52results about How to "Avoid error propagation" patented technology

The invention discloses a time sequence processing method and device for a distributed target InSAR with enhanced space-time coherence. The method comprises the steps of carrying out threshold separation on a permanent scatterer (PS) and a distributed scatterer (DS) on the basis of an obtained single complex image data stack; performing phase stability analysis on separated PS candidate points bymeans of a StaMPS approach to obtain PS points; carrying out covariance matrix estimation on DS candidate points one by one through a maximum likelihood estimator and refining a covariance matrix through a bootstrapped coherence estimator; performing time sequence phase optimization on each DS candidate point; selecting DS points; constructing a two-layer network for the selected PS and DS points;carrying out phase error separation on an unwrapping phase after three-dimensional phase unwrapping is performed on the PS points and the DS points; utilizing the maximum likelihood estimator and a phase model to estimate geophysical parameters. The method can reduce the uncertainty of InSAR time sequence analysis caused by coherence loss and an ill-conditioned matrix, improve the quality controlof time sequence data processing and thus improve the estimation precision and reliability of products.

The invention discloses an end-to-end context-based knowledge base question and answer method and device, and the method and device consider the connection relationship between entities and relationships in a knowledge base, enable the two tasks including entity links and relationship prediction contained in the knowledge base to be mutually promoted, and jointly improve the two tasks. The methodmainly comprises the steps of prreprocessing a problem and removing special symbols; constructing a candidate entity set related to the problem based on the knowledge base, and constructing a candidate relationship set according to the relationship of the candidate entities in the knowledge base; for each entity in each candidate set, extracting the context of the entity in the question; dividingthe candidate relationships into different granularities; predicting a subject entity and a predicate relationship based on the CERM model; and finding the object entity in the knowledge base as an answer to return by using a predicted subject entity and relationship. Entity links and relationship prediction in knowledge base questions and answers are integrated into a unified prediction model. Joint prediction of the subjective entities and relationships is achieved, and accuracy of questions and answers is improved.

The invention provides a knowledge graph-based patient medical record similarity evaluation method and system. The method comprises the steps of obtaining and preprocessing patient medical record textdata; carrying out entity identification and entity relationship extraction on the preprocessed data by adopting a joint information extraction model based on weak supervised learning, and expressingan obtained medical entity type by using a knowledge vector; constructing triples in the knowledge graph according to the obtained entity relationship, and representing the triples by using a joint knowledge vector; combining the knowledge vectors into the same semantic space by utilizing the joint vector, and judging the similarity by calculating the semantic distance between entities; accordingto the method, important knowledge in the electronic medical record is mined and extracted by using the bidirectional recurrent neural network for multiple times, the subject domain knowledge graph concept is expanded to the medical field, entity identification and relationship extraction are performed under the current situation that the medical entity type is relatively limited, and the accuracy of similarity evaluation is improved.

The invention discloses a method and a device for precoding in a multi-user MIMO (multiple input multiple output) system. The method includes receiving pilot signals of users; computing spatial correlation matrixes according to the received pilot signals, and computing guiding matrixes of expected users according to the received pilot signals and pilot signals transmitted by the users; and computing precoding matrixes of the expected users according to inverses of the spatial correlation matrixes and the guiding matrixes. The method and the device have the advantages that computational complexity of precoding can be effectively reduced, and precoding also can be effectively carried out under the condition of co-channel interference which stops acquisition of channel information.

The invention discloses an online measurement device and measurement method for the NO concentration in a high-NO2 and low-NO atmosphere. The online measurement device comprises an ozone generation unit, a connection pipe cleaning unit and a sampling pipe, wherein a connection reaction bin, an NO2 measuring instrument, a second mass flow meter and a first air pump are sequentially connected behind the sampling pipe; and the connection pipe cleaning unit comprises an electromagnetic three-way valve, an activated carbon tube, a current-limiting hole and a second air pump. By adopting a cavity-enhanced absorption spectroscopy, a sampling mode (actual atmosphere of ozone is added) or a zero point mode (actual atmosphere) is set through the electromagnetic three-way valve; absorption spectrum intensity signals of two modes are measured through an NO2 measuring instrument as a sampling spectrum and a reference spectrum respectively, and then NO in the atmosphere is converted into a concentration value of the NO2 through spectral fitting, that is to say, measurement of the NO concentration in the atmosphere environment is achieved, so that the online measurement device and measurement method have the characteristics of being free of NO2 concentration interference, a low detection limit, high sensitivity, high temporal-spatial resolution, high stability and low cost.

The invention discloses a three-dimensional reconstruction system calibration method based on multi-frequency structured light. The calibration method comprises the following steps of step S1, initializing equipment; step S2, preparing a calibration board; step S3, acquiring a calibration image; step S4, calibrating a camera; step S41, determining kp0; step S42, determining kp1 and kp2; step S43,determining kp3 and kp4; step S44, searching the remaining of small circles; step S5, calibrating a projector; step S51, generating a G matrix by using a calibration board white-light chart; step S52,unwrapping phase; step S521, acquiring a grating projection chart; step S522, unwrapping the phase; step S523, iteratively computing to obtain the finest stripe absolute phase; step S53, solving parameter; and step S6, ending. The calibration method provided by the invention has the advantages of being simple in calibration process, high in accuracy, strong in practicability, convenient to use and low in cost.

The invention discloses an intelligent question answering method based on depth learning, comprising the following steps: step 1, collecting original conversation data; Step 2, data preprocess, 3, construct a Seq2seq model, establishing an encode layer (Encoder) and a decoding lay (Decoder), and connecting that Encoder layer and the Decoder layer to obtain the Seq2seq model; 4, model prediction, wherein that model prediction is a Seq2seq model construct according to the above, and aft the Seq2seq model is trained with the original conversation data, the problem data is taken as input, and themodel automatically generates comment data. Through deep learning, the problem is directly mapped to the answer, so as to optimize the whole question answering method, avoid the problem of error transmission, and greatly reduce the complexity of the system.

The invention discloses a quality map weighting based multi-baseline least-square phase unwinding method. According to the method, a first-order phase gradient principal value in the winding phase ofa multi-baseline InSAR is calculated, and is extended in mirror symmetry, a second-order phase gradient is calculated, a quality map is used to determine the quality of the winding phase, the windingphase as prior information is used to carry out phase compensation on residual error points in the winding phase, the second-order phase gradient of the winding phase is optimized and weighted on thebasis of the quality map, and multi-baseline least square phase unwinding is carried out to obtain the unwinding phase. Multi-baseline measurement enriches interference phase information of the targetscene, the problem in transferring the phase error of a residual error point in a traditional single-baseline least square phase unwinding algorithm is solved, and the phase unwinding precision is improved.

The invention relates to the field of face image processing, in particular to a non-inductive three-dimensional face reconstruction method and an acquisition reconstruction system. According to the reconstruction method provided by the invention, a circular ring type double-circle-center calibration plate is used for calibration, an implicit function method is used for global fusion and curved surface reconstruction, and the three-dimensional curved surface model is subjected to smooth processing and optimization, so that the precision of a final output image is greatly improved, and the requirement on equipment is also reduced. Meanwhile, a three-stand-column type instrument structure is adopted, installation is convenient, the size of the instrument is reduced, the stability of the instrument is improved, the designed structure is simpler and more convenient through modularization, rapid production is facilitated, and the cost is greatly reduced; according to the invention, the target human face is acquired through the three acquisition units, and fine matching of homonymous points is realized, so that the resolution of a three-dimensional object is increased, the reconstructionprecision of three-dimensional data is improved, and the speed of human face acquisition and three-dimensional reconstruction is also improved.

The invention relates to a method for extracting an event argument in a text and electronic equipment, and the method comprises the steps: detecting an event type contained in a to-be-processed text, and obtaining a target event type; constructing and obtaining an argument extraction problem according to the target event type and the prior association information of the corresponding target argument role; splicing the to-be-processed text and the argument extraction problem to obtain a target text; the representation vector of the target text is input into a machine reading understanding model, the model comprises two layers of classifiers which are arranged in sequence, the first layer of classifier synchronously identifies and predicts whether an answer exists in the target text or not and identifies and predicts the starting position and the ending position of the answer, and a position prediction result carrying an answer indication label is obtained; the second-layer classifier predicts whether the prediction entity is an event argument or not after the start position and the end position are paired and combined, and an answer prediction result is obtained; and outputting the event argument according to the answer prediction result and the corresponding label.

The invention provides a biomedical event extraction method based on a replication mechanism, belonging to the technical field of natural language processing. The biomedical event extraction method based on the replication mechanism comprises the steps of constructing a model input vector; constructing an Encoder module using a bi-directional LSTM model; constructing a Decoder module based on an Attention mechanism and a replication mechanism to recognize triggers and elements at the same time. The method of the invention can effectively avoid the disadvantages of cascading error and independence between subtasks caused by the staged method, and error transmission caused by the simple use of shared parameters in the joint model, thereby improving the performance of biomedical event extraction.

The invention relates to the field of remote sensing image processing, and in particular to an MODIS-based PM2.5 remote sensing inversion method. The method comprises the following steps of: obtainingan MODIS image and PM2.5 monitoring data at the same time; interpolating the PM2.5 data into a PM2.5 interpolation image; constructing a training set and a test set; using the training set to train amachine learning algorithm, using a trained model to the test set, and calculating a performance index, on the test set, of the model; repeating the steps S3 and S4 to obtain a plurality of performance indexes so as to select an optimum model; and using the optimum model to the whole MODIS image so as to obtain a PM2.5 inversion result of the whole MODIS image. The method starts from data of remote sensing images and directly establishes a relationship between the remote sensing images and measured PM2.5 through the machine learning algorithm, so that error transfer is avoided and inversion results with higher precision are achieved. The method has the effects of avoiding error transfer and being high in inversion precision.

The invention discloses a fine-grained numerical information extraction method based on a joint learning model. The fine-grained numerical information extraction method comprises the following steps:preprocessing an input text; candidate numerical values are recognized according to the numerical value correlation regular expression, and the numerical values serve as trigger words and are splicedwith front and back lexical examples of the trigger words to serve as input of a joint learning model; accessing the bottom layer of the joint learning model to a word embedding layer, and adding theposition information of a numerical trigger word to the output of the word embedding layer to obtain the final distributed representation of each lexical example; accessing a subsequent feature extraction network based on distributed representation, and respectively accessing networks of specific tasks at a high level: accessing a classification network to obtain a semantic type of numerical information, and accessing a sequence labeling network to obtain a plurality of semantic roles related to numerical trigger words; combining outputs of specific tasks to obtain a numerical information unit; for a plurality of numerical information units, the composite relationship between the numerical information units is judged based on a statistical method. The method does not need artificial designfeatures, and is high in accuracy and strong in generalization ability.

The invention discloses a classification method and device for multi-level classification objects, and relates to the technical field of computers. One specific embodiment of the method comprises thefollowing steps: constructing a cascade classification model step by step in a joint training mode by utilizing sample feature data of multi-stage classification objects, and a construction process: constructing a first-stage cascade classification model according to a first-stage sub-classifier; when K is larger than or equal to 2 and smaller than or equal to N, the sample feature data of the multi-level classification objects and K-1-level classification information, output by the K-1-level cascade classification model, of the multi-level classification objects are jointly input into a Kth-level sub-classifier, so that the K-1-level cascade classification model and the Kth-level sub-classifier are jointly trained, and a K-level cascade classification model is obtained; and classifying the to-be-classified data of the multi-stage classification object by utilizing the final N-stage cascade classification model to obtain N-stage classification information. According to the embodiment,the inter-class hydrophilicity and hydrophobicity are considered, the classification effect is good, error transfer can be avoided, the classification accuracy is improved, the overall model complexity is low, and the model development difficulty and workload are reduced.

The invention relates to a differential decoding method in a short-range wireless network, belonging to the technical field of wireless communication. Firstly, the characteristic of cycle repetition of a transmission pilot frequency sequence is utilized to carry out front and back conjugation correlation and accumulation on a received pilot frequency digital signal, thus a reference differential correlation value is obtained; the local spread spectrum modulation sequence is subject to the same operation of front and back conjugation and accumulation so as to obtain a conjugation correlation value; the conjugation correlation value and the reference differential correlation value are related to obtain a complex number correlation value; and according to the characteristics of positive number and negative number of the real part of the complex number correlation value, data decoding is realized. The differential decoding method of the invention lowers decoding calculated amount and improves data decoding performance under the influence of frequency shift, so that a receiver has lower power dissipation and the receiving performance under bigger frequency shift, thus avoiding the error code transmission problem of the traditional differential decoding.

The embodiment of the invention discloses an end-to-end quality enhancement method and an end-to-end quality enhancement device based on a binocular stereo image. The method comprises the steps of respectively inputting low/high-quality images in the binocular stereo images into a feature extraction network for feature extraction; respectively inputting the obtained shallow feature maps of the low/high-quality images into an information distillation network for information distillation; and then inputting the obtained high-level feature map of the low/high-quality image into an information fusion network for information fusion, finally inputting the obtained fusion feature map into a long short-term memory network to learn the visual difference of the three-dimensional image, and reconstructing to obtain a quality enhanced image of the low-quality image. According to the invention, an end-to-end image quality enhancement algorithm is adopted. High-quality images are used for guiding low-quality image reconstruction. The visual difference of the three-dimensional images is learned through the long-term and short-term memory network based on information fusion. The operation speed can be increased. The error transmission can be avoided. The speed and accuracy of image reconstruction are guaranteed, and the consumption of internal storage and computing resources is reduced.

The invention discloses an integrated multi-user detection method and system based on signal virtual processing, wherein the method comprises the steps: calculating the channel ambiguity by a common receiver according to the channel state information corresponding to a transmitter of a to-be-detected signal and the modulation mode of the transmitter, and carrying out the descending sort of the channel ambiguity; if the maximum ambiguity of the channel is not zero, detecting the to-be-detected signal which does not exceed the number of antennas of a receiver in a zero-forcing mode, and eliminating the detected signal from the received mixed signal; if the maximum ambiguity of the channel is zero, enabling all the signals to be detected to be equivalent to a structured equivalent signal, then determining a symbol combination possibly carried by the signals to be detected, recovering a data symbol carried by the signals sent by the transmitter with the maximum channel gain in a zero forcing mode, and taking the data symbol as a reference symbol; and determining a symbol combination carried by the signals to be detected. Therefore, multi-user detection is realized by using a small number of antennas, and the detection efficiency is improved.

The invention provides a voice processing method and device, a storage medium and a program product. The method comprises the following steps: determining corresponding feature information of a to-be-processed voice on a plurality of frequency bands; for a sequence formed by the feature information on the plurality of frequency bands, based on a deep learning model used for processing sequence data, obtaining a processing result corresponding to each piece of feature information; and obtaining the processed voice based on the processing result corresponding to each piece of feature information. According to the invention, denoising and dereverberation can be realized at the same time based on the deep learning model, an error transmission phenomenon caused by series connection of different algorithm modules is avoided, the accuracy of the model is improved, in addition, the network model coefficient of each frequency band is shared, the size of the network model and the calculation amount during processing can be effectively reduced, and the processing efficiency is improved.