221 results about "Confusion matrix" patented technology

Occurrences of one or more keywords in audio data are identified using a speech recognizer employing a language model to derive a transcript of the keywords. The transcript is converted into a phoneme sequence. The phonemes of the phoneme sequence are mapped to the audio data to derive a time-aligned phoneme sequence that is searched for occurrences of keyword phoneme sequences corresponding to the phonemes of the keywords. Searching includes computing a confusion matrix. The language model used by the speech recognizer is adapted to keywords by increasing the likelihoods of the keywords in the language model. For each potential occurrences keywords detected, a corresponding subset of the audio data may be played back to an operator to confirm whether the potential occurrences correspond to actual occurrences of the keywords.

A system and method are proposed for measuring confusability or similarity between given entry pairs, including text string pairs and acoustic model pairs, in systems such as speech recognition and synthesis systems. A string edit distance (Levenshiten distance) can be applied to measure distance between any pair of text strings. It also can be used to calculate a confusion measurement between acoustic model pairs of different words and a model-driven method can be used to calculate a HMM model confusion matrix. This model-based approach can be efficiently calculated with low memory and low computational resources. Thus it can improve the speech recognition performance and models trained from text corpus.

The invention relates to the field of ecological environment monitoring, and discloses a terrestrial plant ecological environment monitoring method based on multi-source remote sensing data fusion, which is used for realizing identification, classification and dynamic monitoring of vegetation types in a target area more efficiently. An unmanned aerial vehicle is used for vertically shooting a target area, and unmanned aerial vehicle remote sensing three-dimensional modeling is generated; Obtaining a spectral remote sensing image of a target area through satellite remote sensing, and fusing thesatellite remote sensing image with a model generated by unmanned aerial vehicle remote sensing three-dimensional modeling; Calculating and segmenting the fused image; Determining characteristic parameter indexes participating in object-oriented classification, and classifying and extracting monitored vegetation types; Constructing a confusion matrix, evaluating the classification precision by using the confusion matrix, and extracting the area size and distribution condition of each type of vegetation; Finally, the classification results of the two stages are subjected to superposition comparison analysis, and the dynamic change ranges and areas of different vegetation types are calculated. The method is suitable for terrestrial plant ecological environment monitoring.

Occurrences of one or more keywords in audio data are identified using a speech recognizer employing a language model to derive a transcript of the keywords. The transcript is converted into a phoneme sequence. The phonemes of the phoneme sequence are mapped to the audio data to derive a time-aligned phoneme sequence that is searched for occurrences of keyword phoneme sequences corresponding to the phonemes of the keywords. Searching includes computing a confusion matrix. The language model used by the speech recognizer is adapted to keywords by increasing the likelihoods of the keywords in the language model. For each potential occurrences keywords detected, a corresponding subset of the audio data may be played back to an operator to confirm whether the potential occurrences correspond to actual occurrences of the keywords.

The invention discloses an improvement method of a Ngram model for voice recognition, comprising the following steps of: converting an original Ngram model for voice recognition into an equivalent WFSA (Weighted Finite-State Automaton) network NET1; optimizing the NET1 by using an RNN (Recurrent Neural Network) to ensure that the output possibility for each sentence in a training text is maximized when the training text is marked by using the NET1; converting the NET1 into a WFST (Weighted Finite State Transducer) pronunciation network NET2 with voice model possibility by utilizing a pronunciation dictionary; optimizing the pronunciation network NET2 by utilizing a phoneme confusion matrix to ensure that the error rate of sentences is minimized; and reversely converting the pronunciation network NET2 into an improved Ngram model, and carrying out voice recognition by using the improved Ngram model.

The invention relates to a Chinese painting image identifying method based on a local semantic concept, which comprises the following steps: 1) collecting the image of a Chinese painting works to be identified by using a scanning device and storing the image into a computer; 2) dividing the collected image of the Chinese painting works into a training sample set and a testing sample set by using a random withdrawal device; 3) respectively extracting an obvious area image from the image of the Chinese painting works from the training sample set and the testing sample set by using a visual attention model; 4) establishing an image word-packaging model of the Chinese painting works for the image of the Chinese painting works and corresponding obvious area image in the training sample set; 5) generating two corresponding spatial pyramid feature column diagrams according to the image word-packaging model and a spatial pyramid model; 6) confusing the two spatial pyramid feature column diagrams generated in step 5) by using a serial confusing method; and 7) identifying the Chinese painting image to be identified in the testing sample set by using more than one classifying method of clustering method, K nearest neighbor method, neural network method and support vector machine method, and outputting an identifying result in the manner of identifying accuracy rate and confusion matrix.

The invention discloses a bimodal emotion recognition method and system based on 3D convolutional neural networks. According to the method, first, two 3D convolutional neural networks for face emotionrecognition and body gesture emotion recognition are constructed respectively, and network model parameters are optimized based on a training set and a verification set of a bimodal face and body gesture database; second, the two neural networks obtained after optimization are tested based on a test set of the bimodal face and body gesture database to obtain a face emotion recognition confusion matrix and a body gesture emotion recognition confusion matrix; and last, priori knowledge of the face emotion recognition confusion matrix and the body gesture emotion recognition confusion matrix isutilized to fuse bimodal recognition results of a newly input face video sequence and body gesture video sequence, and a bimodal emotion classification result is obtained. Through the method, the 3D convolutional neural networks and a bimodal fusion algorithm are adopted, the subjectivity of manual feature design is avoided, the limitation of single-modal emotion recognition is overcome, and the accuracy and robustness of emotion recognition can be effectively improved.

The embodiment of the invention discloses an image recognition method, device and equipment, a computer readable storage medium and a confusion perception convolutional neural network. The confusion perception convolutional neural network comprises a prediction classifier, a confusion perception model, a correction classifier group and a probability average layer, and the prediction classifier, the confusion perception model, the correction classifier group and the probability average layer are trained by using a training sample set and are used as traditional convolutional neural network classifiers. The confusion perception model is constructed based on a confusion matrix obtained by cross validation of a prediction classifier on the training sample set. Each correction classifier is obtained by using a confusion perception model as a decision-making system and training confusion class sample data with fuzzy boundaries in the training sample set. The probability average layer outputsa classification result of the to-be-identified image according to the category probability output by the prediction classifier and the category probability output by the target correction classifier, and the target correction classifier is a correction classifier selected by the confusion perception model according to the prediction category of the prediction classifier. The accuracy of image recognition can be improved.

A system for, and method of, creating generalized tied-mixture hidden Markov models (HMMs) for noisy automatic speech recognition. In one embodiment, the system includes: (1) an HMM estimator and state tyer configured to perform HMM parameter estimation and state-tying with respect to word transcriptions and a pronunciation dictionary to yield continuous-density HMMs and (2) a mixture tyer associated with the HMM estimator and state tyer and configured to tie Gaussian mixture components across states of the continuous-density HMMs and a phone confusion matrix thereby to yield the generalized tied-mixture HMMs.

The invention provides a network intrusion detection method based on an improved convolutional neural network. The method comprises the steps of acquiring a data set, performing numeralization, performing normalization processing, improving convolutional neural network model training, performing iterative optimization of model parameters, acquiring five classifiers, outputting a five-dimensional confusion matrix to serve as a classification result, and evaluating the classification result; the network intrusion detection method based on the improved convolutional neural network provided by theinvention is based on an improved convolutional neural network model, model training is performed by using a preprocessed original sample data set in combination with a cross-layer design manner, themodel achieves a good convergence effect through continuous feature extraction and iterative optimization, and then a classification test is performed by using the trained classifier, so that the feasibility and effectiveness of an intrusion detection effect can be improved by the method.

Embodiments of the disclosure are directed towards pipe leak prediction systems configured to predict whether a pipe (e.g., a utility pipe carrying some substance such as waster) is likely to leak. The pipe leak prediction system may include one or more predictive models based on one or more machine learning techniques, and a predictive model can be trained using data for the characteristics of various pipes in order to determine the patterns associated with pipes without leaks and the patterns associated with pipes with leaks. A predictive model can be validated, used to construct a confusion matrix, and used to generate insights and inferences associated with the determinant variables used to make the predictions. The predictive model can be applied to data for various pipes in order to predict which of those pipes will leak. Any pipes that are identified as likely to leak can be assigned for further investigation for potential repair or preventative maintenance.

The invention relates to an integrated diagnosis method of a gas path fault of an aeroengine, wherein the method is based on an adaptive weighted D-S evidence theory. With the method, problems that the precision is low because of a single fault diagnosis method and the evidence conflict exists during the fusion process of multiple methods can be solved. A non-linear local diagnosis module based on a particle filter and a non-linear local diagnosis module based on an extreme learning machine are constructed respectively; and fusion of local diagnosis results is carried out by using an adaptive weighted D-S evidence theory. A multi-method integrated diagnosis structure for an engine fault; diagnosis reliability degrees of all fault modes by a local diagnosis systems based on a particle filter and an extreme learning machine are obtained in an off-line mode by using a confusion matrix; and on the basis of the reliability degrees, different weight coefficients are assigned for an evidence body. Therefore, the evidence conflict is reduced effectively and the precision of an integrated diagnosis of a gas path component fault of an engine is improved.

The invention discloses a vehicle following safety automatic assessment method based on machine learning. The vehicle following safety automatic assessment method comprises the steps that data are acquired; data cleaning is performed on the acquired data, the data meeting the requirements are reserved, and the data are standardized and normalized in the same data set D; extraction and modeling of the required feature fields are performed on the cleaned data; modeling data M used for machine learning are extracted from the cleaned and normalized data according to accident records and relevant monitoring data; the M set is randomly extracted and divided into two subsets MT and ME according to the given proportion, MT is used for model training, and ME is used for model performance verification testing; supervised classification and machine learning algorithms are adopted, modeling learning is performed by utilizing training data MT, the obtained model performance is verified by ME data and relevant confusion matrix and model classification accuracy is calculated; the results of each time are recorded and compared, and an optimal model is selected; and all the records in the data set D are automatically assessed by using the optimal model, and the results are appended to the data set D and the results are outputted.

The present invention discloses a tree species classification method based on multi-source simultaneous high-resolution remote sensing data. High-resolution and hyperspectral data simultaneously acquired by an integrated sensor is utilized; firstly, coronal breadth identification is carried out based on the high-resolution data and an object-oriented method, then tree species classification is carried out based on spatial details and spectral features, which are extracted by the hyperspectral data, by combining a BP neural network classifier, and finally, the accuracy is verified by a confusion matrix. According to the tree species classification method disclosed by the invention, based on an edge detection multi-scale segmentation method, segmentation grades with different scales are established from multiple levels and multiple patterns and segmentation and information extraction are carried out layer by layer, so that the classification accuracy of tree species and forest types of subtropical natural secondary forests are promoted.

Systems, methods, and articles of manufacture to perform an operation comprising processing, by a machine learning (ML) algorithm and a ML model, a plurality of images in a first dataset, wherein the ML model was generated based on a plurality of images in a training dataset, receiving user input reviewing a respective set of tags applied to each image in the first data set as a result of the processing, identifying, based on a first confusion matrix generated based on the user input and the sets of tags applied to the images in the first data set, a first labeling error in the training dataset, determining a type of the first labeling error based on a second confusion matrix, and modifying the training dataset based on the determined type of the first labeling error.

The invention relates to the field of environmental monitoring, discloses a water and soil conservation monitoring method based on unmanned aerial vehicle low-altitude remote sensing and object-oriented classification, and solves the problems of poor applicability and low working efficiency of an existing method. The method comprises the following steps that: carrying a visible light camera on anunmanned aerial vehicle to shoot a water and soil conservation monitoring target area, and collecting a control point; carrying out coordinate system registration and area integral adjustment and multi-visual angle image dense matching to generate dense point clouds, furthermore, generating a TIN (Triangular Irregular Network), and generating the real three-dimensional model of an area to be detected through texture mapping; calculating and segmenting the three-dimensional model to obtain a segmentation scale which is suitable for the actual optimum of the water and soil conservation monitoring target area; determining a feature parameter index which participates in object-oriented classification, and carrying out supervision classification and extraction; and adopting an accuracy evaluation method and a confusion matrix based on position information to calculate total classification accuracy and a Kappa coefficient, and evaluating the classification accuracy of the model. The method is suitable for water and soil conservation monitoring.

The invention relates to an indoor positioning method based on a hidden Markov model. According to the indoor positioning method, an RSSI fingerprint feature is utilized for positioning, and an offline stage and a real-time positioning stage are included. The offline stage comprises the steps that beacon base stations and reference positions are set; an undirected connected graph of the reference positions is set up, and a position transfer matrix is set according to the undirected connected graph; the RSSI feature values of the beacon base stations are collected at all the reference positions, and an RSSI fingerprint database of a positioned area is set up; the confusion matrix of the reference positions and RSSI fingerprints is set up through a bayes method, and a hidden Markov positioning model is set up. The real-time positioning stage comprises the steps that mobile equipment collects the real-time RSSI feature values and sends the real-time RSSI feature values to a positioning server; the positioning server calculates a track sequence with the maximum continuous motion possibility, and the last position of the track sequence is adopted as a positioning result. According to the method, errors caused by RSSI fluctuation in the continuous motion process to the positioning result can be effectively reduced, and the precision and stability of indoor wireless positioning are improved.

The invention discloses an underdetermined blind source separation method based on density, mainly solving the problems that the computation complexity of the prior art is high, the prior art is easily influenced by an initial value and the number of source signals needs to be given. The method comprises the following steps: after the low energy sampling data of an observed signal is removed, projecting the observed signal onto a right half unit hypersphere; computing the density parameters of all projective points, and deleting the projective points with smaller density; utilizing an improved K-mean value clustering algorithm to cluster the surplus projective points and determining the optimal clustering number and clustering center; removing the cluster with less number of data objects, wherein the number of the surplus clusters is the estimated value of the number of source signals, and the corresponding clustering center is the estimated value of each column vector of a confusion matrix; and according to the observed signal and the estimated confusion matrix, adopting a linear programming method to recover the source signals. According to the method, the computation complexity is reduced, the influence of the initial value on the estimated performance is reduced, the confusion matrix can be estimated when the number of the source signals is unknown, and the estimating precision of the confusion matrix and the source signals can be increased.

The invention discloses an unsupervised domain adaptation method based on an adversarial learning loss function, and the method comprises the steps: (1), generating a high-level feature of a source domain image through a feature extraction network G, carrying out the cross entropy loss with a real label through a classifier C, generating a confusion matrix through a domain discriminator D, and correcting a pseudo label into the real label; and (2) generating high-level features of the target domain image through a feature extraction network G, generating pseudo tags through a classifier C, generating a confusion matrix of the high-level features through a domain discriminator D, and correcting the pseudo tags to be in opposite distribution. (3) confronting and optimizing the loss functionby a feature generator and a discriminator. In addition, for the confusion matrix on the target domain, a correction label is generated and serves as a label of the target domain, and the classifier is optimized. By utilizing the method and the device, the noise of the pseudo tag can be corrected in unsupervised domain adaptation, and the distribution difference between the domains is matched, sothat the classification precision of the target domain is improved.

The present invention provides a method for integrating crowdsourced annotations. According to the method provided by the present invention, a generalized anti-Gaussian distribution is defined by using a regularization hyperparameter, a spacing distance hyperparameter, an annotator voting weight, and a difference between the number of times that annotators annotate a current prediction item as a corresponding estimate and the number of times that annotators annotate the current prediction item as a subcategory; sampling is performed to obtain an auxiliary parameter; and annotator weights are updated by using the auxiliary parameter, thereby significantly enhancing a discrimination capability of a model. Then, a traditional annotation integration majority voting model and a confusion matrix model are integrated, thereby achieving an objective of more comprehensively describing a data generation process. In addition, an accurate prediction item update value is obtained through sampling, and moreover, running efficiency is also improved.

A multi-resolution feature extraction method and apparatus. In the illustrative embodiment, the feature extractor includes circuitry for receiving and transforming a time variant data signal into a multi-resolution data signal. The multi-resolution data signal is compared to each of a plurality of object templates. The system then generates a feature vector based on a correlation of the multi-resolution data signal to one of the object templates. The multi-resolution feature extraction method employs object templates formed by transforming time variant image data for each of a plurality of objects into a respective multi-resolution template and averaging all templates for each respective object. The method includes steps for transforming an incoming time variant data signal into a multi-resolution data signal, comparing the multi-resolution data signal to each of the object templates, and generating a feature vector when the multi-resolution data signal correlates to one of the object templates. In a more specific implementation, the method further includes the steps of calculating a confusion matrix (CM), classifying the feature vectors as one of the objects to thereby produce classified objects responsive to the CM, and selecting a target from the classified objects. A multi-resolution feature extractor according to the present invention employs object templates formed by transforming time variant image data for each of a plurality of objects into a respective multi-resolution template and averaging all templates for each respective object to thereby generate object templates.

The invention discloses a ballistocardiogram signal automatic multi-classification identification method based on LSTM. The method comprises the steps of acquiring a data set training sample; preprocessing the obtained data set training sample by adopting a wavelet threshold denoising method to obtain a pure ballistocardiogram signal; matching the obtained ballistocardiogram signals with a self-adaptive threshold value to complete positioning of IJK waves, and obtaining heart beat interception of the ballistocardiogram signals; constructing an LSTM network model, intercepting a heart beat of the obtained ballistocardiogram signal as input data of the LSTM model, and training and testing the network model; in the training process, a back propagation algorithm is used for carrying out weightoptimization on the constructed LSTM network model, so that the network is converged to global optimum; and outputting an identification rate by the LSTM network model to obtain a classification accuracy rate, calculating a kapaa coefficient according to the correctness of confusion matrix classification, and evaluating the classification accuracy of the model. According to the method, the problems of long-range dependence of RNN and excessive dependence on manual design and low classification precision of a traditional method can be solved, and a good classification effect is obtained.

The invention discloses a method for classifying hyperspectral images on the basis of a combination of unmixing and adaptive end member extraction. The method is used for solving the technical problem of large errors of an existing method for classifying hyperspectral images on the basis of spectral unmixing. The technical scheme includes that the method comprises steps of roughly classifying the images, and extracting end member sets of various categories by the aid of a confusion matrix; linearly spectrally unmixing training samples in the various categories by the aid of the acquired end member sets, and acquiring optimal classification results by the aid of a probability classifier with an abundance value optimized on the basis of multinomial logistic regression; updating the end member sets of the various categories according to the classification results; iterating the procedure, and continuously optimizing the classifier so that the classification accuracy is improved. The method has the advantage that as shown by test results, the average accuracy of tests on a simulated data set, the average accuracy of tests on data of a true hyperspectral data set AVIRIS Indian Pine and the average accuracy of tests on data of a true hyperspectral data set ROSIS Pavia University are 81.98%, 62.19% and 82.38% respectively.

A speech recognition system comprises exactly two automated speech recognition (ASR) engines connected to receive the same inputs. Each engine produces a recognition output, a hypothesis. The system implements one of two (or both) methods for combining the output of the two engines. In one method, a confusion matrix statistically generated for each speech recognition engine is converted into an alternatives matrix in which every column is ordered by highest-to-lowest probability. A program loop is set up in which the recognition outputs of the speech recognition engines are cross-compared with the alternatives matrices. If the output from the first ASR engine matches an alternative, its output is adopted as the final output. If the vectors provided by the alternatives matrices are exhausted without finding a match, the output from the first speech recognition engine is adopted as the final output. In a second method, the confusion matrix for each ASR engine is converted into Bayesian probability matrix.

The present invention relates to a method for modeling a common-language speech recognition, by a computer, under the influence of multiple dialects and concerns a technical field of speech recognition by a computer. In this method, a triphone standard common-language model is first generated based on training data of standard common language, and first and second monophone dialectal-accented common-language models are based on development data of dialectal-accented common languages of first kind and second kind, respectively. Then a temporary merged model is obtained in a manner that the first dialectal-accented common-language model is merged into the standard common-language model according to a first confusion matrix obtained by recognizing the development data of first dialectal-accented common language using the standard common-language model. Finally, a recognition model is obtained in a manner that the second dialectal-accented common-language model is merged into the temporary merged model according to a second confusion matrix generated by recognizing the development data of second dialectal-accented common language by the temporary merged model. This method effectively enhances the operating efficiency and admittedly raises the recognition rate for the dialectal-accented common language. The recognition rate for the standard common language is also raised.

The invention relates to the field of neural network technical methods, in particular to a road network traffic state judgment method based on MFD + spectral clustering + SVM, and the method comprisesthe following specific steps: (1) firstly, carrying out traffic state grade division on a road network MFD by using a spectral clustering algorithm; (2) training an SVM multi-classifier by using thedivided road network MFD parameters, and giving a model classification result precision evaluation method based on a confusion matrix; and (3) finally constructing an Internet of Vehicles simulation platform, selecting a BP neural network classifier as a comparison classifier, and carrying out truth analysis. In the road network traffic state judgment method based on MFD + spectral clustering + SVM, traffic state grade division is performed on the road network MFD by using a spectral clustering algorithm; then an SVM multi-classifier is trained by using the divided road network MFD parameters,a classification result precision evaluation method is given based on a confusion matrix, finally an Internet of Vehicles simulation platform is built, a BP neural network classifier is selected as acomparison classifier, and truth analysis is carried out.

The invention discloses a short-time and long-time feature modeling fusion-based environmental sound recognition method and device. According to the invention, a model cascaded fusion method is adopted, so that short-time and long-time information can be utilized in a whole identification process. According to the technical schemes of the invention, the method includes two stages. According to the first stage, pre-classification is performed on sliding windows based on short-time features and by using the modeling of the Gaussian mixture model (GMM); confidence judgment is performed on the classification results of the GMM; a result with high confidence is directly adopted as a final classification result; and when lower confidence appears, re-classification is carried out based on long-time features. According to the second stage, based on analysis on a GMM classification result confusion matrix, classes easy to be confused are found out; and a support vector machine (SVM) classification model between the classes is trained; and re-classification is carried out by using a support vector machine (SVM). The probability score of the Gaussian mixture model used in the modeling process of the second stage is added to the long-time features, so that the probability score and the long-time features are together adopted as the input of the SVM.

The present invention discloses a method and system of land cover verification considering POI (Point of Interest) data spatial heterogeneity. The method comprises: obtaining a POI data set from a Sina Weibo open platform, performing overlapping o the urban area building data provided by an OSM (Open Street Map) and the POI data set, dividing POIs of an urban region and a non-urban region, and correcting a division result; classifying the determined urban region POI, and obtaining the POI classification result of the urban region; classifying different types of POIs of the non-urban region, and obtaining the POI classification result of the non-urban region; generating grid data according to the POI classification result, employing a majority voting algorithm to determine the land cover model of each grid, and generating a land cover classification result based on the POI; and employing a confusion matrix to obtain the verification precision of GlobeLand30 land cover data products based on the POI, and realizing the land cover precision verification considering POI spatial heterogeneity.

The invention discloses a hierarchical front face bayonet vehicle model recognition method. The hierarchical front face bayonet vehicle model recognition method regards a vehicle front face detectionand positioning technology and a vehicle logo positioning technology as the basis, adopts a plurality of deep convolutional neural networks for integrated learning, realizes the recognition of three tasks including brand vehicle logo, vehicle model and vehicle model year of a vehicle respectively (wherein the vehicle model year contains the vehicle logo and the vehicle model, and the vehicle modelcontains the vehicle logo), then combines output confidence of learning classifiers at different levels, a similar vehicle model confusion matrix obtained through prior statistics and an inclusion relationship of the three levels, and outputs a robust vehicle model recognition result. The vehicle model identified by adopting the hierarchical front face bayonet vehicle model recognition method canbe specific to the brand, model and year of the vehicle; and the vehicle model precision of the vehicle can be further improved through the integration of the three classifiers.

A system for, and method of, combined state- and phone-level pronunciation adaptation. One embodiment of the system includes: (1) a state-level pronunciation variation analyzer configured to use an alignment process to compare base forms of words with alternate pronunciations and generate a confusion matrix, (2) a state-level pronunciation adapter associated with the state-level pronunciation variation analyzer and configured to employ the confusion matrix to generate, in plural states, sets of Gaussian mixture components corresponding to alternative pronunciation realizations and enlarge the sets by tying the Gaussian mixture components across the states based on distances among the Gaussian mixture components and (3) a phone-level pronunciation adapter associated with the state-level pronunciation adapter and configured to employ phone-level re-write rules to generate multiple pronunciation entries. The phone-level pronunciation adapter may be embodied in multiple stages.