272 results about "Model complexity" patented technology

Methods and apparatus for camera calibration based on multiview image geometry. A lens profiling module may estimate two or more mathematical models for correcting aberrations in images in a single pass from a set of calibration images captured with a camera / lens combination. For example, the module may estimate the lens aberrations of geometric distortion, lateral chromatic aberration, and vignette models in a single pass. The module may determine point correspondences, 3D transformations, and camera intrinsics for views of calibration charts captured in the images. The module estimates the mathematical models for the two or more types of aberrations from the information determined from the views of the calibration charts. The module may automatically determine an optimal model complexity when estimating the mathematical models. The estimated models may be written or appended to a lens profile for the camera / lens combination used to captured the calibration images.

The invention relates to the technical field of computer application engineering, in particular to a method for using an improved neural network model based on particle swarm optimization for data prediction. The method includes the steps of firstly, expressing data samples; secondly, pre-processing data; thirdly, initiating the parameters of an RBF neural network; fourthly, using the binary particle swarm optimization to determine the number of neurons of a hidden layer and the center of the radial basis function of the hidden layer; fifthly, initiating the parameters of the local particle swarm optimization. By the method for using the improved neural network model based on particle swarm optimization for data prediction, the number of the neurons of the hidden layer of the RBF neural network model can be determined easily, RBF neural network performance is improved, and data prediction accuracy is increased. In addition, the improved neural network model based on particle swarm optimization is low in model complexity, high in robustness and good in expandability.

The invention provides a pattern generation method for a three-dimensional model, which comprises the following steps: original precise three-dimensional model data is obtained, and the three-dimensional model data comprises polygon mesh vertex data which indicates a three-dimensional object; the three-dimensional model data is compressed, and redundant data of the three-dimensional model is reduced; a user request is received, the compressed three-dimensional model data is downloaded, and a three-dimensional model pattern is generated through rendering. Through the pattern generation method, the model complexity is reduced, the number of the polygons which need to be processed in an image system is reduced, the data size which needs to be transmitted is reduced, more particularly, the data size which is needed for the pattern generation in a rendering manner is reduced, the rendering speed is increased, and the real-time rendering request of the user can be satisfied. A pattern generation device for the three-dimensional model is also provided.

Aspects of the present invention include object detection training systems and methods and using object detection systems and methods that have been trained. Embodiments presented herein include hybrid learning approaches that combine global classification and local adaptations, which automatically adjust model complexity according to data distribution. Embodiments of the present invention automatically determine model complexity of the local learning algorithm according to the distribution of ambiguous samples. And, embodiments of the local adaptation from global classifier avoid the common under-training problem for local classifier.

The invention relates to a vehicle yaw stability predicting model control method and belongs to the technical field of a vehicle control. The disturbance-observer-based vehicle yaw stability predicting model control method uses a model predicting control method to design a linear model predicting controller, considers constraint conditions and can reduce computation time, track an expected value as much as possible and keep vehicle stability. The method includes: designing a reference model; performing linearization on a vehicle two-freedom-degree nonlinear model to obtain a linear model withmodel error disturbance terms; designing error terms in a disturbance observer pair model according to the linear model; using a model predicting control algorithm to build a target function, and solving an optimal problem corresponding to a cost function to obtain control input acting to a system to allow the vehicle system to track the expected value as much as possible so as to guaranteed vehicle stability. By the method, model complexity can be lowered effectively, control precision requirements are satisfied, and the constraint of driving torque is considered.

The invention discloses a real-time multi-scale target detection method based on a lightweight convolutional neural network, and the method comprises the following steps: employing a clustering algorithm to cluster the depth-to-width ratio and the area scale coefficient of a training data set sample target, wherein the clustering center is used for optimizing the setting of an anchor frame; takinga lightweight convolutional neural network as a backbone network of a RetinaNet framework, and reducing the model complexity; enhancing the multi-scale feature map by using a residual error module, performing adaptive cross-layer feature fusion on the partially enhanced multi-scale feature map, and finally, using the generated fusion feature pyramid to replace the feature pyramid, so that the detection precision of the model for small and medium-sized targets is improved. The lightweight multi-scale target detection method disclosed by the invention can reduce the parameter quantity of the model, reduce the operation complexity of the model and improve the detection efficiency of the model under the condition of ensuring a certain accuracy rate.

Methods and apparatus are described for modeling sequences of events with Markov models whose states correspond to nodes in a provided taxonomy. Each state represents the events in the subtree under the corresponding node. By lumping observed events into states that correspond to internal nodes in the taxonomy, more compact models are achieved that are easier to understand and visualize, at the expense of a decrease in the data likelihood. The decision for selecting the best model is taken on the basis of two competing goals: maximizing the data likelihood, while minimizing the model complexity (i.e., the number of states).

The invention discloses a feasible road segmentation method based on an improved Unet network model, and the method comprises the steps: modifying a convolutional layer in a Unet training model into aResNet residual module, and carrying out the back propagation training through employing a cross entropy loss function and a Lovsz hinge loss function; Performing feasible region segmentation on theroad scene without the segmentation label by adopting the trained model; Carrying out scale, contrast and normalization processing on the training picture and the corresponding segmentation mask mark,carrying out enhancement processing on the training sample, carrying out histogram matching processing on the test data by taking an average histogram of the training data as a template, carrying outtesting, and outputting a road segmentation result. According to the method, on the basis of not increasing the model complexity, the road segmentation precision is greatly improved, and real-time accurate segmentation of the feasible region of the road can be basically realized.

A learning agent is disclosed that receives data in sequence from one or more sequential data sources; generates a model modelling sequences of data and actions; and selects an action maximizing the expected future value of a reward function, wherein the reward function depends at least partly on at least one of: a measure of the change in complexity of the model, or a measure of the complexity of the change in the model. The measure of the change in complexity of the model may be based on, for example, the change in description length of the first part of a two-part code describing one or more sequences of received data and actions, the change in description length of a statistical distribution modelling, the description length of the change in the first part of the two-part code, or the description length of the change in the statistical distribution modelling.

The invention relates to a cooperative agent learning method based on multi-agent reinforcement learning. The cooperative agent learning method comprises the steps of 1, resetting a plurality of target environments; 2, initializing a model parameter theta pi of the strategy network pi theta and a model parameter theta f of the global information prediction network f theta; 3, sampling the multipleagents in the multiple environments according to the current strategy pi in the environment, wherein in each step, a plurality of agents in the environment share the same state, and characteristics are extracted from the state by aiming at each agent and then are used as data input by the model; 4, updating the model parameters theta pi and theta f; and step 5, updating until the model convergesor reaches the maximum step number. According to the method, the global feature information is better utilized in the environment that the agents are in the cooperative relationship, and each agent learns to perceive the relationship between the local information and the global information through the model for predicting the global information through the local information, so that the agents canbetter cooperate; therefore, different agents can directly share the model parameters, the model complexity is simplified, and the efficiency is improved.

The invention relates to a 3D MIMO channel modeling method. The method includes the following steps that: (1) simulated scenes and network layout are determined, large-scale parameters are calculated according to the scenes and calculated correlations; (2) small-scale parameters are generated sequentially based on the large-scale parameters, a probability density function and the scenes; (3) a channel coefficient is calculated; (4) the small-scale parameters are updated according to a calculation result, and a drift model is built; and (5) time evolution is carried out according to the drift model, and then, modeling is carried out. According to the 3D MIMO channel modeling method of the invention, short-term time evolution of the channel coefficient is realized through updating time delay, a departure angle, an arrival angle, polarization, shadow fading and a K factor; smooth transition between adjacent channel segments is supported; a visual-range scene and a non-visual-range scene are simulated jointly by a common framework structure, and therefore, the complexity of a model can be reduced, and multi-unit scenes can be configured freely; and an algorithm of position graph generation is expanded, and diagonal angle movement directions are considered, and smoother output is created.

A new word model is trained from synthetic word samples derived by Monte Carlo techniques from one or more prior word models. The prior word model can be a phonetic word model and the new word model can be a non-phonetic, whole-word, word model. The prior word model can be trained from data that has undergone a first channel normalization and the synthesized word samples from which the new word model is trained can undergo a different channel normalization similar to that to be used in a given speech recognition context. The prior word model can have a first model structure and the new word model can have a second, different, model structure. These differences in model structure can include, for example, differences of model topology; differences of model complexity; and differences in the type of basis function used in a description of such probability distributions.

The invention provides a convolutional network-based lightweight face key point detection method and system, and a storage medium, and the method comprises the steps: employing a multi-task network tocomplete the face detection and face alignment parameter calculation in parallel, and carrying out the alignment of an original inclined face; sending the return face into a light weight key point detection network; detecting face key points, for a multi-face key point detection task,using a pre-training scheme of non-frozen transfer learning, training multiple face key points step by step, and using a parallel face aligning mechanism during training; face rotation return original angle. The method has the beneficial effects that the method is improved according to the characteristics of a face key point detection task, an attention mechanism is introduced to score and select the network output of the convolutional network, and the problem of loss function imbalance of face key point detection is relieved; and a face detection task and a face return parameter calculation task are synchronously trained, so that the efficiency of the overall architecture is improved, and the model complexity is reduced.

The invention discloses a radar high resolution range profile (HRRP) target recognition method based on a convolution factor analysis (CFA) model. The radar HRRP target recognition method mainly solves the problem of poor target recognition performance under the condition of small samples in the prior art, and is implemented by the steps of: (1) carrying out framing on HRRPs of various kinds of targets according to angular domains, and carrying out modulus operation on each frame of data to obtain time domain features; (2) carrying out per-processing on each frame of data; (3) constructing a CFA model for each frame of HRRP after preprocessing, and calculating condition posterior distribution of each model parameter; (4) initializing each parameter and performing I-th iterative updating; (5) carrying out intensity normalization on a test sample, and translating and aligning frames of average profiles; (6) calculating frame probability density function values of the test sample according to a posterior mean of parameters of the CFA model; (7) and finding out the maximum probability density function value, and determining a type of the test samples. The radar HRRP target recognition method has the advantages of being low in model complexity, and being capable of applied to radar target recognition under the condition of small samples.

The values X(n) input to a current-steering digital-to-analog converter (49) are modified (41) before the actual conversion to compensate for conversion errors of the digital-to-analog converter. The input values are modified according to a model (43) of the digital-to-analog converter in which each output value of the digital-to-analog converter Y(n) is a sum of a desired value directly proportional to the respective input value and an error. The error is a product of the settled output value, i.e. the difference between the desired value and the previous output value Y(n−1) actually provided by the digital-to-analog converter, and a relative step error that is a function only of the respective input signal and is stored in a table. The relative step error can be a function also of the previous output signal and of the previous input signal. This model has a low complexity and is suitable for on-chip implementation.

The invention discloses a backward swimming control method of a biomimetic carangiform robot fish, based on a CPG (Central Pattern Generator) to control the robot fish to swim. The backward swimming control method comprises the following steps: firstly, building a CPG chain type network topology according to the characteristics of fish swimming to reduce a CPG parameter number and reduce the complexity of a model; secondly, by reasonably regulating parameters of the CPG model, enabling an output signal of the CPG model to satisfy the following condition, in a direction from the head to the tail of the robot fish, phases of the control signal of a steering engine of the robot fish sequentially lag, and amplitudes are sequentially reduced, so that backward propelling force is produced, and backward swimming of the robot fish is realized. The backward swimming control method has the promotion action in deeply understanding a backward swimming motion mechanism of fishes and also provides guidance for enriching motion modals of underwater robots and improving the maneuvering ability of the underwater robots.

The invention discloses a dynamic equivalent modeling method for the cluster of distributed photovoltaic power stations and belongs to the technical field of distributed energy grid-connected modelingand simulation. According to the method, an electric quantity waveform capable of accurately describing the dynamic characteristics of photovoltaic power stations can be selected as a clustering index, and a dynamic time bending distance among the similar clustering indexes of all photovoltaic power stations serves as the measurement of clustering. After that, according to a clustering result, the aggregation equivalent parameters of photovoltaic power stations of the same cluster are calculated, and the equivalent simplification of the power distribution network is carried out. In this way,not only the dynamic similarity among photovoltaic power stations can be accurately captured, but also the model complexity can be greatly reduced. Meanwhile, the simulation calculation time can be greatly shortened.

The invention discloses a method for designing a thin cloth circular antenna array through changing auxiliary grid circle radius. The method comprises the following steps that (1) distance between array elements on the same grid circular ring is constrained, range of the circular ring radius is confirmed according to the formula and S units are randomly generated so that an initial group is formed; and (2) the circular ring radius is optimized by utilizing a genetic algorithm, the optimal unit is iteratively outputted through intersection, mutation and selection, and peak side-lobe level of the optimal unit is calculated. Defects of graded optimization of the track circle radius and the number of the array elements on the track circle in existing circular array integrated methods are improved so that combined optimization of all the array elements is realized, lower peak side-lobe level is acquired, and optimization computation amount and model complexity are reduced.

The invention provides a rapid depth map sequence interframe mode selection fractal coding method. The method includes: firstly, utilizing a fractal video compression method to code a color video, then decoding the color video by a fractal video decompression method to obtain macro block coding modes of the color video, and designing a complexity calculation formula of the depth map sequence frame macro block coding modes by utilizing correlation of depth map sequence frame macro blocks and the corresponding color video macro block coding. A frame coding I of a depth map sequence adopts an H.264 intra-frame prediction coding method; a frame coding P of the depth map sequence adapt an Roberts operator to detect object boundaries, and if the object boundaries are contained in the current depth map sequence frame macro blocks, a traditional full-search coding method is used, all the modes are traversed, and the best coding modes are obtained according to a Lagrangian rate-distortion optimization model; otherwise, coding model complexity of the current depth map sequence frame macro blocks is calculated, the best coding modes are selected in different effective modes to aim at two situations when the complexity is larger than a threshold value T and smaller than or equal to the threshold value T, and then the depth map sequence frame macro blocks are decoded by a fractal method.

The invention discloses a video target detection method based on multi-layer feature fusion, which solves the problems that the existing detection method does not utilize video time sequence information and is poor in detection effect, and adopts the technical scheme of inputting a frame of video image as a current frame, selecting a front frame of image from the front 9 frames, and selecting a rear frame of image from the rear 9 frames; inputting the three frames of images into an improved convolutional neural network to obtain three feature maps respectively; inputting into a sampling network to obtain sampling images of the front and back frame feature images, and calculating sampling coefficients of the front and back frame feature images according to the sampling images; and obtainingan enhanced feature map of the current frame by using the sampling coefficient according to a fusion formula, taking the enhanced feature map as the input of the detection network, generating a candidate region set, and detecting the final target category and position through the classification and regression network. According to the method, video time sequence information is used, the model complexity is low, the parameter quantity is small, the detection effect is good, and the method can be used for traffic monitoring, security protection, target identification and the like.

Interconnected networking systems is becoming a challenge in terms of dependability estimation as two main communication technologies co-exist in today's networks: switching and routing. These two technologies have two different and complementary levels of resilience. Switching is focused on sensitivity to delays and connectivity whereas routing is focused on traffic losses and traffic integrity. The main challenge in modeling these systems dependability is to aggregate the complexity and interactions from various layers of network functions and work with a viable model that reflects the resilience behavior from the service provider and the service user standpoints. The method uses a hierarchical approach based on the Markov Chains and RBD modeling techniques to build a multi-layered model of assuring a multi-services networking system meets its reliability targets dictated by a service level agreement. To cope with modeling complexity the multi-layered model is constructed so that each layer reflects the network resilience required level of details.

The invention relates to a catalytic cracking device optimization method based on a data model, and provides a reaction device model established according to big production history data. The method does not depend on a complicated process mechanism, and the yield and key properties of a product can be accurately predicted. The variable correlation algorithm provided by the invention can intelligently screen out variables which are strongly correlated with target variables from massive DCS bit numbers and Lims variables, so that the model complexity is reduced to the minimum, and meanwhile thereliability is guaranteed. In addition, the multi-neural network integrated learning prediction model constructed by the invention is high in operation speed, high in convergence and wide in adaptability, and the proposed optimization method can quickly calculate a result on the premise of consuming few computing resources and can meet the real-time online optimization requirement of the device. An intelligent algorithm is adopted to determine the time delay effect of different process parameters, so that the calculation is more accurate.

The invention relates to a method for improving the parameter number of a deep-layer convolution neural network, which comprises the following steps: step 1, constructing a deep-layer convolution neural network; 2, improving deep-seated convolution neural network; 3, training and testing the deep-seated convolution neural network; 4, optimizing network parameters. Compared with the prior art, theinvention effectively reduces the complexity of the model through the improvement of the structure of the convolution neural network and the reduction and optimization of the network parameters in themodel, while maintaining the original recognition rate of the model, greatly reducing the training time and reducing the hardware requirement.

The invention discloses a method for establishing a comprehensive energy system random optimization model considering scene simulation. The method comprises the following steps: processing uncertaintyof wind-solar power supply output and thermoelectric load prediction by using a scene analysis technology in random optimization; carrying out latin hypercube sampling according to source load probability distribution to obtain various operation scenes in different time periods through simulation, then carrying out clustering reduction on the scenes through a Kmeans clustering algorithm, and therefore constructing a typical operation scene set used for system operation optimization. Based on operation scene set, the overall economy of system operation is taken as a target; meanwhile, the overall energy efficiency level and the new energy consumption capability of the system are considered, a distributed comprehensive energy system random optimization model is constructed, a random optimization problem is converted into a deterministic optimization problem in different operation scenes, operation strategies in different optimization periods are generated, the model complexity is simplified, and the economy and the safety stability of the system under the influence of uncertain factors are guaranteed.

The invention discloses a genome breeding value estimation method for integrating dominant effect, which relates to the technical field of livestock and poultry genetic selection. The method comprisessteps identifying a reference group and a candidate group, the phenotype of the target traits of the reference population was determined, Genome-wide marker typing of reference population, quality control of gene marker of reference population, statistics of heterozygous marker deviation of reference population, formulation of genome marker re-coding rules, genome-wide marker typing of candidatepopulation, quality control of gene marker of candidate population, re-coding of genome marker and estimation of genome breeding value, etc. Based on the deviation degree between the phenotype of theheterozygous genotype and the phenotype of the homozygous genotype, the invention formulates coding rules, starts from the genomic marker end, re-codes the heterozygous genotype, causes the gene marker coding to include dominant effect, and then estimates the genomic breeding value. The invention is adapted to the needs of livestock and poultry genetic breeding, and can greatly improve the accuracy of genome estimation breeding value without increasing the complexity of the model.

The invention discloses a CNN (Convolutional Neural Network)-based voiceprint recognition method for anti-record attack detection. The CNN-based voiceprint recognition method comprises the following steps: step S101, acquiring to-be-detected voice frequency and establishing a voiceprint recognition data set; step S102, carrying out character extraction on the voice frequency of the voiceprint recognition data set, wherein extracted characters comprise a character MFCC (Mel Frequency Cepstrum Coefficient) and a bottleneck layer character; step S103, establishing a CNN by combining MobileNet andUnet; step S104, inputting the voiceprint recognition data set to the CNN for training; step S105, inputting the bottleneck layer character to the trained CNN by using testing voice frequency, thus obtaining a testing score for judging real talk or record voice frequency. The CNN-based voiceprint recognition method disclosed by the invention combines the characteristics of two models of the Unetand the MobileNet, has lower model complexity, i.e., lower model size, smaller computation resource loss and higher recognition accuracy rate, and can be transplanted and applied to a mobile phone side and an embedded device.

The invention discloses a virtual observation data generation method and device. The method comprises the steps that a base station network element in which a terminal is positioned is determined; a main reference base station, network element global common-view satellites and a reference satellite are determined in the base station network element according to the determined base station network element; double differential atmospheric parameters of a network element baseline in which the main reference base station is positioned are calculated so that the double differential atmospheric parameters of a virtual baseline are obtained through interpolation; and finally the double differential atmospheric parameters are converted into single differential atmospheric parameters through a single differential model, and carrier phase and pseudo-range virtual observation values corresponding to all systems and all frequency points are generated. With application of the method, the problem of redundant matrix transform caused by inconsistence generated by independent selection of all the baseline common-view satellites and the reference satellite in the network element can be solved. Besides, virtual reference base station observation data are constructed by adopting the single differential model equivalent to a double differential model so that result precision is guaranteed and model complexity is reduced.

The invention discloses a speech emotion recognition method for extracting depth space attention characteristics based on a spectrogram, which comprises the following steps : a, preprocessing speech in a standard emotion database marked with specified emotion labels to generate a spectrogram; b, sending the speech spectrogram into an AItti model to extract SEF characteristics; c, sending the speech spectrogram into a speech emotion model to extract shallow features; d, taking the SEF feature and the shallow feature as input, and sending the input to a CSWNet to generate a calibration weight feature; and e, sending the calibration weight characteristics to a network layer behind the CRNN, extracting depth emotion characteristics, and performing emotion classification through a Softmax classifier to generate a final classification result. Compared with the traditional model, the method disclosed by the invention has the advantages that the average recognition rate is improved by 8.43% onthe premise of slightly increasing the complexity of the model, the non-similar emotion distinguishing effect is obvious, and the generalization property is good.

The embodiment of the invention provides an anti-interference control method and device for a dynamic positioning ship, and an electronic device. The method includes steps of in consideration of the environmental interference such as ocean wind, ocean waves, and ocean current, establishing a state space mathematical model of the dynamic positioning ship by analyzing the dynamic characteristic andthe dynamic feature of the dynamic positioning ship; designing an interference observer based on the state space mathematical model and using the interference observer to obtain the external environmental interference estimated quantity of the dynamic positioning ship; designing an anti-interference controller based on the external environmental interference estimated quantity and acquiring an anti-interference control closed-loop system based on the anti-interference controller, the state space mathematical model, and the external environmental interference estimated quantity; and based on the anti-interference control closed-loop system, subjecting the dynamic positioning ship to anti-interference control based on an anti-interference compensation quantity. The method in the embodiment of the invention reduces the complexity of the model and the calculation amount of the control process while ensuring the control precision and the control stability.

The invention belongs to the application field of mutual cross combination of multiple subjects and technologies, and specifically relates to a spatial-temporal data model based city-level PM2.5 concentration prediction method. The method comprises the steps of firstly acquiring city-level PM2.5 concentration data by using a developed online data acquisition system; determining an optimal model prediction factor by comprehensively considering the model interpretability, the data accessibility and the model complexity, and thus determining the model type and the model structure; then determining an interpolation algorithm for the missing PM2.5 concentration data according to missing conditions of the actually acquired PM2.5 concentration data and the determined model structure; and finally determining an algorithm capable of performing real-time online prediction according to the model prediction precision. According to the invention, the model structure can be determined by making full use of readily available data, and thus an effective city-level PM2.5 concentration prediction model is built.