295results about How to "Easy to converge" patented technology

The invention discloses a deep learning-based image high-density population counting method. The method comprises the following steps of S1, establishing a depth complementation convolutional neural network by utilizing a deep learning framework caffe; S2, performing image data enhancement on an image according to operations of angle rotation, image multi-scale zooming, image mirroring and image pyramid zooming; S3, performing Gaussian kernel fussy normalization processing on the enhanced image data to obtain a real crowd density graph, outputting an estimated density graph and the real density graph by the network, and performing continuous iterative training optimization on the whole network structure according to a loss function; and S4, inputting a crowd picture and a tag picture to the network for training, and performing continuous iterative optimization to obtain a trained network model finally. According to the method, the end-to-end convolutional neural network is designed; a picture is given and input, and the estimated density graph corresponding to the picture is output, so that an estimated crowd number is obtained; and by outputting the density graph, more useful information is reserved.

An endoscope system has an insertion portion and a control portion connected to a base end portion of the insertion portion. An illumination system includes a light guide having a front end positioned near the front end of the insertion portion and a rear end positioned rearward of the base end of the insertion portion, phosphors disposed in the light guide toward the front end thereof and an illumination light source which is positioned toward the rear end of the light guide and emits stimulating light exciting the phosphors.

The invention relates to a satellite remote sensing image large-area seamless orthographic image manufacturing method. The method comprises the following steps: 1) constructing a remote sensing imageimaging geometric model; 2) reading a to-be-matched remote sensing image and DEM data, and matching the DEM data with the to-be-registered remote sensing image data; 3) carrying out remote sensing image ortho-rectification to obtain a corresponding DOM product; 4) according to the position relation between the orthographic images obtained in the step 3), determining an overlapping region between the images, performing matching in the overlapping region to obtain homonymous points, establishing an error equation based on a TPS model for each group of homonymous points, and performing iterativesolution to obtain a correction amount of geometric deviation between the images; And 5) carrying out dodging and inlaying on all the corrected ortho-images to obtain large-area ortho-image map products, and introducing a thin plate spline model in the ortho-image adjustment process to solve the problem of low precision of a traditional geometric correction model caused by a small overlapping range between the images.

The invention relates to a pavement crack image detection method. The method comprises the steps of: carrying out graying and filtering processing on a collected pavement image, constructing a pulse coupling neural network PCNN model, utilizing a genetic algorithm to rapidly find advantages of an optimal solution in a non-linear manner in a solution space so as to optimize important parameters of the model, and rapidly and accurately segmenting cracks and a background in the image; then according to the characteristics of the image after the segmentation, carrying out connected domain detection on the whole image, and filtering out the interference of noise and background textures; and finally, extracting a crack skeleton, calculating the maximum widths of the cracks along the normal line of the skeleton, and making marks in the original image. According to the invention, the digital image processing technology is adopted, the genetic algorithm is utilized to optimize the parameters of the PCNN model, optimization searching is accelerated, the iteration times f the PCNN are reduced, and the iteration is more liable to come to convergence, the interference resistance of the segmentation effect is relatively high, and the segmentation is more accurate; in addition, the modes of connected domain rectangularity, circularity filtering and irregular noise filtering are utilized to filter out a large number of irregular patches, and convenience is provided for the crack detection.

The invention discloses a fiber optic gyroscope temperature drift modeling method by optimizing a dynamic recurrent neural network through a genetic algorithm. The fiber optic gyroscope temperature drift modeling method by optimizing the dynamic recurrent neural network through the genetic algorithm comprises the following steps of (1) initializing network parameters, and establishing an improved Elman neural network model; (2) obtaining a training and testing sample; (3) training an improved Elman neural network, and optimizing model parameters through the genetic algorithm; (4) outputting forecasts of an fiber optic gyroscope, and compensating errors. The output of the fiber optic gyroscope processed through a denoising algorithm is trained by introducing the improved Elman neural model with self-feedback connection weight, constant iterative optimization is carried out on the model parameters through the genetic algorithm, and the optimal model is obtained according to the magnitude of the errors of the model under different parameters. According to the fiber optic gyroscope temperature drift modeling method by optimizing the dynamic recurrent neural network through the genetic algorithm, the complexity of the algorithm is taken into consideration, the accuracy of the fiber optic gyroscope temperature drift model is improved, the application of the fiber optic gyroscope temperature drift model in engineering is expanded, and certain practical significance is achieved.

The invention discloses a face detection method and device adopting a cascade classifier, relating to the field of video monitoring. The face detection method comprises the following step of: carrying out face detection on rectangular windows with different positions and sizes in a current image to be detected by adopting a light type cascade-structure classifier to detect face windows, wherein each level of the light type cascade-structure classifier is a weak classifier. For a large quantity of windows to be detected in one image, relative to a scheme that a cascade-structure classifier is bound to eliminate partial non-faces after computing the results of all weak classifiers of a strong classifier, each weak classifier of the light type cascade-structure classifier can eliminate the partial non-faces, and therefore faster detection speed can be achieved, and the real-time detection of faces is further realized on megapixel high-definition videos.

A clothing simulation apparatus precisely determines clothing pressure for bringing a clothing into tight contact with a human body. A fitting part 22 divides a paper pattern model into a plurality of elements, imparts dynamic characteristics shown by a fabric model to each element, deforms the paper pattern model by solving the motion equation of each element using a finite element method, and then fits the clothing virtually to a human body model. The fitting part 22 sets a temporary model to cover a predetermined section of the human body model, deforms the paper pattern model to bring the paper pattern model into contact with the temporary model, and thereafter deforms the paper pattern model to bring the paper pattern model into contact with the human body model.

The invention relates to the technical field of image processing, more specifically, to a depth convolution neural network image denoising method based on an Inception model. Building a training network with one-layer Inception model; setting the parameters of network solution and choosing the appropriate supervisory framework; the parameters of the network structure are optimized, and the obtained network is used for denoising. The depth convolution neural network image denoising method based on the Inception model of the invention is based on the depth convolution neural network and combinesGaussian white noise to simulate unknown real noise and the network containing the Inception model to reduce the parameter quantity on the basis of maintaining the original denoising effect; throughthe way of data expansion, the content of training data is enriched, and the neural network can learn the inner structure of image distribution or noise. By introducing the residual structure, the network converges more easily and the effect is the best.

The invention provides a calculation method of the bearing capacity of a reinforced wall plate and belongs to the field of strength analysis of aircraft structures. According to the technical scheme, the calculation method is an axial compression loading action based calculation method of the effective width of buckling of a skin. Stiffness reduction is performed through modification of the cross section area and the moment of inertia of a stringer in a finite element model which is formed by plate beam units after the buckling of the skin is produced, stress analysis of a structure is performed by a linear iteration solution method of load loading steps, comparison is performed on the working stress of the stringer in a calculation result and the allowable limiting stress of a stringer and skin combined structure, the allowable limiting stress serves as a damage standard and is solved based on an engineering method, and accordingly the bearing capacity of the structure is obtained. Compared with the traditional full housing model, a calculation method model is a plate beam unit model, the geometric nonlinearity and contact problem solution which is large in time consumption and difficult to converge is replaced by the linear solution, and accordingly the modeling time consumption is small, the calculation speed is high, and the requirements for the professional quality of the computer hardware and analysis staff are low.

The invention discloses a high-resolution image generation method based on a generative adversarial network. The method comprises the following steps: firstly, preprocessing a to-be-learned data set image to obtain a training set; constructing a generative adversarial network comprising a generative network and a discrimination network; pre-training the generative adversarial network; obtaining pre-trained model parameters as initialization parameters of the generative adversarial network; then, separately inputting the training set and an image generated by the generative network into the discrimination network, enabling the output of the discrimination network to react on the generative network, carrying out adversarial training on the generative adversarial network, optimizing network parameters of the generative network and the discrimination network, and ending the training when a loss function converges to obtain a trained generative adversarial network; and finally, inputting the random data distribution into the trained generation network to realize high-resolution image generation. According to the invention, the generated image is clearer, the training process is stable,and the network converges quickly.

The invention provides an open set domain adaptation method based on entropy minimization. The open set domain adaptation method comprises the steps of creating an open set domain adaptation network, and initializing network hyper-parameters; inputting a source domain picture and a target domain picture into an open set domain adaptive network, obtaining picture depth features through a networkfeature extraction layer, and then obtaining prediction probability vectors of various categories through a softmax layer according to a feature map; calculating an entropy loss function by using a known category prediction probability vector and a picture label so as to correctly identify a common category; on the basis, calculating binary cross entropy and a binary diversity loss function by using an unknown category prediction probability value so as to correctly classify unknown categories, updating network parameters by using a back propagation gradient until a loss function value is minimum, and stopping training the network; and storing the network model and the training result, and introducing the target domain data set into the network model to obtain a final target domain label.

The invention relates to a wind farm multi-model draught fan optimized arrangement method based on a genetic algorithm. The method includes the following steps that (1) a wind farm region is divided into square meshes which are the same in size according to the diameter of a draught fan, and an integer matrix which is the same in line and row is generated randomly to be used as the initial solution of the algorithm; (2) the individual fitness value of a current generation is calculated; (3) parent individuals participating in crossover are selected through even random selection operators, and then filial generation individuals are generated by the adoption of improved crossover and mutation operators; (4) repairing operators are introduced to the individuals in a population; (5) a Tabu operator is introduced to an optimal solution of the current generation of the population, the optimal solution is used as the initial solution of a Tabu algorithm, and the neighborhood solution of the optimal solution is searched for; (6) whether the biggest number of iterations is reached or not is judged, if yes, the multi-model draught fan optimized arrangement is completed, and if not, the step (2) is executed again. Compared with the prior art, the wind farm multi-model draught fan optimized arrangement method based on the genetic algorithm has the advantages of being visual in coding mode, good in performance index, high in local search capacity, high in expansibility, high in practicability and the like.

The invention discloses an agent learning method based on knowledge guidance-tactical perception. In the environment of a two-party fight game, human knowledge is used to train an agent which fights with a specific tactical strategy through two stages of training. The method comprises the steps that 1) a game screen and the state of the agent are encoded to form a state feature; 2) a script with aspecific fight strategy is constructed by manually writing a script; 3) the agent fights with the script to collect the state and motion data of an opponent to carry out training in the first stage;4) the trained agent fights with a computer built-in AI, and training in the second stage is carried out through reinforcement learning; and 5) the learning framework is used to train the agent to participate in fight with specific tactics. The method provided by the invention is applicable to agent training in a two-party fight mode in a micro-operation environment, and acquires a good winning rate in the face of various micro-operation fight scenes.

The invention discloses a DDPG-based optimization method with distributed energy participating in distribution network voltage regulation. A deterministic depth deterministic strategy gradient algorithm is applied to distributed energy participating in distribution network voltage regulation. As the adopted DDPG algorithm is a model-friendly algorithm, the original distributed energy participatingin distribution network voltage regulation optimization strategy applies a Markov decision process method to be converted to a strategic decision problem, and the versatility of the algorithm is improved. The sample training stability is improved through applying a target network, easier convergence is achieved, and the method has the advantages of higher feasibility and lower operating cost.

The invention provides a method for precisely solving the length of an unstressed cable of a cable-stayed bridge. The length of the unstressed cable is solved by substituting the average value of the nodal force at the end of the stayed cable with a precise calculation method for the force at the end of the stayed cable, and by building a precise expression s0 of the length of the unstressed stayed cable and a characteristic parameter beta constraint equation of the stayed cable when a tensile force at the end of a tower is known without selecting the primary value of the length of the unstressed cable. The method comprises the following steps of: building the precise expression of the length of the unstressed stayed cable and the characteristic parameter beta constraint equation of the stayed cable when the tensile force Tj at the end of the tower is known, and solving the length of the unstressed stayed cable. The method has the advantages of solving the problem on the selection of the primary value of the length of the unstressed cable, with high convergence rate and small calculation error. Compared with the linear search, the dichotomy and the improved chord secant method for iteratively solving the length of the unstressed cable, the method of the invention greatly improves the solution precision and solution efficiency of the length of the unstressed stayed cable, so the method is of relatively high application value in practical projects.

The invention discloses a training method and device of a voiceprint recognition model, electronic equipment and a storage medium. The method incudes: acquiring a training set which comprises user audio data; inputting the acoustic feature vector corresponding to each audio data into a voiceprint recognition model to obtain the acoustic feature vector corresponding to each audio data in the training set and predicted probability distribution; calculating cross entropy loss according to the predicted probability distribution corresponding to each audio data in the training set and the real probability distribution corresponding to each audio data; calculating central loss according to the voiceprint feature vector corresponding to each audio data in the training set and the reference voiceprint feature vector corresponding to user identification; updating the model parameters of the voiceprint recognition model according to the cross entropy loss and the central loss. By the training method, the voiceprint feature distance of the same user can be lowered while the voiceprint feature distance of different users can be increased.

The invention discloses a structural finite element model modifying method based on displacement time history area under vehicle load. According to the method, objective functions based on the displacement response time history area are in one-to-one correspondence with local rigidity of a structure, and accordingly, the local rigidity and boundary conditions can be selected as modifying parameters, the objective functions and modifying variables are definite in meaning, improvement in computing efficiency is benefited and convergence is easy. The method does not require bridge traffic interruption and extraction of dynamic characteristics of the structure (namely, time-frequency transformation), thereby being capable of avoiding shortcomings of existing methods.

The embodiment of the invention discloses an advertisement delivery method and an advertisement delivery device. The method comprises steps: advertisement click rate characteristic parameters are acquired from a group user history search record; the advertisement click rate characteristic parameters are classified to obtain multiple feature sub groups; factorization is carried out on the multiple feature sub groups to acquire a sub group factorization model; the sub group factorization model is subjected to iteration to acquire a predicted advertisement click rate; and according to the predicted advertisement click rate, advertisement delivery is carried out. The embodiment of the invention reduces the training calculation amount and improves the predicted click rate. A service object is displayed according to the predicted click rate, the possibility of displaying the service object meeting user requirements is improved, multiple times of page turning operation or search in other platforms by the user is avoided, the operation convenience is improved, time consumption of the user is reduced, the user experience is improved, and the use satisfaction of the user is enhanced. Resource occupancy and resource waste by a server and a client are reduced.

The invention discloses a cognitive wireless sensor network spectrum access method based on deep Q learning. The cognitive wireless sensor network spectrum access method comprises the following steps:step 1, constructing a Q neural network: selecting a training sample to update a weight parameter of the Q neural network by taking the state values of all channels of a t-2 time slot as an input layer and taking the q values of all channels of a t-1 time slot as an output layer; obtaining an experience sample before the t-1 time slot, calculating the priority, summing the arrangement of the binary tree according to the priority accumulation, and performing sampling to obtain a training sample; step 2, taking the state values of all channels of the t-1 time slot as an input layer, inputting the state values into a Q neural network to obtain q values of all channels of the t time slot, and selecting a channel corresponding to the maximum q value; and step 3, sensing the channel energy, accessing if the state value is idle, and not accessing if the state value is busy. The method has the beneficial effects that the energy consumption is low, the convergence speed is high, the diversityloss of experience samples is avoided, the overfitting phenomenon is avoided, and the prediction accuracy is high.

The invention discloses a scale estimation-based human face detection method. According to the method disclosed in the invention, a size of a human face on an image is estimated via human face scale estimation, the image is zoomed based on a human face scale, a suggestion frame extraction is performed rapidly via a full convolution network, and a human face detection result can be obtained via twotimes of cascaded classification and regression. According to the method, human face detection is performed based on combination of human face scale estimation and a cascaded convolution neural network, an overall calculation amount of human face detection can be reduced, overall time consumption for human face detection can be reduced, and human face detection effects can be ensured.

The invention discloses a method and apparatus for eliminating multichannel acoustic echoes. The method comprises the following steps: obtaining I microphone signals and J loudspeaker signals; when the I microphone signals only comprise echo signals, obtaining time delays between echo signals received by a microphone and the J loudspeaker signals; according to the time delays, performing delay compensation on the loudspeaker signals so as to realize alignment with the echo signals in the microphone, and obtaining loudspeaker signals after I*J delay compensation; performing sub-band analysis on the microphone signals to obtain I*N first sub-band signals; performing sub-band analysis on the loudspeaker signals after the delay compensation to obtain I*J*N second sub-band signals; performing acoustic echo elimination on the I*N first sub-band signals in each sub-band by use of the I*J*N second sub-band signals to obtain I*N third sub-band signals after the acoustic echo elimination; and performing sub-band integration on the I*N third sub-band signals to obtain I sound signals after the echoes are eliminated, and such that the effect of eliminating the multichannel acoustic echoes is improved.

The invention provides a method for calculating micro dendritic growth in continuous casting solidification on basis of a cellular automaton, and relates to the technical field of micro dendritic growth process simulation. According to the method, improvement is carried out on the basis of an eight-cell capture model in the cellular automaton; aiming at problems in the eight-cell capture model, aright adjacent cell of a captured adjacent unit cell is set as an interface cell to be calculated, so as to solve solid phase rate increase and residual liquid phase concentration for the captured adjacent unit cell; when no problems exist, calculation is directly carried out through a dendritic growth model; and through judging whether the solid phase rate of the captured adjacent unit cell is greater than or equal to 0 or not, users can determine whether to convert the captured adjacent unit cell from the interface cell to a solid phase cell or not. According to the method, the capture manner of the eight-cell capture model is improved, so that the calculation process is easy to converge, the calculation efficiency is higher, the success rate is higher, and isometric crystals growing atpreferential directions are basically identical with set preferential angles.

The invention discloses a road image semantic segmentation method based on a hybrid automatic encoder. The road image semantic segmentation method comprises the following steps of (10) sample set collection of dividing sample set images into a training set and a test set; (20) sample image preprocessing of performing size transformation and contrast normalization processing on the sample set images, and converting the sample set images to be processed into a standard form; (30) acquiring a hybrid automatic encoder network model, respectively training a sparse automatic encoder and a de-noisingautomatic encoder by utilizing the preprocessed training sample, extracting an intermediate encoding weight value and an intermediate decoding weight value, and constructing the hybrid automatic encoder network model by establishing a reasonable model arrangement sequence and a stacking form; and (40) performing road semantic segmentation, performing semantic segmentation on a road image shot bythe vehicle-mounted camera by using the hybrid automatic encoder network model. According to the invention, through a stacking form of the hybrid automatic encoder, the optimization description of image semantics is realized, a concise and effective semantic segmentation model is established, and better road semantic segmentation performance is obtained.

The invention provides a method and a system for constructing a simulation model for electromagnetic interference prediction. The method comprises the following steps: acquiring a motor static common-mode port network scattering parameter characteristic curve based on the bandwidth of a test system, and converting the motor static common-mode port network scattering parameter characteristic curveinto a port impedance parameter characteristic curve; performing rational fitting on a function relationship corresponding to the port impedance parameter characteristic curve based on a vector fitting method to obtain a vector fitting function; the vector fitting function being equivalent to a plurality of sub-circuits according to a set rule, and connecting the sub-circuits to obtain a circuit which is a simulation model for electromagnetic interference prediction; wherein the common mode port network scattering parameters comprise amplitude and phase angle. The simulation modeling method provided by the invention has the advantages of high modeling speed, high precision and easiness in convergence. The simulation modeling method provided by the invention has the advantages of high modeling speed, high precision and easiness in convergence.

The invention discloses a hyperspectral image denoising method based on non-convex low-rank matrix approximation and total variation regularization. The hyperspectral image denoising method comprisesthe following steps: 1) acquiring hyperspectral image data to be denoised, wherein M, N and K represent the number of width and height spectral bands of a hyperspectral image respectively; 2) constructing a denoising model based on non-convex low-rank matrix decomposition and TV regularization; (3) optimizing the model by adopting an Augmented Lagrange function (ALM) algorithm, and optimizing themodel by adopting an ALM algorithm; 4) optimizing and solving the model by adopting an augmented Lagrange function algorithm; and 5) outputting the hyperspectral image after the mixed noise is removed.