757 results about "Vector generation" patented technology

The embodiment of the invention discloses a method and a device for establishing a picture search correlation prediction model, and a picture search method and device. The method for establishing the picture search correlation prediction model comprises the following steps: using a training sample to train a pre-constructed original deep neural network, wherein the training sample comprises a query and picture data, and the original deep neural network comprises a representation vector generation network and a relevant computational network; and taking the original deep neural network which finishes training as the picture search correlation prediction model. The technical scheme of the invention optimizes the traditional picture search technology, and is better than the traditional technology and various fusion and variation capabilities on multiple aspects including the semantic matching of the query and a picture text, the semantic matching of the query and picture contents, click generalization and the like, and relevancy between a picture search result and the query input by the user can be greatly improved.

Data obtaining means repeatedly obtains acceleration data. Acceleration vector generation means generates first acceleration vector in accordance with first acceleration data obtained by the data obtaining means, and generates second acceleration vector in accordance with second acceleration data time-sequentially obtained by the data obtaining means following the first acceleration data. The cross product direction calculation means calculates a direction of a cross product between the first acceleration vector and the second acceleration vector. The swing direction identification means identifies a swing direction in which the input device is swung in accordance with the direction of the cross product.

A method for generating a permutation index vector includes receiving a condition vector and performing an index generation function using the condition vector in order to generate the permutation index vector. An index vector generation circuit is also disclosed.

A system for generating a word sequence includes one or more processors in connection with a memory and one or more storage devices storing instructions causing operations that include receiving first and second input vectors, extracting first and second feature vectors, estimating a first set of weights and a second set of weights, calculating a first content vector from the first set of weights and the first feature vectors, and calculating a second content vector, transforming the first content vector into a first modal content vector having a predetermined dimension and transforming the second content vector into a second modal content vector having the predetermined dimension, estimating a set of modal attention weights, generating a weighted content vector having the predetermined dimension from the set of modal attention weights and the first and second modal content vectors, and generating a predicted word using the sequence generator.

The invention discloses an implementation method of a fusion network question and answer system based on a multi-attention mechanism, which comprises the following steps of constructing a question andanswer system network model, preprocessing an original data set to obtain a standby data set, and performing text length distribution analysis; subjecting text in standby data set to one-hot vector representation, using a CBOW model to train one-hot word vector and forming a word2vec word list; adjusting the sequence length of each sentence in the text, and adding a sentence end mark; training the word2vec vector by using an ELMO language model to obtain an ELMO word vector; encoding the ELMO vector to obtain a sentence vector; performing coarse-fine granularity attention on the sentence vectors respectively to obtain memory vectors and attention vectors based on each word; carrying out vector splicing to obtain expression vectors based on words and sentences; and decoding an answer representing the vector generation question sentence. According to the method, the representation ability of sentences is improved through an ELMO language model; and various attention mechanisms are fused, so that the decision making accuracy of the system is improved, and the interpretability of the system is enhanced.

Motion vectors from one video frame to another are detected by segmenting a present frame of video data into plural blocks and then comparing a block in the present frame to a corresponding block in a preceding frame to detect rotational and zoom movement of the present block relative to the preceding block, in addition to rectilinear movement.

The present invention provides a method (30) for training an artificial neural network (NN). The method (30) includes the steps of: initialising the NN by selecting an output of the NN to be trained and connecting an output neuron of the NN to input neuron(s) in an input layer of the NN for the selected output; preparing a data set to be learnt by the NN; and, applying the prepared data set to the NN to be learnt by applying an input vector of the prepared data set to the first hidden layer of the NN, or the output layer of the NN if the NN has no hidden layer(s), and determining whether at least one neuron for the selected output in each layer of the NN can learn to produce the associated output for the input vector. If none of the neurons in a layer of the NN can learn to produce the associated output for the input vector, then a new neuron is added to that layer to learn the associated output which could not be learnt by any other neuron in that layer. The new neuron has its output connected to all neurons in next layer that are relevant to the output being trained. If an output neuron cannot learn the input vector, then another neuron is added to the same layer as the current output neuron and all inputs are connected directly to it. This neuron learns the input the old output could not learn. An additional neuron is added to the next layer. The inputs to this neuron are the old output of the NN, and the newly added neuron to that layer.

N sets of feature vectors are generated from a set of observation vectors which are indicative of a pattern which it is desired to recognize. At least one of the sets of feature vectors is different than at least one other of the sets of feature vectors, and is preselected for purposes of containing at least some complimentary information with regard to the at least one other set of feature vectors. The N sets of feature vectors are combined in a manner to obtain an optimized set of feature vectors which best represents the pattern. The combination is performed via one of a weighted likelihood combination scheme and a rank-based state-selection scheme; preferably, it is done in accordance with an equation set forth herein. In one aspect, a weighted likelihood combination can be employed, while in another aspect, rank-based state selection can be employed. An apparatus suitable for performing the method is described, and implementation in a computer program product is also contemplated. The invention is applicable to any type of pattern recognition problem where robustness is important, such as, for example, recognition of speech, handwriting or optical characters under challenging conditions.

An apparatus for coding a moving picture includes a motion vector generating unit that generates a motion vector based on a first predicted motion vector stored; and a coding information generating unit that generates coding information used to code a target block, based on the motion vector generated by the motion vector generating unit. The apparatus also includes a second predicted motion vector generating unit that generates a second predicted motion vector for the target block; and a coding unit that codes an image of the target block based on the second predicted motion vector.

The invention discloses an intelligent commodity recommending method based on word vector data driving. The method includes the following steps: data pre-processing, word vector generation, score prediction constructing, model training and score prediction. The method includes the following steps: using a word vector method to take user numbers, commodity numbers and commodity scores as semantic words, changing the semantic words to a sparse vector by conducting one-hot encoding, then multiplying the sparse vector and a weight matrix and mapping a high-dimension and sparse original vector to a dense, continuous and fixed and low-dimension feature space, then inputting the original vector to a deep model for carrying out training to obtain weight parameters of each layer of a model, predicting and scoring a new user's favor to the commodity by using a well-used model so as to complete intelligent recommendation of the commodity to the user. According to the invention, the method applies the word vector method which classifies texts to score prediction and commodity recommendation which is based on favor of the user of an e-commerce platform to the commodity. The method can ensures precision and provide better explanation.

The invention provides a traffic flow prediction system and method, and a model training method. The traffic flow prediction system comprises a first space-time convolution network, a second space-time convolution network, and an output layer, wherein the first space-time convolution network is used for generating a first space-time feature vector according to the graph structure data; the secondspace-time convolution network is used for generating a second space-time feature vector according to the graph structure data and the first space-time feature vector; and the output layer is used forgenerating a prediction value according to the second spatial-temporal feature vector. In the present invention, the traffic flow prediction system processes the graph structure data; and on the premise of ensuring the time attribute of the traffic data, the graph structure data gives the spatial attribute to the traffic data, and the spatial-temporal dependence relationship is obtained by superposing the two spatial-temporal convolutional networks, so that the traffic flow prediction system provided by the invention truly realizes prediction of the traffic flow by combining the spatial feature and the temporal feature of the traffic data.

The invention relates to a license number matching algorithm based on digital image processing. The license number matching algorithm based on the digital image processing comprises the following steps of: twice shooting a same license through shooting equipment to obtain two images, and carrying out scale space extreme value detection, accurate characteristic point position positioning, characteristic point principal direction determination and characteristic vector generation on the two images to obtain image characteristics; and then matching the characteristic vectors of the two images. The license number matching algorithm based on the digital image processing, which is disclosed by the invention, can still complete the matching and keep good stability and robustness under the conditions of scaling, rotation, brightness variation, affine transformation and noise influence through a particular experiment verification algorithm and is suitable for carrying out fast and accurate matching in a great number of characteristic databases.

Provided are a flexible bit rate code vector generation method and a wideband vocoder employing the same. This invention implements a flexible bit rate by getting three code vectors which are composed of 24, 16, and 8 pulses, at a time in a search process, through improvement of an algebraic codebook search process in a wideband AMR-WB vocoder. The method includes the steps of: performing a preprocess, wherein the preprocess divides a sub-frame by tracks and decides a pulse position having a maximum value in each track; among a plurality of pulses to be searched, fixing a same number of pulses as the tracks to the position with the maximum value of each track sequentially, and searching optimal positions having a minimum error with a target signal by combining two pulses in two consecutive tracks for the remaining pulses; and creating a code vector with flexible bit rate.

The invention discloses a comment word emotion analysis method and system based on deep learning. The method includes the steps: receiving a text to be commented, and segmenting the text to be commented to obtain a word sequence; transforming the word sequence into a corresponding word vector sequence by a word vector generation model; judging whether the word sequence comprises emotion tendency words or not according to a domain emotion dictionary, acquiring a corresponding expansion word vector according to the emotion tendency words and adding the expansion word vector into the word vector sequence; inputting the word vector sequence with the added expansion word vector into an emotion judgment model and outputting judgment results. A traditional word vector is expanded based on the domain emotion dictionary, the information intensity of domain emotion words is enhanced, emotion tendency of the domain emotion words in a specific domain can be accurately recognized, and emotion tendency analysis accuracy is effectively improved.

The invention discloses a foreign language writing automatic error correction method and system. The method comprises the steps of pre-building an error correction model used for foreign language statement automatic error correction; obtaining a foreign language statement written by a user; extracting word vectors of words in the statement and sentence vectors of the statement; inputting the word vectors of the words in the statement and the sentence vectors of the statement to the error correction model, obtaining corrected word vectors of the words output by the error correction model, and generating an error correction text according to the corrected word vectors of the words; and displaying the error correction text. By utilizing the method and the system, automatic error correction of statement errors in foreign language writing can be realized and the writing and learning effects of students can be improved.

Systems and methods for Bloom filter compaction are described herein. A method embodiment includes reading a data corpus, inserting a plurality of data items from the data corpus into a Bloom filter, determining a number of the data items inserted, calculating a desired filter size based on the number, and constructing a compact Bloom filter based on the desired filter size. Another method embodiment includes generating a second bit vector from a first bit vector associated with a first Bloom filter, generating an empty second Bloom filter of a size based on the first bit vector, dividing the first Bloom filter's bit vector into a plurality of slices, and setting the second Bloom filter's bits based on an evaluation of the corresponding bits in each slice of the first Bloom filter's bit vector.

To provide a map generation device according to the present invention which extracts a polygon shape of a building having a complex upper portion structure from a wide area image. The map generation device includes an image appointment unit that receives appointment of at least one position in a building existing within an aerial photograph, a polygon extraction unit that extracts a building region based on a result of discriminating a color around the appointed position and extracts a polygon line of the building region, and a vector generation unit that generates a vector of the polygon line of the building region.

The invention discloses a biological characteristic cryptographic system based on fingerprints and an error correcting code. The biological characteristic cryptographic system is constructed by utilizing an image acquisition unit, a feature extraction unit, a two-value fixed-length feature vector generation unit, a template encryption unit, a template storage unit and a template decryption unit and the like. The method provided by the invention comprises the following steps: adopting a triangle which is formed by detail points of three fingerprints and conforms to a certain condition; taking a six-dimensional feature vector formed by a smaller included angle between the direction of the directional field in which the side length of the triangle and the detail points are located, and a connecting line of the detail points as a fingerprint feature; carrying out training and dimension reduction, so as to obtain a two-value fixed-length feature vector; correspondingly encrypting the two-value fixed-length feature vector to be used as a template to be stored; correspondingly transforming a query fingerprint feature, so as to obtain the corresponding two-value fixed-length feature vector; and carrying out decrypting and authenticating operations on the template fingerprint by the two-value fixed-length feature vector.

The invention discloses a voice synthesis method based on voice vector textual characteristics. The voice synthesis method comprises the following steps: receiving an input text by a text analyzing module; carrying out regular processing on the textual characteristics and transmitting obtained text data to a text parameterization module; obtaining a parameterized text by adopting a single-bit heat code encoding method; receiving the parameterized text by a voice vector training module, and training a linguistic model based on voice vectors; then transmitting to a linguistic parameter training model to train a mapping model from the text to voice parameters; receiving the output text of the text parameterization module and the voice vector training module through a voice vector generation module, so as to generate the voice vectors of the text data; and transmitting the voice vectors of the text data and the mapping model from the text to the voice parameters to a linguistic parameter predication module to obtain the voice parameters corresponding to the voice vectors; and finally, synthesizing voices by a voice synthesis module. According to the voice synthesis method based on the voice vector textual characteristics, the accuracy of modeling of a voice synthesis system is improved; and the complexity and the manual participation degree of system realization are greatly reduced.

The embodiment of the invention provides a multi-round conversation semantic analysis method based on a long-term and short-term memory network. The method comprises the steps that current conversation information is acquired; generating a current conversation representative vector according to the current conversation information; generating a knowledge code representation vector according to thecurrent conversation representation vector and a plurality of historical conversation code vectors acquired in advance; inputting the knowledge code representation vector and the word vector of eachsegmented word in the current conversation information into a first long-term and short-term memory model to obtain a prediction sequence label of the current conversation information; and obtaining corresponding semantic information according to the prediction sequence label, and executing corresponding operation according to the semantic information. The embodiment of the invention provides a multi-round conversation semantic analysis method and system based on a long-term and short-term memory network, computer equipment and a computer readable storage medium. According to the embodiment ofthe invention, the conversation information can be accurately understood, and the problems of ambiguity of multiple rounds of conversations and poor prediction capability for new conversations can besolved.

The invention relates to a Chinese multi-keyword fuzzy sort encryption text search method based on local sensitive hash. After Chinese keywords are converted into the corresponding Chinese phonetic alphabet strings, the Chinese phonetic alphabet strings are segmented based on consonants, vowels, tones, and unigram; a vector generation algorithm of three types of Chinese keywords is designed, the Chinese phonetic alphabet strings are mapped to keyword vector; fuzzy matching of keywords is achieved by utilizing the attributes of locally sensitive hash and bloom filters. The encryption index of the document adopts the method of one document corresponding only to one bloom filter, at the addition of a new document (or the deletion of an old document), the encryption index of an original data set does not need to be changed, only the encryption index of the new document needs to be built (or the encryption index of the old document is deleted), and dynamic updating of the document can be achieved. A domain weighted scoring method is introduced into the method in order to improve the accuracy of the sort results, the Euclidean distance between the keyword vectors, the weight of keyword frequency and domain weighted scoring are combined, more accurate three-factor sorting is achieved, and documents which meet user needs more are returned.

The invention discloses a texture feature extraction method fused with visual significance and GLCM (gray level co-occurrence matrix). The method comprises: (1) a initialization step of determining the size of a detection window, a basic block and a super block for a certain image; (2) calculating a significance factor and a texture structural feature vector of an image by selecting the basic block as a unit; (3) generating a significance texture structural feature description operator with a certain number of dimensions by two-dimensional histogram according to the significant factor and the texture structural feature vector by selecting the super block as a unit; and extracting one significance texture structural feature description factor for each super block according to the significance factor operator as well as the size of the detection window, the basic block and the super block, and describing the significance texture structural feature description factor by a one-dimension feature vector. The texture feature extraction method provided by the invention can simulate the human eye to observe the divergence and significance characteristics of things, and has the advantages of simple calculation, low redundancy degree, and high real-time performance.

For testing a device under test (“DUT”) a test specification is converted in a computer system by a pin vector generator process, which includes generating test vectors. The DUT has numerous input pins and such a pin vector is for a signal to drive one such pin. The pin vectors are compressed and saved. Ones of the pin vectors are loaded, upon initialization of a test, into a pipeline having a series of memory stages and extending from the computer system to channel cards in a test head. The pipeline is operated in data transfer cycles, delivering W bits per cycle. The pin vectors are decompressed at the respective channel cards in decompressor read cycles. X bits are read per decompressor cycle, W being greater than X, so that the pipeline may perform its data transfer cycles less frequently than the decompressor performs its read cycles.