2091 results about "Semantic feature" patented technology

The invention provides an unsupervised domain-adaptive brain tumor semantic segmentation method based on deep adversarial learning. The method comprises the steps of deep coding-decoding full-convolution network segmentation system model setup, domain discriminator network model setup, segmentation system pre-training and parameter optimization, adversarial training and target domain feature extractor parameter optimization and target domain MRI brain tumor automatic semantic segmentation. According to the method, high-level semantic features and low-level detailed features are utilized to jointly predict pixel tags by the adoption of a deep coding-decoding full-convolution network modeling segmentation system, a domain discriminator network is adopted to guide a segmentation model to learn domain-invariable features and a strong generalization segmentation function through adversarial learning, a data distribution difference between a source domain and a target domain is minimized indirectly, and a learned segmentation system has the same segmentation precision in the target domain as in the source domain. Therefore, the cross-domain generalization performance of the MRI brain tumor full-automatic semantic segmentation method is improved, and unsupervised cross-domain adaptive MRI brain tumor precise segmentation is realized.

The invention discloses a human behavior recognition method integrating the space-time dual-network flow and an attention mechanism. The method includes the steps of extracting moving optical flow features and generating an optical flow feature image; constructing independent time flow and spatial flow networks to generate two segments of high-level semantic feature sequences with a significant structural property; decoding the high-level semantic feature sequence of the time flow, outputting a time flow visual feature descriptor, outputting an attention saliency feature sequence, and meanwhile outputting a spatial flow visual feature descriptor and the label probability distribution of each frame of a video window; calculating an attention confidence scoring coefficient per frame time dimension, weighting the label probability distribution of each frame of the video window of the spatial flow, and selecting a key frame of the video window; and using a softmax classifier decision to recognize the human behavior action category of the video window. Compared with the prior art, the method of the invention can effectively focus on the key frame of the appearance image in the originalvideo, and at the same time, can select and obtain the spatial saliency region features of the key frame with high recognition accuracy.

The invention discloses a short text classification method based on a convolution neutral network. The convolution neutral network comprises a first layer, a second layer, a third layer, a fourth layer and a fifth layer. On the first layer, multi-scale candidate semantic units in a short text are obtained; on the second layer, Euclidean distances between each candidate semantic unit and all word representation vectors in a vector space are calculated, nearest-neighbor word representations are found, and all the nearest-neighbor word representations meeting a preset Euclidean distance threshold value are selected to construct a semantic expanding matrix; on the third layer, multiple kernel matrixes of different widths and different weight values are used for performing two-dimensional convolution calculation on a mapping matrix and the semantic expanding matrix of the short text, extracting local convolution features and generating a multi-layer local convolution feature matrix; on the fourth layer, down-sampling is performed on the multi-layer local convolution feature matrix to obtain a multi-layer global feature matrix, nonlinear tangent conversion is performed on the global feature matrix, and then the converted global feature matrix is converted into a fixed-length semantic feature vector; on the fifth layer, a classifier is endowed with the semantic feature vector to predict the category of the short text.

Initially, a client requests a specific document and provides the preferences, including readability level preferences of the document, locale preferences, content filtering instructions preferences, governmental regulations preferences, natural language preferences, and document syntactic format preferences. The transcoding proxy requests and receives the document from the origin server, with the document having origin semantic characteristics. The document from the origin server has an origin readability level and origin locale, is conformant with origin content filtering instructions and origin governmental regulations, and is in origin natural language and in origin document syntactic format. Using the client semantics preferences, the transcoding proxy revises the document in a sequential or parallel fashion. The origin semantics characteristics of the document are, thus, revised to the semantic preferences specified by the client.

he invention discloses a medical information extraction system and method based on depth learning and distributed semantic features. The system is composed of a pretreatment module, a linguistic-model-based word vector training module, a massive medical knowledge base reinforced learning module, and a depth-artificial-neural-network-based medical term entity identification module. With a depth learning method, generation of the probability of a linguistic model is used as an optimization objective; and a primary word vector is trained by using medical text big data; on the basis of the massive medical knowledge base, a second depth artificial neural network is trained, and the massive knowledge base is combined to the feature leaning process of depth learning based on depth reinforced learning, so that distributed semantic features for the medical field are obtained; and then Chinese medical term entity identification is carried out by using the depth learning method based on the optimized statement-level maximum likelihood probability. Therefore, the word vector is generated by using lots of unmarked linguistic data, so that the tedious feature selection and optimization adjustment process during medical natural language process can be avoided.

A system and method for determining image similarity. The method includes the steps of automatically providing perceptually significant features of main subject or background of a first image; automatically providing perceptually significant features of main subject or background of a second image; automatically comparing the perceptually significant features of the main subject or the background of the first image to the main subject or the background of the second image; and providing an output in response thereto. In the illustrative implementation, the features are provided by a number of belief levels, where the number of belief levels are preferably greater than two. The perceptually significant features include color, texture and / or shape. In the preferred embodiment, the main subject is indicated by a continuously valued belief map. The belief values of the main subject are determined by segmenting the image into regions of homogenous color and texture, computing at least one structure feature and at least one semantic feature for each region, and computing a belief value for all the pixels in the region using a Bayes net to combine the features. In an illustrative application, the inventive method is implemented in an image retrieval system. In this implementation, the inventive method automatically stores perceptually significant features of the main subject or background of a plurality of first images in a database to facilitate retrieval of a target image in response to an input or query image. Features corresponding to each of the plurality of stored images are automatically sequentially compared to similar features of the query image. Consequently, the present invention provides an automatic system and method for controlling the feature extraction, representation, and feature-based similarity retrieval strategies of a content-based image archival and retrieval system based on an analysis of main subject and background derived from a continuously valued main subject belief map.

The invention discloses a Chinese question-answering system based on a neural network, which comprises a user interface module, a question word pre-segmentation module, a nerve cell pre-tagging module, a learning and training module, a nerve cell knowledge base module, a semantic block identification module, a question set index module and an answer reasoning module. The system comprises the steps of: firstly adopting an SIE encoding mode to encode the in-vocabulary words of the semantic block according to corresponding position, later converting an identification problem of the question semantic block into a tagging classification problem, and then adopting a classification model based on the neural network to determine the semantic structure of the question, and finally combing the semantic structure of the question to realize the question similarity computation based on the neural network and comparing the weight of various semantic features of the question by extracting the tagged semantic features of the question, thereby providing a basis for final answer reasoning. The Chinese question-answering system integrates the syntax, the semantics and the contextual knowledge of the question and can simulate the process that human beings process the sentence.

The invention discloses a multi-task layered image retrieval method based on a depth self-coding convolution neural network. The method is characterized by mainly comprising a multi-task end-to-end convolution neural network for deep learning and training recognition, a rapid visual segmentation detection and positioning method of a region-of-interest secondary screening module based on an RPN network, a coarse search of a full-graph sparse hash code, an area sensing semantic feature and matrix h accurate comparison and search based on the maximum response, and a region-of-interest selectivitycomparison algorithm. According to the method, the end-to-end training can be achieved, the interest region with higher quality can be automatically selected, the automation degree and the intelligent level of search by images can be effectively improved, and the image retrieval requirements of the big data age can be met by using little storage space at a high search speed.

The invention provides a thyroid ultrasound image nodule automatic diagnosis system based on a multi-scale convolutional neural network. The system includes a thyroid nodule coarse-to-fine classification module, a thyroid nodule region automatic detection module, and a thyroid nodule fine classification module. The size features of different sensing regions are extracted through a multi-scale feature fusion convolutional neural network, and then, the context semantic features of a thyroid nodule can be extracted according to local and global information, and the thyroid nodule can be automatically located. Through multi-scale coarse-to-fine feature extraction based on a neural network and the design of a multi-scale fine classification AlexNet of a pyramid structure, the position of a focus and the probability that the focus is benign or malignant can be accurately predicted, doctors can be assisted in diagnosing a thyroid focus, and the objectivity of diagnosis is improved. The systemhas the characteristics of good real-time performance and high accuracy.

One embodiment of the present method, program, and apparatus for natural language generation enables a language generation system to generate a grammatically correct natural language sentence by retrieving and adapting one or more stored sentences having semantic features similar to semantic features of a system-generated semantic representation. The retrieval of stored sentences is guided at least in part by the ease with which the stored sentences may be adapted to produce a grammatically correct sentence, enabling rapid generation of accurate output.

The invention discloses a short text clustering method based on deep semantic feature learning. The method includes the steps that dimensionality reduction representation is performed on original features under the restraint of local information preservation through traditional feature dimensionality reduction, binarization is performed on an obtained low-dimension actual value vector, and error back propagation is performed with the binarized vector being supervisory information of a convolutional neural network structure to train a model; non-supervision training is performed on a term vector through an outer large-scale corpus, vectorization representation is performed on all words in text according to the word order, and the vectorized words serve as implicit semantic features of initial input feature learning text of the convolutional neural network structure; after deep semantic feature representation is obtained, a traditional K-means algorithm is adopted for performing clustering on the text. By means of the method, extra natural language processing and other specialized knowledge are not needed, design is easy, deep semantic features can be learnt, besides, the learnt semantic features have unbiasedness, and good clustering performance can be achieved more effectively.

The invention provides a Chinese sentiment analysis method based on BERT, LSTM and CNN fusion. The Chinese sentiment analysis method comprises the steps: performing text preprocessing on a plurality of Chinese corpora in a Chinese corpus data set to obtain a plurality of sequences corresponding to the plurality of Chinese corpora; extracting word embedding of each sequence by using a BERT model; performing feature extraction on each sequence by adopting BERT, LSTM and CNN to obtain text deep semantic features corresponding to each sequence; and classifying the obtained text deep semantic features by using a softmax classifier to train and test the model so as to realize sentiment polarity prediction analysis. According to the technology, the defects in the prior art can be overcome, and the accuracy of Chinese text sentiment analysis is improved.

The invention discloses a method for the classification of natural scene images on the basis of regional potential semantic feature, aiming at carrying out the classification of the natural scene images by utilizing the regional potential semantic information of the images and the distribution rule of the information in space. The technical proposal comprises the following steps: firstly, a representative collection of the classification of the natural scene images is established; secondly, sampling point SIFT feature extraction is carried out to the images in the representative collection of the classification of the natural scene images to generate a general visual word list; thirdly, the regional potential semantic model of an image is produced on the representative collection of the classification of the natural scene images; fourthly, the extraction of the regional potential semantic feature of the image is carried out to any image; finally, a natural scene classification model is generate, and classification is carried out to the regional potential semantic feature of the image according to the natural scene classification model. The method inducts the regional potential semantic feature, thus not only describing the regional information of image sub-blocks, but also including the distribution information of the image sub-blocks in space; compared with other methods, the method of the invention can obtain higher accuracy, and no manual labeling is needed, thus having high degree of automation.

Document comparisons can be performed at a semantic level by utilizing a rules base in which groups of rules are applied sequentially. In one implementation, (1) syntactic rules are applied to a document to form a tagged sequence in which individual words are tagged with their syntactic categories, (2) ambiguity rules are applied to the tagged sequence to resolve ambiguities, thereby providing a resolved tag sequence, (3) grammar rules are applied to the resolved tagged sequence to determine semantic roles of individual tagged words, thereby providing a role-specific resolved tagged sequence, and (4) property rules are applied to match properties (e.g., adjectives) with the words they modify, thereby providing a semantic feature structure. The semantic feature structure is then compared to at least one other structure.

The invention discloses a video content description method based on semantic information guidance. The method comprises the steps that (1) a video format is preprocessed; (2) semantic information for guidance is established; (3) the weight of each semantic feature vector [A,XMS<(i)>] is calculated; (4) the semantic feature vectors [A,XMS<(i)>] are decoded; and (5) a video description model is tested. According to the method, by use of a faster-rcnn model, key semantic information on each frame of an image can be quickly detected, and the key semantic information is added into original features extracted through a CNN, so that the feature vector input into an LSTM network at each time node has semantic information; thus, in the decoding process, video content space-time relevancy is guaranteed, and language description accuracy is improved.

A system and method for determining image similarity. The method includes the steps of automatically providing perceptually significant features of main subject or background of a first image; automatically providing perceptually significant features of main subject or background of a second image; automatically comparing the perceptually significant features of the main subject or the background of the first image to the main subject or the background of the second image; and providing an output in response thereto. In the illustrative implementation, the features are provided by a number of belief levels, where the number of belief levels are preferably greater than two. The perceptually significant features include color, texture and / or shape. In the preferred embodiment, the main subject is indicated by a continuously valued belief map. The belief values of the main subject are determined by segmenting the image into regions of homogenous color and texture, computing at least one structure feature and at least one semantic feature for each region, and computing a belief value for all the pixels in the region using a Bayes net to combine the features. In an illustrative application, the inventive method is implemented in an image retrieval system. In this implementation, the inventive method automatically stores perceptually significant features of the main subject or background of a plurality of first images in a database to facilitate retrieval of a target image in response to an input or query image. Features corresponding to each of the plurality of stored images are automatically sequentially compared to similar features of the query image. Consequently, the present invention provides an automatic system and method for controlling the feature extraction, representation, and feature-based similarity retrieval strategies of a content-based image archival and retrieval system based on an analysis of main subject and background derived from a continuously valued main subject belief map.

The invention discloses a short text classification method based on a topic word vector and a convolutional neural network, which comprises the following steps: 1) a data acquisition stage: acquiringshort text data according to requirements, and labeling the short text data as a training set; 2) a data preprocessing stage: performing word segmentation, stop word removal, useless text filtering and the like on the text; 3) representing short text features, namely respectively representing a theme level and a word vector level; 4) carrying out subject term vector joint training; 5) optimizing and iterating parameters of the convolutional neural network classification model; and 6) performing category prediction on the new sample. According to the invention, short text data characteristics are combined; in the feature representation stage, a topic vector and a word vector are combined for representation; semantic feature expansion is carried out on the data characteristics of the short text, text semantic information is further mined by utilizing the local sensitive information extraction capability of the convolutional neural network in the classification model training stage, and indexes such as short text classification task category prediction accuracy can be improved.

Trajectory information of objects appearing in a scene can be used to cluster trajectories into groups of trajectories according to each trajectory's relative distance between each other for scene activity analysis. By doing so, a database of trajectory data can be maintained that includes the trajectories to be clustered into trajectory groups. This database can be used to train a clustering system, and with extracted statistical features of resultant trajectory groups a new trajectory can be analyzed to determine whether the new trajectory is normal or abnormal. Embodiments described herein, can be used to determine whether a video scene is normal or abnormal. In the event that the new trajectory is identified as normal the new trajectory can be annotated with the extracted semantic data. In the event that the new trajectory is determined to be abnormal a user can be notified that an abnormal behavior has occurred.

A method semantically labels an image acquired of a scene by first obtaining a local semantic feature for each local region in the image. The local semantic features are combined recursively to form intermediate segments until a semantic feature for the entire image is obtained. Then, the semantic feature for the entire image is decombined recursively into intermediate segments until an enhanced semantic feature for each local region is obtained. Then, each local region is labeled according to the enhanced semantic feature.

The invention provides a Mongolian-Chinese machine translation method for enhancing semantic feature information based on a Transformer model. The method comprises the following steps: firstly, starting from the language characteristics of Mongolian, finding out the characteristics of the additional components of the Mongolian in terms of stem, affixes and lattices, and merging the language characteristics into the training of a model; secondly, distributed representation for measuring the similarity between the two words is taken as a research background, and the influence of depth, density and semantic coincidence degree on the concept semantic similarity is comprehensively analyzed; in the translation process, a Transformer model is adopted, and the Transformer model is a multi-layer encoder which performs position encoding by using a trigonometric function and is constructed on the basis of an enhanced multi-head attention mechanism. A decoder architecture, which completely dependson the mechanism of attention to draw the global dependency between the input and the output, eliminates recursion and convolution.

An initial image is subjected to a plurality of segmentations into uniform components by using various criteria to identify and isolate several series of objects constituted by segmented thumbnail images associated with geometrical descriptors for boxes that cover said objects, the results of the segmentation methods are fused, numerical values are used to characterize each object thumbnail image that results from fusion, and each of the thumbnail images is compared with indexed thumbnail images in an indexed database containing a set of textual descriptions for everyday elementary objects. The list of textual descriptions of everyday elementary objects matching the object thumbnail images that result from the segmentation and fusion operations is provided together with pertinence levels.

The invention discloses a cross-age face verify method based on characteristic learning; the method comprises the following steps: 1, obtaining to-be compared two face images; 2, using a face characteristic point positioning method to carry out align operation for the two face images; 3, respectively carrying out feature extraction for each image, wherein the extraction method includes the following steps: a, automatically extracting high-level meaning characteristics through a depth convolution nerve network; b, calculating LBP histogram characteristics of the image; c, fusing the characteristics obtained in the a and b steps, and expressing characteristic vectors; 4, using a cosine similarity method to calculate a distance between the characteristic vectors obtained by the step3, and determining whether the two images are from a same person or not. The method firstly uses the depth network to cross-age face verify, and creatively combines the handwork design LBP histogram characteristics with depth network autonomous learning characteristics, thus realizing complementation between high-rise meaning characteristic and lower characteristics, and providing better accuracy.

The invention discloses a target tracking method and system of a convolution twin network based on multi-layer feature fusion. The method comprises the following steps of: according to the target position and size of the images, cutting out the target template images and the search area images of all images in the image sequence training set, and forming a training data set by image pairs composedof the target template images and the search area images; constructing a convolution twin network based on multi-layer feature fusion; training the convolution twin network based on the multi-layer feature fusion based on the training data set to obtain a trained convolution twin network based on the multi-layer feature fusion; using the trained convolution twin network based on multi-layer feature fusion for target tracking. In the process of tracking targets, the invention integrates scores of different layers, combines high-level semantic features and bottom-level detail features, better distinguishes the interference of similar or similar targets, and prevents the problems of target drift and target loss in the tracking process.

The invention belongs to the technical field of multimedia information retrieval, and specifically discloses a deep learning-based freehand sketch image retrieval method. According to the method, and edge contour detection technology and a non-maximum value suppression technology are utilized to realize the conversion from colored images to similar sketch images, a deep learning technology is utilized to construct distinguishing feature expressions for querying deep features of sketches and similar sketches, the deep features fuse the high-level semantic features and low-level visual features of images, and the deep features are more distinguishing in sketch retrieval. Through deeply mining visual information of a first retrieval result, uncorrelated images placed at the front in the retrieval result are rejected and a more correlated result is returned to the users. The method is high in correctness and strong in adaptability. On the basis of large-scale image data, the method is significant in carrying out efficient image retrieval which considers semantic information of sketches, so that the influences of fuzziness of freehand sketches can be decreased, the retrieval correlation can be improved and the user experience can be enhanced; and the method has an extensive application value in the field of multimedia image retrieval.

The invention relates to a knowledge graph and discloses a knowledge graph representation learning method for integrating text semantic features based on an attention mechanism. The method solves theproblems that semantic features are insufficient due to the fact that a translation model does not utilize description texts of entities and relations, semantic features cannot be fused into entitiesand relations at the same time by a multi-source information embedding method, and the text extraction effect is poor. The method comprises the steps of firstly obtaining and processing description texts of entities and relationships to obtain text semantic features of the entities and the relationships, then constructing a projection matrix of the entities by utilizing the semantic features of the entities and the relationships, projecting entity vectors into a relationship space, modeling in the relationship space by utilizing a translation thought, and carrying out representation learning,so as to model a many-to-many complex relationship. The method is suitable for representation learning of the knowledge graph.

The invention provides a natural scene text detection method and system. The system comprises two neural network models: a text detection network based on multi-level semantic feature fusion and a detection screening network based on an attention mechanism, wherein the text detection network is an FCN-based image feature extraction fusion network, and is used for extracting multi-semantic level information of input data, carrying out full fusion of multi-scale features, and finally predicting the position and confidence degree of text information in a natural scene by carrying out convolutionoperation on the fused multi-scale information. According to the detection screening network, the trained convolutional recurrent neural network is used for discriminating and scoring an initial detection result output by the convolutional neural network of the first part, so that a background which is easy to confuse with foreground characters is filtered out, and the accuracy of natural scene text recognition is further improved.

A given set of videos are sequenced in an aesthetically pleasing manner using models learned from human curated playlists. Semantic features associated with each video in the curated playlists are identified and a first order Markov chain model is learned from curated playlists. In one method, a directed graph using the Markov model is induced, wherein sequencing is obtained by finding the shortest path through the directed graph. In another method a sampling based approach is implemented to produce paths on the digraph. Multiple samples are generated and the best scoring sample is returned as the output. In a third method, a relevance based random walk sampling algorithm is modified to produce a reordering of the playlist.

The invention discloses a method and system for news event extraction based on neural network. The method comprises the steps of: conducting data pre-processing on original text of training corpus; introducing an event sentence sequence represented by a word vector into a bidirectional long and short memory network, using the bidirectional long and short memory network to train and obtain semantic features of each candidate trigger word; introducing the event sentence sequence represented by a word vector into a convolutional neural network, using the convolutional neural network to train and obtain global features of the event sentences where the candidate trigger words are in; according to the semantic features of candidate trigger words and the global features of the event sentences where the candidate trigger words are in, using softmax as a classifier to classify each candidate trigger word, and therefore finding out the trigger words for news events, and according to the trigger word type, judging the type of the event. The method and system can quickly and accurately extract news events and deal with news events contained in non-standard statements, and has the advantages of high efficiency and universal applicability.

The invention discloses a text classification method based on a bidirectional cyclic attention neural network, and belongs to the technical field of learning and natural language processing. The method comprises the following steps: step 1, preprocessing data; Step 2, according to the preprocessed data, generating and training a word vector of each word through a Word2vec method; Step 3, performing text semantic feature extraction on the word vector according to the word vector, fusing an attention mechanism and a bidirectional recurrent neural network, calculating the overall weight of each word, and converting the weight into an output value Y (4) of a model; And step 4, taking the feature vector Y (4) as the input of a softmax classifier according to the feature vector Y (4), and carrying out classification identification. According to the method, the attention mechanism is fused in the text feature learning model, the effect of keywords can be effectively highlighted, the performance of the model is greatly improved, and the text classification accuracy is further improved.

The invention discloses an individualized recommending method and system based on an element path. The recommending method comprises the following steps: building a heterogeneous network graph through collecting article content information, social relationship among users and user historical behavior information; utilizing a relevance measurement method based on heterogeneous bi-directional random walk to measure the preference degrees of a user on articles based on semantics; utilizing machine learning to synthesize the preference degrees of the user under different semantics and predict the preference degrees of the user on unselected articles so as to recommend favorable articles to the user. According to the system, the accuracy, novelty and diversity of the recommending effect are effectively improved through comprehensively utilizing the additive information of other article types to a target article type based on the content, cooperation filtering and the social network relationship and digging the semantic features of the information.