45results about How to "Reduce labeling workload" patented technology

The invention provides a pedestrian image generation method and a pedestrian image generation system based on a cyclic generation type countermeasure network. The method comprises the steps of learning a migration function and generating a person migration generation type countermeasure network; based on the person migration generation type countermeasure network, completing the character migration in a pedestrian image, migrating the pedestrian image from one database to another database, and generating a new pedestrian image. The migration function comprises style loss and identity loss. Compared with a traditional method, the method is good in migration effect for public data sets based on the re-recognition of a plurality of pedestrians. Therefore, on the premise that no extra data labeling is required, a robust pedestrian re-recognition model is trained in a target application scene and the accuracy is higher.

The invention belongs to the virtual character attitude control in the field of speech synthesis, and particularly relates to a modeling and controlling method for synchronizing the voice and the mouth shape of a virtual character. The object of the invention is to reduce the mouth shape animation data annotation amount and to achieve accurate and naturally smooth mouth motion synchronized with the voice. The method comprises: generating a phoneme sequence corresponding to the to-be-synchronized voice; converting the phoneme sequence into a phoneme category sequence; converting the phoneme category sequence into a static mouth shape configuration sequence; and converting the static mouth shape configuration sequence distributed on a time axis into dynamically changing mouth shape configuration by a dynamic model; rendering the dynamically changing mouth shape configuration into an attitude image of the head and neck of the virtual character, and displaying the attitude image in synchronization with a voice signal. The method can realize efficient and natural virtual character mouth shape synchronous control without mouth shape animation data and with a phonetic prior knowledge anddynamic model.

The invention discloses a segmentation method for unconventional cells in a pathological section. The method comprises the steps that cells in the pathological section are processed into separate to-be-segmented cell images with a transparent background, and a pixel tag is allocated to each pixel point of each to-be-segmented image; a mapping method is used to randomly distribute the to-be-segmented cell images on a white background, the cells are superposed according to a certain probability to form pseudo-input images, and corresponding full-image pixel tags are acquired and marked as real-value tags; the pseudo-input images and the real-value tags are used as training data to train a Mask-RCNN, so that the Mask-CNN has the abilities of detecting an unconventional cell boundary box and predicting the pixel tags in the box; and a new pathological section not marked is input into the converged Mask-RCNN, the unconventional cells in the non-segmented pathological section are detected, and a final segmentation result is obtained through postprocessing. Through the segmentation method, marking time can be effectively shortened, marking cost can be effectively lowered, a large amount of training data can be generated in a short time, and fitting can be well performed on a large amount of data.

The invention relates to a remote sensing image airport detection method based on a weak supervised learning frame. The method comprises the steps that firstly, the saliency of image blocks in a positive sample, the similarity of image blocks in a positive sample set and the inter-class difference between the positive sample and a negative sample are obtained, then, the Bayesian frame is utilized for fusing the three classes of information to obtain an initial positive and negative training set, then, iteration training of the training set is utilized for obtaining a final stable airport detector, the airport detector is used for detecting an airport of a tested image, and finally the airport detection result with better accuracy and robustness is obtained.

The invention discloses a tongue image multi-label classification learning method based on a graph convolution network, and the method comprises the following steps: S1, carrying out the tongue body detection of an original image, and obtaining a tongue body image through extraction; S2, performing image preprocessing on the tongue body image extracted in the step S1, wherein the preprocessing comprises reflection point removing processing, sharpening processing and straightening processing; S3, for each label, performing semi-automatic labeling on the preprocessed tongue body image to obtaina large-sample multi-label data set; and S4, training and inferring the large-sample multi-label data set obtained in the step S3 by using a graph convolution network to obtain a tongue body multi-label classification model based on the graph convolution network. According to the invention, the plurality of tags of the tongue image are classified and diagnosed at the same time through one graph convolution network, and the dependency relationship among the tags is fully learned, so that the tongue diagnosis process of the machine becomes more efficient and accurate.

The invention relates to an image processing method and device, electronic equipment and a storage medium, and the method comprises the steps: carrying out the feature extraction of a plurality of first images obtained in a current clustering period, and obtaining the first features of the first images; performing clustering processing on the first features, so that the category of each first image and the first clustering center of each category can be obtained; respectively determining a first feature similarity between each first clustering center and a second clustering center of each category in the reference feature library; and adding the first feature of the category to which the first clustering center of which the first feature similarity meets the threshold condition belongs tothe reference feature library. According to the image processing method provided by the embodiment of the invention, the first features meeting the threshold condition can be added to the reference feature library, and all the images do not need to be loaded into the memory for clustering processing, so that the memory can perform clustering processing with a relatively small data scale, the typesof all the first images can be automatically distinguished, and the annotation workload is reduced.

The invention discloses a named entity identification method and device, equipment and a storage medium, and relates to the fields of natural language processing, semantic analysis and understanding,artificial intelligence and the like. The method comprises the steps of performing entity identification on new domain text data to obtain new domain seed entity words; labeling the new domain text data according to the new domain seed entity words to obtain labeled new domain text data; training a named entity recognition model by using the labeled new domain text data to obtain a named entity recognition model suitable for the new domain; and identifying entity words in other text data of the new domain by utilizing the named entity identification model suitable for the new domain. Accordingto the embodiment of the invention, the data annotation workload can be reduced, the model migration training threshold is reduced, and the field universality of the algorithm is improved.

The invention provides a furniture layout drawing generation method and device, computer equipment and a computer readable storage medium. The method comprises the steps of obtaining a house drawing of a room; reading the house drawing by using a CNN network and carrying out feature extraction to obtain quantized data of the house, wherein the CNN network is obtained by training based on collecteddecoration drawings; inputting the quantized data into an RNN sequence model for sequence arrangement, and outputting an arrangement vector of furniture, wherein the RNN sequence model is obtained bytraining furniture arrangement information based on the decoration drawing; and searching furniture information of the corresponding furniture from a furniture library storing the decoration drawingaccording to the arrangement vector, and generating a furniture arrangement drawing of the room according to the furniture information. According to the scheme, the efficiency of generating the furniture layout drawing can be improved, so that the use effect of the furniture layout drawing is better.

The invention discloses a network attack traffic generation method based on an auxiliary classification type generative adversarial network, and the method can generate a malicious traffic sample which can cheat and escape from the detection of a defense system according to an existing network attack traffic data set sample by utilizing the principle of the generative adversarial network. The system comprises: a multi-source heterogeneous data fusion processing module which is responsible for defining a unified data format; a generator network which is responsible for generating a network statistical flow sample according to Gaussian noise and feedback from the discriminator; a discriminator network which is responsible for analyzing the attack traffic sample generated by the generator and the original network traffic sample, including authenticity analysis and attack traffic category analysis; and a classification fine tuning module which is responsible for debugging the performance of the generation model for generating specific types of traffic samples. According to the method, the network attack traffic generation model based on the auxiliary classification type generative adversarial network is constructed, the network attack traffic sample of a specific type can be generated according to the type of the network attack when the network traffic is generated, and the network attack can be simulated by generating the adversarial sample to detect the robustness of the existing intrusion detection system, and a new thought is provided for the existing traffic generator.

The invention provides a fundus image segmentation model training method and device. The method comprises the steps of obtaining a first segmentation model and a second segmentation model which are pre-trained; segmenting the fundus image in the training data by using the first segmentation model and the second segmentation model respectively, wherein the first segmentation model outputs first confidence information about the first target of interest, and the second segmentation model outputs second confidence information about the first target of interest and the second target of interest; obtaining integrated confidence information according to the first confidence information and the second confidence information; and optimizing parameters of the second segmentation model according to the integrated confidence information, a second confidence, the annotation data and a loss function constructed by a segmentation result determined by the second confidence.

The invention discloses a method for obtaining a fetal ultrasound four-cavity tangent plane cardiac cycle video based on a hybrid convolutional network. A complete period of the fetal four-cavity heart is defined as a period from the end of four-cavity tangent plane contraction to the end of next four-cavity tangent plane contraction, or a period from the end of four-cavity tangent plane relaxation to the end of next four-cavity tangent plane relaxation, and other intermediate periods are included in the period; wherein a video is used as input data; an image sequence after down-sampling is used as model input; features of time and space are extracted through a 3D convolutional network. By extracting 2D convolution network enhanced spatial dependence features, the features are fused. Finally, the classification probability and category of the next frame are predicted as three categories. A fetal four-cavity tangent plane cardiac cycle video which is complete and obvious in features iscounted through the joint probability under the condition that the category conforms to the complete cycle. The technical problem that an existing fetal heart detection method is difficult to guarantee the detection accuracy is solved.

The invention discloses an annotation method for the structured analysis of a Chinese electronic medical record text, and belongs to the technical field of big data. The method comprises the steps: establishing an annotation system, wherein a person participating in the annotation to perform Chinese word segmentation, word property and named entity annotation of the original table of the medical record annotation through a Web page to generate an annotation result table, thereby solving the problem of the simple annotation of the word segmentation, word property and named entity of the electronic medical record text. The invention simultaneously stands at the perspectives of a labeler and an algorithm designer. On the one hand, the annotation system is simple and easy to use, the annotation work intensity of the labeler is minimized, and the error rate is reduced; on the other hand, the annotation system is designed to be in seemliness butt joint with a knowledge base management systemand a core algorithm system, i.e., directly injecting original electronic medical record data into the labeling system after the original electronic medical record data is preprocessed, and the output of the labeling system directly serves as the input of the core algorithm system.

The invention discloses a dependency syntax model optimization method, device and equipment and a readable storage medium, a dependency syntax model to be optimized comprises a pre-training model at the bottom layer and a dependency relationship prediction network at the upper layer, and the pre-training model is obtained by training by adopting a domain-independent text training set. The dependency syntax model optimization method comprises the following steps: performing underlying vector extraction processing on training statements in a text training set in a target domain by adopting a pre-training model to obtain word vectors corresponding to words in the training statements; and performing upper-layer prediction processing on the word vector by adopting the dependency relationship prediction network, and optimizing a processing result to optimize the dependency syntax model. According to the method, the annotation workload can be greatly reduced, the annotation cost is reduced, and the model optimization efficiency is improved.

The invention discloses a lane line automatic detection method based on AI technology in a traffic law enforcement image, and belongs to the field of image identification. Therefore, the method only predicts the two endpoints of the lane line, the endpoint feature point pairs are combined and analyzed according to the specificity of the lane line, and finally the beneficial effects of being smallin labeling workload, low in requirement, high in prediction speed and accurate in prediction are achieved. The method is high in detection speed and high in accuracy. The detection result can be applied to vehicle illegal lane change detection (compacted line detection), and can also be applied to actual scene applications needing to use image data to detect lane line positions, such as lane number judgment and the like.

The invention belongs to the virtual character attitude control in the field of speech synthesis, and particularly relates to a modeling and controlling method for synchronizing the voice and the mouth shape of a virtual character. The object of the invention is to reduce the mouth shape animation data annotation amount and to achieve accurate and naturally smooth mouth motion synchronized with the voice. The method comprises: generating a phoneme sequence corresponding to the to-be-synchronized voice; converting the phoneme sequence into a phoneme category sequence; converting the phoneme category sequence into a static mouth shape configuration sequence; and converting the static mouth shape configuration sequence distributed on a time axis into dynamically changing mouth shape configuration by a dynamic model; rendering the dynamically changing mouth shape configuration into an attitude image of the head and neck of the virtual character, and displaying the attitude image in synchronization with a voice signal. The method can realize efficient and natural virtual character mouth shape synchronous control without mouth shape animation data and with a phonetic prior knowledge anddynamic model.

The invention discloses a pig drinking water detection imaging method based on BIM and artificial intelligence, and the method comprises the steps: carrying out the automatic marking of a pig house sample image set, and training a drinker target detection network; splicing and fusing the pig house images to obtain a pig house panoramic image; carrying out drinker detection on the hog house panoramic image; carrying out dirt detection on the detected water dispenser; carrying out heat map superposition on the dirty thermodynamic diagrams for a period of time; inputting the superposition resultinto a waterer dirt analysis network to obtain a waterer dirt grade, and generating a cleaning signal; visualizing various kinds of information in the pig house building information model. According to the method disclosed in the invention, automatic and intelligent hog house drinker cleaning condition monitoring is realized, the detection efficiency is high, the detection accuracy is high, the result is more real-time and objective, and hog house management is facilitated.

The invention discloses an image annotation method. The method comprises the following steps: determining at least one image block in a to-be-annotated image; respectively determining feature information of each image block, wherein the feature information is used for uniquely representing the corresponding image block; when target feature information matched with reference feature information exists in the feature information, performing annotation on the image block corresponding to the target feature information, so that labeling of a labeling object in the to-be-labeled image is achieved, and the reference feature information corresponds to the labeling object. According to the image annotation mode, the image blocks are automatically annotated, so that the problem that in the prior art, manual annotation is low in efficiency is solved, and a foundation is laid for efficiently obtaining a training sample.

The invention provides a marketing data processing method and a marketing data processing device. The marketing data processing method comprises the steps of establishing a positive and negative sample set of a marketing training model according to marketing business basic data; establishing a marketing training model according to the positive and negative sample sets and a random forest algorithm; determining initial marketing data according to the marketing business data to be predicted and the marketing training model; updating the marketing training model by utilizing the initial marketingdata; and performing marketing data processing by using the updated marketing training model to generate marketing list data. According to the marketing data processing method provided by the invention, manual marking is changed into automatic marking, so that the marking workload of business personnel is effectively reduced. The marking accuracy is improved. The model updating strategy is changed from a single coverage updating strategy to an intelligent updating strategy. Therefore, the difficulty of setting the model parameters by the business is reduced, intelligent updating is realized,and the marketing success rate is improved.

The embodiment of the invention discloses a point cloud data labeling method and device. The method comprises the following steps: obtaining a to-be-labeled point cloud data frame and corresponding acquisition equipment pose information; obtaining a to-be-displayed point cloud data frame corresponding to the labeled point cloud data frame, and displaying the to-be-displayed point cloud data frame; after a first selection operation is detected, determining first labeling box information corresponding to a to-be-labeled target; determining the current superposition frame of a superposition frame number information frame after or before the current display frame; displaying the current display frame and the current superposition frame in a superposition manner based on the current display frame and the acquisition equipment pose information corresponding to the current display frame; after a second selection operation is detected, determining second labeling box information corresponding to a to-be-labeled target in a target frame of the current superposition frame; and based on the first labeling box information and the second labeling box information, determining third labeling box information corresponding to the to-be-labeled target in each current superposition frame between the current display frame and the target frame, so that simple, convenient and effective labeling of the point cloud data is realized, the burden of labeling personnel is reduced, and the efficiency is improved.

The invention provides a liver case image classification method and system based on a deep neural network. The method comprises the following steps: obtaining a plurality of case images belonging to the same time period set by a user; resampling the plurality of case images to form a single three-dimensional case image; inputting the single three-dimensional case image into a deep neural network model to extract image features, wherein the image features comprise an image color channel and position information; assigning weights to the image features by adopting an attention mechanism, whereinthe weights of the image features related to the liver are greater than the weights of other image features; and inputting the image features endowed with the weights into a classifier to obtain theclassification probability of a single three-dimensional case image, the classification which comprises lesion and normality. According to the method and the system, position information in the case image does not need to be marked.

The invention provides a medical equipment performance index anomaly detection method and device based on PU learning, and the method comprises the steps: taking a historical key performance index (KPI) flow as training data, carrying out the clustering of the training data according to a similarity degree, obtaining a centroid curve of each cluster, and carrying out the marking of the centroid curve of each cluster, obtaining first abnormal annotation data and first non-annotation data; based on the first abnormal annotation data and the first unannotated data, constructing a binary classifier through positive example unannotated PU learning, and obtaining an abnormal label and a normal label of a centroid curve of each cluster in combination with active learning; and obtaining a label on the centroid curve of the cluster corresponding to the to-be-detected KPI flow, training an anomaly detection model corresponding to the to-be-detected KPI flow through semi-supervised learning, and detecting the to-be-detected KPI flow through the anomaly detection model. According to the method, the accuracy of medical equipment performance index anomaly detection is improved while the labeling workload is reduced to the maximum extent.

The invention discloses a power transmission line defect identification method based on a saliency map and a semantic embedding feature pyramid. The method comprises the steps of 1, performing data cleaning and division on a data set; 2, carrying out the super-resolution image generation of the small target of the power transmission line through an EL-ESRGAN super-resolution augmentation algorithm; 3, performing image saliency detection on the data set by constructing a nested U-shaped network; step 4, carrying out data augmentation based on a saliency graph on the data set through a Gridmask and random erasure (Cut Out) algorithm, and generating a classification data set; and 5, carrying out picture classification on the normal set and the defect set by utilizing a ResNet34 classification algorithm through a feature pyramid classification network embedded by deep semantics. According to the method, image saliency detection and data augmentation are combined, the feature pyramid classification network embedded through deep semantics is used as a supplement of ResNet34 classification, the method is used for fault identification in the unmanned aerial vehicle power transmission line inspection image, and the method has high system robustness.

The invention provides a deep learning-based van door state recognition device and system. Side-looking, side-looking and overlooking multi-view images acquired by the image acquisition unit are fusedinto an image sample input into the deep learning network; a deep learning network is established in the identification controller; the feature convolution extraction depth is increased, a feature sharing layer is added in front of an output layer to achieve voting judgment of the state of the compartment door, multi-view compartment door area feature map fusion and sharing layer multi-feature fusion are matched with one another, and the robustness and accuracy of the network for compartment door state recognition are improved. According to the invention, automatic identification of the stateof the cargo compartment door is achieved, hidden dangers caused by opening of the compartment door during transportation can be effectively prevented, and the logistics transportation management level is improved. The license plate area and the vehicle type are recognized in the same network, so that the number of recognition networks is reduced, and the recognition efficiency is improved; and through automatic preliminary frame selection of the compartment door area in the training sample image, the annotation workload is saved.

The invention discloses a method and device for identifying verbal skill intention, and relates to the technical field of computers. One specific embodiment of the method comprises the following steps: matching an interventional verbal skill for a user verbal skill from an interventional verbal skill set, and calculating verbal skill similarity between the user verbal skill and the matched interventional verbal skill; if the verbal skill similarity is greater than the intervention threshold, identifying the verbal skill intention of the user according to the matched verbal skill intervention rule; if the verbal skill similarity is less than or equal to the intervention threshold, identifying the verbal skill intention of the user by utilizing a template rule verbal skill set or an identification model; and updating the intervention verbal skill set, the template rule verbal skill set or the identification model according to the identified verbal skill intention. According to the embodiment, the verbal skill intention can be quickly and accurately recognized, and a quick response can be made to common verbal skill; the data is updated while the user verbal skill is identified, so that the data is fully utilized in the iteration process, and the data annotation workload is reduced.

The embodiment of the invention relates to a detection model training method and device, equipment and a storage medium, and aims to improve the generalization ability of a model. The method comprises the steps of obtaining a plurality of source domain sample images and a plurality of target domain sample images, wherein each source domain sample image comprises annotation box information of a pre-marked source domain object; inputting the source domain sample image and the target domain sample image into a feature extraction model in pairs to obtain a first feature map and a first detection frame of the source domain sample image and a second feature map and a second detection frame of the target domain sample image; judging the domain category of each pixel point in each feature map to obtain a plurality of first judgment results; judging the domain category of each detection box to obtain a plurality of second judgment results; and updating the feature extraction model according to the first detection box and annotation box information and according to the plurality of first judgment results and the plurality of second judgment results.

The invention discloses a method for constructing a spatial topological relation in a subway station based on an SVG (Scalable Vector Graphics) map. And a spatial object semantic element definition part in the subway. On the basis of indoor path planning and navigation requirements, space objects are necessarily simplified and omitted. According to the space topological relation construction part, a hidden layer is newly built on an original SVG map, and the object range on the map is marked through rectangles or line segments. Map object semantic information is described by an XML label id prefix, and space information of the map object semantic information is described by custom attribute information of XML label elements. And based on the annotated layer, writing a program to analyze the annotated layer of the SVG graph to obtain spatial object information in the subway station, a spatial topological relation data model and connection information between three-dimensional space floors. The method has the characteristics of light weight, simplicity and strong operability, and provides an enough data model basis for 2D or 2.5 D views, same-floor and cross-floor path planning and path navigation problems.

The invention discloses an automatic image annotation model construction method and system and application. The method comprises the steps that (1) a plurality of images are collected and annotated to obtain an original sample set; constructing an original annotation model according to the to-be-annotated image; (2) training the original labeling model by using samples in the original sample set to obtain an original automatic labeling model; and (3) automatically annotating the to-be-annotated image by using the original automatic annotation model, judging whether the annotation result of the annotated image is accurate or not, if so, not updating the original automatic annotation model, if not, correcting the annotation of the annotated image, storing the corrected annotated image into the original sample set, and returning to the step (2). According to the method, a small-order-of-magnitude sample size is adopted, heavy-sample-free expansion and model optimization are carried out on a labeling model in detection, the model accuracy and the environment applicability are improved, the accuracy of a model judgment result is ensured, and the model training time / workload can be greatly shortened.

The embodiment of the invention discloses a system, method and device for labeling continuous frame data. The system comprises a cloud end and a labeling end, wherein the cloud end reads continuous frame data, and performs target detection on each frame of data in the continuous frame data according to a labeling task to obtain a detection result of a to-be-labeled object in each frame of data; according to the detection result and the time sequence information of each frame of data, an association relationship between the same object to be labeled in each frame of data is established as a pre-labeling result; an extensible pre-labeling file is generated according to the pre-labeling result, and the pre-labeling file and the continuous frame data are sent to the labeling end; and the labeling end receives the continuous frame data sent by the cloud end and the corresponding pre-labeling file, and after a correction instruction for the pre-labeling file is received, the labeling file is corrected according to the correction instruction, and a target labeling result is obtained. By adopting the scheme, the manual time for marking the continuous frame data is shortened, the marking efficiency of the continuous frame data is improved, and the marking cost is reduced.

The invention provides an ore classification and particle size grading method and device based on a deep learning network in order to solve the problem that in the prior art, an image recognition technology for analyzing ore categories and particle size distribution at the same time needs to be improved, and belongs to the technical field of ore dressing. According to the method, the improved Faster-RCNN target detection network in computer vision is applied to extract features, classification and positioning of ores are completed at the same time, and an artificially designed feature extractor is replaced; through the ore position extracted by the target detection network, segmenting the ore by using an FCN semantic segmentation network to obtain granularity information of the ore; moreover, the high-quality image information is obtained by carrying out the data preprocessing work of contrast-limited adaptive histogram equalization, white noise removal and the like on the image; and meanwhile, the ore is classified and segmented by combining target detection and semantic segmentation technologies, so that the efficiency of the crusher is improved, the crushing energy consumption is reduced, and guidance is provided for subsequent processes.