59results about How to "Reduce training complexity" patented technology

An optical proximity corrected mask design is generated from a given a target mask design by processing the target mask design through a feature trained neural network, configured to perform an optical proximity correction of geometric features, to obtain a representation of a first corrected mask design. The target mask design is processed in parallel through a rule processor, configured to perform placement of sub-resolution geometric features relative to geometric features in the target mask design, to obtain a representation of a second corrected mask design. A layout reassembler operates to generate a corrected mask design through an overlaid composition of said first and second corrected mask designs.

The invention discloses a word vector model based on point mutual information and a text classification method based on CNN. The method comprises the following steps: (1) training a word vector modelthrough a global word vector method based on point mutual information; (2) determining a word vector matrix of the text according to the trained word vector model; (3) extracting features from word vector matrix by CNN and training classification model; (4) extracting input text features according to the trained word vector model and CNN feature extraction model; (5) according to the text featuresextracted from CNN feature extraction model, calculating the mapping distance between text and preset categories by softmax and the cross entropy method, wherein the nearest one is the correspondingcategory of text. This method overcomes the shortcomings of Glove word vector in the semantic capture and statistical co-occurrence matrix, reduces the training complexity of the model, can accuratelymine the text classification features, is suitable for text classification in various fields, and has great practical value.

The invention provides a training method for a vehicle identification model, comprising the following steps of: collecting image samples, transforming the collected image samples, representing a transformed image sample by a feature matrix with constant size but transformable features, extracting partial features in the feature matrix with constant size but transformable features of each image sample to compose a feature space, performing sparse coding on the feature space to obtain a sparse coding feature basis matrix, computing the maximum multi-scale feature vector of each image sample, and determining the parameter of a linear classifier. By the training method for the vehicle identification model, the complexity, occupied memory space and computing time of an algorithm can be reduced, so that the vehicle identification is implemented fast; and meanwhile, the vehicle identification precision is improved. The invention further provides a vehicle identification method and apparatus based on the training method for the vehicle identification model.

The present invention discloses a method and a system for melanoma image tissue segmentation based on a deep neural network. A deep neural network technology is employed to solve the tissue segmentation problem of dermatoscope images of melanoma, and the method and the system for melanoma image tissue segmentation based on a deep neural network are mainly a medical image analysis processing technology. An improved deep neural network structure to perform modeling, dermatoscope images having segmentation tags are employed to train the model, and the trained model has an ability of suspicious tissues to new dermatoscope images. The method and the system are to locate suspicious areas in the melanoma dermatoscope images and to perform pixel segmentation. The method and the system for melanomaimage tissue segmentation based on the deep neural network employ a new deep learning technology, fully develop the capacity of collection of various hierarchical features of the image data, can be applied to the modeling process, can perform location and segmentation of the suspicious skin tissues and can provide good reference for further analysis by dermatologists.

The invention discloses a cognitive radio frequency spectrum perception method based on compressed sensing and a BP (back-propagation) neural network. The method of the invention comprises a training process and an actual detection process. The training process comprises: carrying out compression sampling whose rate is lower than a Nyquist sampling rate on an original signal of cognitive radio; carrying out 1-bit quantification processing on sampling data; taking the compression sampling data after the quantification processing as training input, taking an actual frequency band occupancy situation as training output, and training a BP neural network detector. The actual detection process comprises: carrying out the compression sampling on the original signal of the cognitive radio and carrying out the 1-bit quantification processing on the sampling data; inputting the compression sampling data after the quantification processing into the trained BP neural network detector so as to obtain the output, wherein the output is the frequency spectrum occupancy situation of the cognitive radio. The invention also discloses a cognitive radio frequency spectrum perception system which uses the above method. By using the method of the invention, algorithm complexity is low and real-time performance of the frequency spectrum detection can be greatly increased.

The invention relates to an ultra-short-period wind speed prediction method based on a spectral clustering type and genetic optimization extreme learning machine. The method comprises the steps that S1. data are prepared; S2. the prepared data are preprocessed; S3. the preprocessed data are subjected to wavelet transformation; S4. the data obtained after wavelet transformation are subjected to normalization processing; S5. through correlation analysis, the data obtained after normalization processing are selected, so that an input variable is determined; S6. through main component analysis, the input variable generated in the S5 is subjected to dimension reduction processing; S7. through a spectral clustering type method, the data obtained after dimension reduction processing in S6 are subjected to clustering analysis, and an extreme learning machine sample space is formed with the data obtained after normalization processing in the S4; S8. through the extreme learning machine and a genetic algorithm, the data of the extreme learning machine sample space formed in the S7 are subjected to hierarchical prediction; and S9. hierarchical prediction values are added, and an ultra-short-period wind speed prediction value is obtained. Ultra-short-period and multi-step prediction on wind speed are achieved, prediction accuracy is improved, the computing amount is greatly lowered, and prediction efficiency is improved.

The invention relates to a method for detecting and identifying targets based on a component structure model. The method comprises the following steps: extracting a target and the gradient direction histogram characteristics of different module dimensions in the parts of the target; respectively training the target and each part of the target to generate a boost cascade classifier, wherein the weak classifiers in the cascade classifier comprise the direction members of gradient direction histogram characteristic vectors; determining the position of the target in a manual labeling mode by adopting a semi-supervised training mode, wherein the position of each part of the target is determined by the position of modules at which the multiple weak classifiers with strong separating capacity are selected and the multiple weak classifiers are selected in the course of training the integral cascade classifier; training the space relation model between the target and each part of the target byadopting a star structure; respectively detecting the target and each part of the target through the boost cascade classifier to obtain a part detection cost graph; and then, realizing the detection and identification positioning of the target by utilizing the range conversion and the relevant position relation among the parts of the target.

The invention discloses a CNN multipoint regression prediction model for traffic flow prediction. The CNN multipoint regression prediction model comprises the following steps that a first perception input layer; a second convolutional layer that performs convolution on the input layer data and outputs the data by an activation function; a multi-layer convolutional layer that performs convolution on the output of the last layer as an input and outputs data by an activation function; a fourth full-link layer; a fifth discard layer: a random discard layer discards some redundant neurons and retains 40% to 70% of the full-link nodes of the last layer; a sixth output layer that subjects the effective node output of the discard layer to regression calculation, obtains a regression value as the output of the entire network, sets a total of m output nodes, that is, maps the full-link layer to the output layer to achieve weight combination. Compared with a traditional statistical regression model, the CNN multipoint regression prediction model has a feature extraction ability associated with a data space, has the advantages of local receptive field and weight sharing, and makes a better balance in time complexity and feature selection.

The invention discloses a network traffic anomaly detection method based on a small amount of annotation data, and the method comprises the steps: carrying out the dimension reduction of a feature vector through employing double auto-encoders, and then carrying out the supervised training through employing a deep neural network; dividing the network traffic into two types of positive samples and negative samples, and finally screening out a part of important samples in unlabeled data and submitting the samples to experts for labeling, increasing the number of labeled samples, iteratively updating an auto-encoder and a classifier, and then employing the trained classifier for detecting network traffic abnormality. According to the invention, a double-auto-encoder architecture is proposed, pure positive and negative samples are used for respectively training the auto-encoders, and the stability of the classifier is improved. Meanwhile, the loss function of the deep neural network is improved, the sample weight is adjusted in a finer-grained manner, the problem of overfitting caused by imbalance of positive and negative samples and small training sets is solved, a new method for calculating the marking value of the unmarked data is provided, the samples with high marking value are selected to be delivered to experts, and the marking cost is reduced.

The invention provides a convolutional neural network seismic signal denoising method based on attention guidance. The method comprises the following steps: synthesizing effective signals in a seismic record by adopting Ricker wavelets; preprocessing the synthesized seismic data set, and constructing a training data set containing a noise set and a signal set; and inputting training data into an ADNet network model composed of four modules including a sparse block (SB), a feature enhancement block (FEB), an attention block (AB) and a reconstruction block (RB) for training, and after training is completed, suppressing the noise of seismic signals by using the ADNet network. According to the denoising method provided by the invention, the noise in the seismic signal can be suppressed, detail information is reserved, and the processing effect is relatively good.

The invention relates to a training method of an abnormal region image generation network and a related product. The method comprises the steps of acquiring a training sample image, wherein the training sample image has an abnormal area; inputting the training sample image into an initial abnormal region image generation network to obtain an initial abnormal region image, and fusing the initial abnormal region image with the training sample image to obtain a fused image; wherein the fused image comprises a preset simulation mark; inputting the fused image and the real sample image into an initial discrimination network to obtain a discrimination result of the fused image; calculating the loss between the discrimination result and the simulation mark by adopting a loss function, and training the initial discrimination network and the initial abnormal region image generation network according to the loss; and when the value of the loss function reaches convergence, completing training ofthe initial abnormal region image generation network to obtain an abnormal region image generation network. The method can improve the precision and training efficiency of the abnormal region image generation network obtained by training.

The invention relates to a deep inspection method for automatic identification of a ship corrosion area, wherein the method comprises the steps: building a pre-training image identification model and a corrosion image automatic identification model, then labeling a large number of ship images, and training the corrosion image automatic identification model through a part of images with labels; inputting a plurality of to-be-detected ship images into the trained model, reserving the images marked with corrosion labels, and marking a corrosion area bounding box, wherein the marked images form a corrosion area positioning data set; establishing a corrosion area target detection model, and training the model by using part of the images in the corrosion area positioning data set; and finally, inputting a to-be-positioned rusted ship image into the trained corrosion area target detection model, outputting bounding box parameters of a corrosion area, and marking the position of the corrosion area by using a bounding box. According to the method, an integrated process from ship image acquisition to corrosion area automatic positioning is realized, a deep learning image recognition algorithm and a target detection algorithm are fused, and the method is very intelligent.

The invention provides a filling method of a sensor missing value fused with spatio-temporal information. The method comprises the following steps of: inputting N pieces of historical data X and M pieces of missing data Xmissing, wherein M and N are greater than the input time sequence length T; filling a threshold eta, and inputting historical data into the trained LSTM-AES, then eta = std (X-X ') so as to obtain the trained model LSTM-AES and the repaired data Xrepaired; dividing the original data into a time series data set; initializing LSTM-AES; then, using the Tensorflow for carrying outnetwork initialization; updating the weight W of the LSTM-AES by using a common back propagation algorithm of the neural network; and carrying out missing value filling. Space-time information is considered at the same time, robustness can be achieved when a large number of sensors are missing at the same time, a single model can be trained to process different types of missing, and the real-timerequirement of sensor missing value filling can be met.

The invention discloses a target node key information filling method and system based on an association network, and belongs to the technical field of data mining, machine learning and graph theory. The problem of low accuracy of the filled key information of the target node in the prior art is solved. According to an application scene, the method comprises steps of establishing a relational network of a large number of nodes; based on network relation, obtaining an association network of a target node with key information, integrating the association network into a data structure comprising atarget node, a label, an association node, an association node weight and an attribute vector, performing multiple three-dimensional sampling on the data structure based on an improved random forestmethod to obtain a subset of a plurality of training decision trees, giving a plurality of decision trees to perform training, and performing integration after training to obtain a final model; and based on the associated nodes of the to-be-filled target node, performing prediction through the final model, and performing weighted average on multiple results after prediction to obtain final fillinginformation. The key information of the target node is filled based on the association network.

The embodiment of the invention provides a weak supervision target detection method and device. The method comprises the steps of extracting image features of a to-be-detected image; wherein the to-be-detected image comprises at least one to-be-detected object; determining an initial detection result for each to-be-detected object based on the extracted image features, each initial detection result being an initial detection frame containing one to-be-detected object; based on the extracted image features, determining a salient region of each to-be-detected object; and optimizing each initialdetection result based on each salient region to obtain a target detection result for each to-be-detected object. Compared with an existing method for detecting a supervised target, a segmentation model does not need to be trained additionally, the salient region serves as the auxiliary information for optimizing the detection result, so that the training complexity of the target detection task isreduced, the salient region serves as the auxiliary information of the optimization process, errors in the optimization direction can be avoided, and the training efficiency is improved.

The invention discloses an image quality enhancement system and method based on meta learning, and a storage medium. The system comprises a denoising processing network, a deblurring processing network and a super-resolution processing network which are connected in sequence. The denoising processing network is used for de-noising an input image; the deblurring processing network is used for deblurring the input image; and the super-resolution processing network is used for performing super-resolution processing on the input image. According to the system, the image enhancement technology is integrated, the joint task of denoising, deblurring and super-resolution is realized, and through a meta transfer learning deblurring algorithm, the training speed of the deblurring network is increased, the training complexity is reduced, and the image enhancement network has better universality. According to the method, important information in the image is highlighted according to specific needs, unnecessary information is weakened or removed, the efficiency of network training is improved through meta learning, and the effect of the network can be improved.

The invention discloses a hybrid neural network fault prediction method and system for a high-performance computer, and the method comprises the steps: collecting the log data of the high-performance computer, wherein the log data comprises a log event id, a timestamp when a corresponding log event occurs, and a log event level; performing data cleaning and feature selection on the collected log data to obtain initial feature data; constructing a fault prediction model by using a random forest algorithm, inputting the obtained initial feature data into the fault prediction model, calculating feature importance by using the random forest algorithm, and performing feature selection to obtain feature sample data; and inputting the obtained feature sample data into an LSTM network model, and predicting whether a fault event exists in the feature sample by using the LSTM network model. According to the method, the log data features are scored and selected through the random forest, and the dimensionality is reduced, so that the training complexity can be reduced, and the training degree is accelerated.

The invention relates to POI information classification based on a convolutional neural network, and belongs to the technical field of computers. According to the POI information classification methodand device based on the convolutional neural network, word segmentation processing is carried out on POI information, the POI information is converted to generate a vector matrix corresponding to thePOI information, and the vector matrix is processed through a pre-trained convolutional neural network model to determine the classification corresponding to the POI information. The convolutional neural network model used in the method is easy to train and obtain, the preprocessing process in the classification process is simple, and the classification efficiency is high and accurate.

The invention relates to the technical field of iris recognition, and discloses a method for iris recognition. The method comprises the steps: collecting an iris image of a to-be-recognized object, and carrying out preprocessing of the iris image; performing iris positioning on the preprocessed iris image, and determining an iris effective area; sampling the iris effective area to obtain an iris matrix image corresponding to the iris effective area; performing feature extraction on the iris matrix image to obtain a plurality of iris features corresponding to the to-be-recognized object; and constructing a random forest classifier according to a plurality of decision trees, and performing iris recognition on the plurality of iris features through the random forest classifier to obtain an iris matching result. The technical problem of single feature of an iris recognition system is effectively solved, and the recognition rate and generalization ability of the iris recognition system are improved, so that the algorithm is more stable. The invention further discloses a device for iris recognition, terminal equipment and a storage medium.

The invention provides a license plate character identification method based on a refined character classifier. The license plate character identification method comprises the steps of performing preprocessing on a to-be-trained license plate character image and obtaining a plurality of training samples; coarsely classifying the training samples into 34 classes according to the character meaning; refining each same-class training sample which is obtained through coarse dividing to a plurality of subclasses; training the refined training samples by means of an SVM classifier, and obtaining the refined character classifier; performing preprocessing on a to-be-detected license plate character image and obtaining a testing sample; and identifying the testing sample by means of the refined character classifier, and obtaining the characters in the to-be-detected license plate character image. According to the license plate character identification method, through performing refined classification on the characters of the license plate, not only is training complexity of the classifier reduced, but also the refined character classifier is used for identifying the characters of the license plate, thereby preventing a problem of obscure same-class characteristic caused by character fuzzy or character inclination or different font, and improving a recognition rate of the characters of the license plate.

The invention discloses a transformer state quantity data prediction method based on kernel principal component analysis optimization; the method comprises the following steps: S100: acquiring transformer state quantity data within a period of time, and converting the transformer state quantity data into a transformer state quantity matrix in a matrix form, wherein the transformer state quantity comprises relevant data of the transformer state quantity; S200: screening the transformer state quantity matrix X based on the kernel principal component analysis algorithm, and reserving the main state quantity to obtain a transformer main state quantity matrix; S300: constructing a transformer state quantity data prediction model and training the prediction model based on the transformer main state quantity matrix; and S400: predicting the state quantity data of the transformer based on the transformer state quantity data prediction model trained in the step S300. According to the method, the model training complexity can be reduced, and therefore high prediction accuracy of the transformer state quantity data is guaranteed. In addition, the invention further discloses a corresponding prediction system.

The invention discloses a method for training a target detection model and a target detection method and device, and relates to the field of artificial intelligence, in particular to the fields of computer vision, deep learning and the like. The specific implementation scheme is as follows: obtaining a sample picture set and a labeling area of each sample picture in the sample picture set; determining a first intersection-parallel ratio of each sample picture according to the labeling area and a preset target anchor point box; determining a second intersection-combination ratio of each samplepicture according to the labeling area and an auxiliary anchor point box corresponding to the target anchor point box; and according to the sample picture set, the first cross-parallel ratio and the second cross-parallel ratio, training the to-be-trained target detection model to obtain a trained target detection model. The process can balance the model training complexity and the model training accuracy.

The invention relates to the field of artificial intelligence, and discloses a Bayesian network learning method, an intelligent device and a storage device. The method comprises the steps of obtaininga training sample which comprises the continuous node data; discretizing the continuous node data to obtain the discrete sample data; performing structure learning by using the discrete sample data to obtain a topology of the Bayesian network; and performing parameter learning by using the training sample and combining the topology of the Bayesian network to obtain the parameters of the Bayesiannetwork. In this way, the speed and the accuracy of the training process can be balanced.

The invention relates to the field of block storage systems and artificial intelligence, and discloses a risk prediction model training method and a related device, and the method comprises the steps:obtaining a first financial data set which comprises M pieces of first financial data corresponding to a plurality of first fields; for the first financial data set, vectorizing a plurality of piecesof first financial data associated with each of the plurality of first fields to obtain a plurality of first vectors; determining the correlation between every two first vectors in the plurality of vectors by adopting a preset feature selection algorithm; determining a second financial data set from the first financial data set according to the correlation between every two first vectors; and training a risk prediction model by adopting the second financial data set. By implementing the embodiment of the invention, the training period of the risk prediction model is shortened, and the training complexity is reduced.

The invention relates to a speech translation method, device and equipment and a translation machine; and the method comprises the steps: obtaining speech information, and determining the voiceprint features of a speaker through the speech information; judging whether the speaker is a first user or not according to the voiceprint characteristics of the speaker; if the speaker is the first user, determining a first language of the first user, identifying the speech information based on the first language, and converting the identified information into information of a second language based on apreset second language; and if the speaker is not the first user, determining a preset second language, identifying the speech information based on the second language, and converting the identifiedinformation into information of the first language based on the first language of the first user. Thus, a language recognition model does not need to be established, the bilingual recognition problemamong multiple languages and the automatic voice translation selection problem among the bilingual classes can be solved through voiceprint recognition, the translation training complexity is reduced,and the translation efficiency is effectively improved.

The invention relates to a voice recognition method and device, computer equipment and a storage medium. The method comprises the steps: acquiring training data including standard mandarin training data and non-standard mandarin training data; inputting the standard mandarin training data into the main neural network for training to obtain a mandarin acoustic model; adding a branch neural network in the mandarin acoustic model; inputting the standard mandarin training data and the non-standard mandarin training data into a mandarin acoustic model for multi-task training; updating the mandarin acoustic model to generate an initial acoustic model; inputting the training data into the initial acoustic model for training to obtain a target acoustic model; and inputting the voice information into the target acoustic model to obtain semantic information of the voice information. Not only is the ASR recognition efficiency improved, but also the ASR identification accuracy is prevented from being influenced by accent identification errors. The invention further relates to a block chain technology, and the target acoustic model can be stored in a block chain node.

The invention relates to a blockchain technology and a detection model technology in artificial intelligence, and discloses a training method of a quality detection model and a related device. The method comprises the steps: acquiring a first financial data set, wherein the first financial data set comprises M pieces of first financial data corresponding to a plurality of first fields; for the first financial data set, determining a maximum value and a minimum value in a plurality of pieces of first financial data associated with each first field in a plurality of first fields; according to the maximum value and the minimum value in the multiple pieces of first financial data associated with each first field, mapping the multiple pieces of first financial data associated with each first field to a preset interval to obtain a second financial data set; and training a quality detection model by adopting the second financial data set. By implementing the embodiment of the invention, the training period of the quality detection model is shortened, and the training complexity is reduced.

The invention discloses a method and system for real-time prediction of intelligent shortwave frequency across frequency bands. The method comprises the steps of: using historical sampling data to train a prediction model based on a neural network, and the prediction model is used to predict the communication quality of other frequencies according to the frequency characteristic information of a certain frequency; The range of the next available frequency; Obtain the frequency characteristic information at the short-wave communication interruption moment; Input the frequency characteristic information at the short-wave communication interruption moment into the prediction model to obtain the communication quality of each frequency in the range; Select the communication within the range The frequency with the best quality is selected as the next available frequency after shortwave communication is interrupted. The present invention takes other frequency points with high reliability of link establishment near the communication interruption frequency point as the prediction target, which can reduce the training complexity of the neural network when faced with a large amount of nonlinear data, and the prediction target is more suitable for the actual working equipment Operational needs.

The invention relates to a method for detecting and identifying targets based on a component structure model. The method comprises the following steps: extracting a target and the gradient direction histogram characteristics of different module dimensions in the parts of the target; respectively training the target and each part of the target to generate a boost cascade classifier, wherein the weak classifiers in the cascade classifier comprise the direction members of gradient direction histogram characteristic vectors; determining the position of the target in a manual labeling mode by adopting a semi-supervised training mode, wherein the position of each part of the target is determined by the position of modules at which the multiple weak classifiers with strong separating capacity are selected and the multiple weak classifiers are selected in the course of training the integral cascade classifier; training the space relation model between the target and each part of the target byadopting a star structure; respectively detecting the target and each part of the target through the boost cascade classifier to obtain a part detection cost graph; and then, realizing the detection and identification positioning of the target by utilizing the range conversion and the relevant position relation among the parts of the target.

The invention discloses a circuit path level NBTI aging prediction method and device based on a key gate. The method comprises the following steps: constructing a training set and a test set; the training set is used for training a linear regression model, in the training process, the input of the model is aging delay data of the 0th year and the 1st year of the key sub-circuit, the output of the model is aging delay data of the key sub-circuit from the 2nd year to the Nth year, and N is a positive integer larger than 2; after training is completed, a test set is used for verifying the trained linear regression model, in the verification process, input of the model is aging delay data of the zero year and the first year of the key path, and output of the model is aging delay data of the key path from the second year to the Nth year of the key path; after the test is completed, inputting the aging delay data of the 0-th year and the first year of the key path to be predicted into the final linear regression model, and predicting the aging delay data of the second year to the Nth year; the method has the advantage that the prediction precision and speed are guaranteed at the same time.