677 results about "Manual annotation" patented technology

A manual annotation system of multi-modal characteristics in multimedia files. There is provided an arrangement for selection an observation modality of video with audio, video without audio, audio with video, or audio without video, to be used to annotate multimedia content. While annotating video or audio features is isolation results in less confidence in the identification of features, observing both audio and video simultaneously and annotating that observation results in a higher confidence level.

The invention belongs to the technical field of medical information data processing, and particularly relates to an electronic medical record text named entity recognition method based on a pre-training language model, which comprises the following steps: collecting an electronic medical record text from a public data set as an original text, and preprocessing the original text; labeling the preprocessed original text entity based on the standard medical term set to obtain a labeled text; inputting the annotation text into a pre-training language model to obtain a training text represented bya word vector; constructing a BiLSTM-CRF sequence labeling model, and learning the training text to obtain a trained labeling model; and taking the trained labeling model as an entity recognition model, and inputting a test text to output a labeled category label sequence. According to the method, text features and semantic information in the deep language model are obtained through training in the super-large-scale Chinese corpus, a better semantic compression effect can be provided, the problem that manual annotation is tedious and complex is avoided, the method does not depend on dictionaries and rules, and the recall ratio and accuracy of named entity recognition are improved.

The invention relates to a surface defect detection method based on positive case training. The method comprises two steps of image reconstruction and defect detection, image reconstruction is to reconstruct an inputted original image into an image without defects, reconstruction steps are as follows, artificial defects and noise are added to the positive case image during training, a self-encoderis utilized for reconstruction, the L1 distance between the reconstruction result and the noise-free original image is calculated, the distance is minimized as a reconstruction target, in cooperationwith the generative adversarial network, the reconstruction image effect is optimized; defect detection is performed after image reconstruction, LBP features of the reconstructed image and the original image are calculated, after difference between the two feature images is made, the two images are binarized based on the fixed threshold, so the defects are found. The method is advantaged in thatthe depth learning method is utilized, the method can be sufficiently robust to be less susceptible to environmental changes when positive samples are enough, moreover, based on regular training, themethod does not rely on a large number of negative samples and manual annotation, the method is suitable for being used in real-world scenarios, and the surface defects can be better detected.

The invention relates to a multi-feature fusion phishing webpage detection method, which comprises two parts such as a training process and a detection process. The multi-feature fusion phishing webpage detection method integrates three views of phishing webpage characteristics by combining a semi-supervised learning tri-training method, and mainly solves a problem that the existing phishing webpage detection methods mostly need to perform classification model training by using supervised learning through a large amount of annotation data. The method provided by the invention mainly combines a coordinated training algorithm, starts from webpage URL characteristics, webpage information characteristics and webpage search information characteristics, applies the idea of multiple views and multiple classifiers to phishing webpage detection, and achieves the purposes of reducing the total numbers of manual annotation training samples and timely recognizing a phishing webpage through coordinated training and learning of different classifiers.

The invention discloses a method and system for intention recognition based on a context. The method comprises the steps of carrying out manual annotation on problems provided by a user and included in a user historical consulting log on the basis of the user historical consulting log; extracting features of each problem to generate a training corpus; carrying out model training on the generated training corpuses according to a supervised learning algorithm to obtain a training model; estimating the current problems and obtaining an estimated user intention recognition result on the basis of the obtained training model.

A method and system for on-line learning of landmark detection models for end-user specific diagnostic image reading is disclosed. A selection of a landmark to be detected in a 3D medical image is received. A current landmark detection result for the selected landmark in the 3D medical image is determined by automatically detecting the selected landmark in the 3D medical image using a stored landmark detection model corresponding to the selected landmark or by receiving a manual annotation of the selected landmark in the 3D medical image. The stored landmark detection model corresponding to the selected landmark is then updated based on the current landmark detection result for the selected landmark in the 3D medical image. The landmark selected in the 3D medical image can be a set of landmarks defining a custom view of the 3D medical image.

A system, a method and a computer readable media for providing model-driven feedback to human annotators. In one exemplary embodiment, the method includes manually annotating an initial small dataset. The method further includes training an initial model using said annotated dataset. The method further includes comparing the annotations produced by the model with the annotations produced by the annotator. The method further includes notifying the annotator of discrepancies between the annotations and the predictions of the model. The method further includes allowing the annotator to modify the annotations if appropriate. The method further includes updating the model with the data annotated by the annotator.

The invention is a method and apparatus for quickly marking object trajectories in an image sequence. The traditional method of manually tracking an object is by marking the location of the object in each image in the image sequence. In the present invention, the object is first tracked in one direction by one quick stroke of a spatial input device across the subsampled image sequence that is laid out in that direction. The object is then tracked in an orthogonal direction by another quick stroke of the spatial input device across the same subsampled image sequence that is laid out in the orthogonal direction. The intersections of the two tracks will then form the points of the 2-D neo-trajectory, which is then interpolated onto the image sequence. The invention is useful not only for quick manual annotation of tracks, but also for performance evaluation and optimization of automatic tracking algorithms. Example use of this invention is for behavior annotation and development of tools for automatic behavior analysis.

The invention provides an entity identification method based on Chinese electronic medical records, and relates to the technical field of medical entity identification. In order to overcome the defects of the lack of a public Chinese electronic medical record annotation corpus in China currently, by constructing and managing a medical dictionary, a semi-automatic corpus annotation method is put forward, and the complexity of manual annotation is reduced. Secondly, the problems are solved that existing electronic medical record entity recognition methods based on characteristics mostly aim at ordinary texts or general electronic medical record texts, and unique characteristics of the Chinese electronic medical records are not considered. By means of the method, besides basic characteristicsof the general text, the unique chapter information characteristics of the Chinese electronic medical records are also extracted; core word characteristics obtained by counting character frequenciesand word frequencies are added into extension characteristics after the collected dictionary is subjected to single-character and word segmentation, a relationship of words is also added to the extension characteristics by clustering word vectors, and the accuracy of the entity identification of the Chinese electronic medical records is effectively improved.

The invention discloses an adjudication document annotation method and device based on a machine learning algorithm. The method comprises the following steps that: collecting the text set of adjudication documents to be annotated; carrying out structure segmentation on texts in the text set; establishing a semantic tag library; on the basis of the semantic tag library, carrying out manual annotation on the adjudication documents to be annotated; selecting parts of adjudication documents subjected to the manual annotation as a standard dataset to be handed for machine learning, and training andoptimizing a preliminary annotation model; selecting residual parts of adjudication document models subjected to the manual annotation to obtain a mature adjudication document annotation model; and after a target adjudication document to be annotated is segmented, inputting into the mature adjudication document annotation model to obtain an annotation result. Through the method, the problem in the relevant art that adjudication document legal elements are not completely extracted and case information extraction accuracy is low is solved.

The embodiment of the invention discloses a training method of a semi-supervised learning model, an image processing method and equipment, and can be applied to the field of computer vision in the field of artificial intelligence, and the method comprises the steps: firstly carrying out the prediction of the classification types of a part of unlabeled samples through a trained first semi-supervised learning model, and obtaining a prediction label; judging whether each prediction label is correct or not in a one-bit labeling manner, if the prediction is correct, obtaining a correct label (namely, a positive label) of the sample, otherwise, eliminating an error label (namely, a negative label) of the sample, and then, in the next training stage, reconstructing a training set (namely, a firsttraining set) by utilizing the information, and training the initial semi-supervised learning model again according to the first training set, so that the prediction accuracy of the model is improved, and due to the fact that the annotator only needs to answer 'Yes' or 'NO' to the prediction label in one-bit annotation, the annotation mode can relieve the manual annotation pressure that a large amount of correct label data is needed in machine learning.

The invention provides a cross-domain and cross-category news commentary emotion prediction method. According to the method disclosed by the invention, under the condition that a target domain is provided with a small amount of annotation data only and another related but different source domain is provided with a large amount of annotation data, knowledge transfer among different domains is realized through simulating the relationship between the emotion category collections of the source domain and the target domain, and a cross-domain and cross-category news commentary emotion prediction model is built, so that the problem of difficulty in emotion prediction of news commentaries of the target domain is solved; under the situation that the emotion category collections of the source domain and the target domain are different, the method disclosed by the invention is significantly better than other alternative cross-domain and cross-category online news commentary emotion prediction methods, and high cost resulting from manual annotation work and energy consumed through training more classification models are greatly reduced. The method can be applied to user sentiment analysis and public sentiment supervision.

The invention discloses an image quality detection method, device and equipment. The method comprises the following steps of: extracting image features from training samples so as to determine featurescores of the image features; training an image quality scoring model on the basis of the feature scores; and extracting an image feature from a to-be-detected image, so as to determine an image quality score of the to-be-detected image according to the image feature of the to-be-detected image and the image quality scoring model, wherein when the image quality scoring model is trained, originalimages acquired under different environments and extended images obtained through processing the original images are taken as the training samples. According to the method, feature scores (not artificially marked scores) are determined according to image features of images, and models are trained on the basis of the feature scores, so that image quality scores detected by the models are more objective, visual and correct, the operation speed of scoring carried out through computers is considerably higher than that of manual annotation, and the processing cost of the training samples can be reduced.

The present invention provides a training method/system of an intelligent model, a computer readable storage medium and a terminal. The training method includes the following steps that: initial model training is performed on an inputted first data set and annotation information related to a training task, so that a reference model can be obtained; new data are added and are merged in the first data set, so that a second data set can be formed; data testing and value assessment are performed on the second data set, so that data of which the annotation values are larger than a preset annotation value are selected to form a third data set; unannotated information in the third data set is annotated, and the annotated information is merged into the third data set; the reference model is ret-rained, so that an updated reference model is obtained; and the third data set is defined as a new first data set, and new data are added into the new first data set, and the above steps are executed circularly until the precision of the iteratively-trained model is greater than preset accuracy. With the training method/system of the intelligent model, the computer readable storage medium and the terminal of the method adopted, the number of manual annotations can be deceased; it does not need to annotate all the data, and therefore, annotation costs can be saved, and the training efficiency of the model can be improved.

The invention discloses a semi-automatic image annotation sample generating method based on target tracking. The method comprises processes of target tracking and semi-automatic annotation. A serial of samples are generated through a target tracking mechanism, a template learning mechanism is designed for tracking and detecting target areas, detection on videos or images is performed by means of learned templates, manual annotation is utilized to help to perform determination, and therefore, annotation samples are generated. The semi-automatic image annotation sample generating method based on the target tracking has the advantages of being capable of obtaining a large amount of image annotation samples by means of less labor consumption.

The invention provides an entity relationship extraction method and system integrated with dynamic word vector technology. According to the system, an existing knowledge base corresponds to rich unstructured data by using a remote supervision method so as to generate a large amount of training data, so that the problem of insufficient manual annotation corpora is relieved, and the system can reduce the dependence on annotation data, thereby effectively reducing the labor cost. In order to obtain feature information between entities as much as possible, the basic architecture of the model adopts a segmented convolutional neural network; and the semantic information of the example sentences is further extracted by integrating a dynamic word vector technology.

The invention discloses a deep learning pedestrian detection method based on an embedded terminal. The method comprises the steps that first, at a sample preparation stage, an existing automatic driving dataset is acquired, or manual annotations obtained after a fixed camera and a mobile camera shoot videos are collected; second, at a training stage, a large quantity of training images are used totrain parameters of a constructed convolutional neural network so as to complete detection feature learning; third, at a test stage, a large quantity of test images are input into the trained convolutional neural network, and a detection result is obtained; and fourth, at a porting stage, code level optimization and porting into the embedded terminal are performed. According to the method, the 18-layer convolutional neural network is adopted to perform pedestrian feature learning, and the method has an innovative advantage compared with a traditional machine learning method; and an optimization strategy for the embedded terminal is also proposed, the network scale and algorithm complexity are further reduced, and the method is suitable for ADAS function application.

The invention brings forward a voice data annotation method and apparatus. The voice data annotation method includes the steps of receiving voice data, performing voice recognition of the voice data to obtain a recognized text; obtaining a user-confirmed text after the user confirms the recognized text; and performing analyzing of the recognized text and the user confirmation text to extract automatic annotation features; and performing annotation of the voice data according to the automatic annotation features and a pre-constructed automatic annotation model. The method can solve problems existing in manual annotation, raises the efficiency of voice data annotation and reduces the cost.

The invention discloses a video-based weather phenomenon recognition method. The method achieves classification recognition of common weather phenomena such as sun, cloud, rain, snow and fog. The method comprises the steps of training an off-line classifier, wherein an image sequence is sampled for a given training video; on the one hand, video characteristics of the image sequence are extracted; on the other hand, key frame images and image characteristics of the key frame images are extracted from the image sequence, the AdaBoost is adopted for conducting learning and training on the extracted video characteristics, the extracted image characteristics and manual annotations to obtain the classifier; recognizing the weather phenomena in an online mode, wherein a plurality of sets of image sequences are sampled for a testing video, video characteristics and image characteristics of each set of image sequences are extracted, the characteristics are sent into the classifier for classification to obtain a corresponding recognition result, then decision fusion is carried out in a voting mode, and the voting result is used as the weather phenomenon recognition result of the testing video.

The embodiment of the application discloses a training method, a matting method, a device, a medium and a terminal of an image processing model. The training method of the image processing model includes obtaining a trichotomy of the original image; Generating a training sample set according to the original image and the trichotomy; The preset depth learning network is trained based on the training sample set to obtain an image processing model, wherein, the image processing model is used for labeling the original image to obtain a trichotomy. By adopting the technical proposal, the input original image can be automatically annotated to obtain a trichotomy, and a large amount of hairline level data annotation can be carried out without manual annotation, the annotation workload can be reduced, and the image annotation efficiency can be improved. In addition, the matting effect can be optimized by using image processing model to label the original image, avoiding the error that may be introduced by manual labeling.

The invention discloses an automatic identification method for a milking sow gesture on the basis of a depth image. The method comprises the following steps that: collecting original depth image data,carrying out preprocessing, and carrying out manual annotation to form a milking sow gesture identification dataset; designing and training a milking sow hybrid deformable component model based on animproved HOG (Histogram of Oriented Gradient) feature; constructing a milking sow gesture identification deep convolutional neural network, utilizing an annotation frame and annotated gesture category training set information, and combining with a random Dropout method to train the network; inputting the test set into the milking sow hybrid deformable component model to obtain the target area ofthe milking sow; and inputting a target area result into the milking sow gesture identification deep convolutional neural network to identity the milking sow gesture. By use of the automatic identification method for the milking sow gesture on the basis of the depth image, the problem that an RGB (Red, Green and Blue) image is likely to be affected by the changes of factors, including outside illumination, shades and the like is overcome, the problem that the milking sow gesture is difficult in identification at night is solved, and the practical application requirement of all-weather milkingsow gesture monitoring can be met.

The invention relates to an audio annotation method, in particular to a deep-learning-based automatic audio annotation method. The deep-learning-based automatic audio annotation method comprises the following steps that original audio files are input, and by means of audio preprocessing, multiple original sound spectrograph segments are obtained; the original sound spectrograph segments are inputinto a convolutional neural network for training, and a deep-learning model is built; to-be-annotated audio files are input, and by means of audio preprocessing, multiple to-be-annotated sound spectrograph segments are obtained; on the basis of the deep-learning model, the to-be-annotated sound spectrograph segments are subjected to audio annotation. Accordingly, the convolutional neural network is utilized for training an audio deep-learning network, and the automatic audio annotation method is realized; compared with a traditional manual annotation mode, the annotation accuracy is improved,and the audio annotation efficiency is improved.

The invention relates to the field of natural language processing, in particular to a bidirectional recursive neural network-based enterprise abbreviation extraction method. The method comprises the steps of serializing to-be-processed texts through word segmentation processing; selecting a certain number of the to-be-processed texts to perform manual annotation, segmentally annotating enterprise names in the to-be-processed texts as starting parts, keyword parts, industrial parts and organization form parts, and annotating data except the enterprise names as unrelated parts; inputting annotated training samples into a bidirectional recursive neural network to train the bidirectional recursive neural network; extracting word sequences belonging to the enterprise names through prediction of the bidirectional recursive neural network, and further extracting fields belonging to the keyword parts of the names as enterprise abbreviations; and establishing a corresponding enterprise abbreviation database. Therefore, powerful technical support is provided for related information analysis of informal texts.

The invention relates to a device and method for acquiring road information from map data. The method includes: acquiring point cloud map data and semantic map data; extracting road information in a semantic map from the semantic map data; acquiring a primary road boundary from the point cloud map data; matching a road boundary in the road information in the semantic map with the point cloud map data to obtain road boundary information of a to-be-generated semantic map; matching lane line information in the road information in the semantic map with the point cloud map data to obtain lane lineinformation of the to-be-generated semantic map; obtaining road information of the to-be-generated semantic map according to the road boundary information and the lane line information of the to-be-generated semantic map. By means of matching of data in a traditional data map with point cloud in high-precision map data, manual annotation is greatly saved, and the map drawing speed is increased.