38 results about "Regular constraint" patented technology

The design of typable devices, in particular, touch-typable devices embodying ambiguous codes, presents numerous ergonomic problems. Solutions for these problems are herein disclosed. This invention teaches methods for the selection of ambiguous codes from the classes of strongly-touch-typable ambiguous codes and substantially optimal ambiguous codes for touch-typable devices such as computers, telephones, pagers, personal digital assistants, smart cards, television set-top devices and other information appliances, given design constraints such as the size, shape, and computational capacity of the device, the typical uses of the device, and conventional constraints such as respect of alphabetic ordering or Qwerty ordering.

The invention relates to a turbulence-degraded image blind restoration method based on edge prediction and sparse ratio regular constraints. The method is technically characterized by comprising the steps of predicating an effective edge in a current image to be restored, combining edge predication information with sparse prior information of a natural image edge to guide restoration of a point spread function, restoring the current target image according to a non-blind restoration algorithm, regarding the restored image as input of edge predication of the next time, and carrying out the iterative cycle in this way till a clear restored image is obtained. According to the method, by combining the prior information of an image with effective information contained in the degraded image, artifacts generated in the image restoration process can be effectively restrained, more details can be restored, and the restoration effect is better.

The invention discloses a design system for a human dynamic bone, comprising an initialization module, a node generation module, a bone constraint module and a collision constraint module, wherein the initialization module is used for initializing model data; the node generation module is used for generating and storing bone node data; the bone constraint module is used for performing regular constraint on a node in the change process; and the collision constraint module is used for performing regular constraint on a bone model in collision. The invention further discloses a design method for a human dynamic bone, comprising the steps of: initializing model data; generating and storing bone node data; performing regular constraint on a node in the change process; and performing regular constraint on a bone model in collision. The system and the method have the advantages of greatly shortening the time required for adjusting lattice vertex parameters by adopting a bone animation or a fabric system in the old scheme, physically simulating the motion and relevant constraints of fabrics and hair more truly and vividly, and efficiently processing the collision inside a fabric and between the fabric and other object.

The invention provides a diffracted wave field extraction method and device, and relates to the technical field of diffraction field extraction. The method includes: obtaining a pre-stack common-offset gather data in an area to be processed, wherein the pre-stack common-offset gather data carries the stratum interface information in the area to be processed; on the basis of plane wave decomposition on the pre-stack common-offset gather data, transforming the local dip angle of a reflected wave by curvelet transform and then performing regular constraint by using the L0 norm so as to obtain thefirst objective function of the diffracted wave field to be extracted relative to the local dip angle of the reflected wave; and solving the target reflection wave dip angle by a trust region algorithm, wherein the target reflection wave dip angle is the reflection wave dip angle when the first objective function reaches the minimum; and determining the diffraction wave field to be extracted by combining the target reflected wave dip angle, the pre-stack common-offset gather data and the first objective function. The method and device alleviate a technical problem that the diffraction wave extracted by the traditional diffraction wave extraction method is poor in accuracy.

The invention relates to a semi-supervised audio event identification method based on depth mutual information maximization. A semi-supervised neural network model is used as a backbone, a depth mutual information maximization consistency-based regular constraint and a cross entropy classification constraint are designed, a semi-supervised learning model is constructed, a mutual information discriminator is designed to estimate mutual information between deep representation vectors of the model, the model mines potential relations between samples through global mutual information so as to enhance consistency and nonlinear correlation between global representations, and a semi-supervised audio event classification model with high robustness is obtained; and neural network model parameters are optimized by using a gradient descent method, and the audio event samples are classified. The method has the advantages of being small in error, high in robustness, high in precision and the like,the requirement for sound event classification can be met under the condition that label data is insufficient, and high application value is achieved.

The invention relates to the technical field of social network services, and discloses a symbol network community discovery method based on structural balance constraint. The method comprises the steps of representing a symbol network in a graph form, constructing a symbol network adjacent matrix, setting structural balance constraint information, constructing a community discovery model based onsemi-non-negative matrix decomposition, solving the community discovery model, and obtaining a community discovery result. According to the method, the structure balance constraint relation of nodes is obtained based on the structure balance theory; and then the structure balance constraint relation serving as a regular constraint term is incorporated into a symbol network community discovery model based on semi-non-negative matrix decomposition to solve the node community indication matrix, and finally, a community division result is obtained by judging the node community attribution intensity, so that the community discovery quality is further improved.

The invention relates to a regular auto-encoding text embedded expression method for local topic probability generation and belongs to the field of natural language processing and machine learning. The method comprises the steps of firstly, implementing construction of a text set neighbor graph, which includes calculation of similarity weight of any text word pair, search of a maximum weighted matching distance of the text pair, calculation of the similarity of an averaged maximum weighted matching distance (NMD) and selection of a k-nearest neighbor according to an NMD result and constructionof the neighbor graph with the NMD result as an edge weight; then, constructing a sub-space through a transductive multi-agent random walk process through the neighbor graph to determine the sub-space; and finally generating a pseudo-text by use of an LDA (Latent Dirichlet Allocation) model of the sub-space, taking the pseudo-text as a regular constraint item, taking the pseudo-text and a real text as a reconfiguration object of an auto-encoding network, guiding the encoder network to confront the change of a local neighbor text topic probability generation structure so as to construct smoothaffine mapping. According to the regular auto-encoding text embedded expression method for local topic probability generation, the smoothness of the local neighbor text topic probability generation structure can be effectively kept, thereby constructing a smooth affine mapping function, enhance intra-class compactness and inter-class separation of an out-of-sample text embedded representation vector and improving application effects such as text classification and clustering.

The invention discloses a traffic big data filling method based on a tensor train decomposition model. The method comprises the following steps: constructing a five-dimensional tensor model containingfive traffic data dimensions; constructing an initial filling model based on a tensor train decomposition model through L2 regular constraint; performing conjugate gradient optimization on the filling model to obtain an optimized filling model of each kernel vector; or performing trace norm optimization on the filling model to obtain a final filling model; and performing traffic big data fillingthrough the first filling model and/or the second filling model. According to the method provided by the invention, the data filling precision can be improved, and the filling stability can be maintained at a high loss rate.

The invention relates to an adaptive target tracking method based on PSPCE. A target to be tracked is determined in a tracking sequence, the current frame of the target is the tth frame, and the method comprises the following steps: S1, defining a search box according to the area position of the target in the previous frame, and calculating the feature map of an image block in the search box; s2,calculating the response of the filter and the feature map, and determining the target position of the current frame; s3, updating a PSPCE confidence sample filter queue according to the PSPCE coefficient to obtain a PSPCE confidence sample of the current filter; s4, adaptively updating the parameters of the current filter based on PSPCE regular constraints; and S5, repeating the steps S1 to S4 frame by frame until tracking of the tracking sequence is completed. According to the method, the correctness of the adaptive filter in the updating process can be ensured, the high-confidence filter can be selected on the time axis to update the PSPCE confidence sample, the situation that the filter suddenly changes under various changes of the appearance of the target is effectively prevented, andthe accuracy of target tracking is ensured.

The invention discloses a fine-grained image recognition method based on multi-target Lagrange regularization, and the method comprises the steps: extracting an image feature X of an input image through a neural network, and obtaining a corresponding bilinear image A; obtaining a regularization feature Y from the bilinear image A, and constructing a target function containing a matrix square root,a low rank and a sparse constraint item; introducing two auxiliary variables to weaken the relevance among the three constraint items, converting the target function into an augmented Lagrange form,and alternately optimizing each matrix constraint item to obtain a global approximate optimal solution; and utilizing the global approximate optimal solution to carry out image identification and classification. The method only comprises matrix multiplication, so that the method can be well compatible with a GPU (Graphics Processing Unit), a higher recognition speed is achieved, regular constraints of square root, low rank and sparsity can be effectively carried out on bilinear image expression at the same time, and the recognition accuracy is greatly improved.

The invention discloses a non-negative image data dimension reduction method based on Hessian regular constraint and A optimization. The method includes the steps that step1, a sample feature matrix is constructed; step2, a Hessian regular matrix is calculated; step3, a basis matrix and a coefficient matrix are iteratively output so as be analyzed in a clustered mode. An A optimization regular item and a Hessian regular item are added into a target function so that a data expression obtained through decomposition can keep inherent features included in the manifold-shaped structure of original data while guaranteeing a small prediction error; through dimension reduction, redundancy information in the high-dimension data is removed, the low-dimension expression capable of accurately expressing the semantic structure of the data is extracted, and therefore clustering analysis performed on the high-dimension data becomes simpler and more effective.

The invention provides a neural network processing method and device, a storage medium, and an electronic apparatus, wherein the method comprises: determining an objective function in a convolution neural network, wherein each layer of the objective function corresponds to an initial network weight value, and the convolution neural network is applied to at least one of face recognition, vehicle detection, and object recognition; processing the initial network weights according to preset regular constraint coefficients to obtain target network weights; using the target network weights and input values to an output value in the objective function. The invention solves the technical problems of slow convergence and low precision of convolution neural network in the prior art.