44results about How to "Good classification effect" patented technology

The invention discloses a patent document abstract-based automatic patent classification method. The method comprises the following steps: dictionary construction, generation of various layers and classes or feature vectors of an IPC, patent text feature selection, patent text vectorization, SVM-based classification model construction and to-be-classified patent classification. In the patent text feature selection, the information contained in the patent titles and abstracts are fully utilized so that the calculation amount of the patent feature selection is remarkably decreased; in the calculation process of the patent text vectorization, a TF-IDF calculation method is improved, and the part of speech weights and position weights of words are added, so that the classification effect is further improved; and in the patent classification, a layer-based classification method is adopted, and the method fully utilizes the advantages of an SVM classifier and a KNN classifier, so that the disadvantages that the SVM classifier is overmuch in training and the KNN classifier is large in calculation amount are overcome.

The invention discloses a multispectral remote sensing image terrain classification method based on deep and semi-supervised transfer learning. A training data set and kNN data are extracted according to ground truth; the training data set is divided into two parts to be trained respectively; a multispectral image to be classified is inputted, and two classification result images are obtained from two CNN models; two kNN nearest neighbor algorithm images are constructed according to the training samples; the tested data are extracted by using the two classification result images, and the data are classified by using the kNN nearest neighbor algorithm; the classification result images are updated; the training samples and the kNN training samples of cooperative training are updated; and two cooperative training CNN networks are trained again, and the points having the class label of the test data set are classified by using the trained model so that the class of partial pixel points in the test data set is obtained and compared with the real class label. The k nearest neighbor algorithm and the sample similarity are introduced so that deviation of cooperative training can be prevented, the classification accuracy in case of insufficient training samples can be enhanced and thus the method can be used for target recognition.

The invention discloses a quick evolution method for an optimized deep convolution neural network structure. The method comprises the following steps that: 1) utilizing an evolution algorithm based onGNP (Gene Network Coding) to effectively construct a nonlinear CNN (Convolutional Neural Network) structure, and carrying out mutation on various hyper-parameters of the CNN structure to search an optimal CNN hyper-parameter combination; 2) in an evolution process, designing a multi-objective network structure evaluation method, simultaneously taking classification accuracy and the complexity degree of the classifier as an optimization objective so as to aim at generating the CNN classifier with high classification accuracy and a simple structure; and 3) putting forwarding an incremental training method, and carrying out filial generation CNN structure training on the basis of a previous generation of CNN structure. By use of the method, the training frequency of the model can be reduced,and the time complexity of an algorithm is lowered.

The invention discloses a power equipment classification method based on deep learning under a small sample in the field of image processing. The power equipment classification method comprises the following steps: step 1, obtaining a standardized power equipment infrared image through a substation equipment detection device; step 2, establishing an electric power equipment infrared image sample library, and making a training set, a verification set and a test set; step 3, establishing a small sample learning network, training the established convolutional neural network by using a training set of a sample library, verifying the model through a verification set, and obtaining a connection weight and a bias parameter of the network model after training; and step 4, classifying the infrared images in the test set by using the trained network model to generate a classification result of the infrared images of the power equipment, obtaining a good effect under the condition that the sample size is relatively small, and not needing a large amount of training time, and being applicable to management and control of the power equipment.

The invention discloses a hyperspectral image classification method based on a three-dimensional convolutional neural network and relates to a hyperspectral image classification method. The inventionaims to solve a problem that an existing two-dimensional convolutional neural network coarsely rearranges a three-dimensional signal into a two-dimensional signal, spatial information in a hyperspectral image which can not be fully utilized, and the spatial information and spectral information in an original three-dimensional hyperspectral image are destroyed. The method comprises the steps of (1)inputting a hyperspectral image data set into a MATLAB platform and obtaining a processed data set, (2) taking a new hyperspectral image as a training set, (3) building the three-dimensional convolutional neural network according to the training set of a three-dimensional matrix form, (4) using the training set of the three-dimensional matrix form to train the three-dimensional convolutional neural network and obtaining a trained three-dimensional convolutional neural network, and (5) using a test set of a three-dimensional matrix form to input the trained three-dimensional convolutional neural network and obtaining a test set classification result. The method is used in the field of image classification.

The invention discloses an unbalanced classification method based on deep reinforcement learning, which comprises the following steps: 1) constructing classification tasks and interaction rules of anagent; 2) constructing the action space of the agent; 3) constructing an external environment; 4) constructing a depth neural network model; 5) training the unbalanced classification model based on depth reinforcement learning, that is to say, learning a Q function with the deep neural network model. The present invention applies deep reinforcement learning to the problem of unbalanced classification in supervised learning, to improve the role of minority sample features in classification modeling by giving higher rewards and punishments to minority sample classification actions through rewardfunction. The method makes the agent learn the correct classification strategy in different types of data environment and data of different unbalanced degree, and is applicable to unbalanced binary classification and multi-classification problems. Therefore, the method has practical application values and is worth popularizing.

The invention discloses a strong convection wind power grade prediction method based on weather radar data. The weather radar monitoring data is used as prediction information source, and the prediction of strong convection wind power is processed as "classification problem under supervised learning". According to the corresponding relation between the monitoring data of the strong convection wind power by the weather radar and the wind speed monitoring data of an automatic weather station, a classifier is constructed by means of serial SWM method, and the three-level prediction of wind power can be realized. Through the down-sampling method, the influence of the imbalance between low wind speed data sample and high wind speed data sample on the module classification can be reduced. The wind power prediction model can conduct quantitative prediction on ground wind brought by strong convection weather, and the deficiency that a conventional wind forecast of the meteorology department can not cover "small-scale, emergent and easily-passing-away" strong convection weather can be made up. The wind power prediction model can be the important support of the early warning of the risk of power transmission line wind deflection discharging under the weather of strong convection.

The invention discloses a relation classification method based on a convolutional neural network and relates to the relation extraction and classification technology field, wherein a forward instance and a reverse instance are combined in the method. The method comprises the following steps of S1, for a sentence text entity to be classified, according to a front and back linear sequence of words in a sentence, dividing into the forward instance and the reverse instance; S2, using a CNN sentence encoder to code a forward instance entity and a reverse instance entity, and constructing a coding characteristic vector of the forward instance of the sentence and a coding characteristic vector of the reverse instance; and S3, according to the coding characteristic vector of the forward instance and the coding characteristic vector of the reverse instance, using a softmax layer to carry out relation classification and acquiring a classification result ri. Compared to the other methods, by using the method of the invention, on a CNN, forward and reverse conditions of a same entity pair are combined so as to carry out integration classification, and a final classification effect is increased.

The invention discloses a non-rigid face detection and tracking positioning method. The method comprises the following steps that: a camera shoots a photographic image or a video image; and face detection and tracking are carried out. According to the method of the invention, a shape information mechanism of an object is created through using a model similar to an active appearance model (AAM); an active shape model (ASM) adopts a parametric sampling shape to form an object shape model; a PCA method is adopted to construct a motion model of control points for describing a shape; and a group of parameters is utilized control the position change of the shape control points, so that the shape described by the shape control points can be approximate the shape of a current object. The method simply uses the shape of the object and the shape-based training model, so that the method can be implemented more easily.

The present invention provides a finger vein encryption method. According to the finger vein encryption method, a multi-scale block centrally symmetric local binary mode is used to extract the features of an image, and under a fuzzy commitment framework, BCH codes and binary codes are combined with a key. According to the multi-scale block centrally symmetric local binary mode, the mean gray value of pixel block regions is adopted to replace the gray value of a single pixel in a CSLBP operator, so that comparison is performed, and therefore, the multi-scale block centrally symmetric local binary mode is insensitive to the influence of noises; the extracted features contain the microstructure and macro structure of the image, and therefore, the information of the image can be expressed more completely, and a classification effect is better; after MB-CSLBP operator encoding, obtained binary codes are adopted as the features of the finger vein image; and the features are combined with the key which has been subjected to BCH encoding, so that the key can be encrypted.

The invention discloses a method for classifying hyperspectral images based on a semi-supervised conditional random field. The method comprises the following steps of: (1) initializing; (2) inputting a training sample set corresponding to the object classification of the to-be-classified hyperspectral images; (3) training parameters of a unitary potential information amount; (4) inputting to-be-classified test data; (5) classifying in advance; (6) confirming an average value of all spectral vectors in the training sample set and the to-be-classified test data; (7) updating a training sample; (8) updating the parameters of the unitary potential information amount; (9) training the parameters of a binary potential information amount; (10) confirming a posterior probability of the unitary potential information amount; (11) confirming the posterior probability of the binary potential information amount; (12) updating the information; and (13) confirming classification. A semi-supervised theory is adopted in the invention; the effect for the condition of insufficient training sample is especially obvious; and the method has the advantages of low calculation complexity, high classifying accuracy and wide adaptability.

The invention belongs to the technical field of machine learning, and discloses an audio classification and segmentation processing method based on a support vector machine. The audio automatic classification and segmentation are the important means for extracting structural information and semantic content in audio, and are the basis for understanding, analyzing and searching the audio content; the method comprises two aspects including audio classification and audio segmentation, and the classification method based on the support vector machine is adopted as the classification method; the support vector machine SVM is the main achievements of machine learning in recent years; and the SVM can solve the practical problems including small samples, nonlinearity, high dimension and the like and becomes one new hot spot for neural network research. In the segmentation method, the audio segmentation method with the bayesian information criterion is used for carrying out segmentation point confirmation. For the audio segmentation, different audio types are extracted from audio stream of audio classification, namely, the audio stream is divided according to the category of the timer shaft. The experiment proves that the audio classification algorithm based on the SVM has the good classification effect, and the smooth audio segmentation result is more accurate.

The invention discloses a hepatocellular carcinoma differentiation level grading method based on machine learning. The method comprises the following steps: establishing a prepositive server, a pre-processing server and a data center, carrying out region-of-interest division on the MRI image; selecting stable features of the feature set through Pearson correlation coefficients; carrying out Redundancy removal of features, adopting a support vector machine to train a classification model, obtaining an optimal model. The technical problem of providing graded image data for hepatocellular carcinoma is solved, and a good classification effect is obtained on image classification of hepatocellular carcinoma through feature engineering such as feature construction, feature selection and feature redundancy removal in combination with an SVM linear classifier.

The invention belongs to the image processing technology field and especially relates to a sparse image classification method based on low-rank supervision. The method comprises the following steps of determining a known sample and a sample to be classified; calculating a low-rank expression coefficient matrix Z of the sample to be classified; according to the low-rank expression coefficient matrix Z, solving a supervision matrix W; establishing a sparse coding model based on the supervision matrix; carrying out iteration solving on the sparse coding model; according to a reconstruction error, carrying out classification; and calculating and analyzing a classification correct rate Rate. In the invention, through establishing the sparse coding model based on low-rank supervision and providing a solving method of the sparse coding model based on low-rank supervision, simultaneously sparsity of an expression coefficient and an approaching degree of similar samples are restrained; a weight W is expressed and solved by a low rank so that a sample characteristic can be well captured; and then the model possesses a good discriminating ability among samples so that a good classification effect is acquired.

The invention provides a micro-blog user relationship based user gender identification method and system. The method includes the following steps of S1 collecting micro-blog user information and classifying different users according to an API (application program interface) provided by a micro-blog website; S2 respectively obtaining userids of followers and fans of classified users according to userids of the classified users and organizing the userids of the followers and the fans into text; S3 performing feature extraction on training samples through information gain and classifying the samples to be classified through a maximum entropy classifier. Compared to user gender identification methods and systems using micro-blog text, the micro-blog user relationship based user gender identification method and system has a better micro-blog user gender classification effect.

The invention relates to a maximum entropy classification model and Thai grammar rule correction-based Thai sentence segmentation method, and belongs to the technical field of natural language processing. The method achieves a very good classification effect for classification of space characters in a Thai language, and achieves a good promotion effect for research work of Thai sentence segmentation and Thai sentence boundary identification; the method achieves a very good sentence segmentation effect for Thai sentence segmentation research, and provides powerful support for research work of machine translation, named entity identification, sentence similarity calculation, a large corpus library quick construction technology, information extraction, information retrieval and the like.

The invention discloses a stacking sewage treatment fault diagnosis method based on weighted integration of a plurality of meta-classifiers. A two-layer overlapping type frame structure is constructed, three classification algorithms with good classification effects on unbalanced data are selected, namely, SVM, RVM and WELM are used as base classifiers, a prediction result of the base classifiersfor the original training set is obtained as input of a second-layer meta-classifier, similarly SVM, RVM and WELM are selected as meta-classifiers, weighted integration is performed on the meta-classifiers to obtain a final diagnosis model, and a classification result is output, namely a fault diagnosis result corresponding to the to-be-tested data, through the final diagnosis model. Experiments prove that the diversity and stability of the algorithm and the classification accuracy of sewage treatment faults are improved by performing weighted integration on a plurality of meta-classifiers, sothat the overall performance of fault diagnosis in the sewage treatment process is effectively improved.

The invention relates to a method for identifying financial advertisements in text advertisements. The method solves the technical problem that the existing advertisement analysis model cannot effectively identify the financial advertisements, and comprises the following steps: (1) obtaining crawled advertisement text data from a database; (2) preprocessing the text data in step (1), performing word segmentation, and removing useless information; (3) expressing the text preprocessed in step (2) in different modes as a mode which can be processed by a computer, namely text expression; (4) selecting an appropriate classification algorithm for different text representation modes in step (3), and then extracting semantic information represented by the text into category information; and (5) integrating the classification models represented by different text representation modes in step (4) to obtain a final financial advertisement recognition model. The method can be widely applied to occasions for identifying financial advertisements in text advertisements.

The invention discloses an airborne multispectral LiDAR data land coverage classification method based on a super voxel, and the method comprises the steps: firstly carrying out the abnormal data elimination and multiband LiDAR point cloud fusion of multispectral LiDAR data, and obtaining the spatial position of a fused multispectral LiDAR point cloud and the single point cloud data of multiband spectral information corresponding to the fused multispectral LiDAR point cloud; then, on the basis of the principle of minimum information loss, carrying out voxelization on the data, and assigning values to the voxels; then, by utilizing a simple linear iterative clustering algorithm SLIC, merging voxels which are close in space and spectrum into super voxels, and performing feature extraction and standardization processing on the super voxels; and finally, adopting a support vector machine (SVM) classifier training data set to construct a one-to-many super-voxel-oriented SVM classification model, and completing the classification of ground features. The method has the advantages of being visual in principle and easy to implement, the better and more efficient classification effect is achieved, and a good foundation is laid for application such as urban basic geographic space information obtaining and updating.

The invention discloses a human face living body detection method. The method comprises the steps of obtaining a visible light image and a depth image of a target object; detecting five key points ofthe face area and the face by using a face key point detection algorithm, and cutting out local images of areas corresponding to the visible light image and the depth image according to the coordinates of the key points; respectively inputting the cut local images into local feature network branches, and respectively inputting visible light and depth global images into global feature network branches; after different network branch features are superposed, performing feature fusion, and performing dichotomy by utilizing a classification function to obtain a first processing result and a secondprocessing result; and finally, determining whether the target object is a human face living body or not according to the first processing result and the second processing result.

The invention provides a time-space domain model modeling method for ecological environment monitoring, and the method comprises the following steps: S1, collecting pollutant data at regular time, andobtaining a plurality of data frames; S2, superposing the plurality of data frames in a spatial superposition mode according to a time sequence to obtain a time-space domain data structure; S3, performing feature extraction on the time-space domain data structure by taking the change of the time-space domain data as a feature quantity to obtain a data structure without a null value; S4, performing dimension reduction processing on the data structure without the null value by using a data dimension reduction method to obtain a time-space domain model; and S5, inputting the sparse monitoring data into the time-space domain model, and converting the sparse monitoring data into dense data as input of deep learning training. The invention further provides a time-space domain model modeling system for ecological environment monitoring. The time-space domain model modeling system comprises a data input unit, a data superposition unit, a data feature extraction unit and a data dimension reduction processing unit.

The invention discloses an analog signal identification method based on a complex neural network and an attention mechanism. The identification method is characterized by comprising a model design method based on the complex neural network and the attention mechanism, and a complex meta-learning training method thereof. According to the method, the complex neural network and a multi-head attention mechanism are applied to a meta-learning model, so that the model can better capture category features of training samples, and a better categorization effect and stable performance are achieved; meanwhile, the model can be applied to most signal categorization tasks, and a better test result can be obtained with a small number of training samples.

The invention discloses a malware detection method based on an HTTP behavior graph, belonging to the technical field of network security. The malware detection method based on the HTTP behavior graphcomprises the following steps: building collected traffic into an HTTP behavior tree graph according to the collected HTTP traffic of known malicious or benign soft-wares; then extracting features ofeach node in the behavior tree graph to generate a feature tree graph; then converting the feature tree graph to a feature vector by using the Graph Embedding algorithm; then inputting the feature vector into a model for training and testing; and finally, detecting by a detection model and outputting a test result. The malware detection method based on the HTTP behavior graph solves the problem that many malware can generate legal HTTP traffic and generate a request periodically, which causes that the difficulty of distinguishing the normal software and the malware is increased and the classification effect is poor.

The invention discloses a domain adaptation method based on Fredholm learning and adversarial learning, and the method comprises the steps: feature extraction: extracting features of source domain data Xs and target domain data Xt through the same feature extractor; domain identification: after the extracted features are obtained, identifying which domain the features belong to; wherein the domainidentification is divided into two stages: (1) obtaining Fredholm characteristics; (2) identifying by a domain identifier; and sample classification: using two classifiers, a classification module receiving the features extracted by the feature extractor, inputting the features into a full connection layer for calculation, and outputting classification results from a source domain classifier Cs and a target domain classifier Ct after softmax. The domain adaptation method based on Fredholm learning and adversarial learning provided by the invention has the beneficial effect that a better classification effect can be achieved in image classification.

The invention provides a fuzzy reasoning-based single-class classification method and classifier, which fuzzifies a feature vector, generates a fuzzy rule set and corrects the fuzzy rule set. The specific method is as follows: after a target rule set is generated, rule correction needs to be carried out if the same judgment condition appears but the judgment results are different. And the single-class classifier can accurately identify abnormal samples. Firstly, data processing is carried out, then fuzzification is carried out on the data, then a fuzzy rule set is established, and rule correction is carried out. After the rule set is established, a test sample can be used for testing the rule set, and when the rule generated after the sample of the test set is subjected to data processingof the same mapping is different from the existing rule, the sample is considered as an abnormal sample and then classified into an unknown category. According to the method, the number of data set samples in the aspect of single-class classification is increased. Meanwhile, a better classification effect can be achieved by applying a more optimized algorithm, and the training speed is increased.

The invention discloses a method for extracting ground object spatial distribution positions based on multi-source remote sensing data. The method comprises the following steps: utilizing a hyperspectral image with Orbita "Zhuhai No. 1" 10-meter resolution and a high-resolution image with GF2 0.8-meter resolution; a multi-source remote sensing data fusion technology is adopted to discover that anincluded angle between spectral information of the same ground object in an image is slightly changed by less than 0.05 degree, and the problems of geometric distortion, geographic position offset, coordinate system error reporting and the like are not discovered after the image is compared with an original high-resolution image. The multi-source remote sensing method is adopted to extract the spatial distribution positions of the tea, and experiments find that the multi-source remote sensing method generates a more ideal effect on the extraction of the spatial positions of the tea.

The invention provides a remote sensing image classification method based on a capsule network. The remote sensing image classification method comprises the following steps: inputting an original remote sensing image; processing the original remote sensing image into a sample set with consistent pixels; dividing a training set and a test set from the sample set; building a VGG16-capsule network model; training the VGG16-capsule network model; and utilizing the trained VGG16-capsule network model to classify the test set. According to the method, the hybrid network model of the VGG16 network model and the capsule network is adopted, the hierarchical structure of the remote sensing image can be effectively captured, and appropriate advanced features are extracted, so that the remote sensingimage is classified, and the classification effect is better.

The invention relates to rock slice image recognition, and discloses a rock slice recognition method, device and equipment and a storage medium. The method comprises steps of inputting original rock slice image data into a trained spatial transformation scale stretching network, and obtaining the attribute classification features of a rock; an effective method being provided for attribute determination and identification name identification of rock slice images of unknown categories; and obtaining an identification name corresponding to the original rock slice image data through the attributeclassification characteristics of the rock, thereby providing a method with high feasibility for actual rock slice image classification and identification.

The invention provides a diabetes syndrome prediction system, and belongs to the field of disease syndrome prediction systems. According to the invention, a convolutional neural network model is constructed by using a deep learning algorithm, and the model can predict diabetes syndrome characteristics by using simple information such as gender, age and course of disease of the diabetes patient, and the accuracy is relatively high. The algorithm is built in a diabetes syndrome prediction system, so the method is convenient to apply and has a good prospect.

The invention discloses an image classification method based on differential privacy and hierarchical correlation propagation, belonging to the technical field of data security. The idea is as follows: determining a gray image data set D, wherein the gray image data set D comprises m gray image data sets; calculating the correlation matrix R of the grayscale image data set D and the noise averagecorrelation matrix R(bar) of the grayscale image data set D; setting the convolution neural network comprising num_conv convolution layers and num_FC full connection layers, wherein theta denotes allparameters of the convolution neural network, theta= {theta <Conv>, theta <FC>}, theta <Conv> denotes parameters of num_conv convolution layers of the convolution neural network, and theta <FC> denotes parameters of num_FC full connection layers of the convolution neural network; further obtaining the optimal parameter theta (hat), theta (hat)={theta (hat)<conv>, theta(hat)<FC>} of the convolutionneural network, wherein theta (hat)<conv> denotes the optimal parameters of convolution neural network num_conv convolution layers, and theta(hat)<FC> denotes the optimal parameters of num_FC full connected layers of a convolution neural network; taking the optimal parameter of convolution neural network num_conv convolution layers {theta (hat)<conv> and the optimal parameter of convolution neural network num_FC convolution layers theta(hat)<FC> as an image classification result based on differential privacy and hierarchical correlation propagation.