35results about How to "Improve precision and recall" patented technology

The performance of traditional speech recognition systems (as applied to information extraction or translation) decreases significantly with, larger domain size, scarce training data as well as under noisy environmental conditions. This invention mitigates these problems through the introduction of a novel predictive feature extraction method which combines linguistic and statistical information for representation of information embedded in a noisy source language. The predictive features are combined with text classifiers to map the noisy text to one of the semantically or functionally similar groups. The features used by the classifier can be syntactic, semantic, and statistical.

The invention discloses an audio and video processing method, a device, equipment and a medium, which relate to the computer technical field. The method comprisesseparating image frame information andaudio information from the video file; extracting image feature information and audio feature information from the image frame information and the audio information, respectively; fusing the image feature information and the audio feature information into video content feature information; determining a classification result corresponding to the video file according to the video content characteristic information. The invention combines the audio characteristic information in the video and the image characteristic information of the video frame for video classification, improves the accuracyand recall rate of the video classification, thereby reducing the labor cost of the video classification examination.

The invention provides a satellite image road intelligent extraction method and device base on deep learning. The method comprises the following steps: selecting a part of satellite image pictures, marking roads in the pictures with specific colors and constructing a mark database; carrying out autonomous learning through a convolutional neural network, constructing a learning model and utilizingthe learning model to carry out road extraction on newly-input satellite image pictures; judging the road extraction result, if the result is correct, carrying out road grid vectorization, or otherwise, returning to the mark database to carry out remarking and learning; and for extracted roads, traversing each road and carrying out road centerline extraction, and connecting central coordinates inseries to form a vector road line. The method carries out autonomous learning by providing a lot of mark data, then, carries out road extraction on the newly-input satellite image pictures and judgesthe extraction result, thereby correcting errors constantly and improving accuracy and recall rate of remote sensing image road extraction.

A method for detecting a rectangular target, comprising the following steps: extracting sample features; obtaining a root filter; obtaining a component filter; model detection; and performing gradient search calculations to precisely locate a detection frame. Compared with the prior art, the present invention uses a machine learning method to use pictures containing rectangular objects as positive samples, and any pictures without objects as negative samples to form a training set. During the training process, only the target outline of the sample is required Label information, the information is easy to read and efficient. The training of support vector machine and hidden variable support vector machine ensures the high accuracy of detection, and the use of hidden variable support vector machine to train component filters also makes some detailed information not easy to be lost, which greatly improves the accuracy of detection. rate and recall. Using the gradient search calculation to adjust the detection frame also compensates for the fact that the detection frame does not completely frame the target, making the position of the target more accurate and complete.

The invention relates to a method for calculating the similarity of patent texts. The method comprises the following steps that patent data is extracted from the two patent texts and preprocessed; word weights are calculated by combining part-of-speech weights, word position weights and a TF-IDF algorithm; the two patent texts are expressed in a vector space model, and two distributed word vectorsare obtained; the similarity of the texts is calculated, when the similarity of the obtained patent texts is larger than a set threshold, it is considered that the two patents are similar, and otherwise, the two patents are not similar. According to the method, patent structure characteristics and inter-word semantic relations are comprehensively considered, and special structures, such as IPC classification numbers, abstracts and claims, of the patent texts are fused into the method for calculating the similarity of the texts; compared with common text similarity calculation methods, the method is more targeted, high accuracy and a high recall rate can be guaranteed, and requirements of actual application can be well met.

The invention provides a character recognition method and device for an image, electronic equipment and a readable storage medium, wherein the method comprises the steps: carrying out multi-scale detection of a to-be-processed image, to obtain a pixel-level thermodynamic diagram; carrying out instance segmentation based on the pixel-level thermodynamic diagram, and extracting textbox information corresponding to the to-be-processed image; according to each piece of textbox information, determining a textbox image corresponding to each piece of textbox information in the to-be-processed image;recognizing each textbox image, to obtain corresponding character recognition results. According to the invention, prediction and instance segmentation based on a pixel level are adopted; text of various angles and various curved shapes can be segmented in the to-be-processed image, interference of factors such as noise, wrinkles and non-uniform light in a complex scene is effectively resisted, then the whole text is recognized, recognition errors possibly caused by single-character segmentation are avoided, and the character recognition accuracy and recall rate of the image are remarkably improved.

The invention is applicable to the field of information processing, and provides a method and a device for extracting a webpage title and an information processing system. The method includes: extracting characters of a title tag and a meta tag in a webpage source file; extracting feature points of each text block of the webpage source file; calculating probability value of being the title of each text block according to the extracted feature points of the text blocks and the characters of the title tag and the meta tag; and extracting the text block with the highest probability value as the title. By the method, characters which are irrelevant to webpage contents and piled up in the title tag and the meta tag by a webpage designer can be filtered effectively, characters which most describe webpage subject contents or main ideas are extracted as the webpage title, and accuracy and recall value of webpage research are improved.

The invention discloses an audio detection method and device, and a storage medium. The method comprises steps that pre-screening processing of a to-be-tested audio is performed to filter out candidate audio segments in the to-be-tested audio; the candidate audio segments are inputted into a trained local identification model to perform audio breakpoint detection to generate the detection result of the candidate audio segments, according to the detection result of the candidate audio segments, whether the to-be-tested audio has breakpoints is determined. The method is advantaged in that the candidate audio segments possibly having the breakpoints in the to-be-tested audio are screened out through pre-screening processing, the trained local identification model is then utilized to perform audio breakpoint detection, accuracy and the recall rate of audio breakpoint detection are improved, robustness of rhythm type pseudo breakpoints is improved, and thereby the audio breakpoints can be effectively detected.

The invention relates to the technical field of data analysis. The invention provides an abnormal user identification method. The method comprises the following steps: constructing a weighted fully connected graph for to-be-detected user group data; finding out risk nodes in the connected graph and finding out nodes associated with the risk nodes; and taking the nodes associated with the risk nodes and the risk nodes as suspicious nodes, determining communities where the suspicious nodes are located based on a community discovery algorithm, performing risk scoring on the communities where thesuspicious nodes are located, and finally determining abnormal user groups from the to-be-detected user groups. According to the embodiment of the invention, when the artificial intelligence technology is used for abnormal user identification, the recall rate and accuracy of risk identification are effectively improved, and the application field range of the risk monitoring method is expanded.

The invention relates to the technical field of computers, in particular to an industrial control system anomaly detection method and device, a server and a storage medium. The method comprises the steps of obtaining a historical traffic data set, and performing noise reduction processing on any historical traffic data in the historical traffic data set to obtain a feature vector of any historicaltraffic data; inputting the feature vector into a data traffic prediction model, and outputting a prediction feature vector of subsequent data traffic; and adjusting parameters of the data traffic prediction model according to the prediction feature vector of the subsequent data traffic and the real feature vector of the subsequent data traffic, so that the updated subsequent prediction feature vector meets a preset requirement. According to the invention, the technical problem of low detection accuracy of an industrial control system anomaly detection method in the prior art is solved.

The invention relates to the technical field of oil-gas exploration and development, in particular to a rod-pumped well fault diagnosis method and system based on a residual neural network. The method comprises the following steps: carrying out normalization by using an actual load extreme value of a current indicator diagram, and for a current indicator diagram data set, obtaining k normalization scales by using a clustering algorithm so as to obtain k + 1 normalized indicator diagrams; and secondly, binarizing the indicator diagram, and taking an 18-layer residual convolutional neural network of a k + 1 input channel as an image recognition network model based on pytorch. According to the method, an indicator diagram multi-scale normalization method is combined, an indicator diagram classification model of a multi-channel deep residual convolutional neural network is constructed, deep learning neural network training technologies such as BN and Relu are used, training and testing are carried out under a data set obtained through the multi-scale normalization method (k = 10), the convergence robustness of the model is good, and the testing accuracy reaches 95.6%.

The invention relates to a query result pushing method and apparatus, a computer readable storage medium and a computer device. The method comprises the steps of obtaining input query information; determining target knowledge points corresponding to the query information through a preset knowledge point model, wherein each preset knowledge point is provided with a corresponding preset knowledge point model; and outputting a target page according to the target knowledge point. According to the scheme provided by the invention, accuracy and recall rate of the query information are improved, thenavigation ofa target page is simplified, and the target page is output more visually.

The invention discloses a multi-dense block detection and extraction method for big data, and aims to solve the technical problems of low detection accuracy and low recall rate of a dense block detection method in the prior art. The method comprises the following steps: acquiring K-dimensional tensor data D, the number m of dense blocks to be extracted and the size range of the dense blocks; performing suspicious degree measurement on the K-dimensional tensor data D by using a density tracking coefficient based on a piecewise function, and generating a snapshots list according to the suspicious degree and a dense block size range; and extracting m dense blocks from the K-dimensional tensor data D according to the snapshots list. According to the method, the accuracy and recall rate of the dense blocks can be effectively improved while the detection efficiency is ensured.

The embodiment of the invention discloses a portrait file gathering method and device based on multi-dimensional data, electronic equipment and a storage medium. The method comprises the steps that a to-be-processed data stream in a target time period is acquired; performing target detection on an image frame obtained by decoding the data stream to obtain a target face feature data set and a target human body feature data set; performing file gathering on the target face feature data set to obtain each face file, and performing file gathering on the target human body feature data set to obtain each human body file; based on the incidence relation data, combining each face file and each human body file to obtain each portrait file; wherein the association relationship data is an association relationship between each piece of face feature data in the target face feature data set and each piece of human body feature data in the target human body feature data set. And the file gathering accuracy and recall rate are improved.

Embodiments of the invention provide a resume processing method and device, electronic equipment and a storage medium. The method comprises the steps of obtaining a to-be-converted resume; performing format conversion on the to-be-converted resume to obtain a resume after format conversion; preprocessing the resume after format conversion to acquire a preprocessed resume, wherein preprocessing comprises at least one of text segmentation, text replacement and text splicing; classifying each text included in the preprocessed resume to obtain a category corresponding to each text; and according to the category corresponding to each text, importing each text into a preset resume template to obtain a target resume. The information of the finally generated resume can be more perfect and more accurate. The invention relates to a block chain technology, for example, key information in a target resume can be written into a block chain.

The invention provides an entity linking method and device, equipment and a storage medium, and relates to the field of natural language processing and knowledge maps. The specific implementation scheme comprises the steps: obtaining a plurality of candidate entities related to mention based on a knowledge base, and obtaining a plurality of mention candidate entity pairs; respectively carrying out feature extraction processing on the mention and the candidate entities to obtain mention features and candidate entity features of each mention candidate entity pair; based on the mention feature and the candidate entity feature of each mention candidate entity pair, constructing a dynamic hypergraph, and performing disambiguation processing by using the dynamic hypergraph to obtain a target entity corresponding to the mention. According to the technical scheme provided by the invention, the training time of the neural network is shortened, and the high-level semantic association between the mention and the candidate entity is deeply mined, so that the accuracy and recall rate of the entity linking task are improved.

The invention discloses a dynamic information association method, which comprises the following steps of: 1) receiving a newly added information record and determining a similar information entity set according to a static attribute, and 2) if no similar information entity exists, creating a new information entity, ending a process, if the similar information entity exists, and acquiring a historical information record of the dynamic attribute of the similar information entity. 3) constructing a prediction model for each dynamic attribute of the information entity, determining a reasonable value range of the dynamic attributes, constructing correlation characteristics of the prediction model for each dynamic attribute of the information entity, and calculating a dynamic attribute prediction value range of the entity with similar information; according to the scheme, the information from different data sources and / or different moments is associated by utilizing the change trend of the dynamic attributes and the correlation between the dynamic attributes, so that the association accuracy and recall rate are improved.

The embodiment of the invention provides a query processing method and device and electronic equipment, and the method comprises the steps: generating an explicit query statement according to a querymessage input by a user; If a query result cannot be obtained in a knowledge graph according to the explicit query statement, generating an implicit query matrix according to the query message; Matching query results corresponding to the implicit query matrix in a knowledge set, wherein the knowledge set comprises the knowledge graph; And outputting the query result to the user. According to the method, the query accuracy and the recall rate can be greatly improved, and the cost of intelligent questions and answers can be greatly reduced.

The invention discloses a target detection model training method based on deep learning, and the method comprises the steps of testing training images containing target marks, and obtaining a target contained in each training image; comparing the target with the target label to obtain an unlabeled target and a labeled target contained in each training image; obtaining a total loss value accordingto the loss of the labeled targets and the classification loss of the unlabeled targets; and adjusting network parameters of the target detection model according to the total loss value. According tothe invention, the unlabeled targets in the training images are found out, then the classification loss of the unlabeled targets is calculated, other losses of the unlabeled targets are ignored, the trained target detection model has better performance, and the accuracy and recall rate of target detection are improved. In addition, the invention further discloses an electronic device and a computer readable storage medium.

The invention discloses an oil spill area detection method without human interaction. According to the scheme, SAR image oil spill area detection is carried out based on image significance analysis and an adaptive iterative threshold method. In the scheme, an image saliency detection method is introduced into SAR oil spill detection, and then accurate extraction of an oil spill area is realized byusing the size relationship between an adaptive iterative threshold and saliency.

The invention relates to the technical field of artificial intelligence and image detection, and provides a traffic signal lamp identification method and device based on artificial intelligence, equipment and a medium. Target feature information of a target image subjected to size processing can be extracted by utilizing a darknet53 network and is input into a traffic signal lamp recognition modeltrained by adopting a Mixup algorithm and a residual attention network; the target feature map is output to accurately extract detail features, the recall rate and accuracy are improved; the recognition of the target anchor box on each target feature map is obtained; the target anchor box coordinates and the target score are output, the target anchor box coordinate with the highest score servesas the predicted position coordinate and is mapped to the to-be-recognized image, the recognition result is obtained, automatic recognition of the traffic signal lamp is achieved based on an artificial intelligence means, and the recognition accuracy is higher. The invention also relates to a blockchain technology, and the recognition result and the traffic signal lamp recognition model can be stored in the blockchain. The method can also be applied to a smart traffic scene, thereby promoting the construction of a smart city.

The invention discloses a topic mining-based water conservancy field text retrieval method and system. The method comprises the following steps of collecting a data set; preprocessing the data set; the method comprises the following steps: constructing a topic attention model GAN-BiGRU Topic Attention Model in which a bidirectional adversarial neural network based on topic mining and a bidirectional GRU are combined; verifying and testing the test set; performing theme-related sorting; performing theme retrieval; the system comprises a data crawling module, an index construction module and a data retrieval module. According to the method, the accuracy and recall rate of theme diversity detection, theme coherence detection and downstream classification tasks are obviously higher than those of an existing method, and a new solution is provided for research in related fields; based on sufficient field data reserve support, the network model formed by combining the bidirectional adversarial neural network and the bidirectional GRU is used in the field of water conservancy retrieval for the first time, and contribution is made to application of topic model retrieval in the field of water conservancy.

The embodiment of the invention provides a picture archiving method and device and electronic equipment. The method comprises the following steps: extracting a feature value of a to-be-archived picture; obtaining a first similarity with the highest similarity between the characteristic value of the to-be-archived picture and the characteristic value of each representative picture of an existing file, wherein each file in the existing file comprises a plurality of representative pictures; judging whether the first similarity reaches a first preset similarity threshold value or not; if the first similarity does not reach the first preset similarity threshold, calculating a first average similarity of the first similarity and the second similarity; judging whether the first average similarity reaches a second preset similarity threshold value or not; and if the first average similarity reaches the second preset similarity threshold, classifying the to-be-archived picture into the target file. According to the invention, the recall rate and accuracy of picture archiving can be improved.

The invention discloses an entity linking method and system, a storage medium and electronic equipment, and the method comprises the following steps: obtaining a document entity list: carrying out the entity recognition of an online document through an entity recognition model, and obtaining the document entity list; a user name obtaining step: distributing an ID for each login user to obtain an identifier i d, and obtaining a user name list through the identifier i d; an abstract vector acquisition step: acquiring an abstract vector of the online document through a TextRank algorithm according to the online document; and a linking step: obtaining word vectors according to the document entity list and the user name list, obtaining similarity according to the word vectors and the abstract vectors, and carrying out entity linking according to the similarity and a set threshold value. According to the invention, the text abstract, the entity in the graph, the entity attribute and the relevancy of the relationship are utilized, so that the accuracy and recall rate of entity linking are improved.

The invention relates to a keyword extraction method and device, equipment and a storage medium. The keyword extraction method comprises the following steps: acquiring an input sample; obtaining a first score of each candidate word in the input sample according to a first preset algorithm; obtaining a second score of each candidate word in the input sample according to a second preset algorithm; performing setting operation on the first score of each candidate word and the second score of each candidate word to obtain a third score of each candidate word; and determining the keyword of the input sample according to the sorting result of the third score of each candidate word. According to the scheme provided by the invention, the accuracy rate and the recall rate of keyword extraction can be improved, so that the extracted keywords can better reflect the meaning of the text content.

The invention provides a cell image detection method based on transfer learning, and the method comprises the steps: collecting a cell image through a Fourier laminated microscopic imaging system, carrying out fusion through frequency spectrum iteration, obtaining a large-view-field and high-resolution cell image, constructing a cell density estimation network through VGG and FPN network models, marking a cell center position in the cell image to obtain a cell density map, inputting the cell density map into a training model, constructing a cell detection network model by taking the trained network model as a backbone network, carrying out transfer learning, obtaining a cell detection map, inputting the cell detection map into the cell detection network model, extracting a candidate region by adopting an RPN, and carrying out position regression and classification on cells through a regression device and a classifier to finally obtain a cell prediction result. According to the method, based on transfer learning, the network model of common features of the similar data sets can be extracted through transfer learning training, the problem of insufficient training samples is solved, and meanwhile, the accuracy of model output is ensured.

The invention provides a target recognition method and device, and the method comprises the steps that: a to-be-recognized image or a to-be-recognized video is obtained; and target object identification is carried out on a to-be-identified image or a to-be-identified video through a preset target detection network model to obtain an identification result, wherein a network structure of the target detection network model comprises a backbone network, a guide anchor generation network GA-RPN, an RoI Align and BBoxHead, an output of the backbone network is connected with an input of the GA-RPN and an input of the RoI Align, the output of the GA-RPN is connected with the input of the RoI Align, and the output of the RoI Align is connected with the input of the BBoxHead, so that the accuracy and recall rate of target image recognition are improved.

The invention discloses a Chinese named entity recognition method and device based on vocabulary enhancement and multiple features, and belongs to the technical field of information extraction. The method comprises the following steps: extracting character features of an input sequence in combination with a bidirectional long-short-term memory network and a convolutional neural network, introducing vocabulary information corresponding to characters in a character string mode matching manner, extracting vocabulary features in a word frequency weighted average manner, and extracting pre-training features by using a pre-training model; using a gating mechanism to control vocabulary enhancement of the vocabulary features to the character features; linearly splicing the character features subjected to vocabulary enhancement and the pre-training features to construct multiple features; obtaining context features based on context correlation of multiple features; and combining label decoding with context features to predict an optimal label sequence of the input sequence. Therefore, the character features of the Chinese sequence are extracted more fully; the extracted vocabulary features are richer, and the influence of Chinese word segmentation errors is avoided; and the entity identification index is improved by using a multi-feature combined strategy mode.

The invention relates to the field of disease and insect pest recognition, in particular to an image recognition method for solving the problem of crop disease and insect pest sample imbalance. The method comprises the following steps: performing model training by utilizing a current labeled data set, selecting a current optimal model through model verification, performing multiple times of image enhancement on pictures of a non-labeled data set, reasoning and screening the enhanced images to obtain an identification result of the non-labeled image, and inputting the identification result into a sample selection strategy. Whether the result is reserved or not is judged according to a sample selection strategy, if yes, a pseudo label is generated and moved to the current labeled data set, a new labeled data set continues to be trained, and iterative learning is conducted according to the process till the accuracy is not improved any more. According to the method, the influence of long-tail distribution can be reduced, the head category recognition effect is not influenced while the recall rate and the accuracy rate of the tail category are improved through iterative learning, reasoning is carried out only by adopting a single model, an additional network layer is not introduced, and the reasoning speed is not influenced.

The invention discloses an event tracing method aiming at a news website. The event tracing method comprises the following steps of: using an event seed report set and an event unrelated report set to train a group of SVM (Support Vector Machine) binary classifiers as an event tracing model; using each SVM binary classifier to classify VSM (Vector Space Model) vectors of main information of a non-processed target news website captured in the news website, so as to obtain a corresponding classifying result; carrying out event related judgment on the target news website according to the classifying result; if the target news website and an event are related, adding the event seed report set, and re-training the event tracing model; and otherwise, continuously processing the new target news website. With the adoption of the event tracing method provided by the invention, the defects of the event tracing method in the prior art of less event related information in an initial stage, data inclination and high computation complexity can be overcome; and the accuracy and the recall rate of the event tracing can be effectively improved.