163results about How to "Rich semantic information" patented technology

The invention provides a machine reading understanding method based on a multi-head attention mechanism and dynamic iteration, and belongs to the field of natural language processing. The constructionmethod of the machine reading understanding model comprises the following steps: constructing an article and problem coding layer; constructing a recurrent neural network based on bidirectional attention flow; a self-attention layer is constructed and answer output is predicted based on a dynamic iterative decoder. According to the method, answer prediction can be carried out on questions in a machine reading understanding task text; according to the invention, a new end-to-end neural network model is established, and a new idea is provided for exploring a machine reading understanding task.

The invention discloses a real-time gesture interaction method based on computer vision, which comprises the following steps: acquiring a digital image from an image input device; detecting hands by a statistical learning method; initializing a hand contour tracker according to the detecting result, and calculating a skin color model of a specific hand; tracking the hand contour by a combined method of a conditional probability density transmission algorithm and a heuristic search technology; analyzing the moving track of the hands by a Hidden Markov Model to obtain the gesture identifying result; and applying the gesture analyzing result to the interaction of various application programs. The real-time gesture interaction method of the invention expands the interactive mode of the traditional mouse and keyboard, realizes automatic hand detection, tracking and gesture identification by the computer vision and image processing techniques, has real-time performance, robustness and easy realization and operation, and can enable computer users to interact with the computer more naturally, visually and intelligently by hand gestures.

The invention relates to a patent literature similarity measurement method based on ontology, and relates to the technical field of natural language information processing for the ontology. The method comprises the following steps: extracting a core technical scheme according to the structural features, the position features and the keyword features of patent literatures; constructing a model for the relation between thematic terms of patent classes; constructing a field dictionary according to the model for the relation between the thematic terms of the patent classes and segmenting terms and removing stop terms for the core technical scheme; extracting keywords and weight by combining the relation between the thematic terms to TF-IDF as TextRank term initial weight; training a FastText model, and generating a term vector; and calculating an EMD distance to obtain a semantic distance according to keywords, term weight and term vector. Compared with the prior art, the patent literature similarity measurement method based on the ontology solves the problem that the similarity is low due to the fact that the structural features, the field features, the term relation features and the semantics approximate expression of the patent literature are not fully considered.

The invention provides a remote sensing image semantic segmentation method based on regional attention multi-scale feature fusion. The remote sensing image semantic segmentation method comprises the following steps of: S1, constructing a network model for a remote sensing image semantic segmentation network; S2, constructing a training data set, and preprocessing the collected data set for training; and S3, inputting the data set for training into a network model for training, and predicting a result after acquiring training parameters. According to the remote sensing image semantic segmentation method, the idea of an image cascade network is introduced, and the model parameter quantity is greatly reduced; meanwhile, coding features and decoding features are optimized by using an attentionmechanism, a regional attention module and a multi-scale group fusion module are constructed, feature maps of different scales are extracted and fused, training is guided by using multi-scale semantic tags and boundary tags, and the model performance is effectively improved under the condition that the parameter quantity of the model is only 8.4 M.

The invention discloses a semantic search method based on multi-semantic analysis and personalized sequencing, and belongs to the field of information search. The semantic search method adopts the technical scheme comprising the following steps: firstly, by a crawler technology and other technologies, acquiring webpage documents from the Internet, classifying the webpage documents by using a support vector machine, establishing a word vector library by a multi-semantic analysis method, and writing multi-classification results into an index to form an index library; secondly, based on the word vector library, forming search keywords input by a user into a query vector, performing class matching query with the index library to obtain an initial sequencing result; and finally, according to personalized information and history access information of the user, optimizing the initial sequencing result, and returning the optimized result to the user. By the semantic search method based on the multi-semantic analysis and the personalized sequencing, the word vector library and the index library with rich semantemes are formed; and through the personalized information and the history access information, a search result can meet a search demand of the user better and search satisfaction of the user can be improved.

The invention discloses a text classification method based on feature information of characters and terms. The method comprises the steps that a neural network model is utilized to perform character and term vector joint pre-training, and initial term vector expression of the terms and initial character vector expression of Chinese characters are obtained; a short text is expressed to be a matrixcomposed of term vectors of all terms in the short text, a convolutional neural network is utilized to perform feature extraction, and term layer features are obtained; the short text is expressed tobe a matrix composed of character vectors of all Chinese characters in the short text, the convolutional neural network is utilized to perform feature extraction, and Chinese character layer featuresare obtained; the term layer features and the Chinese character layer features are connected, and feature vector expression of the short text is obtained; and a full-connection layer is utilized to classify the short text, a stochastic gradient descent method is adopted to perform model training, and a classification model is obtained. Through the method, character expression features and term expression features can be extracted, the problem that the short text has insufficient semantic information is relieved, the semantic information of the short text is fully mined, and classification of the short text is more accurate.

The invention provides an automatic text abstraction method based on a pre-training language model. According to the method, an ultra-large-scale unsupervised Chinese corpus is used for training a complex deep language model; the low-layer network structure of the model can extract and retain the grammar and structure information of a text, and the high-layer network structure can extract and retain the semantics and context information of the text, so that the richer text features and semantic information are provided for an automatic text summary task; a pre-training language model and an Encoder are combined to realize, the text features and the semantic information in the pre-training language model are fully utilized, so that a better semantic compression effect is provided, and the performance of an automatic text abstract is improved; a pre-training language model and a decoder are combined, not only the semantics in an original text are considered in the text generation process, but also the semantic information of vocabularies is also considered, so that the readability of the generated text and relevance with the original text are improved, and the performance of an automatic text abstract is improved.

The invention discloses a Chinese image semantic description method combined with multilayer GRU based on a residual error connection Inception network, and belongs to the field of computer vision andnatural language processing. The method comprises the steps: carrying out the preprocessing of an AI Challenger image Chinese description training set and an estimation set through an open source tensorflow to generate a file at the tfrecord format for training; pre-training an ImageNet data set through an Inception_ResNet_v2 network to obtain a convolution network pre-training model; loading a pre-training parameter to the Inception_ResNet_v2 network, and carrying out the extraction of an image feature descriptor of the AI Challenger image set; building a single-hidden-layer neural network model and mapping the image feature descriptor to a word embedding space; taking a word embedding characteristic matrix and the image feature descriptor after secondary characteristic mapping as the input of a double-layer GRU network; inputting an original image into a description model to generate a Chinese description sentence; employing an evaluation data set for estimation through employing the trained model and taking a Perplexity index as an evaluation standard. The method achieves the solving of a technical problem of describing an image in Chinese, and improves the continuity and readability of sentences.

The invention discloses a high-resolution remote sensing image weak target detection method based on deep learning. For a remote sensing image with low resolution, a small target size and fuzzy quality, the method comprises the following steps: firstly, improving the resolution of an image by adopting a WGAN-based super-resolution reconstruction method; inputting the image with the enhanced quality into a target detection framework; carrying out deep feature extraction on the image by using a residual network; fusing the extracted low-level features with the extracted high-level features; it is ensured that the fused multi-layer feature map has rich detail information and also contains high-level semantic information; and carrying out region-of-interest coarse extraction on the feature mapby using the fused multi-layer features and the region suggestion network, mapping the extracted region to the same dimension by using a region-of-interest alignment method, and carrying out subsequent target accurate classification and position refinement to obtain a final target detection result. According to the method, the weak and small target detection precision and recall rate under the conditions of low remote sensing image resolution and complex background are effectively improved.

The invention provides a cross-modal hash retrieval method based on triple deep networks. The method is used to solve the technical problem of low retrieval precision existing in existing cross-modalhash retrieval methods, and includes the realization steps of: preprocessing data, and dividing the data into training data and query data; acquiring hash codes of image training data and text training data; using triple supervisory information to establish an objective loss function; carrying out orderly iterative optimization on the objective loss function; calculating hash codes of image querydata and text query data; and acquiring retrieval results of the query data. According to the solution provided by the invention, the triple information is used to construct the objective loss function, semantic information is increased, an intra-modal loss function is added at the same time, discriminability of the method is improved, and precision of cross-modal retrieval can be effectively improved. The method can be used for Internet-of-things information retrieval and image and text mutual-searching services of e-commerce, mobile equipment and the like.

The invention discloses a short text automatic abstracting method and system based on double encoders, belongs to the technical field of information processing, and is characterized by comprising thefollowing steps: 1, preprocessing data, 2, designing a double encoder with a bidirectional recurrent neural network, and 3, arranging an attention mechanism fusing global and local semantics 4, arranging a decoder with empirical probability distribution and using a decoder designed by adopting a double-layer unidirectional neural network; 5, adding word embedding characteristics, 6, optimizing word embedding dimensions, and 7, carrying out preprocessing and testing on the news corpus data from the Sogou laboratory, substituting the news corpus data into a Seq2Seq model with double encoders andaccompanying empirical probability distribution to carry out calculation, and carrying out experimental evaluation through a text abstract quality evaluation system Rouge. According to the invention,traditional weaving is carried out; and the decoding framework is subjected to optimization research, so that the model can fully understand text semantics, and the fluency and precision of text abstracts are improved.

The invention discloses an increment neural network and sub-graph code based image classification method. The increment neural network and sub-graph code based image classification method comprises the following steps of: extracting local characteristics; studying increment code books in a network; performing characteristic coding based on a sub-graph; getting image spaces together; and studying a classifier and predicting a model. The increment neural network and sub-graph code based image classification method greatly reduces time complexity of a traditional algorithm to a great extent as the code book can be studied efficiently, and the space relationship between visual words is also kept; in addition, the characteristic coding based on the sub-graph is specifically for performing characteristic coding by fully utilizing the space relationship between the visual words, so that more abundant semantic information can be extracted, and excellent classification performances are obtained finally while the computational efficiency of a classification system is increased. Therefore, the increment neural network and sub-graph code based image classification method has a relatively high use value.

The invention relates to a multi-modal sentiment analysis method based on the quantum theory. The method comprises the steps of constructing a multi-modal sentiment corpus set; selecting a training set and a test set, and pre-processing the training set and the test set respectively; extracting respective characteristics from pre-processed texts and images, and constructing text density matrices and image density matrices respectively; inputting the text density matrix and the image density matrix in the training set into a random forest classifier to obtain a text sentiment classification model and an image sentiment classification model; inputting the text matrix and image matrix of the test set corpus into the text and image sentiment classification models to classify the sentiment categories and calculating respective prediction probabilities thereof; and subjecting the text prediction probabilities and the image prediction probabilities to weighted fusion by using a multi-modal decision-making fusion method to finally calculate the classification accuracy of each multi-modal sample.

The invention discloses a Chinese electronic medical record named entity recognition method and system based on an attention mechanism, and belongs to the field of text information mining. The technical problem to be solved by the invention is how to identify named entities in an electronic medical record more accurately and conveniently based on a neural network and an attention mechanism. According to the technical scheme, the method comprises the following steps: S1, obtaining word vector and part-of-speech vector representation of Chinese word part-of-speech and splicing the word vector and the part-of-speech vector; S2, splicing the word vector and the part-of-speech vector, and inputting the spliced word vector and part-of-speech vector into a Double-LSTMs neural network model for feature extraction to obtain more accurate implicit strata vector representation; S3, adding an attention layer, and endowing relatively important information in the text with a higher weight; S4, endowing the weight with a hidden layer vector obtained by corresponding forward encoding and a hidden layer vector obtained by reverse encoding, and respectively splicing the hidden layer vectors to serveas feature vectors; and S5, carrying out sequence labeling based on the conditional random field model to realize an identification task of the named entity.

The invention relates to a fashion garment image segmentation method based on depth learning. The fashion garment segmentation method based on depth learning comprises the following steps of the construction of a depth neural network garment model, the loss function design of reverse error propagation and a model training strategy, wherein the depth neural network garment model comprises a featureextraction module, a garment semantic information extraction module and a garment segmentation prediction module, the loss functions comprise a regression function of a key point position, a key visibility loss function, a cross entropy loss function of a garment prediction category with weights and a regression loss function of a garment position; and the model training strategy comprises a weight parameter initialization method, the data preprocessing, an optimization algorithm and a training step. The method has the advantages of being able to automatically segment and recognize the upperbody clothing, lower body clothing and whole body clothing collocation in complex images, and being conducive to the deep learning and network training for fashion clothing design.

The invention discloses a wheeled robot semantic mapping method and system fusing point cloud and images, solves the problem that a map constructed by laser SLAM only supports robot navigation, constructs a map embedded with semantic annotations, and belongs to the field of mobile robot SLAM. The main content of the method is that semantics are extracted from a two-dimensional point cloud, and semantics is stored, updated and optimized in a two-dimensional grid map. The method mainly comprises the steps of target detection based on deep learning on an image, point cloud segmentation, boundingbox and segmentation matching, map unit semantic updating and semantic optimization based on SLAM global optimization and clustering. The method has the advantages of being real-time, simple in device, rich in map information and the like, and aims at achieving intelligent navigation and man-machine interaction of the indoor mobile robot.

The invention provides a visual question-answering method based on structured semantic representation. The visual question-answering method comprises the following steps: extracting image features ofan input image through a convolutional neural network; extracting a word vector of each word from an input problem related to the input image through a pre-trained character embedding model; performing weighting processing on the image feature and the word vector to obtain weighted image feature vectors and text feature vectors; converting the word vector into a structured semantic representationvector through a Tree-LSTM network; performing fusion processing on the image feature vector, the text feature vector and the structured semantic representation vector to obtain a corresponding fusionfeature vector; and taking the fusion feature vector as the input of a prediction model, and outputting an answer corresponding to the input question by the prediction model. According to the method,richer semantic information is extracted from the question, and the performance of the prediction model is improved through multi-layer training optimization so that the accuracy of answers is improved.

The invention discloses a personalized scenic spot recommendation method based on tourist preference modeling, and the method comprises the steps: collecting data, carrying out the preprocessing, and carrying out the numbering of tourists, scenic spots and other objects; converting the display score into an implicit score, and dividing a positive case scenic spot and a negative case scenic spot; constructing a triple and scenic spot knowledge map, and generating a feature vector and a context feature vector of each scenic spot; generating vector representations of historical tourist tour scenic spots and candidate scenic spots through the KCNN; calculating an influence weight of each historical touring scenic spot of the tourist through the attention network to obtain a preference vector of the tourist to the scenic spot; calculating the scenic spot touring probability of the tourists by using the DNN, and generating scenic spot recommendation lists of the tourists according to the probability from small to large. According to the method, when different influences of historical visiting scenic spots of tourists on the candidate scenic spots are depicted and diversification preferences of the tourists are represented, the attention network is used for calculating the influence weights of the historical visiting scenic spots of the tourists on the candidate scenic spots, so that the recommendation result better conforms to the preferences of the tourists.

The invention discloses a sow side-lying posture real-time detection system based on joint partitioning of sow key parts and environment, and the system comprises a delivery room, a camera, a video storage unit, and a server, and the delivery room is used for placing a to-be-delivered sow; the camera monitors and obtains video data of a delivery room, continuously stores the video data to the video storage unit on one hand, and is directly connected with the server on the other hand; the server calls the backup video data and analyzes the monitoring data in real time; the working steps of thedetection system are as follows: monitoring the postures of the sows in real time, simultaneously detecting three areas of an approval area, a lactation area and a delivery area through a convolutional neural network area identification model, identifying the sows as lying postures when more than two areas are simultaneously detected, and outputting an identification result to a database. Comparedwith a method for recognizing the posture of the sow through a sensor technology, the computer vision technology avoids contact with the sow, stress response is reduced, and the method has the advantages of being low in cost and high in efficiency.

An irregular character recognition device based on deep learning comprises a memory for storing instructions and character images, a processor which is configured to execute the instruction so as to input an original character image containing irregular characters to the neural network model, and identify and outputt characters in the character image, wherein the neural network model comprises a correction sub-network used for correcting an original character image containing irregular characters and outputting the corrected character image, and an identification sub-network which is connectedwith the output end of the correction sub-network and is used for receiving the character image output by the correction sub-network and identifying characters in the character image.

The invention provides a model training method and device, an image super-resolution processing method and device, a terminal and a storage medium, and relates to the technical field of model training. The method comprises the following steps: carrying out downsampling on an original high-resolution image corresponding to a sample low-resolution image to obtain a high-resolution image comprising multiple resolutions of the original high resolution; respectively adopting a plurality of feature extraction branches to carry out feature extraction on the sample low-resolution image to obtain imagefeatures of a plurality of levels; a feature fusion module is adopted to perform fusion processing on the image features of the multiple levels to obtain fusion features of the sample low-resolutionimage; reconstructing the fused features by adopting a plurality of reconstruction branches to obtain super-resolution images with a plurality of resolutions; and training the neural network model according to the high-resolution images with the plurality of resolutions and the corresponding super-resolution images. The low-resolution image is recovered through the neural network model, semantic information contained in the generated super-resolution image is richer, and the definition is higher.

The invention provides a micropapilla detection system based on YOLOv5. The micropapilla detection system comprises a lung adenocarcinoma pathological image acquisition module, a detection model training module and a micropapilla detection module; a feature extraction module in the detection model training module adds a layer of feature extraction on the basis of a BottleneckCSP network structure to obtain a shallow feature map, so that the multiple of downsampling of the feature extraction module is reduced by half; the feature fusion module is used for down-sampling the shallow feature map and splicing the shallow feature map with the deep feature map, so that the features of the shallow feature map and the deep feature map are fused, and the small target detection capability of the network is enhanced; and the prediction module replaces the GIoU loss with the CIoU loss, optimizes the prediction frame, improves the regression speed and precision of the prediction frame, and is more suitable for the detection of micronipples.

The invention discloses an intelligent semantic matching method and device based on a depth feature dimension changing mechanism, and belongs to the technical field of artificial intelligence and natural language processing. The technical problem to be solved by the invention is how to capture more semantic context information and interaction information between sentences. Intelligent semantic matching of sentences is realized; the adopted technical scheme is as follows: the method comprises the following steps: constructing and training a sentence matching model consisting of an embedding layer, a depth feature variable-dimension coding layer, a convolution matching layer and a prediction layer; according to the method, deep feature variable-dimension coding representation of the sentences is realized, so that more semantic context information and interaction information between the sentences are obtained, and meanwhile, a convolution matching mechanism is realized, so that the purpose of intelligent semantic matching of the sentences is achieved. The device comprises a sentence matching knowledge base construction unit, a training data set generation unit, a sentence matching model construction unit and a sentence matching model training unit.

The invention relates to a neural machine translation method based on Multi-BiRNN encoding. Multi-BiRNN encoding is adopted at an encoder end, that is to say that on the basis of using source languagesentences as input sequences, one or more groups of BiRNN are added to encode other input sequences associated with the input sequences; on the basis of Multi-BiRNN-encoded neural machine translation, in the encoding process of a source end, the source language sentence sequences and a dependency syntax tree thereof are considered at the same time, and serialization results of the syntax tree areobtained by means of two different traversal modes respectively and serve as Multi-BiRNN encoding input with the source language sentence sequences; at the output end of each BiRNN, a word is formedin a vector splicing mode. According to the method, vectors obtained by encoding contain more abundant semantic information, the source language sentence sequences and other sequences associated withthe source language sentence sequences are considered at the same time, and the disambiguation function is achieved in the semantic representation process of the source language sentences.

The invention belongs to the technical field of computers, and particularly relates to a multi-level hybrid vulnerability automatic mining method, which comprises the following steps: S1, obtaining pointer distribution and assignment processes in a program by utilizing pointer analysis through a static analysis technology, and recording memory address positions pointed by all pointers; obtaining adestination operand of the CMP type instruction by using an alias analysis technology so as to clarify a jump address of the target program; and S2, realizing instrumentation of the monitoring code to the key position of the target program through an intermediate code static instrumentation technology. According to the invention, the binary executable program is translated into the middle-layer representation language of the compiler, and compared with the binary executable program, the middle-layer representation language source code obtained through processing of the method has more semantic information; compared with C / C + + language analysis, the static analysis technology for the LLVM IR intermediate representation language can realize finer-grained static semantic information extraction during compiling.

The present invention relates to method and device for text-enhanced knowledge graph joint representation learning, the method at least comprises: learning a structure vector representation based on entity objects and their relation linking in a knowledge graph and forming structure representation vectors; discriminating credibility of reliable feature information and building an attention mechanism model, aggregating vectors of different sentences and obtain association-discriminated text representation vectors; and building a joint representation learning model, and using a dynamic parameter-generating strategy to perform joint learning for the text representation vectors and the structure representation vectors based on the joint representation learning model. The present invention selective enhances entity / relation vectors based on significance of associated texts, so as to provide improved semantic expressiveness, and uses 2D convolution operations to train joint representation vectors. As compared to traditional translation models, the disclosed model has better performance in tasks like link prediction and triad classification.

The invention discloses an intelligent semantic matching method and device based on deep hierarchical coding, andd belongs to the technical field of artificial intelligence and natural language processing. The technical problem to be solved by the invention is how to capture more semantic context information and interaction information between sentences to achieve Intelligent semantic matching ofsentences; the adopted technical scheme is as follows: the method comprises the following steps: constructing and training a sentence matching model consisting of an embedding layer, a deep hierarchical coding representation layer, a hierarchical feature interaction matching layer and a prediction layer; according to the method, deep hierarchical coding representation of the sentences is realized,so that more semantic context information and interaction information between the sentences are obtained, and a hierarchical feature interaction matching mechanism is realized to achieve the purposeof intelligent semantic matching of the sentences. The device comprises a sentence matching knowledge base construction unit, a training data set generation unit, a sentence matching model construction unit and a sentence matching model training unit.

The invention discloses a pedestrian re-recognition method combining posture and attention based on a double-flow network. The method comprises the following steps of: 1, preprocessing an input imageand inputting the input image into a double-flow network to extract features; and 2, combining middle-layer features and high-layer features and associating global information through an attention mechanism. And 3, fusing the attention flow and the attitude estimation flow through bilinear pooling operation to obtain a final feature map. And 4, model training: training neural network parameters byusing a back propagation algorithm. The invention provides a neural network model for pedestrian re-recognition, particularly provides a network structure combining an attention mechanism and attitude estimation based on a double-flow network, and obtains an effect of competitive power in the field of pedestrian re-recognition at present.

The invention relates to an image saliency target detection method. Region segmentation of image analysis is involved. The method is an image saliency detection method based on multi-graph model priorand short connection network optimization. The method comprises the following steps of: utilizing color and position information for each input image; calculating a KNN graph model and a K regular graph model; obtaining a saliency map S1 under the KNN graph model and a saliency map S2 under the K regular graph model; carrying out pixel-level fusion on the KNN graph model and the K regular graph model; S3, obtaining an initial saliency map of the original image; and S3, optimizing the initial saliency map by using a short connection network to obtain a final saliency map Sfinal of the originalimage, and completing image saliency target detection, thereby overcoming the defects of incomplete saliency target detection and inaccurate algorithm detection when foreground and background colorsare similar in the prior art of image saliency target detection.

The invention provides an online cross-modal retrieval method and system using a three-step strategy. The online cross-modal retrieval method comprises the following steps: acquiring analog stream data formed by different modals; for analog stream data, generating the representation of each class label by introducing a hadamard matrix, the representation of each class label is used as global information for learning a hash code, and the representation of each class label also keeps local similarity information, using the correlation between newly arrived data in the analog stream data and existing data to learn a Hash code with more discriminative power; updating a hash function by using the learned hash code; and calculating a hash code of the to-be-retrieved sample by using the updated hash function, and calculating a Hamming distance of a binary sample based on the hash code, thereby returning a sample of another modal similar to the to-be-retrieved sample according to the Hamming distance. According to the method, the THOR can reserve more semantic information and learn more accurate hash codes.