51results about How to "Improve annotation accuracy" patented technology

The invention discloses an image marking method based on multi-mode deep learning. The method comprises the following steps: firstly, a depth neural network is trained by utilization of images without labels; secondly, each single mode is optimized by utilization of counter propagation; finally, weights among different modes are optimized by utilization of on-line learning power gradient algorithm. The method employs a convolution neural network technology to optimize parameters of the depth neural network, and the marking precision is raised. Experiments of public data sets show that the method can raise the image marking performance effectively.

The invention discloses a performance fragment marking method, a video playing method, a video playing device, and a video playing system, and belongs to the multimedia technology field. The performance fragment marking method comprises steps that a multimedia file corresponding to a role is acquired; according to the multimedia file, the characteristic of the role can be determined; a target video is decoded to acquire a data frame and a plating time stamp corresponding to the data frame; a target data frame matched with the characteristic of the role is identified in the data frame of the target video; according to the playing time stamp of the target data frame, performance fragment information corresponding to the role is marked automatically. According to the invention, a server can be used for the automatic marking of a plurality of target videos at the same time, and therefore problems of operation editors of low efficiency and poor accuracy caused by the fact that the operationeditors can only mark a small part of vides in a limited time period can be solved, and effect of marking the target video effectively in the limited time period can be achieved.

The invention discloses a movie and television play role labelling method based on search matching. The method comprises the following steps of: obtaining the to-be-labelled object set of a labelling scene and all to-be-labelled object information according to a to-be-labelled object list; constructing a text keyword for each of to-be-labelled objects, and obtaining the corresponding image set by virtue of an image search engine; carrying out face detection and visual attribute analysis on the image of the search result, and removing a noise therein to obtain a role face set which is closely related to the labelling scene, of the to-be-labelled objects; carrying out face detection and tracking on the labelling scene to obtain all face sequences therein; carrying out role labelling on the labelling scene on the basis of a visual similarity among the face sequences, and a visual similarity analysis on the face sequences and the role faces of the to-be-labelled objects. According to the method disclosed by the invention, movie and television play role labelling is carried out by virtue of face images related to movie and television play roles in the Internet; the method disclosed by the invention has the beneficial effects that the labelling process is fully-automatic, high in labelling accuracy, and high in method extensibility and universality.

The invention relates to a cross-domain knowledge transfer tag embedding method and apparatus. The method comprises the steps of obtaining text data of a source domain and a target domain, performing model representation, solving word vector parameters of keywords in the source domain and the target domain, and performing transfer of keyword tags from the source domain to the target domain; obtaining nearest neighbors of labeled keywords in the source domain and the target domain, performing weight assignment on keywords of the nearest neighbors by keyword tags of the labeled keywords to obtain extended keyword tags; performing user-level keyword tag labeling according to extracted user-level text data; dynamically optimizing parameters of user-level keyword tag parts according to click and / or access data information of a user based on the word vector parameters of the keywords and the user-level keyword tags; and obtaining new user-level text data from the target domain, performing user-level keyword tag labeling prediction and sorting, and outputting a result. According to the method and the apparatus, the accuracy and high efficiency of tag labeling can be taken into account and business demands of business personnel are met.

The invention discloses a medium-high-end talent intelligent recommendation system and method based on domain self-classification. A mapping knowledge domain technique is adopted , science and technology big data are mined and analyzed; a large-scale multi-field science and technology knowledge graph is constructed; based on matching fields based on the knowledge graph and the like, an automatic association relationship between the middle and high-end talent portraits and the multi-level science and technology knowledge graph is established, thereby achieving domain automatic classification ofmassive talent data and talent retrieval recommendation based on a 'small domain ', and providing an effective tool for talent introduction, talent information aggregation and retrieval of a talent introduction institution.

The invention provides a data annotation method. The method comprises the step of data annotation task allocation, wherein according to a data identification code of to-be-annotated data and an identification code of an annotator, a to-be-annotated data annotation task is matched with the annotator, and according to a matching result, the to-be-annotated data annotation task is allocated to the annotator; the step of data annotation, wherein according to the required annotation form, the to-be-annotated data is annotated; the step of collection and integration, wherein after the annotation results of the to-be-annotated data annotation task are all submitted, according to the annotation scores of the annotator and the annotation results, the annotation result is integrated, and an accuratelabel is doped out.

The invention provides an automatic image annotation method based on unsupervised domain adaptation. The method comprises the following steps: acquiring a source domain image and annotation, and acquiring a target domain image; building a detection framework, and constructing a domain classifier to extract global features and local features; training the existing data by using a PyTorch deep learning framework application algorithm to obtain a trained domain adaptation detection model; detecting the test data set (unlabeled pictures in the target domain) by using the existing latest model to obtain a preliminary detection result; and carrying out secondary processing and extraction by utilizing the preliminary detection result file to generate an xml annotation file in a PASCAL VOC format.On the basis of the domain adaptation method, under the condition that a large amount of target domain data are not labeled, only source domain pictures and labeling data similar to the target domaindata need to be owned, and training can be put into use for automatic labeling of the data. Compared with the prior art, the method is good in flexibility, high in classification precision, simple inmodel and high in practicability.

The invention relates to an image annotation method comprising the following steps: 1, defining a target function of an image annotation model; 2, inputting an image into a CNN model so as to obtain original image characteristics; 3, weighting the original image characteristics; 4, inputting information into a LSTM model; 5, making reversed propagations for errors caused by prediction result. Theimage annotation method firstly uses a convolution nerve network to extract image bottom layer characteristics, then uses a focusing mechanism to extract image characteristics related to image annotation words in an image specific position area, inputting the information into the LSTM network model, forming corresponding prediction annotation words, and finally realizing image annotations; the image annotation method is excellent in annotation performance, high in annotation precision, thus well satisfying real application demands.

The invention discloses an object 6D posture prediction method based on RGB image and coordinate system transformation. The method includes decoupling 6D posture parameters of the object, and realizing 6D posture prediction of the object by solving six parameter solving problems. The 3D translation of the object is predicted by positioning the center of the object in the image and estimating the distance between the object and the camera. 3D rotation of a predicted object is converted into a pose of a predicted camera through coordinate system transformation, pose parameters of the camera aredecoupled into an azimuth angle, an elevation angle and a rotation angle around a main optical axis, and the three parameters are predicted, so that 3D rotation prediction of the object is indirectlyrealized. The invention provides a universal frame for object 6D posture prediction, 2D target detection and 6D posture prediction can be carried out in one RGB image at the same time, and the methodhas good robustness for the conditions of complex illumination conditions, disordered placement, mutual shielding between objects and the like.

The invention relates to a calibration method of a laser radar and a camera, and the method comprises the following steps: enabling the laser radar to obtain N frames of point cloud data comprising a calibration board, eliminating the random error of the laser radar, and finding out the edge points of the calibration board; setting a transformation matrix between the calibration plate and the radar coordinate system as a T1 pseudo camera coordinate system, transforming the edge points of the calibration plate to the pseudo camera coordinate system, and projecting the edge points to an imaging plane of a pseudo camera to obtain a projection image; finding a 2d coordinate of an angular point of the calibration plate in a pseudo camera coordinate system from the projection image; and obtaining 3d coordinates of the angular points of the calibration plate in the camera coordinate system, thereby obtaining a conversion matrix T2 between the camera coordinate system and the pseudo camera coordinate system, and further obtaining a conversion matrix T = T2-1 * T1 between the laser radar coordinate system and the camera coordinate system. According to the method, random errors caused by laser radar scanning can be eliminated, so that the marking precision is improved.

The invention discloses an automatic ink-jet printing system, which can ensure an automatic ink-jet printer to accurately complete automatic marking work, and comprises a positioning ink-jet printing platform, a longitudinal moving device, a transverse moving device, a camera and an ink-jet printing head, wherein the positioning ink-jet printing platform is fixed on a supporting beam between two wall plates of the whole system, the longitudinal moving device is slidably arranged at both sides of the positioning ink-jet printing platform, the transverse moving device is slidably arranged on the connection frame of the longitudinal moving device, the camera is arranged above two diagonal lines of the positioning ink-jet printing platform through a support, and the ink-jet printing head is fixed on the transverse moving device. The automatic ink-jet printing system can complete the automatic marking work about geographic information changes in a map or the automatic map printing work in cooperation with the automatic air draught positioning system, and has the advantages of being high in automation degree and marking accuracy, small in error, standard and regularized in marking work, good in consistency, and high in working efficiency.

A remote sensing multispectral data semi-supervised labeling method based on self-learning relates to the field of data labeling, and comprises the following steps: acquiring remote sensing multispectral image data in a research area, determining category information of a to-be-classified target, and performing image fusion on the remote sensing multispectral image data; selecting a sample plot in the research area, recording category information of a to-be-classified target in the sample plot, determining a pixel corresponding relation between the to-be-classified target in the sample plot and the fused remote sensing multispectral image data in combination with the fused remote sensing multispectral image data, and obtaining pixel category information of the fused remote sensing multispectral image data; taking the seed point data as initial labeled data, and removing redundant information between wavebands by using a principal component analysis method; using the processed data with the labels to construct a classification model by adopting a random forest algorithm; the method comprises the following steps: classifying unlabeled data, removing abnormal points, and obtaining a self-labeling data set after multiple iterations; the method has the advantages of low requirement on manual labeling data volume, high precision, fast classification speed and strong anti-noise capability.

The invention provides a training data set generation method and device, electronic equipment and a storage medium. The method comprises the steps of obtaining a classified source data set and an unclassified target data set; extracting a first feature vector set of the source data set and a second feature vector set of the target data set through a feature extractor; determining a class center feature vector corresponding to the source data set according to the first feature vector set, and determining a clustering label of the target data set and an average feature vector in a clustering cluster according to the second feature vector set; iteratively optimizing the feature extractor, so that the overall difference between the feature vectors of samples in the source data set and the feature vectors of a class center and the overall difference between the feature vectors of the elements in a clustering cluster and average feature vectors in the clustering cluster are made to be minimum; and obtaining a training data set according to the clustering label of the target data set and the elements in the clustering cluster. According to the method, the workload of manual labeling can be reduced, the manual labeling cost is reduced, and the labeling precision is improved.

The invention discloses an lncRNA identification and expression quantification analysis method, which is characterized by comprising a series of a sequencing data filtering step, a sequencing data comparison step, a transcription splicing step, a gene comparison step, an lncRNA filtering step and an lncRNA analysis step. According to the method, the new lncRNA can be discovered, and the annotationaccuracy of the lncRNA is improved.

The invention discloses a related feedback method for actively selecting a multi-instance multi-mark digital image. Multi-instance multi-marking learning is a new machine learning frame which is presented in recent years and is successively applied in many practical problems. Image automatic marking technology based on multi-instance multi-mark input expression can be well applied on real tasks. But along with expression capability improvement, demand of marked training samples rapidly increases along with expression space increase. According to the related feedback method, through combining multi-instance multi-mark learning and active studying technology in machine learning, under a precondition that user marking cost does not increase, more fine and abundant marking information is obtained in a process of each related feedback, thereby improving system marking precision to a higher extent, and effectively reducing participation cost of the user.

The invention discloses a data generation system. The data generation system comprises a terminal, a server and a database, the data generation system organizes three-dimensional scene data, analog data and structured data based on an entity component system architecture; wherein three-dimensional scene data based on entity and component organization is stored in the database; the terminal collects a user-defined three-dimensional scene condition through a three-dimensional scene editing interface, collects a processing instruction edited by a user through a processing instruction editing interface provided based on the system, and sends the three-dimensional scene condition and the processing instruction to the server; and the server generates a three-dimensional scene graph meeting three-dimensional scene conditions according to the three-dimensional scene data, then adjusts the three-dimensional scene graph according to the processing instruction, and performs sensor data simulationbased on the three-dimensional scene graph according to the processing instruction to generate simulation data or / and directly outputs and stores structured data. The data generation system can generate a large amount of three-dimensional data with high quality and accurate annotation, and is low in cost.

The invention discloses a data labeling method and device, and the method comprises the steps: inputting each piece of data in a to-be-labeled data set into K labeling models, and obtaining K labels for each piece of data, wherein the K labeling models are obtained through the training of K sub-training sets, the K sub-training sets are obtained by performing K times of random sampling with replacement on samples in a total training set, and K is an integer greater than 1; dividing the data corresponding to the label into samples with different confusion degrees based on the confidence coefficient of the label, the confidence coefficient being the consistency degree of K labels obtained for each piece of data; and in a preset stage, sequentially labeling the samples with different confusion degrees to obtain a label of each piece of data in a to-be-labeled data set. According to the technical scheme, the samples with different confusion degrees are compared and verified through the K trained labeling models, so that the samples with different confusion degrees are automatically labeled, and the labor and time cost is greatly saved.

The invention relates to a data set acquisition method and device based on artificial intelligence, equipment and a medium. The method comprises the steps of obtaining an initial sample set; labelingthe initial sample set by using an initial language model to obtain a model labeling reference index; filtering the initial sample set according to the model annotation reference index to obtain a corrected set; training the initial language model by using the corrected set to obtain a corrected initial language model; when the precision of the corrected initial language model does not reach a preset threshold, expanding the data volume of the corrected set to update the corrected set, continuing to train the initial language model by using the corrected set to obtain a corrected initial language model, and when the precision of the initial language model reaches the preset threshold, obtaining a target language model; and processing to-be-processed service data according to the target language model to obtain a data set. By adopting the method, the data set acquisition efficiency can be improved. In addition, the invention also relates to a block chain technology, and the initial sample set, the corrected set and the data set can be stored in a block chain.

The embodiment of the invention discloses an image labeling method and device, electronic equipment and a storage medium. The image annotation method comprises the steps that a feature map, first position information and second position information of a to-be-annotated image are acquired, the first position information comprises position information of a first annotation point on the to-be-annotated image, and the second position information comprises position information of a second annotation point on the to-be-annotated image; convolution processing is performed on the feature map, the first position information and the second position information to obtain a plurality of annotation points on the to-be-annotated image; and the image to be labeled is labeled according to the plurality oflabeling points.

The invention discloses a cross-domain image example level active labeling method. Digital image target detection is one of the basic tasks of computer vision, and generally needs a large number of samples with object frame labels to be used for training a machine learning model. However, in real tasks, a large number of training samples of target tasks cannot be obtained due to sensibility and the like, so that the model performance is low, and the model is difficult to promote. According to the method, the unsupervised source domain which is easy to obtain and rich in knowledge is utilized, and efficient example labeling is automatically selected through an active learning technology, so that finer labeling information is obtained, and the data labeling difficulty is greatly reduced; meanwhile, the obtained supervision information is fully utilized, the performance of the model on the target task is efficiently improved, and the participation cost of the user can be remarkably reduced.

The invention provides an inscription labeling, detecting and recognizing system based on a deep neural network, and finally, information such as inscription positions, word meanings and fonts can beaccurately, effectively and automatically extracted, so that a basis is provided for subsequent inscription retrieval work. The whole system is divided into an annotation module group, a training module group and a test module group from the overall structure, wherein the annotation module group comprises a character position annotation module based on pre-positioning, a character annotation module based on pre-recognition and a segmentation annotation module based on connected components; the training module group comprises a detector training module and a classifier training module; and thetest module group is used for detecting, identifying and segmenting an input image, and establishing the retrieval function on the basis.

The invention aims to provide a simple and practical geographic position labeling method. The method has the following advantages that: 1, global codes are unified; 2, the precision is high enough (the current GPS positioning precision can reach 4m level); 3, the number of bits is small (10-bit main encoding + 0-4-bit auxiliary encoding); and 4, the actual position is easy to determine (a simple formula can be directly converted into longitude and latitude). By using the coding system, a geographic position becomes simple to label, an actual position can be accurately positioned through an electronic map and GPS positioning only by providing a dozens of codes, and a simple and practical method is provided for determining the positions of manual express delivery and unmanned aerial vehicledelivery and labeling and positioning a certain area (such as fields, water areas, forests and building lands).

The invention discloses a target labeling method, and the method comprises the steps: obtaining a to-be-labeled image, and identifying a first identification area and a second identification area in the to-be-labeled image; labeling the first identification area and the second identification area to obtain a to-be-filtered target; based on a pre-constructed evaluation function, filtering the to-be-filtered target to obtain a labeled target, wherein the area of the first identification area is larger than that of the second identification area. The invention further discloses a target labelingdevice and equipment and a computer readable storage medium. According to the intelligent labeling method and device, the labeling object is divided into the first recognition area and the second recognition area, then the first recognition area and the second recognition area are recognized and labeled, the labeling speed and accuracy are improved, and in order to further improve the labeling precision, the evaluation function is adopted to filter the labeling result, so that a high-precision labeling target is obtained, and intelligent labeling is achieved.

The invention discloses a movie and television play role labelling method based on search matching. The method comprises the following steps of: obtaining the to-be-labelled object set of a labelling scene and all to-be-labelled object information according to a to-be-labelled object list; constructing a text keyword for each of to-be-labelled objects, and obtaining the corresponding image set by virtue of an image search engine; carrying out face detection and visual attribute analysis on the image of the search result, and removing a noise therein to obtain a role face set which is closely related to the labelling scene, of the to-be-labelled objects; carrying out face detection and tracking on the labelling scene to obtain all face sequences therein; carrying out role labelling on the labelling scene on the basis of a visual similarity among the face sequences, and a visual similarity analysis on the face sequences and the role faces of the to-be-labelled objects. According to the method disclosed by the invention, movie and television play role labelling is carried out by virtue of face images related to movie and television play roles in the Internet; the method disclosed by the invention has the beneficial effects that the labelling process is fully-automatic, high in labelling accuracy, and high in method extensibility and universality.

The invention provides a statement entity labeling model training method and device and electronic equipment. The method comprises the steps: carrying out the word vector processing of sample statements in a plurality of obtained sample texts, and obtaining a plurality of sample matrixes; attention weighting processing is carried out on each sample matrix, and a plurality of processed sample weighting representation matrixes are determined; determining an auxiliary weighted representation matrix corresponding to each sample weighted representation matrix by training a constructed auxiliary classification model; and training a pre-constructed conditional random field model by taking the plurality of auxiliary weighted representation matrixes as input features and taking the obtained samplestatement label corresponding to each sample statement as an output feature to obtain a statement entity labeling model. Therefore, the labeling precision of the model in the use process can be further improved, the noise interference can be reduced, the entity can be accurately labeled, and the generalization ability of the model is improved.

The invention belongs to the technical field of data annotation, and discloses a rule-based picture data annotation method and device and a storage medium, and the method comprises the steps: obtaining picture annotation data; judging whether original data exists in the annotation data or not, wherein the original data comprises existing annotation information; if the original data does not exist, labeling all labeling boxes on the labeling data, and selecting one labeling box as a father box; according to the association degree of the father box and other labeling boxes around, labeling child boxes, and outputting a labeling result; if the original data exists, classifying the annotation data according to the original relationship of the annotation box in the annotation information; calculating the correlation degree between the annotation boxes, and establishing a bipartite graph for matching according to the relationship between the annotation boxes; and outputting an annotation result according to a matching result of the bipartite graph. The effect of improving annotation efficiency and precision can be achieved.

The invention provides an image marking method and system, and the method comprises the steps: monitoring a frame selection instruction of a to-be-annotated image, so as to determine a region of interest; obtaining a minimum pixel value and a maximum pixel value in the region of interest; comparing the pixel value of each pixel point in the set area with the minimum pixel value min and the maximum pixel value max, and setting the pixel value of the pixel point smaller than or equal to the minimum pixel value min as 0; setting the pixel value of the pixel point of which the pixel value is greater than or equal to the maximum pixel value max as 2n-1; setting the pixel value of the pixel point between the minimum pixel value min and the maximum pixel value max to be (value-min) / (max-min) * (2n-1); and generating an annotation file by taking the region of interest as an annotation box. The pixel value of each channel of the region of interest can be expanded to the maximum range through local contrast adjustment, the to-be-labeled object is highlighted to the maximum extent, and the labeling precision is improved; contrast adjusting and marking are carried out at the same time, compared with an existing method of firstly adjusting the contrast and then marking, the contrast adjusting and marking method has the advantages that two steps are shortened into one step, and the marking speed can be increased.

The invention discloses a correlation feedback method for actively selecting multi-instance multi-label digital images. Multi-instance multi-label learning is a new machine learning framework proposed in recent years, which has been successfully applied in many practical problems. The automatic image annotation technology based on multi-instance multi-label input representation can be well applied to real-world tasks. However, with the enhancement of its expressive ability, the demand for marked training samples of automatic image annotation methods increases sharply as the representation space becomes larger. The present invention combines multi-instance multi-label learning and active learning technology in machine learning, without increasing the cost of user labeling, and obtains more detailed and rich label information in the process of each relevant feedback, thereby improving the system to a greater extent Labeling accuracy effectively reduces the user's participation cost.