile, a loss function based on Gumbel softmax is added to high-level output features of the basic semantic segmentation model, so that the model training speed is increased.

The invention discloses a method for optimizing a removing function of a jet flow polishing material of a workpiece to be polished, which comprises the steps of: detecting surface shape data of a workpiece to be polished before polishing by using a jet flow polishing system; placing the workpiece to be polished in a container filled with a polishing solution; regulating the distance between an outlet of a nozzle and the surface of the workpiece to be polished to ensure that the outlet of the nozzle is immersed in the polishing solution; within a unit time, carrying out full-immersion jet polishing on the workpiece to be polished in a fixed point manner by using the nozzle to obtain removal distribution of a removing material for removing the workpiece to be polished; and detecting the removal distribution of the removing material for removing the workpiece to be polished by the nozzle in a fixed point polishing region within a unit time by using an interferometer to generate surface shape data of the workpiece to be polished after the removing material is polished, and subtracting the surface shape data of the workpiece to be polished after the removing material is polished from the surface shape data of the workpiece to be polished before polishing by using a calculating unit to obtain optimization distribution for displaying the removing function of the material.

The invention relates to a mask pooling model training and pedestrian re-identification method for pedestrian re-identification. The method comprises the steps of S1, obtaining an anchor image a, a positive sample image p and a negative sample image n; S2, respectively inputting a, p, n and masks corresponding to a, p and n into the mask pooling model to obtain corresponding three-dimensional tensors Ta, Tp and Tn; S3, respectively carrying out pooling operation and convolution operation on Ta, Tp and Tn to obtain corresponding Ha, Hp and Hn; S4, inputting Ha, Hp and Hn into a classifier respectively to obtain corresponding prediction results Ra, Rp and Rn; S5, calculating loss values according to the prediction results Ra, Rp and Rn; and S6, training the mask pooling model according to the loss value. According to the invention, the non-background information in the image can be enhanced, and the most critical feature of the image can be learned.

The invention belongs to the technical field of Raman scattering, and discloses an asymmetric weighted least squares based Raman spectrum detection baseline correction method. The method includes, fora Raman spectrum to be fitted to a baseline, setting smoothing parameters and an iteration termination condition, initializing weight into unit matrixes during first iteration, and obtaining an initial fitting baseline through minimizing penalized least squares; and performing updating on the weight through a softsign function according to the difference signal between an original spectrum and the fitting baseline, the process being repeated continuously until a termination condition is met. The original Raman spectrum subtracts the final fitting baseline, so that baseline correction of the Raman spectrum can be realized. Correction can be performed on the original Raman spectrum by utilizing the method, so that spectrums after baseline correction can be closer to real Raman signals, theaccuracy of Raman spectrum detection can be enhanced, and the method has important practical significance for the quantitative analysis and applications of the Raman spectrum.

The invention discloses a method for describing a power pole image based on texture clustering, belonging to the technical field of description of the power pole. The method comprises the steps of: firstly, generating codebooks from a training sample; then carrying out convolution on the codebooks and test images to generate correlation surfaces, selecting the correlation surfaces which has the largest similarity with the codebooks to form a prototype distribution diagram of the power pole, and presenting the prototype distribution diagram into a column diagram, i.e., tensor; and finally, quantifying the tensor into a correlation surface set descriptor which has a fixed size and can be used in a classifier. The method has the beneficial effects that a real image containing the power pole, obtained by polling various overhead power lines, is used as experimental data to verify that the method is suitable for the description of the power pole in rich texture; and the description is accurate, and the computation complexity is low.

The invention provides a compressed video acquisition and reconstruction system based on a data driven subspace set. The compressed video acquisition and reconstruction system comprises a subspace set construction module, a sparse basis matrix construction module, a video signal sensing module and a reconstruction processing module, wherein a clustering method is utilized by the subspace set construction module to generate the subspace set, a linear subspace learning method is utilized by the sparse basis matrix construction module to generate a sparse basis corresponding to the subspace set, video signals are projected in a block mode through the sensing module, and obtained data are decoded and reconstructed in the reconstruction processing module at last. Compressed sampling is provided, meanwhile a distributed gradual model structure in a video sampling process agrees with the system, and reconstruction accuracy and efficiency of a sparse basis matrix special structure is also promoted. Sampling efficiency of video signals is greatly improved, reconstruction gains are obtain compared with other methods under different sampling compression ratios, and meanwhile the system has good expandability.

The invention relates to a link convergence accelerating method and device, adopting link state detection protocol to detect link state and obtaining link state variation information; according to the link state variation information, applying spanning tree protocol to make link convergence processing so as to quickly and accurately obtain transmission link state variation information, and thus assuring accuracy of topology calculation results and increasing the speed of link convergence processing.

The invention relates to a method for calculating micro-pressure waves generated in the process of passing through a tunnel by a high-speed train. The method for calculating micro-pressure waves is characterized by comprising the following steps: dividing a flow field of the tunnel into a plurality of areas by using a dynamic mesh; calculating areas containing a first compression wave and an expansion wave and an area before the compression wave and the expansion wave; and simulating to obtain a micro-pressure wave value of a tunnel outlet. The method for calculating the micro-pressure waves generated in the process of passing through a tunnel by a high-speed train better solves the problems of oversize scale, overlong calculating period, and the like of a three-dimensional calculating mesh of a long tunnel and a tunnel group. In the method, the flow field is divided into a plurality of the areas by adopting dynamic mesh technology and the value simulation on the micro-pressure waves of the tunnel outlet are carried out by only calculating the areas containing the first compression wave and the expansion wave and the area before the first compression wave and the expansion wave according to the forming mechanism of the micro-pressure waves. Meanwhile, the invention provides a new method for the calculation of the micro-pressure waves of the tunnel outlet.

The invention discloses a multi-grid processing method in particle image velocimetry (PIV). The method comprises the following steps: setting a first network, obtaining an equations set under the first network (corresponding to a high-resolution image) corresponding to an equations set to be solved; iteratively solving the equations set and obtaining an equation error according to a practical solution; setting a second network with a size thicker than the first network; mapping a solving scale of the equations set under the first network to a solving scale of the equations set under the second network; iteratively solving a high-frequency calculation error under the second network; mapping the scale of the high-frequency calculation error under the second network to obtain a low-frequency calculation error under the first network; and updating the practical solution with the low-frequency calculation error under the first network. The method can greatly accelerate the PIV processing process and can not lower the processing reliability and the processing precision.

The invention is suitable for the technical field of anti-icing and deicing, and provides a wall surface temperature simulation method of an anti-icing system, which comprises the following steps: carrying out grid division on a wall surface and a dry air flow field, the grid division of the wall surface being consistent with the grid division of the dry air flow field; an initial wall surface temperature Ts0 is obtained through a dry air flow field simulation module; calculating the total energy Qi-1 of the surface film through the (i-1) th wall surface temperature Tsi-1, and substituting the total energy Qi-1 of the surface film into the dry air flow field simulation module to obtain the ith wall surface temperature Tsi; judging whether the ith wall surface temperature Tsi converges or not; if the ith wall surface temperature Tsi is not converged, i is made to be equal to i + 1, and the steps are repeated; if the ith wall surface temperature Tsi converges, the ith wall surface temperature Tsi is the final wall surface temperature. The wall surface temperature simulation method provided by the invention is good in convergence, convenient to solve and closer to a real physical process.

Stereo sound signals are reproduced directly from loudspeakers (SP(L), SP(R)). By using a sum signal and a difference signal of the stereo sound signals as a reference signal, and according to a cross spectrum calculation of the reference signal with a microphone-collected sound signal, calculation is performed to obtain transfer functions of four sound transfer systems between the loudspeakers (SP(L), SP(R)) and microphones (MC(L), MC(R)). The transfer functions obtained are subjected to inverse Fourier transform to obtain impulse responses, which are set in filter means (40-1 to 40-4) to create echo cancel signals and perform echo canceling. This solves the problem of an indefinite coefficient in the echo cancel technique of a multi-channel sound signal.

The invention discloses an ancient poem automatic identification method based on deep learning, and the method is characterized in that the method comprises the steps:by including collecting trainingcorpora, carrying out the data preprocessing, carrying out the embedding of feature vectors, carrying out the training of a neural network, and carrying out the automatic identification of ancient poems. The deep neural network model in the text classification form is constructed, whether the text sentences are ancient poetry types or not is automatically recognized, and meanwhile it can be effectively avoided that wrongly written characters are wrong, and the recognition accuracy is reduced. And the requirements of application scenes such as poetry quality detection, literature work classification management and automatic collection of ancient poetries on an ancient poetry automatic identification technology can be met.

A search engine is provided that searches the internet for a word (or set of words) referred to a searched words. This first search may use a commercially available search engine. The results of the first search are used to create correlative words using unique and count procedures. Those correlative words with the highest count (correlation) are displayed first. A subset of the correlative words is inserted in the first search engine and reruns the search, This previous step is repeated recursively or sequentially until the results converge. The search converges faster if a word of high correlation is excluded or a word of low correlation is included.

The invention provides an unmanned aerial vehicle auxiliary resource allocation method based on deep reinforcement learning, and the method comprises the following steps: S1, constructing a deep reinforcement learning model, obtaining a neural network, and initializing neural network parameters; s2, obtaining calculation task information generated by the intelligent device and integrating the calculation task information into a system state St; s3, inputting a system state St to train the neural network to obtain a system action At; s4, calculating to obtain the corresponding total overhead Ctotal according to the system action At; s5, training a neural network according to the total overhead Ctotal to obtain a system action for minimizing the total overhead; s6, completing the training ofthe neural network, and carrying out the resource distribution according to the obtained system action enabling the total overhead to be minimized. The invention provides an unmanned aerial vehicle auxiliary resource allocation method based on deep reinforcement learning, and solves the problems of relatively high calculation task time delay and energy consumption of existing industrial Internetof Things intelligent equipment.

The invention discloses a multi-unmanned aerial vehicle task unloading and resource allocation method for an edge computing system. The method comprises the following steps: combining the current position of an unmanned aerial vehicle, the relative distance of the unmanned aerial vehicle, the relative distance between the unmanned aerial vehicle and intelligent equipment of the Internet of Things, and the service times of the intelligent equipment of the Internet of Things into a system state; constructing a depth deterministic strategy gradient optimization neural network; inputting the system state into the depth deterministic strategy gradient optimization neural network to obtain a system action; enabling the edge computing system to execute the system action and obtain the reward value of the system action according to the instant reward function; and continuously training parameters of the depth deterministic strategy gradient optimization neural network according to the obtained reward value until the reward value tends to be stable, and training to obtain an optimal strategy [pi]. According to the method, the trajectory, the unloading strategy and the computing resource allocation strategy of the unmanned aerial vehicle are optimized through the depth deterministic strategy gradient optimization neural network, and the energy consumption of the system is minimized on the premise of ensuring the service fairness of the intelligent equipment of the Internet of Things.

The invention relates to a non-rinse pregnant woman perineum medical care gel and a preparation method thereof. The gel comprises the following components in percentage by weight: 0.1 to 5 percent of water-soluble chitosan, 0.3 to 1 percent of A.SAP natural organic fungicide, 0.1 to 0.5 percent of aloe extract, 0.3 to 1.5 percent of water-soluble polymer gel material, 0.1 to 1 percent of triethanolamine, 0.05 to 0.5 percent of chelating agent, 0.05 to 0.5 percent of emollient and the balance of distilled water, and is prepared by dissolving the components in the distilled water and uniformly mixing the solution. The chitosan non-rinse pregnant woman perineum medical care gel has the effects of cleaning perineum, removing dirt, promoting blood circulation and enhancing immunity of the perineum part and is antibacterial and antiviral; most importantly, the medical care gel has high using safety performance and does not have any adverse effect on a fetus in a pregnant woman. The medical care gel is easy and convenient to produce, is environmentally-friendly during production and is suitable for mass industrial production.

One virtual rod electrode is composed by a plurality of electrode plain plates arranged in the ion optical axis direction, and four virtual rod electrodes are arranged around the ion optical axis to form a virtual quadrupole rod type ion transport optical system (30). In one virtual rod electrode, the interval between the adjacent electrode plain plates is set to be large in the anterior area (30A) and small in the posterior area (30B). As the interval between electrodes becomes larger, high-order multipole field components increase and therefore the ion acceptance is increased, which enables an efficient acceptance of ions coming from the previous stage. On the other hand, if the interval between electrodes is small, the quadrupole field components relatively increase and the ion beam's convergence is improved. Therefore, ions can be effectively introduced into a quadrupole mass filter for example in the subsequent stage, which contributes to the enhancement of the mass analysis' sensitivity and accuracy.

The invention discloses a method for compressing a computed hologram by adopting a quantum neural network with optimized initial weight, and belongs to the technical field of the compressed transmission of the computed hologram. On the basis of transmitting the computed hologram in the compressed way by the quantum BP neural network, the method comprises the following steps: pre-training by utilizing a computed hologram training set to acquire quantum BP neural network optimized initial weight; accelerating the convergence process of the pre-training network by setting the pre-trained parameter random initialization variance, and performing secondary network fine-tuning training on the optimized initial weight acquired through pre-training for the given holographic compressed data, and dynamically adjusting the network learning rate in the optimization process so as to accelerate the compressed transmission process of the quantum BP neural network. The training on the compressed transmission network structure can be accomplished by using less iteration times without changing the basic structure of the original quantum BP neural network, the compression speed on the computed hologram by the quantum BP network is accelerated, and the quality of the reproductive image of the hologram can be guaranteed.

The invention discloses a current loop delay compensation method for a three-phase permanent magnet synchronous motor driving system. The method comprises steps that a d-q axis stator current of a next period based on a motor model is predicted, and the d-q axis stator current is taken as a PI regulator feedback value for calculation to obtain a stator voltage reference value; aiming at the influence of motor parameter change on prediction precision, the prediction model is further corrected, when actual parameters are inconsistent with motor parameter nominal values, an error voltage item inthe model plays a role in correcting a current prediction value. The method is advantaged in that hardware configuration does not need to be changed or extra logic devices do not need to be added, themethod is simple in operation, low in cost, simple to implement and easy to expand, weakens the limitation of digital delay on the adjustment speed of the current loop, accelerates the current tracking characteristic, still ensures higher control precision even under the condition of larger motor parameter change, and is suitable for permanent magnet synchronous motor driving occasions with higher requirements on dynamic tracking and control precision.

The invention belongs to the technical field of shield tunneling machines, and discloses a shield tunneling machine working posture real-time prediction method based on big data, which comprises the following steps: obtaining overall data: obtaining engineering, equipment and stratum parameters; analyzing and preprocessing the shield data; and establishing a shield attitude prediction model. In order to convert original data into a data format for deep learning, based on a data trend of shield feature engineering, a data standardization preprocessing process is divided into four parts: data segmentation, discrete point processing, missing value processing and data noise reduction. According to the prediction method, the shield data is subjected to standardized preprocessing, the workload of computer processing can be effectively reduced, the working efficiency is improved, and compared with a traditional artificial intelligence prediction model, a Butterworth noise reduction method is provided in a targeted mode, and it is guaranteed that the working posture of the shield tunneling machine is accurately predicted in real time.