39 results about "Class separability" patented technology

The invention provides a traffic flow sequence classification method based on density peak value clustering. The traffic flow sequence classification method utilizes distance variance for measuring similarity degrees of subsequences, and comprises the steps of: calculating local density of each subsequence, measuring mutual cluster degrees between sequences, and searching for a clustering center by combining the local density with sequence spacing, wherein the sequence spacing is used for measuring class separability; and classifying sequences of non-clustering center by utilizing a density value, so as to obtain reasonable groups of the traffic flow sequences, and finally output a clustering result. According to the traffic flow sequence classification method, the subsequences of the same class can adopt the same time period dividing scheme and signal control scheme, and the workload of time period division and signal optimization under a timed control strategy is reduced on the premise of guaranteeing operation efficiency of the traffic flow.

The invention relates to a remote sensing classification method for binary tree multi-category support vector machines, and belongs to a remote sensing classification method. The method comprises the following steps of: dividing all classes into two subclasses; further dividing each subclass into two secondary subclasses; and repeating the process until a separate class is obtained. Therefore, a binary classification tree consisting of a series of support vector machines is finally obtained to complete the construction of a binary tree multi-category support vector machine classifier; in the classification process, classification errors appear in a classifier away from a root node, and a class with the maximum JM distance is separated first; and the established binary tree multi-category support vector machine classifier is utilized to classify remote sensing images. The invention has the advantages that: the classification speed of the method is high; the classification accuracy is superior to that of the common multi-category support vector machine classification method and the traditional classifier; and advantages of class separability and the binary tree support vector machines are integrated, so the remote sensing images can be quickly classified and the classification accuracy can be improved.

A front-end processor that is robust under adverse acoustic condition is disclosed. The front-end processor includes a frequency analysis module configured to compute the short-time magnitude spectrum, a adaptive noise cancellation module to remove any additive noise, a linear discriminant module to reduce the dimension of feature vectors and to increase the class separability, a trajectory analysis module to capture the temporal variation of the signal, and a multi-resolution short-time mean normalisation module to reduce the long-term and short-term variations due to the differences in the channels and speakers.

The invention provides a multi-visual-angle gait recognition method based on optimal discrimination coupling projection. The multi-visual-angle gait recognition method comprises the following steps: extracting a plurality of gait video sequences at a known visual angle in a training set to obtain an optimal discrimination coupling projection matrix pair; extracting and storing a gait video sequence at a standard visual angle in a registration set; and extracting and estimating the gait video sequences in a test set; projecting standard visual angle gait characteristics in the registration set and gait characteristics in the test set into mutual coupling gait characteristic space which exhibits optimal discrimination capability or optimal class separability through an optimal discrimination coupling projection matrix selected by visual angle estimation, and carrying out similarity measurement to obtain a gait recognition result. The test visual angle gait characteristics and registration standard visual angle gait characteristics are projected into joint coupling space which exhibits strongest discrimination capability, so that a multi-visual-angle gait recognition still can obtain a high recognition rate when a test gait visual angle and a standard gait visual angle in the registration set exhibit obvious differences.

The invention discloses a high-resolution remote sensing image building extraction method. Firstly, candidate object sets are extracted by combining technologies such as image segmentation and shadowdetection; secondly, based on a scale parameter self-adaptive extraction strategy, a differential attribute profile is constructed under object boundary constraints, and an initial building set is obtained; and finally, uncertain objects in the set are automatically extracted according to the inter-class separability, and secondary judgment is carried out to obtain a final building extraction result. According to the invention, the performance of remote sensing image building extraction in urban scenes with numerous ground object types is improved.

The invention provides a high-spectrum color visualization method with distance maintaining property. The method comprises the following steps of: (1) reducing the dimensions of each end member spectrum in a high-spectrum image to a 2D space as chromaticity coordinates; (2) optimizing the third-dimensional brightness value to obtain optimal distance correlation so as to determine the color label of each end member; and (3) performing linear mixing of the color according to various kinds of abundance included in the pixel. In the invention, the original high-spectrum data is optimized to a CIELab space with uniform sensing characteristics. The optimization target of the invention is as follows: the distance difference between the pixels of the generated image is related to the distance difference between the spectrums of the high-spectrum data as much as possible, and relatively great inter-class separability of the generated image is guaranteed.

The invention discloses an unsupervised domain adaptive fault diagnosis method, which comprises the steps of obtaining bearing source vibration data, and dividing the bearing source vibration data into a training sample and a test sample; constructing a CACD-1 DCNN model, training the model, determining model parameters, and performing the fault diagnosis; the source vibration data comprises target data without labels and source domain data with labels; the acquired bearing source vibration data is acquired through a sensor; an objective function of the model provided by the invention comprises cross entropy classification loss of a source domain, based on center discrimination loss and related alignment loss based on source domain and target domain features, the last two losses are executed in the last full connection layer, representation learned in a source domain can be applied to a target domain after model training, it can be guaranteed that the extracted domain invariant features have better intra-class compactness and inter-class separability, and meanwhile the extracted features can effectively improve the performance of cross-domain testing.

The invention relates to a method for on-orbit evaluation of remote sensing image change detection performance. The method includes the steps: step S1, taking a sign function of discrete cosine transform of remote sensing image gradation as a significance measurement standard to respectively extract significant areas on a current remote sensing image and a historical remote sensing image, and taking a union set of the significant areas as a significant area of a multi-temporal image; step S2, performing automatic clustering on multi-temporal image spectral signatures generated by the current remote sensing image and the historical remote sensing image, and extracting multi-temporal ultra-pixel areas according to clustering classes and spatial position relation; step S3, respectively extracting significant target surface elements in the multi-temporal ultra-pixel areas of the current remote sensing image and the historical remote sensing image, and describing each significant target surface element with structural features; step S4, carrying out on-orbit evaluation of to-be-evaluated target area change detection result performances according to inter-class separability of change class and non-change class of structural features of each significant target surface element.

The invention discloses an SVM (support vector machine) multi-class classification method based on inter-class separability and adopting a complete binary tree. The method comprises the following steps: the inter-class separability measure is calculated according to a sample distribution variance and an inter-class centre distance; two classes with minimum separability measure are selected from a class set and placed into a left child node set; a class having the maximum sum with the separability measure of all the classes in the left child node set is placed into a right child node set, and then a class having the minimum sum with the separability measure of all the classes in the left child node set is placed into the right child node set; classes are selected and added to another child node set; the process is repeated until the class set is divided completely; the left child node set and the right child node set are divided until each child node set only comprises one class, the complete binary tree is constructed, and SVM classifiers are trained for classification. By means of the method, the complete binary tree can be constructed and the classification accuracy can be improved.

The embodiment of the invention provides an image-to-video face identification method based on distinguish analysis oriented to scenes. The image-to-video face identification method comprises the steps: (1) detecting a face area in an input static image and a face video and normalizing the face area to the same size; (2) extracting the feature of the face area subjected to normalization processing and conducting dimensionality reduction; (3) establishing different mapping matrixes for the extracted static scene face and video scene face and calculating the feature expression of an original feature in mapping space; (4) optimizing mapping matrix parameters of a static scene and a video scene by binding intra-class compactness and inter-class separability according to image video face training data coupled with identification; (5) inputting the face video to be tested and the face image in the data base into the corresponding mapping matrixes, calculating new features, and adopting the nearest neighbor heuristics to identify face identification. According to the image-to-video face identification method based on distinguish analysis oriented to scenes, different mapping matrixes are established for different scenes, modeling is effectively conducted according to the data feature of different scenes, and the mapping matrixes oriented to the scenes are optimized by utilizing linear judgment, analysis and learning, so that the converted feature is well distinguished.

The invention relates to a hyperspectral image waveband selection method based on a quantum evolution particle swarm optimization algorithm, and belongs to the field of image processing. The hyperspectral image waveband selection method comprises the following steps: inputting a hyperspectral image of a waveband to be selected, and setting the scale, dimension and maximum iteration frequency of apopulation; mapping the position occupied by each particle from the unit space to the solution space of the optimization problem, and selecting the combination of the inter-class separability and theoptimal index as a fitness function; and introducing the variation probability into a quantum evolutionary particle swarm algorithm, classifying the output optimal waveband combination image by adopting a maximum likelihood method, calculating the overall classification precision, and calculating the average correlation between waveband combination wavebands adopted by the correlation. The hyperspectral image waveband selection method combines the quantum evolution particle swarm optimization algorithm with the particle swarm optimization algorithm, and can overcome the defect that local optimization is likely to happen, and the quantum evolution particle swarm optimization algorithm has the higher convergence speed, so that the operation time of the algorithm is shortened, and when waveband selection is carried out, the algorithm is more stable, and the classification precision is high, and the application prospect is wide.

The invention discloses a bypass distinguisher method and system based on class separability criteria. The method includes the steps that after plaintext is input for hardware to be measured, an objective function is cyclically executed, and a power consumption curve generated in the execution process is collected; the intermediate value of each conjecture subkey is calculated; after the intermediate values are classified according to different classification structures, the divergence matrix criterion in each class is analyzed to obtain a corresponding statistical magnitude curve, and combination processing is carried out on obtained classification results to obtain comprehensive statistical magnitude curves of the conjecture subkeys; finally, the conjecture subkey corresponding to the comprehensive statistical magnitude curve with the highest peak value is a candidate key. The bypass distinguisher method and system based on class separability criteria are suitable for all application scenes of the DPA method, have higher universality than the DPA method and solve the problem of singular information utilization of the DPA distinguishing method; meanwhile, multi-point information utilization in DPA can be achieved, and analysis efficiency can be improved.

The invention belongs to the technical field of radar target identification and specifically relates to a between-class separability enhanced subspace true and false target characteristic extraction method. According to the invention, the between-class separability level is utilized as an enhancement factor for increasing the effect of a data sample with large between-class separability in a subspace establishment process, so that the separability between different classes is increased further and the target identification performance is improved. Even superposition exists in target areas, a high identification rate can also be achieved. Simulation tests are performed on one-dimensional distance image data of four classes of simulation targets and the test results prove the effectiveness of the method.

The invention discloses a hyperspectral remote-sensing-image active contour segmentation method based on a spectral angle constraint. A spectral angle is used to measure a spectral similarity among index measurement pixels. According to a class separability principle, an optimal wave band suitable for being segmented is selected. And then, a hyperspectral remote-sensing-image active contour segmentation model based on a spectral angle constraint function is designed. A segmentation method which is applied to a two-dimensional image is expanded to and applied to a multidimensional hyperspectral remote sensing image. In the invention, the model can comprehensively use space information and spectrum information of the hyperspectral remote sensing image during a segmentation process; and influences of insufficient spatial resolution, a fuzzy target edge, a heterogeneous area and the like on a segmentation result are reduced.

The invention relates to the technical field of person-face identification and in particular to an identification method for person-face similarity. The identification method comprises the following steps: obtaining person-face images; detecting whether the person-face images include person-faces or not; converting the person-face images into a grey column diagram associated with pixel points; positioning person-face characteristics; calculating matrixes within class scatter, matrixes between class scatter as well as average scatter indexes; calculating a class separability matrix and separability measurement; obtaining separability indexes, thereby obtaining mapping degree indexes. The identification method disclosed by the invention is relatively high in accuracy, safety and reliability, and can be used for effectively realizing automation of staff management.

The invention relates to a hyperspectral image classification multi-kernel learning method capable of maximizing the class separability, and relates to multi-kernel learning model solving. The invention aims to solve problems that multi-kernel model solving is not considered to be combined with subsequent classification application and the solving efficiency is low in an existing multi-kernel learning method applied to hyperspectral image classification. The hyperspectral image classification multi-kernel learning method is implemented according to the following steps: step one, training samples and test samples are acquired from a given input hyperspectral image data set; step two, a base kernel matrix Km in the multi-kernel learning model is constructed by using the training sample set Xtrain={xi}<i=1><N>, and a base kernel matrix set {Km}<m=1><M>={K1, K2,..., Km} is acquired; step three, the within-class discrete degree and the between-class discrete degree of the data set in a Hilbert kernel space are measured by using the base kernel matrix set; and step four, the class separability is measured according to a maximum class interval principle on the basis of the within-class discrete degree and the between-class discrete degree, and the weight of the base kernel is solved by taking the maximum class separability as a solving principle of the multi-kernel learning model. The hyperspectral image classification multi-kernel learning method is applied to the field of pattern recognition.

The invention discloses a multi-target fast fuzzy clustering color image segmentation method based on semi-supervised learning and histogram statistics. The method comprises the following steps: inputting a color image to be segmented; Setting an initial population number of the color image, the Maximum number of iterations, and the Maximum clustering number, obtaining paired constraint information given by a user and processing the paired constraint information to obtain paired constraint matrixes; obtaining a color histogram according to the color image; based on the Color histogram, constructing a fitness function of intra-class compactness based on paired constraint and inter-class separability based on clustering inpurity; secondly, designing self-adjusting crossover and variation rate, utilizing mixed crossover and non-uniform variation to generate filial generations so as to ensure the diversity of the filial generations, finally constructing clustering non-purity effectivenessindexes based on color histograms, selecting an optimal solution from an optimal solution set, and obtaining a final segmentation result. According to the method, under the condition that a small amount of paired constraint information is used, rapid segmentation can be achieved, and the segmentation number and the good segmentation effect are obtained in a self-adaptive mode.

The invention provides a nonlinear process fault identification method and system based on discriminant kernel slow characteristic analysis. The nonlinear process fault identification method based ondiscriminant kernel slow characteristic analysis comprises the steps of extracting fault directions of a to-be-identified fault data set and a historical fault mode data set by utilizing a discriminant kernel slow feature analysis algorithm according to minimization of time change of a normal working condition data set and maximization of inter-class separability between the normal working condition data set and the to-be-identified fault data set at the same time; and comparing the similarity between the fault direction of the to-be-identified fault data set and the fault direction of the historical fault mode data set, and determining the fault type of the to-be-identified fault data according to the similarity coefficient. Class label information of nonlinear process data can be utilized, and the fault identification performance of the nonlinear process is improved.

The invention provides a fuzzy clustering evaluation method based on dichotomy modularity, which integrates intra-class compactness, inter-class separability and dichotomy modularity together and is used for determining an optimal classification result of a fuzzy C-means clustering algorithm. The index is combined with intra-class compactness and inter-class separability, the robustness of the index is enhanced, the optimal cluster number can be accurately detected, and the accuracy of evaluating the clustering result is improved.

The invention relates to an SAR image ship target identification method. According to the SAR image ship target identification method, training data are preprocessed through a proposed inter-class sample imbalance processing technology (including up-sampling processing based on data enhancement and a method for generating batches in proportion), and the training data sent into a network are kept in inter-class balance while the diversity of the training data is improved; through the provided dense residual network used for SAR image ship target identification, reutilization of original features can be realized while more new features are learned; through a proposed loss function based on center loss, network model parameters are adjusted, and simultaneous optimization of intra-class compactness and inter-class separability is realized. A recognition result on OpenSARShip shows that the designed dense residual network has higher accuracy, smaller model size and smaller calculation amount compared with a conventional neural network.

The invention discloses an insect classification scheme design method based on insect biological parameters. The invention can be used for the classification and identification research of multiple migratory insects. This method obtains the multidimensional biological parameters of various migratory insects, constructs the measure of inter-class separability based on the cluster center, and then forms the model of multi-class separability by repeated sorting-based merging according to the measure of inter-class separability. A binary tree composed of one branch is used to obtain the classification scheme. Finally, based on the training samples, l-1 binary classifier is trained, using binary classifier support vector machine to build classification model, to complete the classification scheme design process. A method for classify insects base on biological parameters such as insect weight, wing flapping frequency, body length and that like is also provided, and a data process and construction method are provided. The invention provides a reasonable classification and identification scheme, finally realizes species identification of migratory insects with high probability, and improves identification accuracy of migratory insect species.

The invention discloses a face recognition method based on a mutual exclusion regularization technology, and belongs to the technical field of image processing, and the method comprises the steps: firstly importing a screened and calibrated face image data set into a model; Performing feature extraction on the input image data by using a residual neural network to obtain a feature vector; calculating a classification probability by using the feature vector and the classification parameter matrix to obtain an identification vector; using an A-Softmax loss function calculates a loss difference between the identification vector and the label and using the loss difference as a gradient to carry out back propagation so as to update parameters in the feature extraction network; calculating a partial derivative of the mutually exclusive regularization term with respect to the classification parameter matrix and updating the classification parameter matrix using the partial derivative; considering The intra-class compactness and the inter-class separability at the same time, the method has more practical significance for application of a face recognition model in an open environment, denseclustering of intra-class data and discrete distribution of inter-class data under a face recognition task can be achieved at the same time through the method, and the method is more practical and universal compared with an existing method.

The invention relates to a method for on-orbit evaluation of remote sensing image change detection performance. The method includes the steps: step S1, taking a sign function of discrete cosine transform of remote sensing image gradation as a significance measurement standard to respectively extract significant areas on a current remote sensing image and a historical remote sensing image, and taking a union set of the significant areas as a significant area of a multi-temporal image; step S2, performing automatic clustering on multi-temporal image spectral signatures generated by the current remote sensing image and the historical remote sensing image, and extracting multi-temporal ultra-pixel areas according to clustering classes and spatial position relation; step S3, respectively extracting significant target surface elements in the multi-temporal ultra-pixel areas of the current remote sensing image and the historical remote sensing image, and describing each significant target surface element with structural features; step S4, carrying out on-orbit evaluation of to-be-evaluated target area change detection result performances according to inter-class separability of change class and non-change class of structural features of each significant target surface element.

The invention discloses a rotating machinery fault diagnosis method based on zero trial learning and feature extraction, which belongs to the field of deep learning and fault diagnosis, and comprises a feature refining module which mainly solves the problem of cross-dataset deviation existing in most existing methods and integrates semantic visual mapping into a unified generative model, so that the fault diagnosis efficiency is improved. In order to refine visual features of visible and invisible class samples, adaptive edge center loss is introduced to explicitly encourage intra-class compactness and inter-class separability, and the adaptive edge center loss is combined with semantic cycle consistency constraints, so that a feature refinement module can learn more distinct feature representations related to classes and semantics, and the robustness of the features is improved. According to the method, the problem of cross-dataset deviation is effectively solved, low efficiency and over-fitting risks of fine tuning are avoided, and the method has remarkable performance gain.

The invention provides a high-spectrum color visualization method with distance maintaining property. The method comprises the following steps of: (1) reducing the dimensions of each end member spectrum in a high-spectrum image to a 2D space as chromaticity coordinates; (2) optimizing the third-dimensional brightness value to obtain optimal distance correlation so as to determine the color label of each end member; and (3) performing linear mixing of the color according to various kinds of abundance included in the pixel. In the invention, the original high-spectrum data is optimized to a CIELab space with uniform sensing characteristics. The optimization target of the invention is as follows: the distance difference between the pixels of the generated image is related to the distance difference between the spectrums of the high-spectrum data as much as possible, and relatively great inter-class separability of the generated image is guaranteed.

The invention relates to a method for recognizing a ship target in a SAR image. This SAR image ship target recognition method preprocesses the training data through the proposed inter-category sample imbalance processing technology (including upsampling processing based on data enhancement and proportional generation batch method), while improving the diversity of training data, Keep the training data sent to the network balanced between categories; through the proposed dense residual network for ship target recognition in SAR images, the original features can be reused while learning more new features; through the proposed The loss function of the center loss adjusts the parameters of the network model to realize the simultaneous optimization of intra-class compactness and inter-class separability. The recognition results on OpenSARShip show that the designed dense residual network has higher accuracy, smaller model size and less calculation than the existing neural network.

The invention discloses a classification-oriented hyperspectral image wave band selection method, which comprises the following steps of solving traces of a ratio of an inter-class dispersion matrix to an intra-class dispersion matrix of each wave band of a hyperspectral image, and arranging the traces in a descending order; improving a linear decline convergence factor into a self-adaptive nonlinear decline convergence factor by adopting a grey wolf algorithm; reading the first half of the hyperspectral image waveband sequence, performing random arrangement, and taking the first half of the hyperspectral image waveband sequence as an initial population of an improved grey wolf algorithm; using the trace of the ratio of the inter-class dispersion matrix to the intra-class dispersion matrix of each population as an objective function of the improved grey wolf algorithm, searching the maximum value of the objective function, wherein an individual corresponding to the maximum value is the selected wave band combination. The method can effectively select the waveband subset suitable for classification, considers that the basic grey wolf algorithm is slow in convergence speed and easy to fall into a local extremum, combines the class separability criterion with the grey wolf algorithm, improves the convergence factor, and improves the search performance of the grey wolf algorithm.

A hyperspectral image classification multi-kernel learning method that maximizes category separability, and the invention relates to solving multi-kernel learning models. The present invention aims to solve the problems that the existing multi-kernel learning method applied to hyperspectral image classification does not consider the combination of multi-kernel model solution and subsequent classification application and the solution efficiency is low. Follow the steps below: step 1, obtain training samples and test samples from a given input hyperspectral image data set; step 2, use the training sample set to construct the base kernel matrix K in the multi-kernel learning model m , get the base kernel matrix set. Step 3: Use the base kernel matrix set to measure the intra-class dispersion and the inter-class dispersion of the data set in the Hilbert kernel space; Step 4. Based on the intra-class dispersion and the inter-class dispersion, use The maximum category interval criterion measures the category separability, and the maximum category separability is used as the solution criterion for the multi-kernel learning model to solve the base kernel weight. The invention is used in the field of pattern recognition.

The invention discloses a hyperspectral remote-sensing-image active contour segmentation method based on a spectral angle constraint. A spectral angle is used to measure a spectral similarity among index measurement pixels. According to a class separability principle, an optimal wave band suitable for being segmented is selected. And then, a hyperspectral remote-sensing-image active contour segmentation model based on a spectral angle constraint function is designed. A segmentation method which is applied to a two-dimensional image is expanded to and applied to a multidimensional hyperspectral remote sensing image. In the invention, the model can comprehensively use space information and spectrum information of the hyperspectral remote sensing image during a segmentation process; and influences of insufficient spatial resolution, a fuzzy target edge, a heterogeneous area and the like on a segmentation result are reduced.