36results about How to "Expand the dataset" patented technology

A system and method for enhancing bitmap indexing representation of a dataset, which comprises a plurality of cases and features, each case characterized by one or more values of each feature. Currently, the bins vector for each case in the dataset, is a binary array, which is a bitmap indexing representation of each respective feature of the case. The system and method enhance the bitmap indexing by padding each bins vector. The padding is carried out by identifying all target bit locations with a ‘1’ value and replacing at least one ‘0’ bit adjacent to a target bit location with a non-zero numerical value, thereby creating a padded bitmap index. The padding factor may be based on any mathematical or statistical factor concerning population or subpopulation relevant to each of the features of the dataset.

The embodiment of the invention provides an image processing method, device and system and a computer storage medium. The method comprises the step of generating a generative adversarial network, which comprises a generative model and a discrimination model, through the following way that: inputting an input image into the generative model to obtain an output image, wherein the input image and theoutput image have different styles; inputting the output image into the discrimination model to obtain an output value; according to a pre-defined loss function, training the generative model and thediscrimination mode until convergence is realized; and using the trained generative model to obtain a target image which has different styles with the input image of the trained generative model. Therefore, by use of the embodiment of the invention, the generative model can be obtained through training, on the basis of the generative model of the generative adversarial network, the target image which has different styles with the input image can be obtained, so that a dataset used for vehicle detection can be expanded, and therefore, a vehicle detection effect is guaranteed.

The invention discloses a wind turbine generator main shaft bearing fault prediction method. According to the method, based on historical fault maintenance data of a main shaft bearing of a draught fan, statistics and a machine learning method are combined, multiple monitoring indexes of the draught fan serve as input variables, the state of the main shaft bearing serves as a prediction output variable, statistical analysis is conducted on a prediction value of the output variable, and a threshold value is set for fault prediction. The method has high accuracy and stability, and can predict the fault of the main shaft bearing one week in advance and discover early abnormality as soon as possible.

The invention relates to a method for detecting quick-changing attack domain name based on host group characteristics, which mainly comprises the following steps of 1, capturing a network data package and extracting a DNS (domain name server) message characteristic; 2, detecting the quick-changing attack domain name; and 3, performing misjudgment detection, wherein detection of the quick-changing attack domain name comprises the steps of computing an IP (internet protocol) distribution program of the host group which corresponds to the domain name, assessing service availability, and detecting network wave, and is the core of the invention; and misjudgment detection gets rid of the normal large scale network domain name in the process of detection of the quick-changing attack domain name and a detection result when a local network is not good in the process of detection of online rate. According to the invention, set of the DNS message in a local area network is analyzed; the problem of the accuracy rate of analysis of the single DNS message is avoided based on the characteristics of IP degree of dispersion of the host group which corresponds to the domain name and online rate; and the scale that the domain name corresponds to the host group is considered when an IP distance is computed, so that the misjudgment of a large scale fine quick-changing network is avoided.

The invention belongs to the technical field of named entity recognition, and particularly relates to a Chinese named entity recognition method and device based on a RoBERTa-BiGRU-LAN model, and the method comprises the steps: converting a to-be-processed Chinese corpus into a word vector sequence; inputting the obtained word vector sequence into a first layer of BiGRU-LAN of a RoBERTa-BiGRU-LAN model, and obtaining a coding sequence fused with local information; inputting the obtained coding sequence into a second layer of BiGRU-LAN of the RoBERTa-BiGRU-LAN model, and obtaining attention distribution fused with global information; and obtaining a named entity identification result according to the obtained attention distribution. According to the improved word embedding method disclosed by the invention, Chinese representation is better carried out, and meanwhile, BiLSTM-CRF is improved into BiGRU-LAN, so that the parameters of the model are reduced, the complexity of the model is reduced, and the training time is saved.

The invention discloses a natural language processing model training method, a natural language processing method, natural language processing equipment and a natural language processing system, whichbelong to the field of natural language processing, and the method comprises the following steps: training a teacher model by utilizing a marked original data set; enhancing text sentences in the original data set to obtain enhanced text sentences, and labeling the enhanced text sentences by using a trained teacher model to obtain a labeled enhanced data set; taking the original data set and theenhanced data set as a training data set, training the student model, and taking the trained student model as a natural language processing model, wherein the teacher model and the student model are both deep learning models and execute the same natural language processing task, and the teacher model is more complex and larger in scale. According to the invention, the data set of the natural language processing task can be effectively enhanced in a knowledge distillation scene, and the processing capability of the natural language processing model is improved, so that the execution effect of the natural language processing task is improved.

The embodiment of the present invention provides a method and device for enhancing network flow data based on a generative adversarial net. The method comprises the following steps: acquiring a data set of real network flow in a target scene; training the generative adversarial net model according to the data set; and obtaining final enhanced flow data based on the trained generative adversarial net model. The embodiment of the present invention can be applied to various scenes without the need of expert experience in flow data aspect, flow data enhancement is adaptively implemented, the dataset of the network flow is expanded, and the effect of optimizing network performance by using a machine learning method is improved.

The invention provides a new viewpoint image synthesis method based on a depth-aided full resolution network. The main content comprises the depth-aided full resolution network, the loss function andtraining. The process comprises the steps that the encoder part of the full resolution network firstly extracts important local features from the input image; then a depth predictor performs pre-training on the large image data set after detecting the global image information so as to estimate the depth map of the input image; then the local features and the depth are fed to the decoder and dual channel mapping of the target viewpoint position is indicated; and finally the decoder converts the combined features into the distortion field so as to synthesize the final target image based on flowdeformation. The full resolution network is designed and the local image features are extracted according to the same input resolution so that the blurred ghost in the finally synthesized image can beprevented, and the high-resolution and high-quality image can be acquired.

The invention discloses a depth estimation method based on a light field EPI image, and the method comprises the steps: 1, carrying out the refocusing of light field data, and obtaining the light field data under different focusing parameters; 2, extracting sub-aperture images of a horizontal view angle and a vertical view angle from the refocused light field data; 3, extracting light field EPI images in the horizontal and vertical directions from the sub-aperture images; 4, building a double-branch depth estimation model based on light field EPI image correlation reasoning, and using the extracted EPI images in the horizontal and vertical directions for training; and 5, carrying out depth estimation on the to-be-processed light field data by utilizing the trained depth estimation model. According to the method, the relevance between the central pixel and the neighborhood of the EPI image can be fully utilized, and the data enhancement can be realized by utilizing the light field refocusing principle, so that the accuracy of light field EPI image depth estimation can be effectively improved.

The invention discloses a defective workpiece image recognition method based on a convolutional attention neural network. The method comprises the steps of performing pre-training weight on the convolutional neural network to form a deep feature for input image input, of a feature extraction network; carrying out convolution operation on the depth feature image to obtain an attention graph, and performing normalization; carrying out attention region cutting and discarding on the normalized attention map, and inputting the attention region cutting and discarding into the convolutional neural network again for training; inputting a defective workpiece image to be recognized into the trained convolutional neural network to obtain a feature map, and performing dot product operation on the feature map and the attention map to obtain a new part feature map; performing maximum pooling operation on the part feature map to obtain attention features, stacking all the attention features to obtaina part feature matrix, and completing classification and recognition of defect images. According to the invention, the accuracy of workpiece defect detection is improved, and the method can be applied to various tiny defect detection tasks.

The invention discloses a motor imagery electroencephalogram signal classification method based on a parallel CNN-Transform neural network. The motor imagery electroencephalogram signal classification method comprises the steps of S1, preprocessing motor imagery electroencephalogram signals; S2, adding noise into the preprocessed motor imagery electroencephalogram signals to expand data; S3, performing time-frequency analysis on the motor imagery electroencephalogram signals processed in the step S1 and the step S2 to generate a two-dimensional feature map containing time features, frequency features and position information; S4, constructing a CNN model, setting network parameters, and extracting frequency features and position information in the two-dimensional feature map; S5, constructing a Transform model, setting network parameters, and extracting time features in the two-dimensional feature map; and S6, connecting the features extracted in the step S4 and the step S5 in series, and inputting the features into a classifier to obtain a motor imagery classification result. Through verification on a data set BCI complex IV dataset 2b, compared with a motor imagery classification method with good performance in recent years, an experimental result shows that the motor imagery electroencephalogram signal classification method has better classification performance.

A method for identifying a type of recording medium using a time-varying output signal from a photosensor includes amplifying the time-varying output signal of the photosensor; converting the amplified time-varying output signal of the photosensor to digitized data points using an analog to digital converter thereby creating a first set of digitized data; filtering the first set of digitized data to provide a low pass data set; filtering the first set of digitized data to provide a high pass data set; computing the standard deviation of the low pass data set; computing the standard deviation of the high pass data set; and identifying the recording medium type using values from both the standard deviation of the low pass data set and the standard deviation of the high pass data set.

The invention provides a new crown epidemic situation comprehensive evaluation method based on a classification regression model. The method specifically comprises the following steps: P1, obtaining feature information influencing epidemic situation factors; p2, constructing a classifier to classify the influence factors; p3, fitting a regression model between the influence factors and the severity of the epidemic situation; and P4, calculating factor values of all levels, weight coefficients and epidemic situation risk coefficients, and analyzing important factors influencing the epidemic situation. The computer technology of artificial intelligence and big data analysis is comprehensively utilized, the classification regression model of machine learning is dynamically established, through multiple times of adjustment of the fitting function and the floating module, evaluation of various levels of influence factors on epidemic situation risks is optimized, main and key aspects influencing epidemic situations in a target area can be analyzed, effective warning and guidance are provided for prevention and control of epidemic situations, effective and positive influences are providedfor life, economy and related aspects of people in a target area from optimization of macroscopic regulation and refinement of microcosmic prevention and control, and the effect of epidemic situationprevention and control work is greatly improved.

The present disclosure relates to systems and methods for processing real-time video and detecting objects in the video. In one implementation, a system is provided that includes an input port for receiving real-time video obtained from a medical image device, a first bus for transferring the received real-time video, and at least one processor configured to receive the real-time video from the first bus, perform object detection by applying a trained neural network on frames of the received real-time video, and overlay a border indicating a location of at least one detected object in the frames. The system also includes a second bus for receiving the video with the overlaid border, an output port for outputting the video with the overlaid border from the second bus to an external display,and a third bus for directly transmitting the received real-time video to the output port.

The invention provides an object image augmentation method and device based on multiple modes, equipment and a storage medium, and relates to the technical field of machine learning image recognition. The method comprises the following steps: acquiring a three-dimensional point cloud model of a to-be-augmented object; segmenting the three-dimensional point cloud model to obtain a plurality of composition blocks of the to-be-augmented object; acquiring actual motion video data of the to-be-augmented object, and determining one or more motion axes of the to-be-augmented object according to the actual motion video data; identifying one or more motion blocks of the to-be-augmented object from the plurality of composition blocks according to one or more motion axes of the to-be-augmented object; and setting different motion states for each motion block of the to-be-augmented object to generate image augmentation data of the to-be-augmented object. According to the invention, the image of the three-dimensional object is widened, the data set for training the image recognition model is expanded, and the robustness of the image recognition model is improved.

A system and methods are provided in which an artificial intelligence inference module identifies targeted information in large-scale unlabeled data, wherein the artificial intelligence inference module autonomously learns hierarchical representations from large-scale unlabeled data and continually self-improves from self-labeled data points using a teacher model trained to detect known targets from combined inputs of a small hand labeled curated dataset prepared by a domain expert together with self-generated intermediate and global context features derived from the unlabeled dataset by unsupervised and self-supervised processes. The trained teacher model processes further unlabeled data to self-generate new weakly-supervised training samples that are self-refined and self-corrected, without human supervision, and then used as inputs to a noisy student model trained in a semi-supervised learning process on a combination of the teacher model training set and new weakly-supervised training samples. With each iteration, the noisy student model continually self-optimizes its learned parameters against a set of configurable validation criteria such that the learned parameters of the noisy student surpass and replace the learned parameter of the prior iteration teacher model, with these optimized learned parameters periodically used to update the artificial intelligence inference module.

The invention relates to the field of natural language processing, in particular to a reading understanding choice question answering method based on data enhancement. The method comprises the following steps of: cutting a background material of a reading understanding choice question by utilizing a sliding window method; standardizing background materials, question stems and option data formats of reading understanding choice questions; using a TF-IDF method to extract answer candidate sentences from a word frequency perspective to obtain an answer material sentence set X; extracting answer candidate sentences from the angle of the high-dimensional sentence vector by utilizing Bi-Attention to obtain an answer material sentence set X; combining the sentence sets X and Y obtained in the steps S3 and S4 to obtain a candidate sentence set Z; expanding the candidate sentence set Z by utilizing an EDA strategy adaptive to reading understanding choice questions to obtain a final data enhancement candidate sentence set; and inputting the final data enhancement candidate sentence set into a BERT model for reading understanding choice question answer prediction.

The invention discloses a real-time detection and counting method for solanaceous vegetables in a plant factory. The system comprises an image and video acquisition unit, a cloud computing platform unit, a computer terminal unit, a handheld intelligent terminal unit, a data acquisition module, a data annotation module, a data enhancement module, a data set conversion module, a data set preprocessing module, a detection model training module, a detection model evaluation module and a target detection and counting module. The method comprises the following steps of: labeling image data containing a target by using a modified Labelimg and an automatic labeling and artificial auxiliary supplementary correction method, and constructing a special data set; training the optimized deep learning model in combination with an improved real-time detection method to obtain a target detection model and a parameter file; and the computer and the intelligent terminal are imported for real-time detection and counting of vegetables and fruits in the plant factory, so that the detection precision and speed are improved, a large amount of manual labor is saved, and the production efficiency is improved.

The invention relates to the field of tunnel exploration, in particular to a tunnel face surrounding rock intelligent grading method and equipment based on deep learning. The method comprises the following steps: marking and preprocessing tunnel face surrounding rock photo samples, and establishing a face image database for training; and then, a surrounding rock grading prediction model based on deep learning is established, and the model is trained according to the tunnel face image database, so that a prediction model with more accurate identification is constructed. According to the method, the to-be-detected tunnel face picture is segmented by adopting image segmentation, each sub-image is predicted, and finally, comprehensive evaluation and prediction are performed according to the prediction result of each sub-image, so that the data on which the prediction result is based is more comprehensive, and efficient and accurate discrimination of the tunnel face surrounding rock is also realized.

The invention provides a three-dimensional velocity geologic modeling method with a randomly arranged structure and wave velocity, which comprises the following steps: determining a base point in a three-dimensional space, establishing an equation according to the base point to determine a planar layered model, complicating an inclined layer of the planar layered model, and constructing a wrinklelayer model of a curved surface in the three-dimensional space; establishing a three-dimensional fault wrinkle model based on the three-dimensional curved surface wrinkle layer model in combination with the fault surface of the random reference point and the displacement of each point in the global coordinate system; based on the three-dimensional fault wrinkle model, constructing a speed model containing the salt dune, and simulating upward invasion of the salt dune in a geologic body with a certain depth; and carrying out random wave velocity amplitude according to the set stratified category, the set velocity range and the wave velocity difference range between the geology of layers, and achieving three-dimensional velocity modeling. According to the invention, the model data volume ofthe deep learning method used for geophysical inversion is improved, and the inversion effect of the deep learning method is improved.

A system and method for enhancing bitmap indexing representation of a dataset, which comprises a plurality of cases and features, each case characterized by one or more values of each feature. Currently, the bins vector for each case in the dataset, is a binary array, which is a bitmap indexing representation of each respective feature of the case. The system and method enhance the bitmap indexing by padding each bins vector. The padding is carried out by identifying all target bit locations with a ‘1’ value and replacing at least one ‘0’ bit adjacent to a target bit location with a non-zero numerical value, thereby creating a padded bitmap index. The padding factor may be based on any mathematical or statistical factor concerning population or subpopulation relevant to each of the features of the dataset.

The invention discloses a malicious code data imbalance processing method based on a swarm intelligence algorithm and a cGAN. A malicious code generation model is constructed. And calculating an acceptable optimal initial sample proportion of the malicious code by adopting a swarm intelligence algorithm. Generating malicious codes of each family, and constructing a relatively balanced malicious code data set. According to the method, an acceptable optimal sample proportion of each malicious code family is obtained by utilizing a swarm intelligence algorithm, cGAN is introduced to learn data distribution of different families of malicious codes and perform sample generation, and finally, an unbalanced data set is processed to construct a malicious code data set with relatively balanced samples, so that the malicious code data set is optimized. Different types of malicious codes reach an ideal proportion during selection, positive and negative samples have the same status in the training process, and the problem of data imbalance is more effectively solved.

The invention discloses a medical image segmentation method and system based on a distributed generative adversarial network, and the method comprises the steps of setting discriminators at all hospitals, setting a generator on a central server, and constructing the generative adversarial network between each discriminator and the generator; acquiring a medical image of each hospital; training the generative adversarial network through the medical image of each hospital; and segmenting the medical image to be segmented through the trained generative adversarial network. The generative adversarial network is trained through the medical images of the hospitals, the data set during network training is expanded, and the network training effect is improved.

A technique for simplifying structure data for representing an environment. Polyhedrons can make up structure data used in an application such as modeling, visualization, and navigation. Consequently, the operations that are performed on the data often involve determining, for each polyhedron that defines a structure such as a building, whether the polyhedron obstructs a line-of-sight line between a first point in space being considered in the application and a second point. In order to determine whether a polyhedron obstructs a line-of-sight line, a data-processing system operating on the structure data must determine whether any walls of the polyhedron intersect the line. Thus, the more polyhedrons there are or the more vertices that are in each polyhedron, the more walls there are, and the more intersection checks are required, thereby adding to the computations. The disclosed technique reduces the number of walls by simplifying objects that make up the structure data.