33results about How to "Improve recall rate" patented technology

A photo retrieval method, system, and medium using multimodal information which prevents a photo of an irrelevant category to be retrieved having a high ranking due to an unfavorable side effect from a query expansion, by reflecting a level of consistency between an object category of a query and a category of a photo retrieved by expansion retrieval using ontology, to a photo retrieval result ranking is provided. The photo retrieval method using the multimodal information includes: assigning an object category with respect to a query; retrieving photos associated with an expanded query term extracted from the query; determining a ranking of the retrieved photo by reflecting the assigned object category; and arranging and displaying the photos in response to the determined ranking.

The invention relates to the technical field of big data, in particular to a question and answer library knowledge management system based on deep learning and an implementation method thereof. The system comprises a global setting module, a knowledge management module, a knowledge learning module, a knowledge statistics module, a session simulation module and a terminal docking module, each module has setting functions of adding, deleting, modifying, checking and the like; the method comprises the following basic steps: (1) carrying out global setting; (2) performing knowledge management, including knowledge creation; (3)carrying out knowledge learning, wherein knowledge mining and problem clustering are carried out through an iteration process of big data information / historical sessions,and user annotation feedback is carried out; (4) carrying out session simulation: simulating a session of a real environment; (5) carrying out knowledge statistics, and high frequency and sudden increase problems in the application field are understood for improvement and decision making; and (6) carrying out multi-terminal adaptation and distribution on the packages by the modules. According tothe system and the method, the natural language habit of human beings can be known, the complex problem of a user can be known, and manual intervention is not needed.

The invention discloses a fast pedestrian detection method with the combination of static bottom characteristics and motion information. For a single image frame in a video sequence to be detected, amulti-scale sliding window method is used to extract a detection window of a specified size, a fast pyramid characteristic calculation method is used to calculate the static layer characteristics in each detection window, the static bottom layer characteristics comprise a direction gradient, a gradient amplitude and a color channel, the pre-classification of a pedestrians and no pedestrian is carried out on each detection window based on the static bottom characteristics of an image in each detection window. The specific motion speed characteristics of the pedestrians are used, through a detection window pixel-difference-averaging method, a detection window which is mistakenly detected to be a pedestrian in the pre-classification is removed, and finally the position of an image in a detection which is detected to be a pedestrian is the position of the pedestrian in the image frame. According to the method, the accuracy of detection is improved while the false detection rate is reduced.

The invention discloses a fast pedestrian detection method with the combination of static bottom characteristics and motion information. For a single image frame in a video sequence to be detected, amulti-scale sliding window method is used to extract a detection window of a specified size, a fast pyramid characteristic calculation method is used to calculate the static layer characteristics in each detection window, the static bottom layer characteristics comprise a direction gradient, a gradient amplitude and a color channel, the pre-classification of a pedestrians and no pedestrian is carried out on each detection window based on the static bottom characteristics of an image in each detection window. The specific motion speed characteristics of the pedestrians are used, through a detection window pixel-difference-averaging method, a detection window which is mistakenly detected to be a pedestrian in the pre-classification is removed, and finally the position of an image in a detection which is detected to be a pedestrian is the position of the pedestrian in the image frame. According to the method, the accuracy of detection is improved while the false detection rate is reduced.

The invention relates to a student behavior detection method based on target detection, and the method comprises the following steps: S1, building a data set comprising student behavior information which comprises hand raising, standing and sleeping; S2, establishing a student behavior detection model, wherein the model is an improved Faster R-CNN model based on a residual network ResNet-101; S3,training the student behavior detection model based on an online hard sample mining method; S4, detecting a to-be-detected video by using the trained student behavior detection model to obtain a student behavior result, and visualizing the student behavior result; wherein the residual network ResNet-101 adopts a multi-layer feature fusion strategy and the fifth convolution stage of the residual network ResNet-101 includes multiple branches with different receptive field sizes.. Compared with the prior art, the method has the advantages of high precision and the like.

The invention relates to a convolutional neural network-based face detection method. The method comprises the following steps that: 1) a face detection model is established, the model adopts an RFCN network structure, wherein the RFCN network structure comprises a feature fusion-based feature extraction layer; 2) a sample set is obtained; 3) the face detection model established in step 1) is trained; and 4) the trained face detection model is adopted to perform face detection on a picture to be detected. Compared with the prior art, the method of the present invention has the advantages of high accuracy, high recall ratio, good adaptability for complex scenes, and the like.

The invention discloses an automatic electroencephalogram epilepsy recognition method based on multi-view depth feature fusion, and belongs to the field of biomedical engineering and machine learning.According to the automatic electroencephalogram epilepsy recognition method disclosed by the invention, through expressing an electroencephalogram time-frequency spectrum and using deep learning to train a multi-view depth feature fusion model, the fusion feature thereof has robustness while improving algorithm performance. Compared with the prior art, the automatic electroencephalogram epilepsyrecognition method disclosed by the invention effectively combines signal processing and deep learning techniques to extract different angle information of electroencephalogram data in time and channels, thereby jointly expressing epileptic characteristics and improving automatic recognition accuracy.

The invention discloses an event extraction method and system based on an attention mechanism. The extraction method comprises the following steps: obtaining a to-be-processed text and an event type thereof, extracting a vector of each character in the to-be-processed text, obtaining text vector data, extracting a vector of each character in the event type, and generating event type vector data; inputting the text vector data and the event type vector data into a pre-constructed extracted word prediction model to obtain a label sequence formed by labels corresponding to each character in the to-be-processed text; and performing event extraction on the to-be-processed text based on the label sequence to obtain corresponding event trigger words and event arguments. By introducing the event type and combining the feature data corresponding to the event type to perform joint extraction on the event trigger word and the event argument, the accurate calling rate of event extraction is effectively improved.

The invention provides a lane line detection method and device, an electronic device and a storage medium. The method comprises the following steps of inputting an acquired image to be detected into a preset target detection model, and acquiring the first detection information and the second detection information of each grid in the image; performing non-maximum suppression processing on the first detection information of each grid, and obtaining the position of each lane line boundary point in the image; performing the non-maximum suppression processing on the second detection information of each grid, and obtaining the position of each lane center point and the corresponding lane width in the image; and determining the lane line in the image according to the position of each lane line boundary point, the position of each lane center point and the corresponding lane width in the image. Therefore, through the lane line detection method, the interference of noise on lane line detection is reduced, and the accuracy and the call rate of the lane line detection are improved.

The invention relates to the technical field of automatic driving, and provides a parking space recognition method and device, a vehicle-mounted terminal and a storage medium. According to the invention, parking stall corner points are taken as density clustering nuclear points, and adjacent parking stall corner points are connected by parking stall lines, i.e., detection points of the parking stall lines exist between the adjacent parking stall corner points, so that the adjacent nuclear points of each nuclear point can be determined based on the density, and the adjacent parking stall cornerpoints of the parking stall corner points are obtained. By utilizing the characteristic that adjacent parking stall corner points are connected through the parking stall lines, the distorted parkingstall / special-shaped parking stall in the image can be identified, and the parking stall identification precision is improved. Moreover, each parking stall corner point is used as a density clusteringnuclear point, so that the parking stalls corresponding to the parking stall corner points can be identified under the condition that parking stall lines near the parking stall corner points are shielded and detection points of few parking stall lines exist, the parking stall detection rate is improved, and the missing detection of the parking stalls is avoided.

The invention discloses a resume search method, device and system, a resume delivery method, device and system, a synonym dictionary construction method and device and electronic equipment. The resumesearch method comprises the steps of receiving a talent resume screening request sent by a client for target position description; determining a first key word described by the target position; determining a second talent keyword of the plurality of to-be-screened talent resumes; according to a pre-generated synonym dictionary, determining a second unable keyword matched with the first unable keyword; and returning the talent resume corresponding to the matched second talent keyword to the client. By adopting the processing mode, the first talent keyword described by the target position and the second talent keywords of the plurality of to-be-screened talent resumes are matched according to the synonym dictionary, and the talent resumes matched with the target position description are automatically determined according to the matching result; therefore, the call completion rate and the call accuracy rate of the resume search result can be effectively improved.

The invention discloses a similar video detection method and device, and belongs to the technical field of artificial intelligence. The method comprises the following steps: acquiring a feature embedding vector of a first video; performing comparison recall processing on the feature embedding vector of the first video and the feature embedding vector of at least one second video to obtain a recalled video associated with the first video; performing re-checking processing on the first video and the recalled video, and determining a feature checking result of the first video and the recalled video on at least one checking dimension; and determining video feature data of the first video and the recalled video on the target verification dimension. In the technical scheme provided by the embodiment of the invention, the video recall process and the video similarity verification process are decoupled, the video recall is performed based on the feature embedding vector of the video, and after the recalled video is obtained, the first video and the recalled video are more accurately rechecked and verified from the plurality of verification dimensions, and therefore, similar video judgment is completed, and the video duplicate removal efficiency is effectively improved.

The invention provides a people flow counting method and system, and belongs to the technical field of image processing. The method comprises the following steps of: detecting the object of the imageset with the frame sequence according to the frame sequence; when a first image with a first object set is detected in the image set, obtaining a first local vector corresponding to the position of the first object set relative to the first image in the first image, and then obtaining a second local vector corresponding to the position of the second object set relative to the second image in the second image with the second object set in the image set; and obtaining a tracking value set according to the first local vector and the second local vector in combination with a preset tracking mapping relationship, and obtaining a tracking state set of objects in the first object set according to a relationship between each tracking value in the tracking value set and a preset tracking thresholdcondition. According to the invention, detection and tracking of high on-call rate are realized, and a high-precision and high-robustness people flow counting system is realized based on the detectionand tracking of the high on-call rate.

The invention belongs to the technical field of computers, and relates to an event recognition method and device, a computer readable storage medium and electronic equipment. The method comprises thesteps of obtaining a target statement used for identifying a target event, and determining the type of the target statement according to a statement sequence of the target statement; if the type is asentence pattern type, matching the sentence sequence with a corpus sequence in a corpus to determine an identification result of the target event according to a matching result; and if the type is the paragraph type, obtaining configuration information generated based on the preset vocabulary, and determining an identification result of the target event according to the configuration information.On one hand, excessive manpower and material resources do not need to be input, the manpower cost for identifying the target event is reduced, and the identification efficiency of the target event isimproved; and on the other hand, the situation of missing report of the target event is reduced, the accurate calling rate of target event identification is improved, and the integrity of the targetevent identification system is improved.

The invention discloses a commodity search system, a commodity search method and device, a head word extraction model construction method and device and electronic equipment. The method comprises thesteps that a search server in a commodity search system receives a commodity search request sent by a client for a search term specified by a user, if the search words are zero and few result words, extracting a central word of the search word through a central word extraction model, determining a commodity object matched with the central word, and returning the commodity object to the client, thecentral word extraction model being a model obtained by learning from a corresponding record set between the commodity title and the central word annotation data through a machine learning algorithm.By the adoption of the processing mode, the central words of the search words are automatically extracted through the central word extraction model, and then commodity search is conducted according to the central words; therefore, the head word extraction efficiency and accuracy can be effectively improved, the call completion rate and the call accuracy rate of the search result are improved, anduser conversion and enterprise benefit improvement are brought.

The invention provides a method for automatically detecting a magnetic material, electronic equipment and a storage medium, and relates to the technical field of magnetic material detection.The method comprises the steps of collecting a to-be-detected image of the to-be-detected magnetic material, preprocessing the to-be-detected image, extracting image features of the preprocessed to-be-detected image, comparing the image features with comparison data, and judging whether the to-be-detected magnetic material is qualified based on a comparison result. The magnetic material is detected according to the preset standard, so that the detection efficiency is effectively improved, the full detection rate and the accurate detection rate of the magnetic material are improved, and errors or mistakes caused by manual operation during manual detection are avoided.

The invention discloses a junk short message new word identification method and device, and electronic equipment. The method comprises the following steps: acquiring a short message set; determining candidate word sets corresponding to the plurality of junk short messages; determining short message category tendency related indexes of the candidate words according to the short message category information; obtaining document rare related indexes of the candidate words; determining junk short message new word scores of the candidate words according to the short message category tendency relatedindexes and the document scarcity related indexes; and determining the junk short message new words from the candidate word set according to the scores of the junk short message new words. By the adoption of the processing mode, most new words recalled according to the short message category tendency related indexes have representative significance for junk short messages, and recalling of many normal vocabularies is avoided; therefore, the call accuracy can be effectively improved. Meanwhile, by means of the processing mode, low-frequency junk short message new words can be recalled according to related indexes of document scarcity; therefore, the calling full rate can be effectively improved.

The invention provides a data processing method and device. The method comprises the following steps of: obtaining a search word input by a user when a search instruction is received; determining expanded words of the search word on the basis of the search word, and carrying out search in a document database by utilizing the search word and the expanded words so as to obtain a plurality of target documents; respectively extracting feature information from each target document, wherein each target document corresponds to one piece of feature information; and predicting a hotspot technology by utilizing a preset prediction rule on the basis of the feature information corresponding to each target document, and feeding the predicted hotspot technology back to a user. According to the data processing method and device, the document recall rate is improved through expanding the search word; research hotspots can be predicted through processing a search result, so that scientific reference basis can be provided for the researches carried out on topics by researchers.

The application discloses a similar video detection method and device, belonging to the technical field of artificial intelligence. The method includes: obtaining a feature embedding vector of the first video; comparing and recalling the feature embedding vector of the first video with at least one feature embedding vector of the second video to obtain a recall video associated with the first video; The first video and the recalled video are checked and verified, and the feature verification results of the first video and the recalled video in at least one verification dimension are determined; the video feature data of the first video and the recalled video are respectively determined in the target verification dimension. In the technical solution provided by the embodiment of the present application, by decoupling the video recall process and the video similarity verification process, the video recall is first performed based on the feature embedding vector of the video, and after the recalled video is obtained, the first video is checked from multiple verification dimensions. The first video and the recalled video are reviewed and verified more accurately, and then the similar video judgment is completed, which effectively improves the efficiency of video duplication.

The invention discloses an EEG seizure detection method based on depth channel attention perception, which belongs to the fields of biomedical engineering and machine learning. The present invention introduces an attention mechanism into multi-channel EEG epileptic seizure detection, and trains an end-to-end deep channel attention perception model. The model can not only extract the deep features of the EEG signal, but also learn the contribution scores of each channel to the detection of epilepsy at the same time, so as to realize the dynamic selection of the most relevant EEG channel. Compared with the prior art, the present invention dynamically selects the most relevant EEG channel and expresses epilepsy features cooperatively by combining deep feature extraction and attention mechanism, so that its fusion features have channel-aware ability, and can improve the detection rate of epilepsy while improving the epilepsy detection rate. explanatory.

The invention provides an identifier classification method, and belongs to the technical field of internet content recognition and classification. The method comprises the steps of obtaining a labeledsample set, the labeled sample set having vectors of known class digital content of known class objects and vectors of known class identifiers of different hierarchies of the known class digital content; and obtaining a target class sample set and a target set with identifiers to be classified, and identifying the identifiers to be classified in the target set as target classes the same as the target class sample set according to a transfer learning method in combination with the labeled sample set.

The invention discloses a construction system and method of a knowledge map, and belongs to the technical fields of natural language processing (NLP) and computer information processing. The system includes: a crawler module, which carries out crawling and data cleaning on text; a basic labeling module, which is used for carrying out basic labeling work including subject completion operations; a candidate relationship extraction module, which is used for extracting candidate relationships including candidate relationship sentences and / or relationship entity pairs, a feature extraction module,which is used for carrying out feature extraction; a relationship classifier training module, which is used for carrying out model training according to candidate relationship extraction results and feature extraction results to construct a relationship classifier; and a relationship audit module, which is used for carrying out audit determination on candidate sentence relationships obtained by the relationship classifier, and accordingly adjusting the relationship classifier according to results of audit determination. The system realizes higher relationship extraction capability, reduces costs of manual participation, and improves efficiency of constructing the knowledge map.

The present application proposes a lane marking detection method, device, electronic equipment and storage medium. Wherein, the method includes: inputting the acquired image to be detected into a preset target detection model, obtaining first detection information and second detection information of each grid in the image; Second, the detection information is subjected to non-maximum value suppression processing to obtain the position and category of each lane line boundary point in the image, the position of each lane center point, the corresponding lane width and the lane line category; according to each lane line boundary point in the image The location and category of each lane center point, the lane width corresponding to each lane center point, and the lane line category determine the lane line in the image and the category of each lane line. Therefore, through this lane line detection method, not only the accuracy and accuracy of lane line detection are improved, but also the robustness and accuracy of lane line type detection are improved.

The present application relates to the technical field of automatic driving, and provides a parking space recognition method, device, vehicle-mounted terminal and storage medium. In this application, the corner points of parking spaces are used as the core points of density clustering. Since adjacent corner points of parking spaces are connected by parking space lines, that is, there are detection points of parking space lines between adjacent corner points of parking spaces. Therefore, based on density The adjacent core points of each core point can be determined, and the parking corner points adjacent to each corner point of the parking space can be obtained. Using the characteristics that the adjacent corner points of the parking space are connected by the parking space line, the distorted parking space / special-shaped parking space in the image can be identified, and the parking space can be improved. recognition accuracy. In addition, using each corner point of a parking space as the core point of density clustering can ensure that when the parking space line near the corner point of the parking space is blocked and there are fewer detection points of the parking space line, the parking space corresponding to the corner point of the parking space can also be detected. If it is identified, the parking space inspection rate can be improved and the parking space missed inspection can be avoided.

The invention provides a method, device, terminal server and storage medium for searching for pets. The method includes: receiving the pet-finding information sent by the first client; displaying the pet-looking information sent by the first client to the second client and the third client; receiving the pet-finding information sent by the second client Send the information of the suspected lost pet corresponding to the pet you are looking for; send the information of the suspected lost pet to the third client, so that the catcher can capture the suspected lost pet according to the information of the suspected lost pet . The technical solution of the embodiment of the present invention receives and displays the information of finding pets, and receives the information of suspected lost pets, so that the catcher can catch the suspected lost pets according to the information of suspected lost pets. Increases the retrieval rate of pets.

The invention discloses a mining method, a production method and a system for high-quality data. The mining method of the high-quality data comprises the following steps: based on periodic characteristics and / or based on the same cluster or approximate cluster of a current hotspot, estimating to obtain related data information of high-quality test questions. According to the mining system, the mining method is combined with the crowdsourcing platform. By estimating related data information of high-quality test questions, the high-quality test questions are produced in advance, and the accuratecalling rate of photographed test questions is increased; and the A / B-level commissioning mode is more suitable for production of high-quality test questions, crowdsourcing user resources can be better concentrated, and the return on investment of production is better guaranteed.

The invention discloses experimental device and method for researching mouse and plant seed diffusion, and belongs to the technical field of seed diffusion. The experimental device comprises a plurality of electronic tags, the plurality of electronic tags is arranged in seeds in a one-to-one correspondence manner, the electronic tags are in signal connection with a monitoring device, the electronic tags are RFID radio frequency tag tags, and the monitoring device is a handheld radio frequency detector. The invention also provides a method for carrying out an experiment by adopting the experiment equipment. According to the experimental device and the experimental method, scattered seeds can be accurately and quickly found, the seed recovery rate is increased, and meanwhile, behavior characteristics of different mice and different treatment conditions of the seeds can be known in detail.

The application discloses a method, a device and an electronic device for identifying new words in spam short messages. Wherein, the method includes: obtaining a short message collection; determining a plurality of candidate word sets corresponding to spam short messages; determining the short message category tendency related indicators of the candidate words according to the short message category information; and obtaining the document rarity related indicators of the candidate words; According to the related indicators of SMS category tendency and document rarity, the new word score of spam message is determined; according to the new word score of spam message, the new word of spam message is determined from the candidate word set. With this processing method, most of the new words recalled according to the related indicators of SMS category tendencies are representative of spam messages, avoiding the recall of many normal words; therefore, the recall rate can be effectively improved. At the same time, this processing method makes it possible to recall low-frequency neologisms in spam text messages according to relevant indicators of document rarity; therefore, the recall rate can be effectively improved.

The invention relates to a media resource data processing method and device, electronic equipment and a storage medium, and the method comprises the steps: obtaining full sample data of media resources, the full sample data comprising labeled sample data and unlabeled sample data; training the current media resource alignment model through the labeled sample data to obtain a trained media resource alignment model; based on a processing result of the full sample data by the trained media resource alignment model, performing clustering processing on each group of media resource pairs in the full sample data to obtain a plurality of clusters; and based on the labeled media resource pairs in the plurality of clusters, labeling unlabeled media resource pairs, and adjusting the trained media resource alignment model based on newly added labels to obtain a target media resource alignment model so as to perform alignment identification on the media resource pairs. According to the method, the accuracy of the alignment recognition result of the media resource pair can be improved.