44 results about "Global local" patented technology

An “Image Segmenter” provides a variational energy formulation for segmentation of natural objects from images. In general, the Image Segmenter operates by adopting Gaussian mixture models (GMM) to capture the appearance variation of objects in one or more images. A global image data likelihood potential is then computed and combined with local region potentials to obtain a robust and accurate estimation of pixel foreground and background distributions. Iterative minimization of a “global-local energy function” is then accomplished by evolution of a foreground/background boundary curve by level set, and estimation of a foreground/background model by fixed-point iteration, termed “quasi-semi-supervised EM.” In various embodiments, this process is further improved by providing general object shape information for use in rectifying objects segmented from the image.

A system and method for capturing images is provided. In the system and method, preview images are acquired and global local and local motion are estimated based on at least a portion of the preview images. If the local motion is less than or equal to the global motion, a final image is captured based at least on an exposure time based on the global motion. If the local motion is greater than the global motion, a first image is captured based on at least a first exposure time and at least a second image is captured based on at least one second exposure time less than the first exposure time. After capturing the first and second images, global motion regions are separated from local motion regions in the first and second images, and the final image is reconstructed at least based on the local motion regions.

The invention provides a significant detection algorithm based on a multistage convolution neural network. The algorithm comprises the following steps of using a global estimation model of a large receptive field to carry out global significance estimation; during a global estimation model training process, using a full connection layer as an output layer to train and initialize parts of convolution layer parameters; using a plurality of alternative convolution layers and liter sampling layers to replace the full connection layer and training and acquiring an optimal global significance estimation graph; using a local convolution neural network with a small receptive field and a large output picture size to fuse global and local information to acquire a high quality significance graph. Through processing of the local convolution neural network, an original image is served as input of the model. A final output result possesses a same size with an original input image and is clear. By using the significant detection algorithm based on the multistage convolution neural network, compared to a traditional method, high accuracy is possessed; a significant object can be accurately found and simultaneously a contour is clear.

The invention provides a method and device for constructing a legal text syntax tree based on a deep neutral network. The method comprises the following steps: acquiring a training text of the legal service; performing word vector sparse representation on the training text; performing semantic annotation on the training text through the DNN according to the word vector sparse representation; performing dependency description on the training text according to the semantic annotation to construct the legal text syntax tree. The invention provides a systematic method to construct the syntax tree about technical language understanding in the legal service field, and the global-local training is used for solving the semantic bias problem in the legal language understanding.

In a semiconductor memory device, a memory cell is connected with a local sense amplifier and a global sense amplifier via a local bit line and a global bit line. The local sense amplifier is a single-ended sense amplifier including a single MOS transistor, which detects a potential of the local bit line which varies when reading and writing data with the memory cell. The threshold voltage of the MOS transistor is monitored so as to produce a high-level write voltage and a low-level write voltage, which are corrected and shifted based on the monitoring result so as to properly perform a reload operation on the memory cell by the global local sense amplifier. Thus, it is possible to cancel out temperature-dependent variations of the threshold voltage and shifting of the threshold voltage due to dispersions of manufacturing processes.

The invention, which belongs to the field of image processing and identification, discloses a pedestrian re-identification method and device based on global features and coarse granularity local features. The method comprises: detecting a pedestrian image in a query image as a global image, detecting a body key point of the pedestrian, and dividing the body of the pedestrian to obtain a local partregion; extracting global feature description of the global image and local feature description of the local part region, carrying out fusion of the global feature description and the local feature description to obtain global-local feature description; and carrying out associated analysis and combined indexing on all images in a pedestrian database, carrying out pedestrian retrieving on the processed images from the coarse granularity to fine granularity according to the global-local feature description, and determining the identity of the pedestrian in the query image. According to the invention, coarse granularity division is carried out on the body in the image, so that the good robustness is realized; and because the global features and regional local features are combined, accuratematching of the pedestrian images and pedestrian identity identification are realized.

The invention discloses a global-local RGB-D multimode-based gesture recognition method. An inputted gesture video is expressed mainly through data modes comprising a bone position, an RGB image, a depth image and an optical flow image; after the multimode gesture data expression is obtained, a method of a convolution neural network and a recurrent neural network is used to carrying out feature expression on gesture data in different modes, and the features obtained in different modes are used for gesture classification; and finally, different classes of gesture scores obtained in different modes are fused to obtain a multimode-based gesture classification result finally. The method can be applied to a client or a cloud for recognizing a gesture video inputted by the user, and through gesture input, a computer or mobile phone software or hardware makes a corresponding response.

The invention provides a pedestrian image feature classification method and system, and the method comprises the steps: carrying out the data expansion of a pedestrian image sample in a sample dataset; carrying out the grouping of the pedestrian image sample in the sample dataset after expansion, and obtaining a plurality of pedestrian sample groups; selecting samples, building a multi-channel convolution neural network, and extracting the global and local features of the sample data through the multi-channel convolution neural network; setting a loss function, calculating a loss value of the multi-channel convolution neural network, and optimizing the multi-channel convolution neural network; carrying out the feature classification of each global-local feature through the optimized multi-channel convolution neural network, and obtaining the feature class of each pedestrian sample group. The method enables the sample data to be expanded, meets the condition that triple loss exerts strict requirements for an input sample, can guarantee the robustness through employing multi-loss to optimize the multi-channel convolution neural network, and is suitable for the processing of pedestrian image features of a plurality of scenes.

The invention discloses an image description method, and the method comprises the steps: extracting global image features in a picture through a VGG convolution neural network, and extracting local image features in picture by using Faster R-CNN network, fusing global image features and local image features through a global-local feature fusion algorithm to acquire the image fusion features; processing image fusion features through a bidirectional long-short-term memory network with an attention mechanism; and generating a preliminary image description statement, performing WordNet word vectorsimilarity calculation based on the image target information obtained during local image feature extraction and nouns in the preliminary image description statement, correcting the image descriptionstatement, and generating a final image description statement. According to the method, the influence of useless information is reduced, the expression of key information is enhanced, the fault tolerance and generalization ability of the model are enhanced, and the accuracy of statement description is improved.

The present invention provides a high-speed train real-time monitoring oriented mass data management method. According to the method, a two-layer storage architecture is formed by a distributed file system and a relational database, and a mileage triggered oriented window data management mechanism is established on an upper layer based on the relational database to realize the organization of monitored data; and by utilizing the distributed file system at a lower layer, balanced storage management of mass data is implemented. Ordered organization of the data can be realized in combination with a storage model of object dynamic coupling on the basis; and meanwhile, by using a way of global-local property delivery management and establishing mileage accumulated triggering for trains, demands of train monitoring and analyzing are met. Aiming at high-speed train real-time monitoring, the high-speed train real-time monitoring oriented mass data management method provides core technologies of massive and cross-discipline data management under the large data background and has active application value in the fields of transportation, logistics, smart city and the like.

The invention discloses a mesh surface curve design method based on distance constraint. The designed curve passes through a given interpolation point and is smooth (in the sense of discreteness) andstrictly located on the mesh surface. The method transforms the complex manifold constraints into distance constraints and describes them as optimization problems together with smooth constraints andinterpolation constraints. The local surface is approximated by the tangent plane, and the distance constraint is relaxed to the distance from the point to the tangent plane. Because the points on thecurve used to calculate the distance are interdependent with their corresponding tangent points, a global-local iterative strategy is adopted and Gauss- Newton idea method is used to control its convergence behavior: in the whole stage, it is relaxed into a convex optimization problem by distance approximation to solve the iterative step size; In the local phase, a robust and efficient projectionmethod is used to map the optimized curves to the surface to update the tangent points. Finally, all the relaxed polygons are mapped to the mesh surface by using the cutting plane method. Compared with the existing methods, this method has many advantages in efficiency, robustness and application range.

The invention belongs to the field of computer vision and computer graphics, and aims to ensure the universality of the application of the technology and generate a relatively accurate three-dimensional model. The technical scheme adopted by the invention is as follows: the sparse multi-view three-dimensional reconstruction method for the indoor scene comprises the following steps of: obtaining depth information and semantic information based on a deep learning method, and then realizing model fusion of each sparse view angle by using a global-local registration method, wherein the global-local registration method specifically comprises the steps of filtering and smoothing the depth map by using a method of establishing a patch by using a single view angle, converting the depth map under each view angle into point cloud and carrying out fusion. The method is mainly applied to image processing occasions.

The invention provides a freehand interactive three-dimensional model retrieval method based on high-level semantic property comprehension. According to the method, first, a data driving mode is utilized to extract global-local semantic properties of freehand sketches under different styles and categories, and a semantic property space of the freehand sketches is defined; second, on the basis of the defined semantic property space, automatic annotation of semantic properties is performed on a three-dimensional model in a database according to content characteristics corresponding to the three-dimensional model; and last, the semantic properties of the freehand sketches and the semantic properties of the three-dimensional model in the database are mapped to the same measurement space by means of constructing a semantic property tree for comparison, and if the similarity between the semantic properties of the freehand sketches and the semantic properties of the three-dimensional model reaches a set similarity, information of the three-dimensional model is fed back, and retrieval is completed. Through the method, deviation brought by an existing freehand interactive three-dimensionalmodel retrieval algorithm through which a three-dimensional model cannot be compared with freehand sketches unless the three-dimensional model is projected into a two-dimensional view is avoided.

The invention discloses a video description method based on complementary attention mechanism cyclic convolution decoding. The video description method is used for generating short and accurate description for a given video clip. The method specifically comprises the following steps: obtaining a video data set for training a video description generation model, and defining an algorithm target; modeling a global time sequence feature and a local semantic feature in the video data set; establishing a complementary attention mechanism on a circular convolutional decoder based on global-local features; using the model to generate a description of an input video. The method is suitable for video description generation of a real video scene, and has a good effect and robustness for various complex conditions.

The present invention provides an aggregation node location method based on an improved discrete difference algorithm. The influence of the isomerism of nodes and the reliability of the routing path in real engineering on aggregation node location is fully considered, an adaptive zoom factor is introduced into the improved discrete difference algorithm to allow the algorithm to maintain high global search capability at the initial stage and maintain high global local search capability at the later period; an adaptive variation mechanism is introduced, an appropriate variation strategy is selected according the trend of the population evolution in the evolution process to maintain the diversity of the population and avoid falling into the local optimum so as to improve the global optimization capability of the discrete difference algorithm, rapidly converging the algorithm and solve the technical problems of the precocity convergence and falling into the local minimum of the original difference algorithm at the discrete variable optimization, and therefore the optimized aggregation node location disposition is obtained, the data communication reliability is enhanced, and the network service quality is improved.

The invention provides a method for extracting a complete periodic waveform, belonging to the signal processing field. The invention adopts the technology of combining global local extremum, successfully overcomes the interference of waveform glitch, intercepts the complete periodic waveform from the starting position of the rising edge, and facilitates the acquisition of the periodic signal waveform in scientific research.

The invention discloses a panoramic light field splicing method based on global local adaptive optimization. The panoramic light field splicing method comprises the following steps: inputting a plurality of light fields to be spliced; extracting, matching and screening feature points of the to-be-spliced light fields to calculate a global homography transformation matrix of all the to-be-spliced light fields; respectively carrying out 4D meshing on the plurality of to-be-spliced light fields, evaluating the global registration precision of each grid according to the global homography transformation matrix, calculating a local homography transformation matrix of the grid of which the global registration precision is lower than a preset threshold, and evaluating the local registration precision of the corresponding grid according to the local homography transformation matrix; and comparing the global registration precision with the local registration precision, taking the homography transformation matrix with the highest precision for each grid, and fusing the light fields to obtain a splicing result. According to the invention, the overall and detail splicing results of the light field are most accurate, so that an accurate multi-light-field panoramic splicing method is realized.

The invention provides a global local time representation method for video-based pedestrian re-identification. A network provided by the invention is composed of a DTP convolution model and a TSA model. The DTP is composed of time convolution expanded in parallel and used for simulating short-term time clues between adjacent frames. The TSA captures global time cues using relationships between non-contiguous frames. Experimental results on five reference data sets show that the proposed GLTR method is superior to the most advanced method at present.

The invention discloses a method and system for acquiring global rainfall information at different places by using a rainfall data set, and the method comprises the steps: collecting and summarizing the global rainfall data set, and obtaining global rainfall data taking the world time as a reference; converting the global rainfall data taking the world time as the benchmark into global rainfall data taking the designated local time as the benchmark; taking the global rainfall data taking the world time as the benchmark of any three adjacent days as an original data set, and obtaining global rainfall data of one hour interval of the next day by adopting adjacent point interpolation, so as to obtain gridding global rainfall data of one hour interval taking the designated local time as the benchmark; and calculating annual average hourly rainfall of the local time according to the global rainfall data of all days of the local time. According to the method, the nearest neighbor interpolation method is adopted, the method is simple and convenient, the operation speed is high, and the global local rainfall result can be obtained in a short time.