111results about How to "Reduce the cost of manual labeling" patented technology

The present invention relates to a method and an apparatus for processing lane line data, a computer device and a storage medium. The method includes: acquiring and dividing three-dimensional point cloud data of a to-be-processed road; and processing the three-dimensional point cloud data of each segment after the segmentation to obtain a two-dimensional gray image of the three-dimensional point cloud data of each segment; using a pre-trained deep neural network model, respectively extracting a lane line region and a lane line attributed such as a dotted or full lane line in each two-dimensional gray image to obtain a corresponding lane line region map; and according to the three-dimensional point cloud data corresponding to the lane line region map and the lane line attribute such as thedotted or full lane line, splicing the lane line region map to obtain lane line data of the to-be-processed road. By adopting this method, the processing efficiency is improved, and compared with ordinary machine learning, it is not easily affected by interference items such as characters and cars in the three-dimensional point cloud data, and the accuracy of lane line region extraction is improved.

The invention discloses an image classification model training method and image processing method and device. The method comprises the steps of obtaining a to-be-trained image; when the first model parameter of the offset network to be trained is fixed, obtaining first prediction class marking information of the image to be trained through the image classification network to be trained; determining a second model parameter by adopting a classification loss function according to the image content category information and the first prediction category annotation information; When a second modelparameter of the to-be-trained image classification network is fixed, obtaining second prediction class marking information of the to-be-trained image through the to-be-trained offset network; Determining a third model parameter by adopting a classification loss function according to the image content category information and the second prediction category annotation information; And obtaining animage semantic segmentation network model according to the second model parameter and the third model parameter. The invention further discloses an image processing method and device. According to themethod and the device, manual pixel level marking is not needed, so that the manual marking cost is reduced, and the model training efficiency is improved.

The invention discloses a fundus image focus area labeling method based on deep learning, and the method comprises the steps: selecting a sample, and carrying out the preprocessing: carrying out the cutting of an image, and carrying out the horizontal and vertical overturning and normalization processing; constructing a convolutional neural network and a deconvolutional neural network as an imagegenerator, inputting the preprocessed color fundus image, and outputting a corresponding focus probability graph; constructing a convolutional neural network as a discriminator, inputting and generating a focus image and a real focus image, and outputting a probability that the focus image is judged to be a real image; alternately training the generation network and the discrimination network until a satisfaction result can be generated; and marking a focus area in the fundus image according to the generated focus probability graph. The fundus image focus probability graph is generated by using the deep convolutional neural network, and the fundus image focus area is automatically labeled. The automatic labeling can provide an auxiliary basis for the diagnosis of doctors, and meanwhile, the cost of manual labeling can be greatly reduced.

The invention requests to protect a remote sensing image classification method based on active learning and a convolutional neural network, and the method comprises: carrying out the waveband processing of a hyperspectral remote sensing image through a principal component analysis method, and then processing the image into blocks; dividing the data into a training set, an unmarked sample set, a verification set and a test set according to a certain proportion; and training the convolutional neural network by using the training set, predicting the category to which the sample in the unmarked sample set belongs, and introducing active learning to evaluate the sample; and then sorting the evaluation results, selecting samples with low confidence, assigning labels to the samples by experts, and automatically assigning labels to the samples with high confidence by a computer. A high-quality training sample set is constructed by adjusting a prediction label coefficient, and a classifier model is iteratively optimized by using the selected training sample set. And stopping iteration when a stop condition is met, and outputting a final classification result.

The invention provides an emotion analysis method and system of an active learning framework based on committee query. The method comprises: acquiring training corpus, selecting sample data from the training corpus and marking the sample data; forming a training set, constructing an initial model according to the training set, carrying out data classification according to the user emotion representation; selecting an algorithm according to the committee voting; selecting sample data from the unlabeled training corpus, labeling the sample data, supplementing the sample data into the training set, carrying out update iterative learning on the initial model according to the supplemented training set, obtaining an emotion analysis model, carrying out emotion analysis on an input text through the emotion analysis model, and obtaining an emotion analysis result; According to the method, active learning is introduced, in the emotion analysis task, under the condition that training corpora isreduced, the cost of manual marking is reduced, and meanwhile the performance of the model reaches the expectation.

The invention discloses an associated question aggregation model generation method and device, a question-and-answer mode aggregation method and device as well as equipment. The methods include the following steps: obtaining a first quantity of basic training samples according to network behavior data of at least two users, and training a first machine learning model by using the basic training samples to obtain a basic semantic matching model; migrating a semantic representation layer in the basic semantic matching model to a second machine learning model, and training the second machine learning model according to a second quantity of pre-labeled associated question pairs to obtain an associated question aggregation model. According to the embodiment of the invention, the associated question aggregation model that aggregates answers of questions with consistent meanings can be obtained, the basic semantic matching model trained by the network behavior data of the users is adopted toperform migration learning to further generate the associated question aggregation model, so that the number of manually labeled samples and the manual labeling costs can be greatly reduced, and the answer satisfaction rate of the questions in a question-and-answer community can be optimized.

The invention provides a semi-supervised remote sensing image target detection and segmentation method based on a class activation graph. The method comprises the following steps of firstly, generating a classification annotation data set by utilizing given remote sensing image annotation data, training a global average pooling (GAP) classification convolutional neural network model, and constructing a convolutional neural network model capable of generating a class activation map (CAM) by utilizing a weight superposition principle of a feature map; secondly, performing semi-supervised training on the target detection and segmentation model by taking the class activation graph and the real label as training targets through data enhancement; then, verifying the target detection and segmentation model by using a test set with a real label to obtain a model with higher detection and segmentation precision. and finally, under the condition that only a small amount of annotation data is used for training, the method has good remote sensing image target detection and segmentation effects.

The invention provides a test paper correcting method and a test paper correcting device. The test paper correction method comprises the steps of obtaining a test question image and an answering arearule; determining an answering area image according to the answering area rule and the test question image; identifying answer text information in the answer area image; comparing the answer text information with standard answer information corresponding to the test question image; correcting the test paper according to a comparison result of the answer text information and the standard answer information. Therefore, intelligent correction of paper test paper objective questions is realized; the problem that in a common examination process, if no answer sheet exists, paper can only be manuallymarked by a teacher is avoided; the time of teachers can be greatly saved, the test paper correcting efficiency is improved, errors are reduced; compared with the prior art, correction can be more flexibly carried out in a fine-grained mode, meanwhile, the problem that multiple sub-answers exist in one question is solved, the correcting accuracy is greatly improved, and the correcting efficiencyand subject flexibility are improved.

The invention discloses an interest mining method for user browsing behaviors. For users u1,u2 and u3, within an appointed time, the user u1 accesses tags t1, t2 and t3, the user u2 accesses tags t2, and the user t3 accesses tags t2 and t3. The method comprises the following steps: (1), certain typical websites inside each interest label are labeled, and the default of weight of interest corresponding to the labeled websites tag->interest is 1.0; (2), according two graph models established between users and website tags accessed by users within an appointed time and through n turns of random walk, wherein n refers to positive integers, the results of n turns of random walk are collected, and the weight of user-> tag can be calculated; (3), user->interest is obtained through the product of the user->tag obtained in the step (2) and the user->interest obtained in the step (3), and user->interest refers to a confidence coefficient of each user to the interest tag; (4), a threshold value a is arranged, and when the confidence coefficient of the user->interest is larger than a, the interest tag for the user is predicted.

The invention relates to a retrieval method, in particular to a keyword-to-enterprise retrieval method based on semi-supervised learning, and belongs to the field of information retrieval. The self-training method comprises the following steps: firstly, training a model by using initial annotation data, then identifying part of unannotated data by using the model, and adding the identified part ofunannotated data into an annotation data set to serve as new training data; and obtaining a final model through multiple rounds of automatic data annotation and iterative training learning. The semi-supervised learning method can greatly reduce the manual annotation cost and improve the retrieval matching efficiency.

Embodiments of the invention provide a video data processing method and apparatus. The method comprises the steps of obtaining comment information added for video data by a user; identifying emotion information expressed by the comment information; according to the emotion information, marking feature video frames in the video data; and generating feature video clips in the video data according tothe density of the feature video frames. The comment information of the user is objective existence relative to a technical staff, so that the subjectivity of marking the video data can be greatly reduced, the identification accuracy of the emotion information is improved, and the manual marking cost is greatly reduced; in addition, the feature video clips are clustered by using the density, so that the discreteness and fragments of time can be avoided, the clustered feature video clips can be ensured to have a certain continuity under the condition of ensuring the wonderfulness degree of thefeature video clips, and the watching quality is improved; and moreover, due to the clustering by using the density, complex distance calculation can be avoided and the processing speed is increased.

The invention provides a remote sensing image target detection method and system based on semi-supervised iterative learning, and the method comprises the steps: collecting a remote sensing image target detection data set which comprises a label data set and a mass label-free data set; training a target detector model based on the labeled data set, and obtaining a trained initial target detector model; dividing the massive label-free data set into a plurality of label-free data subsets; and performing iterative optimization training on the initial target detector model based on the labeled data set and the plurality of unlabeled data subsets to obtain a final target detector model. According to the invention, a small amount of tagged remote sensing image data is used, and the target detection precision is greatly improved and the manual tagging cost is reduced under the combined auxiliary optimization of massive non-tagged data.

Embodiments of the invention disclose an information generation method and apparatus. A specific embodiment of the method comprises the steps of obtaining to-be-analyzed information according to a target keyword; inputting the to-be-analyzed information to a pre-built emotional analysis model to generate an emotional tendency information of the to-be-analyzed information, wherein the emotional analysis model is obtained through the following training steps of obtaining unlabeled sample data and labeled sample data; generating label information corresponding to the unlabeled sample data by using a pre-built label generation model, wherein the unlabeled sample data and the generated label information are taken as expanded sample data, and the label generation model is used for representinga corresponding relationship between the unlabeled sample data and the label information; and performing training by using the labeled sample data and the expanded sample data to obtain the emotionalanalysis model. According to the method and the apparatus, the expanded sample data is automatically generated; sample set data is expanded; the manual labeling cost is reduced; and the accuracy of the information generated by the emotional analysis model is improved.

The invention discloses a model training method and device, which can be used for determining a training sample containing any entity from each segment of text related to a target field based on each entity in a constructed knowledge base, and carrying out sequence labeling on each training sample according to the position of the entity in each training sample and the entity attribute of the entity. Then, the pre-trained language model is further trained through each training sample and the label thereof. According to the method, the training samples are determined on the basis of the entities in the constructed knowledge base, and the training samples are automatically labeled, so that the manual labeling cost is saved.

The invention discloses a network model training method and device and a computer readable storage medium, and the method comprises the steps: sequentially carrying out the self-supervised pre-training, domain data fine tuning and knowledge distillation of a pre-training model, i.e., carrying out the unsupervised pre-training of a super-large-scale neural network model through employing mass data, and carrying out the fine tuning of the pre-training model through employing a limited labeled sample, and compressing the fine-tuned super-large model into a target model by using a knowledge distillation method, so as to meet the deployment requirements of target equipment. On the basis, the dependence on annotation data can be reduced, the manual annotation cost can be reduced, the problem that the manual annotation data cost is high can be solved, the universality and generalization of the model can be improved, and the target task precision of the output target model exceeds that of an original customized model.

The invention discloses a method and a device for detecting whether a pedestrian wears a safety helmet or not. The method comprises the steps that a pedestrian safety helmet detection model is trained; in the training process, a sample image used for training a pedestrian safety helmet detection model is processed in a secondary marking mode, so that sample features with missing labels in a training sample do not participate in network weight parameter updating when participating in neural network training. According to the method, a human body target rectangular frame and a safety helmet target rectangular frame are detected through the pedestrian safety helmet detection model, then human body features in the human body target rectangular frame are extracted through a feature extraction network, and target pedestrians are tracked and matched with safety helmets in continuous video frames based on the human body features. According to the scheme, the high-accuracy detection model can be trained on the basis of label missing unbalanced data, and multi-frame detection is carried out on the safety helmet wearing condition of each pedestrian by adopting a target tracking method, so that the false alarm rate is greatly reduced.

The invention relates to a hip joint X-ray image segmentation method and system based on a local vision clue. The method comprises the following steps: preprocessing the image data in the data set; learning a first-layer convolution block feature of the standard U-net network; obtaining a rough label by using a Sobel operator; extracting a local visual clue LVC through the first convolution blockfeature and the rough label; combining LVC with an S-loss loss function to guide a U-net network to output a preliminary segmentation result graph; utilizing LVC to generate LVC local visual guidance,outputting a sampling offset field with the preliminary segmentation result image through a deformable spatial transformation network, and resampling the preliminary segmentation result image to obtain a final segmentation result of the image. According to the method, the problem that a supervised learning model is easily influenced by label noise in the medical image segmentation process is solved, and the complexity of medical image label generation is reduced.

The invention discloses a commodity label labeling method and device, equipment, a medium and a product. The method comprises the following steps: acquiring an image feature vector of a commodity picture of a commodity object; the image feature vector and a preset template text vector are spliced to form image-text coding information, the template text vector comprises a first mark and a second mark, the first mark is used for indicating redundant semantic information generated by a mark model which is trained to be converged in the pre-training process, and the second mark is used for indicating redundant semantic information generated by the mark model which is trained to be converged in the pre-training process; the second mark is used for indicating the sequence position information of the commodity label produced by the labeling model; performing task word prediction according to the image-text coding information by adopting the labeling model to obtain a commodity label corresponding to the second label; and labeling the commodity object with the commodity label. The commodity label is output according to the semantic information of the commodity picture by means of the labeling model, and the labeling efficiency of the commodity label is improved.

The invention discloses a data-augmented deep semi-supervised extreme learning image classification method and system. The method comprises: performing feature extraction on a training image by adopting a deep convolutional network model; finely adjusting and optimizing the deep convolutional network model based on part of artificial label data, and generating a pseudo label for the training image without the label; fusing the high-level semantic features extracted from the training image with the low-layer and shallow-layer structural features to obtain fused image features; augmenting the fused image features and labels of the training images by adopting a random linear interpolation technology; and training a single hidden layer feedforward neural network for the augmented fusion image features and labels, and replacing a full connection layer in the deep convolutional network model to obtain a final image classification and recognition network model. The invention has the advantages of low manual marking requirement, robust noise interference resistance, good classification and recognition performance, high task expansibility and data augmentation.

The invention discloses an iterative construction method and device for a military scenario text event extraction corpus. The method comprises the following steps of 1, preprocessing, and obtaining anoriginal data set represented by a word sequence; 2, constructing a seed data set, defining an event template, constructing an event trigger word dictionary, forming the seed data set through manualannotation, and dividing the seed data set into a seed training set and a test set; 3, training a model, training a machine learning model by using the seed training set, testing the model by using the test set, and optimizing the model parameters according to a test result to obtain a first learning model; 4, selecting an unlabeled training corpus, and inputting the unlabeled training corpus intothe first learning model to obtain a prediction result set; 5, correcting the prediction result set to form a new annotation corpus; and 6, through the continuous iteration, generating the training sets in sequence to form the event extraction corpus. According to the iterative construction method for the military scenario text event extraction corpus, the corpus construction efficiency is improved, the manual annotation cost is reduced, and the relatively higher corpus annotation accuracy is obtained.

The invention belongs to the technical field of machine learning, and discloses a model dynamic training, checking, updating maintenance and utilization method under a cloud platform. The resource manager obtains a workflow table according to different service requests and historical model training results; The model is verified by the verification data, and the result is notified to the resourcemanager; The service manager releases resources; And the resource manager re-issues the service to the scheduler of the service pool, and starts a new computing module for the service module. According to the invention, a lot of manual labeling cost is reduced; A large amount of model monitoring statistical data is obtained through the resource management module and used for solving the problem ofexploring and utilizing balance of the model monitoring statistical data and the original data, the model trained in the process and the original data are multiplexed to a certain extent, and after alarge amount of data is accumulated, a set of efficient workflow can be completed through excellent intelligent arrangement of the model monitoring statistical data. According to the method, hardwareresources are virtualized by utilizing the characteristics of a cloud platform, the characteristics of all functional modules are fully utilized, and the resources are utilized to the maximum extent.

The invention discloses an intelligent question and answer interaction method and system based on machine reading understanding, and the system comprises: a first text word segmentation module which segments an original text into a sequence taking words and characters as units; a statistical mining module, a data translation module, a data layering module, a vector representation module, a first model training module and a statistical rule mining module. The legal provision pushing system comprises a second text word segmentation module, a new word discovery module and a second model trainingmodule. The interaction method comprises the following steps: respectively inputting an original text and a questioning text into a reading understanding system and a legal provision pushing system, wherein the original text and the questioning text are default Chinese texts; preprocessing the input original text and questioning text; inputting the preprocessed text into a model training module for training; and outputting a prediction result. By establishing an intelligent question-answer interaction system, the public service capability in the judicial field is improved, and the manual customer service cost is reduced.

The invention discloses a text annotation method and device, and the method comprises the steps: obtaining a to-be-annotated target text, determining a specific field to which the target text belongs,and carrying out the semantic slot annotation of each entry in the target text through the structured data in the specific field. The adopted annotation basis is the structured data of the specific field to which the target text belongs. Since the structured data comprises each field and the value under each field, and each field generally represents the semantic slot in the specific field, the structured data can be used for carrying out semantic slot marking on each entry in the target text without manual marking, so that the manual marking cost is reduced. Besides, the corresponding relation between the fields and the field values of the structured data is fixed, so that semantic slot annotation is carried out based on the structured data, and the consistency of annotation results canbe ensured.

The invention provides an image search method and system, and the method comprises the steps: carrying out the matching in a database according to a search statement and / or a search word of a search instruction under the condition that the search instruction is obtained, wherein the database stores a target image and a label generated according to the target image; and outputting the target imagecorresponding to the matched label, and searching the target image through the description statement with similar semantics by the user because the database contains the label of the target image description statement and the description statement contains more complete semantic description for the image scene. The method supports statement retrieval, enriches image search modes, improves image search efficiency and quality, and enhances user image search experience.

The embodiment of the invention relates to the field of intelligent insurance, and discloses a method and device for generating a vehicle appearance part recognition sample, a medium and a server. The method comprises the following steps: acquiring three-dimensional model information of a to-be-processed vehicle; obtaining an ID map of the vehicle part according to the three-dimensional model information; setting rendering data, wherein the rendering data comprises internal and external parameters, light source parameters and background images of the observation camera; rendering the vehicle three-dimensional model into a two-dimensional vehicle image according to the rendering data and the ID chartlet of the vehicle part, and obtaining a vehicle part mask on the vehicle image; and training a vehicle image style converter by using a CycleGAN network, and further migrating the rendered vehicle image into a real style vehicle image while keeping the position of the component mask. According to the implementation of the invention, the vehicle picture sample with the accurate mask-level part label is automatically generated, and the manual label cost of a vehicle part identification task can be remarkably saved.

The invention provides an artificial intelligence data labeling method and device. The method comprises: acquiring a to-be-labeled data set; obtaining an AI label with the highest probability score of each piece of to-be-labeled data and a probability score based on the established AI model; for any to-be-labeled data, determining whether the probability score is greater than a first preset threshold; when it is determined that the probability score is larger than a first preset threshold value and sampling inspection is carried out on the to-be-labeled data, or when it is determined that the probability score is not larger than the first preset threshold value, labeling an artificial label on the to-be-labeled data; and when it is determined that the probability score is greater than a first preset threshold and it is determined that the to-be-labeled data is not sampled, labeling the to-be-labeled data by using the acquired AI label with the highest probability score. According to the method, the manual marking cost and the implementation time cost are saved, and marking errors caused by human subjective factors and marking personnel technical backgrounds are reduced.

The invention discloses a news emotion entity extraction method based on remote supervision. The method comprises the steps: crawling official news website news corpus and caching the corpus to a local warehouse; preprocessing the crawled news corpus to obtain news corpus segmented into sentences; constructing a key entity knowledge base, and automatically labeling the news corpus segmented into sentences according to the knowledge base; training an emotion sentence extraction model by using the labeled news corpus to enable the model to have the capability of performing automatic emotion judgment on the input sentences; using the extracted sentiment sentences, and taking the sentiment sentences as a training set of a sentiment entity extraction model for training; and crawling the news corpus, segmenting the news corpus into sentences, inputting the news corpus segmented into sentences into the trained emotion sentence extraction model to extract emotion sentences, and inputting the extracted emotion sentences into the trained emotion entity extraction model to obtain emotion entities. According to the method, a data set with noise is generated for a large number of samples in a remote supervision mode for model training, and the model training efficiency is improved.

The invention provides a target tracking method and system based on unmarked video training, and the method comprises the steps: carrying out the unsupervised optical flow prediction of an original video, extracting a candidate frame of each frame in the original video, and obtaining a candidate frame sequence; constructing a pseudo calibration frame sequence of a moving object in the original video based on the candidate frame sequence; constructing a training sample based on the pseudo calibration frame sequence, inputting the training sample into a naive twinning network to train the naive twinning network, and generating a preliminary tracking model; performing storage loop training on the preliminary tracking model to obtain a target tracking model; and tracking a target in a to-be-tracked video by using the target tracking model. Meanwhile, the invention provides a corresponding terminal and a medium. The cost of manual annotation of the video data is greatly reduced, and available video data is enriched and trained; and under the condition of no annotation, a target tracking model based on calibration frame regression is trained from an unannotated video.

The invention discloses an online service reputation measurement method based on semi-supervised learning and belongs to the online reputation measurement and on-line service field. The method comprises the following steps of firstly, carrying out normalization processing on the attribute scoring matrix R of a service and analyzing a principal component, and carrying out dimension reduction on a service attribute; then, synthesizing service multi-dimensional attribute information, manually labeling a training set and training a classifier model, based on an improved semi-supervised cooperationtraining algorithm, using an acquired classifier to carry out reputation classification on services, and adding the classified services and classification tags to the training set so as to retrainingthe classifier; and finally, using the new and acquired classifier to classify the online services so as to realize reputation measurement. In the invention, the reputation measurement of the services is realized by establishing the multi-classifier model of the services, and simultaneously, the semi-supervised learning algorithm is used to add the unlabeled services to the training set so as toretrain the classifier when the classifier is modeled, and classifier model classification performance is increased and simultaneously the cost of manually labeled samples is reduced.

The invention provides a classification model training method, system and device and a storage medium. The method comprises the steps of obtaining a training set which comprises a first text sample and a second text sample, wherein the first text sample is provided with a classification label, and the second text sample is not provided with a classification label, and training the classification model based on semi-supervised learning by adopting the training set, carrying out supervised learning on the classification model by adopting the first text sample, and carrying out unsupervised learning on the classification model by adopting the second text sample. According to the method, existing data including label data and data without labels are fully utilized, the model training effect isimproved, on one hand, the problem that in the prior art, when a classification model is trained, no enough training data exists, and consequently the classification model is prone to over-fitting issolved, on the other hand, manual labeling is not needed, and the manual labeling cost is reduced.