Patents
Literature
Hiro is an intelligent assistant for R&D personnel, combined with Patent DNA, to facilitate innovative research.
Hiro

39 results about "Product quantization" patented technology

Product quantization (PQ) is an effective vector quantization method. A product quantizer can generate an exponentially large codebook at very low memory/time cost.

Large-similarity image similarity retrieval method and system based on deep product quantization

The invention puts forward a large-similarity image similarity retrieval method based on deep product quantization. The method comprises the following steps that: inputting a picture to be retrieved into a trained deep neural network, and obtaining feature representation corresponding to the picture to be retrieved, wherein the deep neural network comprises that the last fully connected layer before a multiple-term relative rate regression classifier in AlexNet is replaced with a fully connected quantization layer with a plurality of units; on the basis of the feature representation corresponding to the picture to be retrieved and the feature representation corresponding to each picture in a retrieval library, calculating an asymmetric quantization distance between a picture to be retrieved and each picture in the retrieval library; and in the retrieval library, selecting a preset quantity of pictures which have the shortest asymmetric quantization distance with the picture to be retrieved from the retrieval library as a retrieval result. By use of the method, a quantization error is minimum on the basis of deep representation learning, the quantifiable property of deep features can be obviously improved, and therefore, retrieval accuracy and time efficiency are greatly improved.
Owner:TSINGHUA UNIV

Image processing method and device

The invention discloses an image processing method and device. The method comprises the following steps: determining a first clustering center corresponding to a first feature vector; wherein the first feature vector is a feature vector of a query sample, the first clustering center is obtained according to a product quantization distance between the first feature vector and the second clusteringcenter, the second clustering center is a clustering center of each feature sub-vector in N feature sub-vectors included in the second feature vector, and the second feature vector is a feature vectorof a sample library; calculating a Hamming distance of a second clustering center of the first clustering center; determining a sample set with high similarity with the query sample from a sample library according to the Hamming distance; determining a target sample from the sample set according to the product quantization distance of the first feature vector and the third feature vector; whereinthe target sample is the sample with the highest similarity with the query sample, and the third feature vector is the feature vector of each sample in the sample set. By adopting the method, the query efficiency of the query sample can be improved.
Owner:SHENZHEN INTELLIFUSION TECHNOLOGIES CO LTD

Compression method and system used for neural network language model (NN LM)

The invention discloses a compression method and system used for a neural network language model (NN LM). The method includes: inputting training data into the neural network language model for pre-training; respectively carrying out base decomposition and clustering quantization operations on word vector matrices of input and / or output of the language model to compress the word vector matrices; and inputting the training data again into the neural network language model after compression of the word vector matrices to finely tune the language model. The invention provides a novel and efficient structured word embedding framework based on product quantization, the framework is used for compressing the input / output word vector matrices, and a significant memory reducing rate can be obtainedin a case of not damaging NN LM performance.
Owner:AISPEECH CO LTD

Medical image retrieval method based on deep learning and Radon conversion

The invention discloses a medical image retrieval method based on deep learning and Radon conversion, and relates to the field of computer vision and image retrieval. In the crude retrieval stage, a BING target suggestion algorithm is adopted for detecting a region with a remarkable object, a partial mean value Pooling is introduced into deep convolution network architecture, region-based remarkable differentiation characteristics can be extracted, the characteristic dimensions are reduced, and then polymerization is performed to form global characteristic expression. In the characteristic vector quantization process, and a product quantization algorithm is used for solving the problem that characteristic vector similarity measurement calculation is high in complexity. In the fine retrieval stage, Radon conversion is used for performing integral projection on the image at multiple angles, more detailed information characteristics of the image can be obtained, and the Top 50 images obtained in the crude retrieval are subjected to Radon conversion to generate Radon bar codes, and more accurate retrieval is achieved through similarity measurement. The accuracy of the medical image retrieval is improved, and the medical image retrieval problems that the characteristic differentiation is not high and the characteristic dimension is high due to direct use of a convolution neural network are solved.
Owner:BEIJING UNIV OF TECH

Quick large-scale high-dimensional data retrieval method and system

The invention provides a method and system for approximate nearest neighbor retrieval of large-scale high-dimensional data based on product quantization and multi-reverse indexing. The method comprises the steps that binary codes corresponding to data to be retrieved are obtained based on a trained product quantization unit, wherein the binary codes are used for determining a clustering center nearest to the data to be retrieved; the binary codes are input into a multi-reverse indexing unit matched with the trained product quantization unit, and a set composed of data nearest to the data to beretrieved in a preset database is obtained; according to the distance between each piece of data in the set and the data to be retrieved, all the data in the set is sorted, and all the sorted data serves as retrieval result. The large-scale similarity retrieval method and system based on high-dimensional data can greatly improve the retrieval accuracy and time efficiency.
Owner:TSINGHUA UNIV

Image retrieval method, terminal and storage device

The invention discloses an image retrieval method, a terminal and a storage device. The method comprises the steps of extracting a target feature vector of a to-be-retrieved image; querying a trainedcoarse codebook table, and determining a plurality of target coarse clustering center points closest to the target feature vector and a target sample identification code corresponding to each target coarse clustering center point; querying the trained sub-codebook table by using the plurality of target coarse clustering center points to obtain a plurality of target sub-clustering center points; and performing query and calculation through the target feature vector, the plurality of target coarse clustering center points, the plurality of target sub-clustering center points, the target sample identification code and a pre-calculated distance table to obtain a preset number of retrieval results closest to the target feature vector, the distance table comprising the sample identification codeand a corresponding distance value. According to the method, the data processing amount is reduced to a great extent in a product quantization mode, and the search precision is ensured.
Owner:ZHEJIANG DAHUA TECH CO LTD

Face recognition method and device, terminal and storage medium

The invention provides a face recognition method and device, a terminal and a storage medium. The face recognition method comprises: acquiring a face image to be recognized, and extracting a first high-dimensional feature vector of the face image to be recognized according to a preset deep learning model; quantifying the first high-dimensional feature vector into a first low-dimensional integer code by using a product quantization algorithm; performing similarity matching on the first low-dimensional integer code of the to-be-identified face image and a pre-obtained second low-dimensional integer code of the face template image; and determining a target low-dimensional integer code most similar to the first low-dimensional integer code in the second low-dimensional integer code according to a similarity matching result, and taking the face template image corresponding to the target low-dimensional integer code as a target face image. In the face recognition process, the low-dimensionalinteger codes are used for carrying out similarity comparison, so that the comparison frequency between the low-dimensional integer codes and the low-dimensional integer codes is reduced, the response speed of the system is greatly increased, and the face recognition efficiency is improved.
Owner:PING AN TECH (SHENZHEN) CO LTD

Steel strip defect detection method and device

The invention provides a steel strip defect detection method and device, and the method comprises the following steps: obtaining a sample data set which comprises a plurality of sample product imageswith steel strip defects and a plurality of sample product images without steel strip defects; training a neural network through the sample data set to obtain a steel belt defect detection model, andadopting a QuantitNoise-added product quantization algorithm during training; acquiring an image of a to-be-detected product; and inputting the to-be-detected product image into the steel belt defectdetection model to judge whether a steel belt defect exists or not. The method can be suitable for defect detection of the rapidly rolled steel coil strip, and is high in detection efficiency, low inlabor cost and high in detection accuracy.
Owner:深兰智能科技(上海)有限公司

Intelligent recommendation method and device, computer equipment and readable storage medium

The invention relates to the technical field of big data, and discloses an intelligent recommendation method and device, computer equipment and a readable storage medium, and the method comprises thesteps: obtaining user information, and carrying out the characterization of the user information to obtain a user vector; calling a product quantization process to segment the user vector to obtain aplurality of sub-vectors, identifying the category to which each sub-vector belongs, and summarizing the categories to obtain a user category set; calling a minimum hash process to perform similaritycomparison on the user category set and each reference category set in a preset index library, and setting the reference category set of which the similarity exceeds a preset similarity threshold as atarget category set; and taking the associated information corresponding to the target category set as recommendation information. According to the method, the fineness and the accuracy of user vector category identification are improved, the operation efficiency of the server is improved, the matching speed between the user information and the reference information in the index database is increased, and the data calculation amount and the data storage amount are reduced.
Owner:CHINA PING AN LIFE INSURANCE CO LTD

Multi-modal image retrieval method and system for appearance patent

The invention discloses a multi-modal image retrieval method and system for an appearance patent. Firstly, feature extraction and fusion are carried out on multiple views of an appearance patent, thenfeature extraction is carried out on a text, information of multiple modes is comprehensively considered, and finally deep visual semantic embedding is carried out, so that a good retrieval effect can be achieved in a large-scale appearance design patent database; for a tree structure in an ANN, compact coding representation is not performed on data so that efficiency is not high. Calculation ofthe Hamming distance in the hash method is not an accurate distance calculation problem. According to the invention, distance coding product quantization is provided, in the coding process, data points are coded into series connection of subspace clustering indexes, the distance between each data point and a reconstructed coded representation of the data point is coded, and an effective compact coded representation of each datum is formed; and therefore, the retrieval efficiency and accuracy are improved.
Owner:GUANGDONG UNIV OF TECH

Product quantization method based on semi-supervised learning

The invention discloses a product quantification method based on semi-supervised learning, which is an improved algorithm based on a common Cartesian K-means algorithm, namely a semi-supervised Cartesian K-means algorithm. In the algorithm, a traditional minimum square loss function in a quantization step needs to be replaced with an optimal reverse prediction loss function. According to the traditional semi-supervised learning, the marked data can be directly used for model training, and different from a traditional semi-supervised learning model, the marked data can be used for model training only through Laplacian regularization.
Owner:JIANGNAN UNIV

Companion learning robot for children and self-learning method for early education system of companion learning robot

The invention discloses a companion learning robot for children and a self-learning method for an early education system of the companion learning robot. The self-learning method comprises the following steps: SA10, training a convolution neural network; SA20, carrying out feature vector extraction on an input image by adopting the convolution neural network; SA30, carrying out grouping quantization on the feature vector by adopting the product quantization technology; SA40, generating a standard alphabet according to the Imagenet data set; SA50, acquiring the image and type of unknown new things, extracting the feature vector of the image of the new things and carrying out grouping quantization, and looking up the matched character string of the new things in the standard alphabet, carrying out matching connection on the character string and type of the new things in an associative memory model, and thus the new things are learnt in the early education system; and SA60, acquiring theimage of the to-be-identified things, wherein the early education system carries out identification to obtain the type of the to-be-identified things. With the technical scheme, the learning for the new knowledge with the children can be realized, the joint competition can be realized, and thus the learning interest of the children is improved.
Owner:CENT SOUTH UNIV

Product information recommendation method and device, storage medium and computer device

The invention discloses a product information recommendation method and device, a storage medium and a computer device, relates to the technical field of information processing. The method can improvethe information recognition accuracy. The method comprises the steps of training and initializing a first product recommendation model according to crowd statistical information to obtain a trained first product recommendation model; training and initializing a second product recommendation model according to the product quantization information to obtain a trained second product recommendation model; constructing a product recommendation prediction model according to the first hidden layer of the trained first product recommendation model and the second hidden layer of the trained second product recommendation model; and obtaining product recommendation prediction information matched with the historical purchase information by utilizing the constructed product recommendation prediction model according to the historical purchase information of the user. The method and device are suitable for accurate pushing of products in the mobile internet and electronic commerce.
Owner:PING AN TECH (SHENZHEN) CO LTD

CBIR method based on improved PQ algorithm

The invention relates to a CBIR method based on an improved PQ algorithm, and belongs to the technical field of image processing. The method comprises the steps of extracting image depth features by improving a deep convolutional network; encoding and compressing the image feature data through an index retrieval module of an inverted index-based product quantization IVPQ algorithm adopting a non-linear retrieval ANN search strategy; obtaining a full-index database, generating indexes of the dynamic index database based on a Faiss framework, segmenting a data space of the full-index database through feature vector coding, quickly locking a certain subspace through Hamming distance rearrangement and then traversing during retrieval of a query picture, and outputting a retrieval image. According to the method, the dynamic retrieval of the index database is realized on the basis of the Faiss framework, and the high operation and maintenance cost generated for reconstructing the index database in the practical application occasion is avoided.
Owner:CHONGQING UNIV OF POSTS & TELECOMM

Feature similarity search implementation method and device, electronic equipment and storage medium

The invention provides a feature similarity search implementation method and device, electronic equipment and a storage medium, and the method comprises the steps: establishing a codebook in advance based on a product quantization algorithm, establishing a plurality of user-defined functions, and loading the codebook and the user-defined functions into a database system memory; aiming at all to-be-stored characteristics, establishing an index by adopting a characteristic storage function and a codebook, inserting the index into a database, and inserting other attributes into the database; establishing a distance lookup table by adopting a distance lookup table construction function and a codebook under the condition of receiving the query feature; filtering the features in the database byusing other attributes to obtain an index meeting a preset condition; and obtaining the similarity between the query feature and the index meeting the preset condition according to the distance lookuptable, and obtaining a feature search result based on the similarity. The feature similarity search implementation method facilitates improvement of integral performance of similarity search.
Owner:SHENZHEN INTELLIFUSION TECHNOLOGIES CO LTD

Automatic production line for sausage type food

The invention discloses an automatic production line for sausage type food. The automatic production line comprises a sausage conveyor line, wherein one end of the sausage conveyor line is fixedly equipped with a lifter I; one end of the lifter I is further fixedly equipped with a cooking line; one end of the cooking line is equipped with a lifter II; the lifter II is additionally designed with a turning function; one end of the lifter II is further fixedly equipped with an air-drying line; one end of the air-drying line is further fixedly equipped with a lifter III; one end of the lifter III is further fixedly equipped with a pre-cooling line; and one end of the pre-cooling line is fixedly equipped with a withdrawing rod device. According to the automatic production line for the sausage type food disclosed by the invention, industrial robots such as the sausage conveyor line, a lifting loading machine, the cooking line, the air-drying line, a pr-cooling line and the withdrawing rod device are applied on a large scale, the sausage type food can be cooked, smoked or roasted in a fully-automatic mode on the highly intelligent machine production line. The automatic production line for sausage type food improves production efficiency, so that the production quality is stable, product quantization is uniform and is more precise, labor power, material resources, financial resources and time are reduced, and artificial food pollution probability is reduced.
Owner:FOSHAN VEHICIAN LNTELLIGENT EQUIP CO LTD

Compression sensing video encoding and decoding method based on Gaussian mixture model (GMM)

ActiveCN105578183AImplement codec applicationImproved codec rate-distortion performanceDigital video signal modificationTime domainProduct quantization
The invention brings forward a compression sensing video encoding and decoding method based on a Gaussian mixture model (GMM). First of all, modeling is performed on a compression sensing video by use of the GMM, based on this, a GMM lossy compression method based on a product quantizer is designed, and a compression sensing video encoder and decoder is brought forward. Time-domain redundancy of the compression sensing video is eliminated by use of a DPCM difference coding technology, data redundancy is further removed by use of algorithm encoding, and output code streams are obtained for storage and transmission. According to the invention, the modeling is carried out on the compression sensing video with a random feature by use of the GMM, and at the same time, the time-domain redundancy of the compression sensing video is eliminated, such that system energy and calculation resources can be reduced, and the compression efficiency is improved. The method provided by the invention can satisfy application environments having energy and calculation complexity restrictions on a video encoding system, such as wireless multimedia sensing networks, space video obtaining, mobile terminal video obtaining and the like.
Owner:XI AN JIAOTONG UNIV

Index file generation method and device

The invention relates to an index file generation method and device. The method comprises the following steps of extracting a feature vector of each piece of training data in a training data set; carrying out product quantization processing on the feature vectors of the training data to obtain a class center of the training data set; generating an empty index file according to the class center ofthe training data set; sending the null index file to each task node of a cluster; and obtaining an index file returned by each task node based on the empty index file. According to the method and thedevice, the main node of the cluster generates the empty index file based on product quantification, the empty index file is deployed on each task node of the cluster, each task node processes the data to be processed to obtain the index file, and sends the index file to the main node, so that the method and the device can be suitable for the high-dimensional feature retrieval, and the implementation complexity is relatively lower.
Owner:ALIBABA (CHINA) CO LTD

Data processing device, data processing method, and recording medium

The invention relates to a data processing apparatus, a data processing method, and a recording medium, wherein feature vectors can be quantized efficiently. A data processing device according to an embodiment includes a sub-vector group generating unit, a codebook generating unit, and a converting unit. The sub-vector group generating unit generates, from a feature vector set of N number of D-dimensional feature vectors, M number of sub-vector groups (where M<D holds true). Each of the M number of sub-vector groups includes N number of dimension-variable sub-vectors obtained from the N numberof D-dimensional feature vectors. For each of the M number of sub-vector groups, the codebook generating unit performs clustering of the N number of dimension-variable sub-vectors, and generates a codebook in which the representative vector of each cluster is associated with an index. The converting unit performs product quantization using the codebook and converts each of the N number of D-dimensional feature vectors into a compressed code made of a combination of M number of indexes.
Owner:KK TOSHIBA +1

Adversarial sample generation method based on depth product quantization

The invention provides an adversarial sample generation method based on depth product quantization. The adversarial sample generation method comprises the steps of: inputting an original image into apreset network model, so as to output a quantitative distribution center vector corresponding to the original image through the preset network model; inputting an initial adversarial sample corresponding to the original image into the preset network model to obtain an adversarial feature vector corresponding to the initial adversarial sample; based on the quantitative distribution center vector and the adversarial feature vector, determining a loss function corresponding to the initial adversarial sample; and performing back propagation on the preset network model based on the loss function toobtain an adversarial sample corresponding to the original image. According to the adversarial sample generation method, the derivable loss function is determined on the basis of the quantitative distribution center vector and the adversarial feature vector, so that the mobility and effectiveness of the adversarial sample are improved, and a basis is provided for further researching the robustness of the neural network.
Owner:PENG CHENG LAB +1

Product quantization search method and device, terminal and storage medium

The invention is suitable for the technical field of computers, and provides a product quantization search method and device, a terminal and a storage medium. The product quantization search method specifically comprises the following steps: inputting a reference high-dimensional feature of a reference object into a feature compression network to obtain a reference low-dimensional feature output by the feature compression network, wherein the loss function of the feature compression network is a function obtained based on a high-dimensional neighbor relation of a sample object and a low-dimensional neighbor relation of the sample object; determining a plurality of clustering centers by using the reference low-dimensional features; obtaining a target feature of the target object; and performing product quantization search by using the target feature and the plurality of clustering centers to obtain a reference object closest to the target object. According to the embodiment of the invention, the precision of product quantization search can be improved.
Owner:SHENZHEN INTELLIFUSION TECHNOLOGIES CO LTD

Defect detection algorithm based on product quantization learning strategy

The invention discloses a defect detection algorithm based on a product quantization learning strategy, which is an improved algorithm based on a common Cartesian K-means algorithm, namely a semi-supervised Cartesian K-means algorithm. The method mainly adopts a mode comprising the following three stages: stage 1, determining positions and types of defects, and adopting an optimal reverse prediction algorithm as a detection means; stage 2, carrying out positioning detection on defect positions in structural images detected in the stage 1 also by adopting a mode based on target detection; and stage 3, on the basis of the stage 2 cutting detection images, and sending the cut detection images to a classification model for classification so as to determine final classification and identification results. According to the method, quantization errors of each subspace can be effectively reduced, and the recognition performance is improved.
Owner:WUXI XINJIE ELECTRICAL

Large-scale picture infringement detection method and system based on feature selective aggregation

The invention provides a large-scale picture infringement detection method and system based on feature selective aggregation, and belongs to the field of picture infringement detection. According to the method, a deep residual convolutional neural network pre-trained on a large-scale picture data set is used for extracting high-level semantic feature maps of pictures, a selective feature aggregation algorithm SCDA is used for selectively aggregating the feature maps, and a principal component analysis PCA is used for compressing features obtained through aggregation into a low-dimensional dense feature vector for representation. And an index is constructed for the feature vectors in a product quantization mode so as to realize large-scale picture infringement detection. The method is high in efficiency, and millisecond-level searching in million-level vectors can be achieved.
Owner:浙江阿蚂科技有限公司

Semi-supervision-based central product quantitative retrieval method

The invention provides a semi-supervision-based center product quantization image retrieval method, which comprises the following steps of: segmenting a space into a plurality of subspaces after feature extraction, then carrying out normalization processing on a feature vector, and calculating a cosine distance to enable the subvector to find the closest code word in the corresponding subspace. In the calculation process, a semi-supervised loss module is used for reducing quantization errors, minimizing empirical errors of marked data and embedding errors of unmarked data, finally, code words are used for replacing sub-vectorization to form binary codes, the binary codes are stored in a product quantization lookup table, and image retrieval is carried out through asymmetric distance calculation; the method provided by the invention has higher robustness, and the image retrieval precision is improved.
Owner:XIAMEN UNIV OF TECH

Product quantification method based on fuzzy clustering and asymmetric distance calculation

The invention is suitable for the field of computer vision and big data, and provides a fuzzy clustering and asymmetric distance calculation-based product quantization method, which comprises the following steps of: firstly, carrying out initial clustering on an original vector set, calculating a residual vector set, and carrying out fuzzy clustering on residual sub-vectors to obtain a membership matrix; carrying out sub-vector classification according to the membership degree; carrying out weighted distance calculation on the query vector according to a corresponding mode, and counting the sum of the distance ranking serial numbers of each sub-segment; according to the method, a fuzzy clustering method is applied to product quantization index subspace vector clustering, so that errors caused by non-uniform sample category distribution in hard division are avoided, and vector clustering is more objective; weighted asymmetric distance calculation and distance ranking statistical methods are adopted in similarity measurement, and retrieval errors caused by overlarge calculation of a certain distance of a vector are avoided. Compared with an original product vector method, the method provided by the invention has higher precision in retrieval of complex background pictures.
Owner:中船凌久高科(武汉)有限公司

Image retrieval method, device and storage medium

The disclosure relates to an image retrieval method, device and storage medium, and belongs to the field of data retrieval. The method includes: obtaining the graph index structure corresponding to the image feature posting table, based on the target image feature and the graph index structure, obtaining the nearest neighbor image feature of the target image feature from the image feature posting table, and obtaining the nearest neighbor image feature corresponding Candidate image set; Based on the residual feature codebook, the residual feature between the target image feature and the nearest neighbor image feature is quantized by product to obtain the coding of the residual feature. The residual feature codebook adopts the optimal product quantization method The product is obtained by quantizing the sample residual features of a plurality of first sample images; based on the encoding of the residual features and the encoding of the candidate residual features of the candidate images in the candidate image set, the candidate image is obtained from the candidate image set as an image retrieval result . In this way, the accuracy of image retrieval can be improved on the premise of ensuring the efficiency of image retrieval.
Owner:BEIJING DAJIA INTERNET INFORMATION TECH CO LTD

Artificial intelligence system and method for semantic retrieval

This application relates to systems and methods for semantic retrieval. The system can perform the following methods: obtain query words from the user terminal; use a pre-generated model to transform the query words into query word vectors; use a product quantization algorithm to retrieve a group of POI vectors from the retrieval library according to the query word vectors; And in response to the query, at least one POI is determined based on the set of POI vectors for recommendation to the user.
Owner:DITU (BEIJING) TECH CO LTD

A kind of children's learning robot and its early education system self-learning method

The invention discloses a self-learning method for a learning companion robot for children and its early education system. The self-learning method includes: step A10, training a convolutional neural network; step A20, using a convolutional neural network to extract feature vectors from input images; and step A30 , using the product quantization technique to group and quantify the feature vectors; step A40, generate a benchmark alphabet according to the Imagenet dataset; step A50, obtain the images and categories of unknown new things, extract the feature vectors of the new thing images and quantify them in groups, and Find the matching new thing string in the alphabet; match and connect the new thing string with the category in the associative memory model, so as to learn new things into the early education system; step A60, obtain the image of the thing to be recognized, and the early education system recognizes The category of things to be recognized. The invention can learn new knowledge and compete with children together, and improve children's learning fun.
Owner:CENT SOUTH UNIV

A Method of Image Feature Point Matching

The invention discloses a method for matching image feature points, which comprises the following steps: extracting feature points of stored pictures: extracting features of stored images to form a stored feature vector, and performing dimension reduction on its dimension; vector storage: segmentation and dimension reduction The final input feature vector, and perform product quantization on each part after division, and then vector quantization to form a product quantizer and a vector quantizer, and establish a retrieval tree and hash table; feature point extraction of images to be matched: extraction to be Match the features of the image and form the feature vector to be matched, and reduce its dimension; Vector matching: segment the feature vector to be matched after dimensionality reduction, and find out the cluster center distance between the feature vector to be matched and the product quantizer and vector quantizer For multiple cluster centers in the front, find the pictures corresponding to multiple cluster centers according to the search tree and hash table and form a candidate set, and use floating point vectors to calculate the picture with the closest distance between the candidate set and the feature vector to be matched; Fast speed and high precision.
Owner:UNIV OF ELECTRONICS SCI & TECH OF CHINA

Deep neural network calculation acceleration method and device

The embodiment of the invention provides a deep neural network calculation acceleration method and device, a terminal and a computer readable storage medium. The method includes the following steps: sampling each input vector that needs to be input into a matrix model to obtain multiple sampling vectors; performing product quantization on each sampling vector according to a preset quantization parameter to obtain a plurality of quantization points; dividing the matrix model into a plurality of matrix blocks according to the quantization parameter; performing calculation on each quantization point and each matrix block to obtain a plurality of pre-calculation tables; and calculating each input vector through each pre-calculation table to obtain a calculation result of the matrix model. According to the embodiment of the invention, the pre-calculation table of the same matrix model only needs to be established once, and all input vectors that need to be calculated by the matrix model canuse the pre-calculation table to perform table lookup calculation, so that the calculation process of the input vectors and the matrix model can be effectively saved, and meanwhile, an original calculation effect of the matrix model can also be maintained.
Owner:BEIJING BAIDU NETCOM SCI & TECH CO LTD
Who we serve
  • R&D Engineer
  • R&D Manager
  • IP Professional
Why Patsnap Eureka
  • Industry Leading Data Capabilities
  • Powerful AI technology
  • Patent DNA Extraction
Social media
Patsnap Eureka Blog
Learn More
PatSnap group products