101results about How to "Speed advantage" patented technology

A process and apparatus for implementing parallelization in deblocking filter used in a an H.264 codec are disclosed. In the preferred embodiment, the process is carried out on a parallel architecture consisting of a plurality of groups, each consisting of eight clusters, wherein each cluster is a separate processor capable of tensor operations in SIMD or MIMD or mode on 4×4 matrix data. All eight clusters of one group are used to simultaneously deblock both luma and chroma vertical and horizontal edges between 4×4 blocks of pixels in a macroblock in a total of eight iterations, utilizing in the best way the data dependency between the edges. Processes to deblock these same luma and chroma edges in more iterations on four cluster and two cluster parallel architectures are also disclosed. A comparison of the maximum parallelization achievable with the invention and the amount of parallelization with various species within the prior art is also disclosed.

A parse/build engine that can handle multi-format financial messages. The engine converts the different format messages into a common format, and the common format message is then processed by the business service application. A parser examines the message and determines an appropriate schema for the particular format of message received. The schema is a data structure in a schema registry that includes a grammar structure for the received format as well as pointers to handlers for converting the different fields of the message into the internal message format using the grammar structure (the “grammar” can include field sequence, field type, length, character encoding, optional and required fields, etc.). The handlers are individually compiled. As formats change, new formats or changes to old formats can be dynamically added to the parse/build engine by loading new schema and handlers.

The invention discloses an end-to-end object detection method based on a convolutional neural network. The method includes the following steps: (1) on the basis of a classic basic network structure, removing the last fully-connected layer of the classic basic network, and adding an extra layer so as to establish a convolutional neural network model; (2) randomly selecting, from an original training data set, an original image, obtaining an amplified image through data amplification, and obtaining the position and frame of an object image block, randomly selected from the original image, in the amplified image; (3) through the position and frame of the object image block in the amplified image obtained in the step (2), conducting regression on the convolutional neural network model in the step (2) to obtain model parameters, and thus obtaining a trained convolutional neural network model; and (4) utilizing the trained convolutional neural network model to detect a bounding box and a classification of an object in an image to be detected. Direct regress of a central point coordinate, the width, the height and the classification of the object is adopted by the method. Compared with similar methods, the method has a significant advantage in terms of speed.

Response latencies in a system are reduced by performing a plurality of auctions in parallel using respective processing threads. An ad call associated with an impression for a webpage requested by a user computer is received by the system. Impression information is obtained from the ad call, which is then used to determine a subset of multiple potential impression providers that are to participate in an auction for the impression. Behavior models are used to determine the subset, while floor or reserve price models are used to set a floor/reserve price for each potential impression provider in the subset. Bid requests are sent out in parallel to the subset of potential impression providers using the respective floor/reserve prices, with a thread allocated to each bid request. The models can be further trained on information obtained from the received bid responses and impression information.

A depth image automatic registration process combining with texture information is mainly used to reconstruct the three-dimensional model of various actual objects, which comprises the steps: firstly extracting a texture image from scanning data or generating a texture image according a depth image, secondly extracting interest pixels in the texture image based on the characteristics of SIFT, and finding out candidate sets therein matched with the interest pixels through the crosscheck method, thirdly finding out correct matched pixel pairs in the candidate sets according to geometry information constraint, fourthly finding out matched apex pairs corresponding to the matched pixel pairs in a three-dimensional space, and calculating the rigid replacement matrix between two depth images, fifthly optimizing the result through using the improved ICP algorism, sixthly dividing the input sequences of several depth images into a plurality of bar type subsequences based on the registrations of two depth images, finally combining the subsequences through the strategy of forward searching, and constructing an integral three-dimensional model. The invention can fast register large-scale three-dimensional scanning data to generate a three-dimensional model, and has strong anti-noise capability and excellent universality.

An aqueous acidic polymerization liquid comprising polymerisable monomers or oligomers with carboxylate and/or carboxylic acid moiety or moieties, and basal surface modified clay that is homogenously dispersable in said liquid is provided, and also methods for making such liquids, and also surface-modified-clay linked water-absorbing polymers made with said liquids and methods, and disposable absorbent articles comprising such surface-modified-clay linked water-absorbing polycarboxylate/polycarboxylic acid polymers.

A tool for generating a force downhole comprises a body, a longitudinally movable activating member mounted to the body, and a longitudinally movable driven member also mounted to the body. The driven member is operatively associated with the activating member such that on translation of the activating member in one axial direction, the driven member is translated in an opposite axial direction. The tool may be utilised to convert a pulling action, applied by a spoolable member, to a pushing action, useful in disengaging a downhole lock.

A low-resolution face recognition method and apparatus Aiming at the problem of face recognition at low resolution, Especially for the low-resolution face recognition problem in natural environment, the dual-stream depth learning network structure is adopted to train the high-efficient low-resolution face recognition network by selective knowledge distillation, and the low-resolution face recognition is realized. The low-resolution face recognition network has the precision, the speed and the memory advantages close to the high-resolution face recognition model.

The method includes to steps: (1) constructing exponent matrix of parity check matrix of QC-LDPC codes; (2) using exponent expansion to convert the exponent matrix to parity check matrix H of QC-LDPC codes. When iterative decoding, code error rate of the constructed QC-LDPC codes is lower than the rate of traditional PEG algorithm obviously (specially, in condition of high SNR), and algorithm complexity is mach lower than algorithm complexity of traditional algorithm so that the constructing speed is much fast than speed of traditional method. Moreover, when constructing QC-LDPC codes, the disclosed method can control each parameter for generating code flexibly.

A method and a control sequence determination device for the determination of a magnetic resonance system activation sequence including at least one high-frequency pulse sequence to be transmitted by a magnetic resonance system are provided. A current B₀ map and optionally a target magnetization are acquired. In addition, a k-space trajectory type is determined. An error density is calculated in a k-space based on the current B₀ map and optionally based on the target magnetization using an analytic function. This analytic function defines an error density in the k-space as a function of the current B₀ map and optionally the target magnetization. Taking account of the error density in the k-space, a k-space trajectory of the specified k-space trajectory type is determined. The high-frequency pulse sequence is determined for the k-space trajectory in an HF pulse optimization process.

The invention relates to a filtration method of real-time junk e-mails, and the method is used for filtering a junk e-mail at a network interconnection point. The method comprises the following steps: step one, a path combination P is constructed according to the routing information in the e-mail, and the path combination P is a sending correspondent and a receiving correspondent, and / or a sending correspondent and an information destination; step two, attribute space mapping is performed to the IP address or the domain name of the elements in the path combination P according to IP address mapping or domain name mapping; steps three, whether the current e-mail is a junk e-mail is determined according to the attribute space mapping result of each element in the path combination P, if the e-mail is a junk e-mail, step four is executed, otherwise, the current e-mail is not processed; step four, the current e-mail is filtered. In the invention, only the mail order and the content in the header are analyzed, the processing speed is quick, the network resource utilization rate can be improved, and simultaneously, the filtration pressure at a user terminal can be relieved.

The present invention relates to a multi-layer oriented shrink film comprising a first and a second skin layer comprising a propylene-based polymer having a melting point greater than 130 C and at least one core layer comprising a propylene-based plastomer or elastomer. The films are preferably primarily oriented in either the machine or the cross direction. These films are ideally suited for use as a shrink label which can be applied either as a CD shrink sleeve or applied using a continuous roll process. The films ideally have an overall density of less than 1.0 g/cm3, to facilitate recycling efforts.

The invention discloses a metal identification method based on information entropy and belongs to the image identification technology field. The method is characterized by firstly, carrying out multistage wavelet transformation on a sample image; then, using a K mean value algorithm to carry out identification for one time; for a sample which is not identified, analyzing a transformed multistage wavelet transformation entropy value; and using a decision tree algorithm to carry out the identification for the second time. In the method, through multilevel identification, an identification speed is increased and simultaneously algorithmic accuracy is guaranteed. An algorithm speed is fast, an accuracy degree is high, identification efficiency is high too and a real-time performance requirement can be satisfied. The method supports parallel processing. Equipment needed during identification is simple and environmental-protection. The method has wide applicability. The method is suitable for identification of copper, aluminum alloy and stainless. Under the condition that there are sufficient samples, because hardness and colors of different metal and alloy are different to some extent, the method can be used for distinguishing various kinds of non-ferrous metal.

The embodiments of the present invention relate to the field of communication, and in particular to a mobile communication terminal system and a communication method of the mobile communication terminal system, which are used to improve network utilization. In the embodiment of the present invention, the mobile communication signal transmitted by the base station is received; when it is determined that the strength of the received mobile communication signal is not less than a threshold, it is determined that the mobile communication terminal system communicates through the mobile communication system. Because in the embodiment of the present invention, according to the intensity of the mobile communication signal that the received base station transmits and the size of the threshold value, the communication system of the communication terminal system in motion is flexibly selected; When the value is not less than the threshold, the user can choose the mobile communication system for communication, and the mobile communication system has more advantages in tariff and rate than the satellite communication system. By using the network for communication, network utilization can be increased.

A high-power-efficiency multiplier combines a standard floating-point multiplier with a power-of-two multiplier that performs multiplications by shifting operations without the need for floating-point multiplication circuitry. By selectively steering some operands to this power-of-two multiplier, substantial power savings may be realized. In one embodiment, multiplicands may be modified to work with the power-of-two multiplier introducing low errors that may be accommodated in pixel calculations.

A method is described for serial data transmission in a bus system having at least two subscribed data processing units that exchange messages via the bus, the transmitted messages having a logical structure in accordance with CAN standard ISO 11898-1, the logical structure including a start-of-frame bit, an arbitration field, a control field, a data field, a CRC field, an acknowledge field and an end-of-frame sequence, the control field including a data length code, which contains information regarding the length of the data field. When a first switchover condition is satisfied, the data field of the messages, in contrast to CAN standard ISO 11898-1, may comprise more than eight bytes, the values of the four bits of the data length code being interpreted at least partially in deviation from CAN standard ISO 11898-1 for determining the size of the data field when the first switchover condition is satisfied.

This disclosure concerns an adaptive repair method, for example for aerofoil blade components. The method generates a digitised model of a physical component and uses data from the digitised model to align a part of the physical component with a nominal CAD model of the component. A combination of the digitised data and data from the nominal CAD model are then used to create an adapted model for repair of the component.

The invention discloses a long-time face tracking method based on an intelligent cloud platform of a patrol robot. The method comprises the steps that an adaboost face detection module is used to carry out face detection in a region of interest in an input video image frame; a detected face is initialized; a face tracking module tracks the initialized face, and if tracking is effective, a face template updating module is started to update a face template; and if tracking is ineffective, a template matching face detection module is started to carry out template matching detection of the detected face. According to the invention, the face is detected in higher instantaneity, the modules are tracked and learnt, the speed is higher, a target which is lost in tracking can be relocated, the face can be tracked in long time, and multi-target tracking is in real time.

The invention belongs to the technical field of radar SAR image registration, and discloses an improved Lapras multi-extremum inhibition-based SAR image registration method. The method comprises the following steps of: obtaining a reference SAR image and a to-be-matched SAR image, and respectively establishing SAR-Harris angular point response spaces of the two images; respectively extracting angular point coordinates of the SAR-Harris angular point response spaces; respectively positioning the angular point coordinates of the SAR-Harris angular point response spaces, determining a greyscale statistic value of a local neighborhood area of each angular point, determining a main direction of each angular point, and forming angular point feature vectors of the angular points of the two images; comparing the angular point feature vectors of the angular points of the two images to obtain matching point pairs of the two images; and carrying out image registration on the to-be-matched SAR image according to the matching point pairs of the two images, so as to improve the instantaneity and accuracy of the SAR image matching.

The invention relates to the field of artificial intelligence and medical treatment, in particular to a training method, a diagnosis system and related equipment for a turner syndrome diagnosis model,and aims to improve the accuracy. The training method of the turner syndrome diagnosis classification model comprises the following steps:classifyinginput image samples according to labeling information of data samples, andgeneratingcorresponding training samples and test samples according to a preset proportion; respectively constructing different initial medical classification models based on multiple neural network bases; inputting the training sample into each initial medical classification model for training and parameter adjustment to obtain a corresponding intelligent diagnosis medicalclassification model; and respectively inputting the test samples into each intelligent diagnosis medical classification model for classification, and selecting a turner syndrome diagnosis classification model according to a classification result. The diagnosis system provided by the invention comprises the diagnosis classification model obtained by training through the method, associates the same parts in the photos at different angles, extracts more potential features, and improves the accuracy.