44results about How to "Increase training samples" patented technology

The invention discloses a laser radar 3D real-time target detection method fusing multi-frame time sequence point cloud. Complementing the known data set which contains the continuous frame point cloud and is incompletely labeled by the large-occlusion target by using a projection labeling complementing method; an MADet network structure is constructed; performing registration and voxelization onthe multiple frames of time sequence point clouds to generate multiple frames of aerial views; inputting the multiple frames of aerial views into a feature extraction module to generate multiple frames of initial feature maps; generating corresponding feature description for the multiple frames of initial feature maps, calculating a feature weight map, and performing weighted fusion to obtain a fused feature map; and fusing the multi-scale features of the fused feature map by using the feature pyramid, and returning the position, size and orientation of the target on the final feature map. According to the method, the problem of data sparseness of single-frame point cloud can be effectively solved, high accuracy is obtained in target detection under severe shielding and long distance, theprecision higher than that of single-frame detection is achieved, the network structure is simplified, the calculation cost is low, and the real-time performance is high.

The invention discloses an STAP method based on a wide symmetrical characteristic of a covariance matrix. The method comprises the steps that (1) space-time two-dimensional echo data are received; (2) Doppler filtering processing is carried out; (3) wide symmetrical training samples are acquired; (4) a wide symmetrical covariance matrix is estimated; (5) a self-adaption weight vector is calculated; (6) self-adaption filtering processing is carried out; (7) a result is output. The method mainly solves the problem that the covariance matrix is inaccurately estimated due to the fact that the training samples meeting independent identically distributed conditions are insufficient when an airborne radar is in a heterogeneous clutter environment. The method is based on the wide symmetrical characteristic of the covariance matrix and can effectively improve the use ratio of the echo data, the clutter covariance matrix is estimated more accurately under the condition of few independent identically distributed training samples, and the better clutter retraining and target detection effect is obtained, and the method has better application prospects in the actual extremely-heterogeneous clutter environment.

The invention relates to a characteristic extraction method of imagining action potential in a BCI (Brain-Computer Interface) device and particularly relates to a characteristic extraction method combining a regularization method and a CSSD (Common Special Subspace Decomposition) algorithm. In the method provided by the invention, regularization parameters are led; a covariance matrix of training data of a target experimenter and a covariance matrix of training data of an auxiliary experimenter are combined to form a regularization covariance matrix under the action of the regularization parameters; a regularization space filter is constructed; and characteristic analysis is carried out on the test data of the target experimenter by utilizing the regularization space filter. By using the method in the invention, the problems that characteristic value is unstable, classification accuracy rate is low and the like in the CSSD algorithm are solved when a small-sample problem is processed.

The invention discloses a feasible region training data set expansion method for mobile robot, which comprises the following steps of: acquiring a rated original image containing topography and geomorphology by using a binocular camera, and then performing standardization processing on the image, so as to facilitate subsequent data set expansion and transformation, and then performing image transformation expansion. On the basis, expanding image samples obtained under different weather and shooting conditions through image synthesis, increasing rain and snow marks and simulation of infrared rays, and acquiring image samples which can only be collected under many special conditions through conversion. According to the method, the coverage range of the data set is effectively expanded, moretraining samples are added for subsequent machine learning, the period of constructing the data set is remarkably shortened, the cost of constructing the data set is reduced, the training effect of the mobile robot is improved in an auxiliary mode, and the recognition rate of the robot on feasible regions under various special conditions is increased.

The invention relates to the technical field of power data analysis. The invention discloses a non-invasive household electric equipment online monitoring system and a fault identification method. A non-invasive electrical signal acquisition device and a real-time power utilization information multivariate feature extraction system complete acquisition of waveform signals generated by household electric equipment and extraction of multivariate power utilization features. An autoregressive moving average model ARMA, a multi-objective optimization model and an LSTM classification system analyzeand process the multivariate power utilization features to obtain an abnormal probability and a normal probability of each currently running electric equipment or a line where the electric equipment is located under each multivariate time sequence power utilization feature vector, and finally, whether a fault occurs or not is judged by a joint judgment model according to a joint probability: whenthe joint abnormal probability is greater than the joint normal probability, the current running electric equipment or the line where the current running electric equipment is located has a fault. According to the invention, the technical problem that fault identification is difficult to carry out on household electric equipment according to signals containing various electric appliance componentsis solved, the fault identification cost is reduced, and the identification accuracy is improved.

The invention specifically discloses an automatic driving direction prediction method based on a lightweight neural network. The method comprises the following steps: step 1, a neural network model is trained; and step 2, the neural network model is tested; in the neural network model training process, the obtained images are preprocessed, and data are subjected to horizontal overturning, brightness adjustment, angle adjustment and data screening operation, so that a data set is enriched, training samples are increased, and the network model is better trained. And the EffNet network and the BP neural network propagation algorithm are combined to adjust the error between the predicted steering wheel rotation angle and the actual steering wheel rotation angle, so that the network budget demand is reduced, and the method has actual reference value and a great market prospect.

The invention discloses a RNNLM system based on distributed neurons and a design method thereof. In order to solve problems such as serial problems and inability of simulating biological neuron parallel execution characteristics, large training time cost, and difficulty in containing a plurality of neurons, the structure of the RNNLM system is changed by taking the distributed neurons capable of realizing parallel execution as a center, and then the RNNLM system based on the distributed neurons is designed. The design method comprises the structure based on the distributed type neurons, a distributed type neuron autonomic training method, and a distributed type neuron coordination method. The parallel execution of the biological neurons is simulated, and the training time cost of the RNNLM is effectively reduced, and therefore the neuron number and training samples in the RNNLM are increased under a precondition of reducing the training time cost, and the practicability of the RNNLM is improved.

The invention relates to a characteristic extraction method of imagining action potential in a BCI (Brain-Computer Interface) device and particularly relates to a characteristic extraction method combining a regularization method and a CSSD (Common Special Subspace Decomposition) algorithm. In the method provided by the invention, regularization parameters are led; a covariance matrix of trainingdata of a target experimenter and a covariance matrix of training data of an auxiliary experimenter are combined to form a regularization covariance matrix under the action of the regularization parameters; a regularization space filter is constructed; and characteristic analysis is carried out on the test data of the target experimenter by utilizing the regularization space filter. By using the method in the invention, the problems that characteristic value is unstable, classification accuracy rate is low and the like in the CSSD algorithm are solved when a small-sample problem is processed.

The invention discloses a clutter image generation method for an SAR image, and the method comprises the steps: respectively inputting different random noises into a real part generative adversarial network and an imaginary part generative adversarial network which are trained in advance, and correspondingly obtaining a clutter real part and a clutter imaginary part; combining the clutter real part and the clutter imaginary part to obtain a clutter image, wherein the real part generative adversarial network and the imaginary part generative adversarial network are used for respectively obtaining a real part training set and an imaginary part training set by extracting real parts and imaginary parts from clutter slice images of a large number of SAR images; and training a generative adversarial network by using the real part training set and the imaginary part training set. According to the method, the clutter image with the same characteristics as the real clutter can be generated, and the statistical research requirements of SAR image ground clutter and other clutters are met.

The invention discloses an AI model training method, an AI model calling method, equipment and a readable storage medium. The AI model training method comprises the following steps: acquiring metadatafor training, and loading a to-be-trained AI model, wherein the AI model is realized based on a deep reinforcement neural network; blocking the metadata of which the metadata belongs to the same frame to obtain a plurality of sub-metadata; performing feature extraction on the sub-metadata to obtain a feature vector corresponding to the sub-metadata; training the AI model according to the featurevector, and determining whether the trained AI model is converged or not; and if the trained AI model is not converged, executing a step that if the trained AI model is converged, the trained AI modelis sorted. The training sample of model training is improved, so that the training accuracy and training efficiency of the AI model are improved.

The present invention relates to a social media rumor detection method based on multi-task learning, which specifically includes the following steps: S1: performing data extraction and format conversion on the corpus, obtaining the source, reply and propagation path of the post; S2: extracting the characteristics of the writing style; S3: Extract the features of user confidence; S4: Do text preprocessing on the text part of the source post and the reply post to input the follow-up tasks; S5: Concatenate the features extracted by S2 and S3 with the text representation of S4; S6: Put the spliced ​​vectors into a shared BERT layer; S7: Construct the neural network structure separately; S8: Input the data processed by S5 into the neural network structure, and output position classification and rumor classification. The invention can combine two highly related tasks with a multi-task joint model, improves the tasks of rumor detection and position classification, and improves the performance of rumor detection.

The present application relates to the technical field of robot deep learning models, and provides a method, device and electronic device for obtaining an environment recognition model and control decision-making. The method for obtaining an environment recognition model includes: obtaining an actual data information set and a simulation data information set of a robot; Based on the OPC UA information model, a plurality of the actual data information is converted into a plurality of first nodes, a plurality of the simulation data information is converted into a plurality of second nodes, and by calculating the relationship between the first node and the second node The correlation value of the actual data information set and the simulation data information set are fused to obtain the fusion data information set; the fusion data information set is used as the training set of the environment recognition model, and the Identify the model for training. The training sample of the present invention is large and has high reliability.

The present invention provides a CT image block reconstruction method and system based on deep learning. The system includes: a CT projection data filtering module, which performs local range filtering and full-area filtering on the input CT projection data; an image block reconstruction module, which calculates The projection data position corresponding to the current image block to be reconstructed, and extracting the filtered projection data at the projection data position; and connecting the extracted filtered projection data with each pixel of the image block, and reconstructing the image blocks, and regularize the reconstructed image blocks; then filter the regularized image blocks; the image synthesis module synthesizes all the reconstructed image blocks to obtain the synthesized image; image reconstruction and output module , filtering the synthesized image, obtaining a reconstructed image, and outputting it. The invention improves the image reconstruction ability, saves the projection iteration step, and has fast reconstruction speed, and is suitable for situations such as inconsistent responses of detection units.

The invention discloses a commodity sorting processing method and device, equipment, a medium and a product. The method comprises the steps of extracting a comprehensive feature vector of a to-be-sorted commodity sample of a to-be-sorted commodity object corresponding to a commodity query event; inputting the comprehensive feature vector into a sample dynamic routing layer to drive a routing decision model in the sample dynamic routing layer to determine the commodity sorting difficulty of the to-be-sorted commodity samples according to the comprehensive feature vector; querying a target commodity sorting model corresponding to a sorting difficulty interval in which the commodity sorting difficulty is located in a commodity sorting model pool; and inputting the to-be-sorted commodity samples into the target commodity sorting model, and obtaining sorting scores of the to-be-sorted commodity samples to determine ranking positions of the to-be-sorted commodity objects in the queried commodity ranking. According to the invention, a commodity sorting system suitable for e-commerce shops and independent stations is realized, the commodity samples are pushed to the corresponding models for reasoning through dynamic routing, and the processing efficiency is high.

A data preprocessing method for classifying ECG signals based on a deep learning model, including the following steps: a. Obtain ECG signals marked by experts as training samples, including normal ECG signals and abnormal ECG signals. The measurement time of ECG signals is arbitrary, assuming The longest measurement duration is t seconds, and the sampling frequency is fs; b. Perform noise reduction processing on the original training samples, and use wavelet transform to remove baseline drift; c. Divide the training samples into a training set and a test set, and perform data amplification on the training set; d. Input the training set into the deep learning model for training, and use the test set to optimize the model parameters; e. After the original ECG signal is preprocessed in b and c, take t×fs data points as samples and input it into the model to obtain the ECG signal classification results. The present invention can increase the number of samples, achieve sample balance at the same time, make the model easier to train, and help improve the classification ability and robustness of the model. Sample balance makes the model easier to train, and helps improve the classification ability and robustness of the model. sex.

The invention discloses a hidden channel detection method between Android applications based on the Intent communication behavior semantic graph, which includes the following content: screening suspicious candidate application sets from the target Android platform; obtaining Intent communication events by monitoring Intent related functions to establish candidate applications- Intent function call weight graph; perform relationship matching between sending broadcast message calling behavior and receiving broadcast message calling behavior, and establish sending application-receiving application association graph; decompose sending application-receiving application association graph into multiple Intent communication pairs, and extract Intent communication The semantic description vector of the behavior of the pair, and extract the sensitive permission flag vector of the two applications in the communication pair, combine the two vectors to form the collusion application feature vector, and perform supervised learning on the vector to realize the detection of hidden channels between applications. The invention uses communication features to describe the collusion stealing behavior of a pair of Android application programs, has good applicability, and is suitable for detecting Android collusion stealing applications under the situation of large differences in operating environments and insufficient training samples.