30results about How to "Rich training samples" patented technology

The invention discloses a named entity identification method based on adversarial training. The method comprises the following steps: obtaining relevance features among judicial field characters through RoBERTa model training and Bi-LSTM training; splicing the two relevance features together, and predicting a training sample by using a conditional random field model to obtain a prediction result;according to the method, external word vectors and word vectors of different dimensions can be introduced to be combined with judicial domain text word mixing vectors of different dimensions, confrontation disturbance is carried out on the mixing word vectors in a judicial domain text, and the accuracy of model recognition is improved.

An SAR image change detection method based on a multi-region convolutional neural network comprises the following steps: performing difference analysis on two SAR images with different time phases atthe same geographic position to obtain a difference image; pre-classifying the differential images to obtain a constructed training data set and a constructed test data set; sending the sample training data set into a proposed multi-region convolutional neural network for training; and testing the trained network the test set so as to obtain the change detection result of the whole same-ground multi-temporal SAR image. According to the method, when the data set is constructed, the number of the data set is doubled by adding Gaussian noise, the diversity of samples is enriched, and the overfitting problem is solved; at the same time, the method uses an attention mechanism for channels and spaces to improve the performance of the network, improves the robustness of SAR image change detectionto noise, and has strong generalization ability.

The invention discloses a scientific and technological literature picture extraction method based on a Fast-RCNN. The method comprises the following steps of 1) acquiring the scientific and technological literature data by using a web crawler and preprocessing the scientific and technological literature data; 2) dividing a data set, making a label for the data in the training set, and not processing the data in the test set; 3) inputting the data in the training set into a convolution layer, and extracting the feature mapping of the pictures; (4) mapping and inputting the obtained features into an RPN module to obtain the proposal feature maps with fixed sizes; 5) classifying the specific categories by utilizing the softmax to obtain the accurate position of a target, calculating a loss function, and updating the parameters of the whole network to obtain a training model; 6) utilizing the training model to detect the data in the data set, and outputting the detected pictures. The scientific and technological literature picture extraction method is high in detection speed and high in accuracy, facilitates the further analysis and understanding of the scientific and technological literature pictures, and has the higher practical application value.

The invention discloses a rhythm phrase recognition method and device and electronic equipment, wherein the method comprises the steps: obtaining to-be-recognized target data which at least comprisestext data and audio data corresponding to the text data, wherein the text data comprises at least one statement; obtaining a text feature code corresponding to the text data and an acoustic feature code corresponding to the audio data; processing the text feature code and the acoustic feature code to obtain multi-modal features related to text and audio alignment; inputting the multi-modal features into a pre-trained rhythm recognition model to obtain a rhythm phrase sequence output by the rhythm recognition model, wherein the rhythm phrase sequence comprising a plurality of rhythm phrases, the rhythm phrases are at least segmented by utilizing rhythm symbols, and the rhythm recognition model is obtained by training at least two statement samples with rhythm phrase tags and audio samples corresponding to the statement samples.

The invention discloses a combinable weak authenticator-based named entity identification algorithm architecture. The architecture comprises an entity identification part and a result authentication part, the entity identification part is used for completing an identification task to obtain an identification result; and the result authentication part comprises two or more weak authenticators which are respectively used for checking and authenticating the identification result on the subdivision target corresponding to each weak authenticator. The weak authenticator is a module capable of independently completing a subdivision target, and required training data can be automatically generated on an existing task corpus. The weak authenticator and the entity identification part form an end-to-end network and are used for carrying out optimization learning by using a supervision method. According to the invention, the combinable weak authenticator is used for assisting the named entity identification process, so that the entity identification precision is effectively improved, and the method can be simply and quickly adapted and expanded in an entity identification scene in a specific field.

The invention discloses a training method of a battery state prediction model, a battery state prediction method and device, electronic equipment and a storage medium. The method comprises the following steps: acquiring an electrochemical model, wherein the electrochemical model is constructed by measurement operation data and attribute data of a battery; performing charge and discharge simulation based on the electrochemical model to obtain simulation operation data of the battery under different simulation conditions; and taking the simulation operation data and the measurement operation data as training samples, inputting the training samples into a neural network, adjusting network parameters of the neural network and model parameters of the electrochemical model according to an output result of the neural network, and determining the neural network meeting an iteration stop condition as a battery state prediction model. According to the invention, the training samples of the neural network are expanded and enriched by means of the electrochemical model, and the electrochemical model is optimized based on the output result of the neural network, so that the electrochemical model provides more accurate training samples for the neural network, and the accuracy of model training is improved.

The invention discloses a problem generation method and device, and the method comprises the steps: taking an original problem sample and a target problem sample as training samples, and carrying outthe inverse reinforcement learning training of a discriminator composed of a machine reading model and a scoring function, wherein the target problem sample is a semantic similarity problem generatedby inputting the original problem sample into a problem generation model; and taking the overlap ratio score output by the scoring function as a return; taking the original problem sample as a training sample, and carrying out reinforcement learning training on a generator formed by the problem generation model, so as to further form adversarial training between the generator and the discriminator and taking the target problem sample as an adversarial sample corresponding to the original problem sample, wherein the problem generation model after adversarial training can be used for generatingtarget problems with similar semantics.

The invention discloses a target application program updating method and device based on feedback information. The method comprises the steps that actual scene information input in a target application program and result feedback information corresponding to the actual scene information are obtained; according to the actual scene information and the result feedback information, a target neural network model in the target application program is updated, the target neural network model is a neural network model obtained by conducting model training on a pre-trained neural network model through a training sample set, the training sample set comprises at least two sets of sample scene information, the different groups of sample scene information correspond to different actual sample recognition results of the target object, the loss function used during model training corresponds to different weights in the different groups of sample scene information, and the weights corresponding to the different groups of sample scene information are in negative correlation with the number of the sample scene information in the different groups of sample scene information.

The invention discloses a GPU (Graphics Processing Unit)-and-neural-network-oriented grid quality detection method, thereby solving the problems of high time overhead, low automation degree and the like of a current grid quality detection method. According to the technical scheme, a detection system based on a neural network is built by utilizing the powerful computing power of a GPU and the powerful fitting learning power of the neural network, a grid sample training set is built, the network is trained, the trained neural network is adopted to detect the computing grids, and a quality classification result of the computing grids is obtained. And the detection system analyzes the quality classification result and outputs a final grid quality detection result. The four feature extraction modules are composed of convolution layers with different channel numbers and convolution kernel sizes, the grid high-dimensional features related to the calculation result precision are fully extracted, and the detection accuracy is ensured; the advantage of high calculation speed of GPU data is fully utilized, and the calculation burden of a CPU is reduced; and the grid quality detection processis accelerated by compressing the high-dimensional features.

A gastrointestinal tract endoscope image preprocessing method including information screening and fusion repairing includes the steps that firstly, a key information extractor is used for extracting effective information areas in data set pictures twice, and the results of the two times of extraction complement each other and are stored as a new picture data set; brightness equalization and improvement are carried out on a new picture data set by using an improved EACE method, the overall brightness is improved while the defect of uneven brightness is improved, and more information beneficial to polyp identification is obtained; and light spot detection and light spot region repair are performed on the enhanced result by using a two-image matching fusion repair method. The interference of light spots in endoscope images on polyp detection results is reduced, and the misdiagnosis rate during polyp detection by doctors or computers is reduced.

The invention provides a ship license plate detection method in a natural scene. The method comprises the following steps: 1) data acquisition: collecting videos of ships entering and exiting a port through a camera; 2) data processing: processing the video, and screening out a picture containing a ship license plate; 3) data annotation: carrying out ship license plate annotation on the picture; 4) model training: using a target detection model to train ship license plate detection; and 5) model detection: detecting the picture by using the trained ship license plate detection model to obtainthe position of each ship license plate in the picture. A deep learning model R2CNN is selected, the multi-angle ship license plate can be identified, the detection speed and accuracy of the ship license plate are improved, and the method has high practical application value.

The invention relates to the technical field of information, and provides a customer rating method and device based on rejection inference and a storage medium. The objective of the invention is to speculate the specific performance of a rejected sample by using the overdue performance and rejected condition information of a customer so as to solve the problem of sample deviation. According to the main scheme, the method comprises the steps of obtaining a user passing sample and a rejected sample, and performing processing derivation of related characteristics on the samples; performing vintage analysis and rolling rate analysis on post-loan performance data of the application passing user, determining definition logic of a default target variable, and defining an application pass rejection target label according to a pass rejection user; performing initial modeling on the full sample and the application pass sample by using a rejection pass target label and a default target variable respectively, and calculating a deduced default label of a rejection user according to a KNN thought by using a user vector formed by dividing the full sample by using the two models; and fusing the rejected user deducing the default label and the passing user deducing the real default label to form a final training sample.

The invention provides a method and device for generating a virtual data set. The method comprises the steps of generating a plurality of three-dimensional virtual card models based on a real card; rendering the plurality of three-dimensional virtual card models to obtain a plurality of different three-dimensional virtual cards; and generating a picture of each three-dimensional virtual card, andforming a virtual data set by the pictures of the plurality of different three-dimensional virtual cards. By applying the method, training samples can be enriched, and the recognition accuracy of thetrained intelligent recognition model is improved.

The embodiment of the invention provides a model training method and device and a storage medium, and the method comprises the steps: a terminal device determines first user data generated by a target user on the terminal device, and also can send a data obtaining request to obtain second user data generated by other users and sent by a server, the similarity between the first user data and the second user data meets a preset condition. And then, taking the first user data and the second user data as training samples, and training a user behavior prediction model for the target user. According to the method, through cooperative work of the server and the terminal equipment, the terminal equipment can obtain the first user data and the second user data which are different in source and have high similarity, training samples are enriched, and the accuracy of the model is guaranteed. And the behavior pattern reflected by the second user data and the first user data is similar to the behavior pattern of the target user, so that the prediction result output by the trained prediction model is also close to the behavior pattern of the target user.

The invention discloses a processing method for voice noise reduction of the Internet of Vehicles, and the method comprises the steps of adjusting a signal-to-noise ratio according to the voice and noise energy values, and zooming a noise signal, and obtaining different noise voices; establishing a noise reduction model; and on the basis of a noise reduction result of the noise reduction model, intercepting a part of environmental noise, and removing a mute part which is below -35 dB and lasts for more than 2 seconds by using a voice processing tool according to the fact that the average value of voice energy is greatly concentrated below -35dB. According to the method, an end-to-end deep learning network is used in an actual vehicle-mounted voice processing scene with larger wind dryness, and the step of voice conversion processing in a traditional method is omitted; more training samples are obtained by adjusting different signal-to-noise ratio parameters, the generalization ability of the model is improved, and a sample after model processing is saved after the mute part is removed, so that the storage space is saved.

The invention discloses a photo restoration system construction method, a photo restoration method and system, the photo restoration system construction method comprises the following steps: S1, constructing and training a first adversarial neural network, and converting a first old damage photo into a pseudo old damage photo through the first adversarial neural network; S2, constructing a secondadversarial neural network, and training the second adversarial neural network by taking the pseudo-old damage picture generated by the first adversarial neural network and a preset clear picture as training samples; and taking the second adversarial neural network as a photo restoration system. The method has the advantages that the old and damaged photo can be quickly and efficiently restored, and the restoration effect is good.

The invention discloses a dynamic asset combination transaction method. The method comprises the following steps: S2, establishing a capital asset pricing model; S4, obtaining a total optimization model according to the capital asset pricing model; S6, obtaining a dynamic weight optimization model according to the total optimization model; and S8, outputting an adjustment scheme according to the dynamic weight optimization model. The method has the following beneficial effects: the scheme of the invention performs optimization function calculation in real time according to the market environment, so that the response speed is high enough for the change of all factors influencing the decision; the transaction decision of the dynamic asset combination scheme is carried out all the time, so that the impact cost is low; correlation among different assets, beta risk indexes of total position holding, buying and selling cost functions and different states of risk indexes in a position holding set at different moments are all considered in scheme decisions, and are harmoniously unified together by utilizing a formula, which is also not possessed by a traditional asset combination scheme.

The invention belongs to the technical field of industrial intelligent quality inspection and detection, and particularly relates to a pipeline welding spot visual detection method, system and equipment with angle estimation. The problems that due to the fact an existing pipeline welding spot visual detection method is difficult to adapt to complex scenes with changeable detection scenes, target sizes, shielding and illumination, the pipeline welding spot detection precision is poor and the rotation angle of a local pipeline where a welding spot is located cannot be detected are solved. The method comprises the following steps: acquiring a connecting pipeline scene image of a to-be-detected welding spot as an input image; detecting whether a connecting pipeline in the input image contains a welding spot or not through a pre-trained pipeline welding spot detection model, and if yes, outputting the type, the position, the size and the rotation angle of the welding spot, wherein the pipeline welding spot detection model is constructed based on a deep neural network. The detection precision of the pipeline welding spot is improved, and the problem that the angle of the local pipeline where the welding spot is located cannot be detected is solved.