131results about How to "Guaranteed recall" patented technology

The invention discloses establishment of a virtual case library of cancer pathological images and a multi-scale cancer detection system based on a convolutional neural network. The system is based ona method of a convolution neural network, a cancer mass region is detected on a pathological full-scan section, and the system includes four modules: 1) a pathological section image preprocessing module; 2) a virtual case database construction module; 3) a high-scale cancer mass detection module; and 4) a small-scale cancer mass classification module. The multi-scale cancer detection system provided by the invention can make full use of the multi-scale information of a pathological image, on different scales, according to the characteristics of the image, different strategies are designed to detect a suspected cancer area, and at the same time, under the condition of insufficient training data, the virtual case library method established by the invention can provide more training data setsfor an existing data-driven deep learning method. The multi-scale cancer detection system based on a convolutional neural network has the characteristics of multi-scale detection, driving of a relatively small amount of data and the like, and has the characteristic of reducing computing resources required for one-time recognition and improving time efficiency of an algorithm on the basis of ensuring the overall recall rate and accuracy.

The invention discloses an SIMD (single instruction multiple data) optimization-based webpage duplication elimination and concurrency method, which comprises the following steps of: 1, extracting text information of webpages, namely extracting effective information of webpages; 2, extracting Shingle, namely extracting webpage characteristics and generating a Shingles set; 3, clustering to reduce comparison times and reduce time and space complexity; and 4, comparing fingerprints to find similar webpages and delete the similar webpages. The SIMD optimization-based webpage duplication elimination and concurrency method can ensure the precision rate and the recall ratio, and effectively improves the rate of detecting webpage similarity.

The invention discloses an event warning method based on Internet information and belongs to the technical field of information. The method comprises the following steps. Firstly, crawling of webpage information on each set website is performed, the webpage information is stored, and information relevant to an event is obtained. Secondly, junk information filtering is performed on the information which is relevant to the event and obtained by crawling. Thirdly, an event information main body is set up, the information after filtering is sorted based on example names and attributes in the event information main body by utilizing a mode matching method. Fourthly, the sorted information is clustered according to set characteristics, and events in each kind of information are obtained. Fifthly, all the characteristic parameter values of each event are obtained regularly, and if a difference value of all the characteristic parameter values of one event and an average value of the event in a given previous period exceeds a preset threshold, warning on the event is performed. The event warning method based on the internet information ensures timeliness, accuracy and comprehensiveness of information identification and warning of food safety events, and improves efficiency of risk warning of food safety.

The invention discloses a personalized paper recommendation method and a system thereof. The method includes the following steps: step 1, finding out heterogeneous academic network data based on the behavioral characteristics of researchers writing academic papers in the filed of scientific research, and building a training dataset according to the heterogeneous academic network data, and training as per the training dataset to obtain a ranked study model; step 2, on-line building user configuration, generating candidate collected papers interested by users, generating the result of paper recommendation based on the ranked study model according to the candidate collected papers, generating paper recommendation as per a certain manner and returning the paper recommendation to a user based on the result of paper recommendation; step 3, on-line receiving user feedback, and updating the result of paper recommendation correspondingly as per different user feedback. The method and the system effectively solve the problem of cold start at the initial stage of the recommendation system, and guarantee the accuracy and recall rate of the recommendation result.

The invention discloses an image retrieval system integrating classification with hash partitioning, which comprises a downloading module, a classification module training module, an image classification module, a characteristic extraction module, a recording table building module, a partitioning module, a request processing module, a retrieval module, a similarity acquiring module and a result returning module. The downloading module is used for downloading images so as to build an image library, the classification model training module classifies the images in the image library according tothe shapes and selects representative sample images from the image library to form a sample library firstly, then extracts characteristic descriptors of the classification bottom layer of all images in the sample library and performs trainings on the characteristic descriptors of the classification bottom layer by a support vector machine so as to obtain discriminant of each classification. Classification models are formed according to the discriminants of all the classifications. Precision ratio of the image retrieval system is increased, the problem of low recall ratio during classificationmistakes is overcome, and the retrieval speed of the image retrieval system is increased integrally.

The invention relates to the character string multi-mode matching field, and discloses a sensitive word filtering method. The sensitive word filtering method comprises the steps of performing management on Chinese, English, website sensitive words and excluding words; performing a character normalization processing method; performing a group of filtering policies and realization method for sensitive words in different existence forms, at least comprising a filtering step for Chinese, English, websites, full spelling, pinyin compiling and anagram; setting a group of criterion rules for sensitive words; and performing an approximate matching method for Chinese sensitive words. The invention also discloses a sensitive word filtering apparatus. According to the sensitive word filtering method and apparatus, the requirements of a content administrator and a searcher on issued or searched text filtering sensitive words can be satisfied; filtering for a large amount of sensitive words can be carried out rapidly and accurately; and the sensitive words, the level of the sensitive words and the positions of the sensitive words in the can be returned to the caller.

The invention discloses a video keyframe extraction method, which comprises the steps of: using a ViBe algorithm fused with an inter-frame difference method to perform moving object detection on an acquired original video sequence, so as to obtain a key video sequence containing a moving object; performing keyframe crude extraction on the key video sequence by using a global characteristic peak signal-to-noise ratio to obtain candidate keyframe sequences; and establishing global similarity of the candidate keyframe sequences by using the peak signal-to-noise ratio, establishing local similarity of the candidate keyframe sequences by using SURF feature points, and performing weighted fusion on the global similarity and the local similarity to obtain comprehensive similarity, performing self-adoptive keyframe extraction on the candidate keyframe sequences by using the comprehensive similarity, and finally acquiring a target keyframe sequence. The video keyframe extraction method providedby the invention can effectively extract the video keyframes, obviously reduce the redundant information of video data, and express the main content of the video concisely. Moreover, the video keyframe extraction method has low algorithm complexity and is suitable for real-time extraction of keyframes of surveillance videos.

The invention discloses a mass disturbance warning method and system applied to short texts. The method comprises the following steps of automatically structuring a domain knowledge database applied to mass disturbances; combined with the domain knowledge database, performing local structured extraction and online classification on the short texts to screen out mass disturbance texts with potential risks from mass short texts; combined with the domain knowledge database, performing overall structured extraction and online clustering on identified short texts, and according to whether the number of short texts contained in every cluster exceeds a preset threshold value, determining whether to perform timely alarming. The mass disturbance warning method applied to the short texts has the advantages of fully integrating domain background knowledge, context, shallow semantic expression and deep semantic computation, achieving collaborative analysis and predication of mass disturbances and improving the timeliness, accuracy and recall rate of screening, tracking and alarming of the mass disturbances.

The present invention discloses a voting based classification method for cross-language subjective and objectives sentiments. The method comprises the following steps of: S1, constructing a target language sentiment dictionary according to a source language sentiment dictionary; S2, separately adopting three algorithms of a rule algorithm, a combination algorithm of machine translation and statistical machine learning, and a polarity eigenvalue calculation algorithm to extract words from a sentence of a to-be-tagged text, determining sentiment polarity of the words according to the constructed target language sentiment dictionary, and further determining subjective and objective nature of the sentence; and S3, acquiring subjective and objective nature determination results of the sentence, obtained according to the three algorithms, and determining the subjective and objective nature of the sentence by voting. The method can fully consider contexts and usage habits of a target language on the premise of keeping certain accuracy, effectively solves the problem that a tagged corpus of the target language is scarce, and further improves classification accuracy on the premise of ensuring the recall rate.

The invention discloses a detection method and a detection system for video image text and relates to the technical field of image text detection. The method comprises the following steps of coarse positioning of video image text areas and verification of video image text areas. The specific steps are as follows: (A) coarse positioning of video image text area: the gaussian pyramid multi-scale image of a video image is calculated, the image text area is segmented by threshold segmentation, text areas are combined by the inflation and corrosion technology in mathematical morphology, and candidate text areas are analyzed and positioned by connected domain analysis, text line segmentation and elimination rules; and (B) verification of video image text areas: an adaboost classifier is trained by adopting harr characteristics, and the adaboost classifier is utilized to detect whether text exists in the candidate text areas and the candidate text areas the text of which is not detected are eliminated. The method and the system can improve the accuracy and the recall ratio of video image text detection.

The invention discloses a target field question and answer pushing method combining keyword retrieval and a twin neural network. According to the method, the historical questions are preprocessed based on the Internet and the expert knowledge, and then an answer with the highest matching degree is intelligently pushed to a real-time question proposed by a user in combination with a keyword retrieval method and a semantic similarity model based on the twin neural network. According to the method, the efficiency and the performance can be better balanced, meanwhile, the recall rate and the accuracy of the method are guaranteed, and the method has great practical significance in the field of intelligent question and answer systems.

The invention discloses a multi-strategy combined document automatic classification method which comprises the steps: extracting keywords from a given to-be-classified document, obtaining a vector of the to-be-classified document according to the keywords, and obtaining a class center vector of each class in a standard document base; obtaining a class threshold value of a classified document according to the class centre vector; carrying out comparison of the similarity degree of the vector of the to-be-classified document and the class centre vector with the class threshold value, when the similarity degree of the vector of the to-be-classified document and the class centre vector is larger than the class threshold value, classifying the to-be-classified document through adoption of a space vector model; and if the similarity degree of the vector of the to-be-classified document and the class centre vector is not larger than the class threshold value, classifying the to-be-classified document through adoption of a k-nearest neighbor (KNN) method. Compared with the prior art, through the technical scheme, the multi-strategy combined document automatic classification method can greatly improve efficiency of the document classification and cannot reduce classification precision at the same time.

Embodiments of the invention provide methods for matching text information and pushing business objects, and devices for matching text information and pushing business objects. The method for matching text information comprises: obtaining a to-be-matched first text information set and a to-be-matched second text information set, the first text information set comprising first text information in limited number, the second text information set comprising second text information in limited number; and according to preset rules, inquiring one or more piece of information in the second text information in the limited number, the one or more piece of information being matched with each piece of information in the first text information in the limited number. An open extension thinking of directly finding extended words from the first text information is given up, and the method turns to a closed interval, to search one or more piece of information in the second text information in the limited number, the one or more piece of information being matched with each piece of information in the first text information in the limited number, so as to save unnecessary matching calculation amount, reduce system resource waste, and improve efficiency of matching calculation.

The invention discloses a method for detecting phishing web pages based on a spatial mixed index mechanism, which comprehensively utilizes spatial layout, character features and image features of the web pages. The method relates to a design scheme based on visual layout features of pages and aspatial database, solving the problems of detecting phishing web pages rapidly according to visual similarities of web pages. Combined with a rendering engine of a browser, the method carries out feature extraction of visual layout for appointed suspicious web pages and utilizes spatial database index combined with text and image features of the web pages simultaneously to form a spatial tree i.e., a DIIR tree, wherein the DIIR tree is a reverse index of comprehensive document images in the spatial mixed index mechanism. The DIIR tree improves an R tree of a spatial area in the spatial index mechanism by adding reverse index files of characters and image features of network objects to each node of the R tree of the spatial area. When querying a new network object, not only the spatial layout feature of the object is considered, but also text and image features of the network object are simultaneously combined.

The invention discloses a domain concept extraction method based on Deep Learning. The method includes extracting samples in a training corpus, adopting word frequency, document frequency, inverse document frequency, word length, word frequency variance and domain consensus as feature vectors, training and acquiring a deep network model, which is capable of representing the complex mapping correspondence between the word-type filed concept multi-dimensional feature vectors and class labels, on the basis of the Deep Learning technology, and finally comparing the deep network model established on the basis of the Deep Learning technology, an optimized BP neural network model and mainstream KNN and SVM models in the testing step. According to the tests, the optimal test effect is acquired through the deep network model established on the basis of the Deep Learning technology.

The invention discloses a court electronic file-oriented case information automatic extraction method, which comprises the steps of creating a case information extraction framework, performing statistical analysis on case elements in various case electronic file files, and constructing a basic case body; customizing a basic case information extraction strategy according to the case ontology and the layout characteristics of the electronic file; constructing a case information extraction framework based on a cross-domain information extraction tool TZIE of XML; building an NER model and a training and multi-granularity semantic unit combination mode based on multi-granularity semantic legal documents, employing a Word2vec Skip-gram model and an LDA topic model, obtaining a judicial domain word vector model and a judicial domain word vector model through training respectively, combining domain entity recognition of a BiLSTM-Attention CRF model, adding an Attentio mechanism in a referencemodel BiLSTM-CRF, predicting the probability that each semantic unit belongs to different labels, assisting an optimized model training mode, taking a manual annotation corpus training model as a main task, and taking manual and automatic annotation corpus training models as auxiliary tasks.

The invention discloses an automatic image watermark detection method. The method includes the following steps: S1, establishing a watermark pre-detection model; S2, establishing a watermark re-identification model; and S3, combining the watermark pre-detection model and the watermark re-identification model to realize automatic detection of an image watermark. According to the method, a detectedcandidate watermark region is re-identified, the candidate watermark region is re-identified through manually designing features of the watermark and training the binary model specifically judging whether the watermark exists, false detection of a watermark detector, which is based on a deep convolutional neural network, on of patterns which are similar to the watermark is reduced, and an accuracyrate of watermark detection is increased.

The invention provides a method for quickly detecting a shot boundary, comprising the following steps that: firstly, a video frame inside a lens is removed from an original frame sequence OFS to obtain a recorded frame sequence RFS according to changes of a brightness variance feature of the video frame; meanwhile, a partial lens gradation existed in a video is determined according to a concave shaped parabola of the brightness variance feature; then, a pixel pair difference, a color histogram difference feature, an edge histogram difference feature and so on are selected from the RFS as accordances for a video continuity detection; then, a support vector machine is adopted to perform a shot cut detection; finally, the shot cut detection is performed for the rest video sequence by adopting a method of reducing temporal resolution. The method for quickly detecting a shot boundary improves the processing speed for the shot boundary detection, and guarantees a recall ratio and the accuracy of an operation at the same time.

The invention discloses a precise strike-oriented ground moving target detection and recognition method. The method comprises the steps of: combining space-time information to extract a target candidate region, wherein the time-domain information is utilized to acquire moving images, forward-moving history graphs and backward-moving history graphs are calculated according to the moving images, images obtained by carrying out minimum-value acquisition on the forward-moving history graphs and the backward-moving history graphs according to pixels are used as input images, candidate moving regions are obtained through self-adaptive threshold processing and communication-domain extraction, the space-domain information is utilized to calculate materiality scores of the candidate moving regions , and the target candidate region is obtained through threshold processing; and recognizing the target candidate region, wherein feature learning is completed through local low constraint encoding, a linear support vector machine is utilized to realize recognition for the target candidate region, and a candidate region recognized as specific targets is reserved to obtain a final detection result. The method realizes automatic detection and recognition for the multiple ground moving targets in a complex environment, and improves the strike precision.

A method for extracting information based on a multistage rule base comprises the steps that (1) a URL address of web pages is obtained; (2) the web pages corresponding to the URL address are downloaded; (3) a web page tree-type structure chart is obtained; (4) web page clustering is conducted, web pages are selected from the web pages to be clustered to serve as a training set, and a clustering rule of the web pages is defined according to a robot learning method; (5) a searching result is extracted; (6) information is collected and displayed. After the web page tree-type structure chart is obtained in the step (3) and the web pages are clustered in the step (4), the recall ratio of the retrieved information can be effectively increased, the clustering rule is automatically generated by means of robot learning in a training set mode, manual clustering is not needed, the automation degree of searching is effectively increased, and the condition of large-area use is achieved on the premise that the recall ratio is guaranteed. According to a device for extracting the information based on the multistage rule base, a hardware foundation is provided for an information extraction process, cost is low, and the device is suitable for large-scale use.

Disclosed is a semi-supervised self-learning driven medical text disease identification method provided by the invention, feature classification of the medical text is mainly realized, the features include disease object types, disease progression, whether diseases occur, disease severity, disease condition, disease uncertainty and the like, and by identifying and classifying features, unstructured medical texts are structured and can be directly processed and used in further information mining; based on the features of few medical text annotation texts and many unannotated texts, the invention is developed from the aspects of feature extraction and classification model optimization; experimental results show that the method well makes up for the defects and deficiencies of few tagged texts, and since similar data and wrong classification data are easily introduced into the self-learning method, the effect of the semi-supervised SVM in the aspect of utilizing untagged data is better than that of the semi-supervised SVM, and the feasibility and high efficiency of the method are proved.

The invention provides an image retrieval method and device, and relates to the field of image retrieval. The method comprises the following steps: receiving query images sent by a user; according to an image classification method, obtaining the confidence coefficients, which correspond to each category, of the query images; according to a descending order of the confidence coefficients of the query images, retrieving image characteristics corresponding to the category of each confidence coefficient in a current confidence coefficient segment from an image characteristic library; according to the order of priority of retrieval, loading the retrieved image characteristics step by step; carrying out sorting processing on the loaded image characteristics; and sending a sorting processing result obtained by processing to a display interface step by step so as to bring convenience for the display interface to display. A situation that a result is output after all image characteristics in an established image database are retrieved is avoided, the time promptness and the stability of retrieval result feedback time are guaranteed, the problem that a true target is missed due to classification errors is unlikely to happen, indexes including a retrieval detection rate, a retrieval recall rate and the like are guaranteed, and the image retrieval method and device is more suitable for application scenes including security and protection video surveillance and the like.

The invention relates to an automatic service combination system and method based on semantics. The system comprises a server terminal and a client terminal. The server terminal comprises a database constructing module, an automatic workflow generating module, a service discovering module and a service combination module. The client terminal comprises a requirement semantics inputting module and a display module. The method includes the following steps that according to input requirement semantics, a domain library and a behavior library; workflow is automatically generated; fine adjusting is carried out on the generated workflow; according to the workflow and the semantic information of a service, a candidate service set is obtained through calculation of semantics similarity; according to QoS information of the service, an executing path which can meet the QoS constraint of a user and has the optimal QoS property is chosen. Authentic automatic evolution growth of complicated software is achieved in a large scale, the workflow is provided without need of manual intervention and the service combination efficiency is extremely improved.

The invention discloses a knowledge service system and method suitable for the full life cycle of high-end equipment. The method is suitable for a knowledge service platform which is in communicationwith a high-end equipment manufacturing enterprise information system through a knowledge service system interface. Historical data is extracted from the high-end equipment manufacturing enterprise information system or expert experience is obtained from the brains of technical personnel through a semi-automatic acquisition mode based on a complex knowledge resource acquisition template. Full-lifecycle knowledge of the high-end equipment is obtained through a preset knowledge acquisition process. The knowledge obtained in the above steps is stored into a distributed database according to theexpression form of the knowledge, to form a high-end equipment manufacturing enterprise knowledge base. The knowledge service requirements of technicians of a high-end equipment manufacturing enterprise are obtained in an active and passive combined mode. A corresponding knowledge service engine in the multi-mode knowledge service method library is called according to the knowledge service requirements. The knowledge service engine calls full life cycle knowledge of the high-end equipment in the knowledge base and generates needed knowledge service by utilizing an inference algorithm.

The invention discloses an application searching method and device. The method includes the steps that label systems of applications are built; search words uploaded by a client are received; matching is conducted in the label systems of the applications according to the search words; when the search words are matched with keywords of the label system of one application, relevant information of the application is fed back to the client to be displayed. Therefore, by building the label systems of the applications, when the search words uploaded by the client are received, matching is conducted in the label systems of the applications according to the search words, when the search words are matched with the keywords of the label system of one application, the relevant information of the application is fed back to the client to be displayed, and intelligent searching of the application is achieved. The built label systems of the applications guarantee the recall rate of application engine searching, the searching quality of the application engine searching is improved, and user experience is enhanced.

The embodiment of the application provides an information searching method and device. The method includes the steps that when a searching request of information is received, one piece or multiple pieces of candidate information is searched based on the searching request; the behavior confidence corresponding to user identification where operation is initially conducted on the candidate information is searched; one piece or multiple pieces of targeted candidate information is selected from the candidate information based on the behavior confidence; the targeted candidate information is returned. By means of the information searching method and device, the condition that revision and upgrade are correspondingly carried out with the behavior mode of a user changing is avoided, the recall and precision of cheating behavior detection are guaranteed, and the cost of revision and upgrade detection is greatly reduced; the generalization capability is improved, the limitation of current information is solved, and the recall of cheating behavior recognition is improved.

The invention provides an adaptive cruising system and a vehicular apparatus. The system comprises: a point cloud data collector for collecting a plurality of point clouds in a front road scene during driving of a motor vehicle; a multi-dimensional vehicle candidate area determining device for determining a ground plane in the front road scene based on the plurality of point clouds output by the point cloud data collector, wherein a plurality of multi-dimensional areas are generated on the ground plane according to preset vehicle dimensions and whether the multi-dimensional areas are multi-dimensional vehicle candidate areas or not is determined; a ranging device for detecting spacing between the motor vehicle and the multi-dimensional vehicle candidate areas according to the multi-dimensional vehicle candidate areas determined by the multi-dimensional vehicle candidate area determining device; and a controller for carrying out crushing control for the motor vehicle according to the spacing. The system and apparatus allow vehicles in the front road scene to be recognized and positioned quickly and accurately and allow for adaptive crushing control.

The invention relates to the technical field of automatic question answering, and discloses a medical automatic question-answering system construction method based on a knowledge graph, which comprises the following steps of: pre-defining entities, relationships and attributes in a medical field knowledge graph; acquiring medical field data, and labeling the medical field data by using a data automatic labeling algorithm according to the predefined medical field knowledge graph; utilizing a knowledge graph construction model to construct a knowledge graph for the annotation data to obtain a medical field knowledge graph; encoding a dialogue text of a user by using a dialogue encoder to obtain dialogue text encoded data; and performing path inference on the text encoded data by utilizing aknowledge graph-based reinforcement learning network to obtain an intention path of the user, and taking a knowledge graph result pointed by the intention path as an automatic question-answering result. The invention further constructs a medical automatic question-answering system based on a knowledge graph. Automatic question answering in the medical field is realized.

The invention belongs to the field of data mining, and relates to a digital thesis retrieval method, in particular to a digital thesis retrieval method based on formal concept analyses. According to the thesis retrieval method, concept lattice building and searching scale and time are shortened through the mode of sequencing and selection according to intervals, then other cut theses are attached to a selected thesis, and the effect of result losing is eliminated to a large extent; meanwhile, the problem that retrieval results are too dispersed and too large in the thesis retrieval process is solved through a concept lattice rough and approximate retrieval mechanism, and retrieval result recall rate and precision are also guaranteed. The digital thesis retrieval method based on the formal concept analyses provides a usable retrieval mode based on the formal concept analysis regarding to large-scale data.

The invention relates to a hand raising detection method based on deep learning. The method comprises the steps that (1) samples are collected, wherein the samples are complicated environment samples;(2) a hand raising detection model is established, wherein the hand raising detection model is based on a convolutional neural network structure and is trained through an R-FCN target detection algorithm based on the samples; and (3) the trained hand raising detection model is utilized to perform hand raising detection on a to-be-detected video, and a hand raising box position is obtained. Compared with the prior art, the method has the advantages that a hand raising action in a complicated environment can be detected, and the accuracy rate and the recall rate are high.