64results about How to "Improve annotation quality" patented technology

Embodiments of the present invention disclose a method and a system for optimizing a speech recognition acoustic model, equipment and a storage media. The method includes: obtaining a labeled text ofa sample speech, and acquiring an identification text obtained by the sample speech based on the current acoustic model; comparing the labeled text with the identification text, and determining errorlabeling information of the labeled text relative to the identification text when a comparison result is not matched; updating the labeled text of the sample speech according to a text update decisioncondition corresponding to the error labeling information; and retraining and optimizing the current acoustic model based on sample speeches of the set amount and current corresponding labeled texts.By using this method, the labeling quality of the labeled text corresponding to the sample speech can be effectively improved, thereby achieving the purpose of optimizing the acoustic model.

The invention discloses a data annotation management method and device, electronic equipment and a readable storage medium. The method comprises the steps: obtaining a reference annotation data set according to to-be-annotated data corresponding to a to-be-annotated task and historical annotation behavior data corresponding to a target annotator; obtaining a first annotation result of the target annotator for the evaluation annotation data and first reference annotation data distributed in the evaluation annotation data, the evaluation annotation data being a part of the to-be-annotated data and having a correct annotation answer, and the first reference annotation data belonging to a reference annotation data set; If the accuracy corresponding to the first annotation result is greater than or equal to a preset accuracy threshold, determining whether to allow the target annotator to continue to execute the to-be-annotated task or not according to a second annotation result of the target annotator for second reference annotation data already distributed in the to-be-annotated data, the second reference annotation data belonging to a reference annotation data set. According to the embodiment of the invention, the quality and efficiency of data annotation can be improved.

The embodiment of the invention also provides a marking method of images, a marking apparatus of images, marking equipment of images, and a storage medium. The marking method of images includes the steps: acquiring a plurality of images to be marked for types, and dividing the images into a plurality of sections; allocating each section of divided images to at least two markers; for each section of divided images, acquiring marking result data of the at least two markers; for each image in each section of images, comparing whether the pre-marked type of the image is the same in the marking result data of the at least two markers, and determining the quantity of images with the same pre-marked type in the marking result data of the at least two markers in each section of images; and for each section of divided images, based on the proportion relation between the determined quantity of the images with the same pre-marked type and the total quantity of the section of images, determining the marked type of the images in the section of images. The marking method of images can improve the marking quality of the marking type of images.

The invention discloses a system automatic check method based on multi-person cooperative image annotation. The method can automatically check the overall image annotation quality, avoid the trouble of special annotation by professionals and greatly reduce the check workload. The method avoids the check step by professionals, and for developers, automatic checking improves the work efficiency andreduces the workload. In addition, evaluation of the data quality is stricter, which lies in that the algorithm of the invention evaluates all annotation data. Moreover, annotation errors are greatlyreduced and thus the annotation quality is improved through multi-person joint annotation.

The invention provides an image annotation method and device, and the method comprises the steps: receiving an image display instruction, and obtaining a to-be-annotated image according to the image display instruction; displaying the image to be labeled in a canvas in a target browser; and monitoring a mouse click event in the canvas in real time, if the mouse click event is monitored in real time, obtaining a to-be-annotated position in the to-be-annotated image according to the mouse click event monitored in real time, and annotating the to-be-annotated position by applying a preset annotation element to obtain an annotated image. The image annotation quality and efficiency can be improved, and then the accuracy and efficiency of transaction certificate image classification or intelligent driving image detection achieved by applying the image annotation result can be improved.

The invention relates to an image labeling method and device, an electronic device and a storage medium. The method comprises: recognizing at least one set of collected to-be-labeled images, determining an image area where a target object is located in the at least one set of to-be-labeled images, each set of to-be-labeled images comprising at least two images, and the indication states of the target object in the at least two images being different; and according to the image area where the target object is located, determining labeling information of at least one group of to-be-labeled images. According to the embodiment of the invention, efficient and automatic labeling images of the image can be realized.

The invention discloses a semi-automatic word segmentation corpus labeling and training device, which aims to overcome the defects of the corpora used during the word segmentation corpus labeling and training process. The device of the invention is realized through the following technical schemes of using a text corpus annotation preparation module for managing the to-be-annotated corpora and the segmented word corpora; based on a plurality of word segmentation algorithms, such as the bidirectional maximum matching word segmentation based on an integrated dictionary, CRF, JIEBA, etc., submitting the word segmentation annotation work of the raw corpus to a semi-automatic corpus word segmentation annotation module; creating the segmented word tagging tasks, selecting a labeling applicable algorithm model, carrying out the automatic annotations, on the basis of automatic labeling result fusion, feeding back a training model corpus and a labeling model generated by the text corpus labeling preparation module to the feedback model learning training module; selecting and carrying out model learning training, calling a unified training model interface to generate a core dictionary, updating a word segmentation training model table, establishing a labeling algorithm comprehensive evaluation model to evaluate a model labeling effect, so that a new word segmentation labeling task is completed.

The invention discloses a point cloud entity labeling system, method and device, a point cloud entity labeling task setting method and device and electronic equipment. The point cloud entity labelingmethod comprises the following steps: displaying a three-dimensional scene graph corresponding to point cloud data to be labeled; determining two-dimensional bounding box information of a target entity in the three-dimensional scene graph; determining point cloud data corresponding to the target entity according to the two-dimensional bounding box; determining entity type information of the targetentity; and marking the point cloud data corresponding to the target entity as the entity type information. With application of the processing mode, labeling personnel only need to frame the target entity through the two-dimensional bounding box at a certain view angle of the three-dimensional scene graph, the system can automatically analyze the point cloud position which the two-dimensional bounding box hopes to frame, and the labeling personnel are automatically helped to locate the point cloud data corresponding to the target entity. Therefore, the point cloud entity labeling efficiency can be effectively improved, and the point cloud entity labeling quality can be effectively improved.

The invention discloses a data processing method. The data processing method comprises the steps of collecting data; and performing automatic tagging on the data. According to the data processing method, the collected data is automatically tagged, so that automatic data processing is realized, the manpower cost is reduced, and the tagging quality is high. In addition, the invention discloses a data processing apparatus and a computer readable storage medium.

The invention provides a data labeling method and device and data processing equipment, and the method comprises the steps: carrying out the at least one iteration processing of a classification model, so as to enable the accuracy of the classification model to meet a preset condition; and processing at least one part of the plurality of pieces of to-be-labeled data by utilizing the obtained classification model to obtain an automatic labeling result, wherein each iterative processing comprises the following steps of: respectively inputting other to-be-labeled data, except a target data set, in a plurality of pieces of to-be-labeled data into the classification model to obtain a classification result; selecting at least part of the to-be-labeled data of which the confidence coefficient ofthe classification result is within a preset range from the other to-be-labeled data, and adding the selected to-be-labeled data into a target data set; and training a classification module accordingto the manual labeling result of the to-be-annotated data in the target data set. Therefore, automatic labeling of batch data can be realized under the condition of improving the data labeling quality.

The invention relates to the technical field of artificial intelligence, in particular to a model processing method, a model processing device, a computer readable storage medium and electronic equipment. The model processing method provided by the embodiment of the invention comprises the steps of adding an initial annotation corpus into an initial training set, and performing training by utilizing the initial training set to obtain a language model; obtaining a prediction result output by the language model, and extracting feature information of an error prediction result in the prediction result; when it is judged that the feature information is related to the initial annotation corpus, generating a preset number of simulation annotation corpuses according to the feature information; and adding the simulation annotation corpus into the initial training set, and continuing to train the language model by using the initial training set added with the simulation annotation corpus. According to the model processing method provided by the embodiment of the invention, the data annotation quality can be improved, and the prediction effect of the language model is improved.

The embodiment of the invention discloses a region labeling method, device and equipment in the field of artificial intelligence and a storage medium. The method comprises the steps: acquiring N labeling points labeled for the boundary of a coverage region of a target curved text; determining four corner points of a to-be-labeled target curved region in the N labeling points; according to the fourcorner points, selecting marking points used for fitting a first curve from the N marking points to form a first marking point set, selecting marking points used for fitting a second curve from the Nmarking points to form a second marking point set, wherein the first curve and the second curve are two curve boundaries opposite to the target curved area; fitting a first curve according to the marking points in the first marking point set, and fitting a second curve according to the marking points in the second marking point set; and constructing a target curved region based on the first curveand the second curve. According to the method, the annotation quality of the curved text annotation area can be improved, and the annotation time cost can be reduced.

The invention relates to a power grid image intelligent annotation crowdsourcing platform and a working method, and belongs to the technical field of power grid image data processing. The working method comprises the following steps: collecting to-be-annotated picture collection, performing initial annotation, performing manual adjustment annotation, performing difference re-annotation and data storage. The power grid image intelligent annotation crowdsourcing platform comprises a to-be-annotated image collection module, an initial annotation module, a manual adjustment annotation module, a difference re-annotation module and a data storage module, and is used for executing the power grid image intelligent annotation crowdsourcing platform working method. According to the invention, the preset model is used to carry out initial annotation on the data; meanwhile, platform crowdsourcing is used for manually adjusting annotation; multi-person cooperation is achieved, the annotation efficiency is improved, the unqualified annotation result is modified according to IOU parameters, the annotation quality is effectively improved, meanwhile, the data classification and arrangement functionmeets the requirement for specific model training for a certain hidden danger, and powerful data support is provided for model training and model precision improvement in the future.

The embodiment of the invention provides a three-dimensional information processing method and device. The method comprises the steps of obtaining target image data for a target three-dimensional object; determining target two-dimensional projection information of the target three-dimensional object in the target image data; and obtaining target three-dimensional information of the target three-dimensional object according to the target two-dimensional projection information. According to the embodiment of the invention, the optimization of three-dimensional information labeling is realized, the three-dimensional information is obtained based on the two-dimensional information of the three-dimensional object, the dependence on a laser radar is avoided, and the labeling efficiency and quality of the three-dimensional information are improved.

The embodiment of the invention provides a video target labeling method and device and electronic equipment, and belongs to the technical field of data processing, and the method comprises the steps:obtaining a first labeling box formed for a first type of key frames in a target video; training a target detection module and a target tracking module for executing a labeling function based on the first type of key frames and the first labeling box after whether the number and quality of the first type of key frames meet preset requirements or not; the trained target detection module and the target tracking module are utilized to perform annotation operation on the remaining first type of key frames which are not annotated, and a second annotation box and a third annotation box are generated; after the second annotation box and the third annotation box are confirmed, marking the second annotation box and the third annotation box as a first annotation box again. Through the processing scheme of the invention, the labeling efficiency of the video target can be improved.

The invention relates to an audio annotation error detection method and device, computer equipment and a storage medium. The method comprises the following steps: obtaining a labeling text obtained bylabeling audio data by labeling personnel; carrying out error detection on the annotation text, and generating error detection information when at least one of a word error in the annotation text anda statement error in the annotation text is determined through error detection; and outputting the error detection information. According to the embodiment of the invention, a terminal generates theerror detection information if detecting that the annotation text is wrong in the process of annotating the audio data by the annotation personnel, and can remind the annotation personnel in real time, so that the annotation personnel can correct in time, and the annotation quality is improved.

The invention discloses an automatic voice data labeling method and system for voice recognition, and particularly relates to the field of voice recognition. The system comprises a mute detection module, a volume screening module, a length screening module, a voice recognition module, a recognition result judgment module and a manual proofreading module, the mute detection module splits each voiceinto a plurality of voice segments through a mute detection algorithm, and the volume screening module is used for screening out voices meeting the requirements through a volume threshold value and removing voices not meeting the requirements. The invention discloses a combined system of multiple modules. According to the system, speech preprocessing and speech recognition are carried out, by a public cloud mode, recognition result judgment manual proofreading are carried out, voice data annotation is constructed, after multiple times of iteration of the processes, a new corpus is continuously trained, high-quality corpus data is obtained, manpower is reduced, the voice data annotation quality is improved, and the problems that the manual annotation period is long, the cost is high and the efficiency is low are solved.

The embodiment of the invention discloses a data screening method and device, a storage medium and electronic equipment, and the method comprises the steps: obtaining a sample identification of a datasample marked with a category, and obtaining an identification vector corresponding to the sample identification; then, taking the category number of the labeled categories as a clustering category number to carry out clustering processing on the identification vector; then, for each clustering category, obtaining the similarity between the clustering center identification vector and the non-clustering center identification vector; then, determining a target non-clustering center identification vector of which the similarity with the clustering center identification vector does not reach a preset similarity in each clustering category, and judging the labeled category of the data sample represented by the target non-clustering center identification vector as labeling noise; and finally, filtering out the data sample corresponding to the target non-clustering center identification vector in each clustering category, thereby achieving the purpose of improving the labeling quality of thedata sample, and providing a high-quality data sample for machine learning.

The invention discloses a complex data labeling method and device. The method comprises the steps of splitting a labeling task into sub-tasks according to labeling task requirements; designing an annotation rule for each subtask; designing a logic relationship among the sub-tasks; and executing each subtask and obtaining a final labeling result. The device comprises a splitting module, an annotation rule module, a logic relation module and an annotation result module. According to the invention, the complex data annotation task is used as a continuous data flow processing process, the annotation and auditing standard working module is defined, the data flow combination relation, the logical operation relation and the access operation relation are introduced, the complex data annotation task can be effectively guided to be analyzed and disassembled, and a data task annotation scheme is reasonably planned. By means of the platform tool with the functions, flexible organization of the labeling process can be conducted, the data labeling cost is effectively reduced, the data labeling quality is improved, and management of complex labeling data is effectively conducted.

The invention discloses an information labeling method, device and system for medical image data. The method comprises the following steps: acquiring to-be-labeled medical image data; inputting medical image data to be labeled into an image pre-labeling model, obtaining pre-labeling information, and using the image pre-labeling model for conducting feature extraction on the medical image data to obtain a neural network model of the pre-labeling information; in response to displaying the medical image data containing the pre-annotation information on the target platform, obtaining adjustment information of the target platform for the pre-annotation information; and processing the adjustment information and the pre-annotation information to obtain target annotation information of the to-be-annotated medical image data. According to the method, the medical image data is labeled through the model, and the labeled information can be adjusted in combination with a man-machine interaction mode, so that the information labeling efficiency and quality of the medical image are improved.

The invention provides a point cloud instance labeling method and device, electronic equipment, a computer readable storage medium and a computer program product, relates to the technical field of computers, in particular to the technical field of data labeling and computer vision, and can be applied to a cloud platform. According to the implementation scheme, the method comprises the steps of obtaining to-be-labeled point cloud data; marking one or more instances on the point cloud data, wherein each instance comprises one or more point cloud areas in the point cloud data; and determining attributes of the one or more instances, the attributes being one or more of a group of attributes preset based on the scene corresponding to the point cloud.

The invention relates to a system and a method for revising samples for neural network training. The system comprises a service terminal and a client front end; the service terminal is set to store samples, distribute the samples to the client front end, receive and store a processing result of the client front end, and generate statistical display according to the processing result; and the client front end is configured to receive the sample, execute revision processing and transmit a processing result to the service terminal. The service terminal comprises a storage module and a statisticsmodule; the client front end comprises an annotation module and an auditing module; the annotation module comprises a preprocessing unit and a fine processing unit; the auditing module can also be setto score the revision quality. According to the invention, a plurality of client front ends can revise samples in the same data set at the same time, so the revising progress of the samples is accelerated, and the time cost is saved; an automatic pretreatment unit is arranged, so that the workload of subsequent fine treatment is reduced; and an auditing and scoring mechanism is set, so that the enthusiasm of the annotator is not struck, and the reliability of the data set sample is improved.

The invention provides a sample data processing method and device, a server and a storage medium. The method comprises the steps that multiple pieces of target sample data are acquired, and the target sample data carry annotation information; a marking information entropy of the target sample data is determined according to marking information carried by the target sample data; and according to the labeling information entropy of the target sample data, first target data of which the labeling quality meets a preset quality requirement is determined from the plurality of target sample data. In the embodiment of the specification, the consistency degree of different labeling sources for labeling the same sample data is quantified by firstly determining the labeling information entropy capable of reflecting the uncertainty of the labeling information of the target sample data; therefore, the target sample data with relatively high labeling quality can be screened out according to the labeling information entropy to be used as the first target data, so that the data with relatively high labeling quality can be efficiently and accurately screened out from the plurality of target sample data, and the data error is reduced.

The embodiment of the invention provides a visual small target automatic labeling method and device and electronic equipment, and belongs to the technical field of data processing, and the method comprises the steps: classifying video frames containing a visual small target, obtaining a first type of video frames and a second type of video frames, and enabling one or more second type of video frames to be included between each two adjacent first type of video frames; respectively executing target detection and tracking processing on the first type of video frames and the second type of video frames in a sequential mode to obtain a first detection result; performing target detection and tracking processing on the first type of video frames and the second type of video frames in an invertedorder mode to obtain a second detection result; and performing fusion processing on the first detection result and the second detection result by adopting a maximum suppression mode to obtain a finallabeling result for automatically labeling the small target. Through the processing scheme disclosed by the invention, the automatic labeling efficiency of the visual small target can be improved.

The invention discloses a data labeling method and a related device. The data labeling method comprises the following steps: inputting to-be-labeled data into a hierarchical labeling system, and determining a data label of the to-be-labeled data; if the to-be-labeled data is a non-class data label in the sub-label layer corresponding to the determined data label, searching for a data label meetinga preset condition from other data labels located at the same level as the determined data label; and if the data label meeting the preset condition is found, correspondingly labeling the data labelof the to-be-labeled data by using the sub-label of the data label meeting the preset condition and the data label of the upper layer of the sub-label of the data label meeting the preset condition. According to the scheme, the error annotation processing efficiency can be improved, and the data annotation quality is improved.

The invention relates to the technical field of automatic driving, in particular to a data set establishment method and a vehicle. The method comprises the steps: obtaining a millimeter-wave radar image and a laser radar image; performing space and time calibration on the millimeter-wave radar and the laser radar; constructing a deep neural network for speculating the laser radar image, and generating a target speculation result of the laser radar by using the deep neural network; projecting a target speculation result to the matched millimeter-wave radar image to serve as a pseudo label of the millimeter-wave radar image; generating a radar target confidence coefficient according to the millimeter wave radar local signal of the area where the pseudo label is located; and establishing a millimeter wave radar data set according to the target confidence and the pseudo label. Compared with traditional manual labeling, automatic labeling can be achieved by means of the laser radar, so thatthe labeling efficiency is improved; and target recognition with the high recall rate can be achieved based on the deep neural network of model integration, and a false positive target detection boxis filtered out, so that the labeling quality is improved.

The invention relates to an image labeling method, a classification model training method and computer equipment. The image labeling method comprises the steps of obtaining a to-be-labelled image and an labeled image, wherein the labelled image comprises a target label; determining to-be-labelled image feature parameters corresponding to the to-be-labelled image and labelled image feature parameters corresponding to the labelled image, where the labelled image feature parameter comprises a target label; and determining a label corresponding to the to-be-labeled image according to the feature parameters of the to-be-labeled image and the feature parameters of the labeled image. According to the invention, the to-be-labeled image is labeled through the labeled image, and the labeling quality of the to-be-labeled image is unified with the labeling quality of the labeled image, so that the labeling quality of the to-be-labeled image is relatively high, repeated labeling adjustment is not needed, and the labeling efficiency is greatly improved.

The embodiment of the invention discloses an image recognition method and device and a computer readable storage medium. According to the embodiment of the invention, the method comprises the steps of after an image sample set is acquired, adopting a preset recognition model to perform feature extraction on the image samples in the image sample set to obtain an image feature set, then using the image samples as data nodes to construct the neighbor graph according to the image feature set, and then correcting the basic labels of the image samples based on the neighbor graph; obtaining a corrected image sample set, then training a preset recognition model by adopting the corrected image sample set, and recognizing the to-be-recognized image through the trained recognition model. According to the scheme, the image recognition accuracy can be improved.

The invention discloses a data labeling system and method based on an intelligent distribution algorithm, and specifically relates to the field of data processing. The system comprises a data analysismodule, a feature acquisition module and an intelligent distribution module. The output end of the data analysis module is connected with the input end of the feature acquisition module, and the output end of the feature acquisition module is connected with the input end of the intelligent distribution module. The method comprises the following specific processing steps: screening small-scale representative and instructive key data as advanced data by using a data analysis module; carrying out trial labeling, accurate labeling and analysis on the 'leading data' by a labeling person to obtaina 'standard answer', dynamically matching, and then taking the exclusive labeling feature of each labeling person; and using an intelligent distribution module to intelligently distribute the remaining data. According to the method, the manual processing error rate of the data can be reduced by utilizing an intelligent distribution algorithm, and in the manual processing task of the text type data, the manual error rate can be reduced by about 20-30%.