248 results about "Linear classifier" patented technology

A traffic classifier has a plurality of binary classifiers, each associated with one of a plurality of calibrators. Each calibrator trained to translate an output score of the associated binary classifier into an estimated class probability value using a fitted logistic curve, each estimated class probability value indicating a probability that the packet flow on which the output score is based belongs to the traffic class associated with the binary classifier associated with the calibrator. The classifier training system configured to generate a training data based on network information gained using flow and packet sampling methods. In some embodiments, the classifier training system configured to generate reduced training data sets, one for each traffic class, reducing the training data related to traffic not associated with the traffic class.

The invention relates to a rapid target detection method based on a convolutional neural network, and relates to the computer vision technology. The rapid target detection method comprises the following steps: training convolutional neural network parameters by utilizing a training set; solving the problem of max-pooling losing feature by using an expander graph and generating a discriminative complete feature graph; regarding the full-connection weight of the convolutional neural network as a linear classifier, and estimating the generalization error of the linear classifier on the discriminative complete feature by using a probable approximately correct learning framework; estimating the required number of the linear classifiers according to the generalization error and the expected generalization error threshold value; and finally, completing the target detection on the discriminative complete feature graph by using the linear classifiers on the basis of a smooth window. The detection efficiency and the target detection precision are obviously improved.

A computerized system and method for large scale semi-supervised learning is provided. The training set comprises a mix of labeled and unlabeled examples. Linear classifiers based on support vector machine principles are built using these examples. One embodiment uses a fast design of a linear transductive support vector machine using multiple switching. In another embodiment, mean field annealing is used to form a very effective semi-supervised support vector machine. For both these embodiments, the finite Newton method is used as the base method for achieving fast training

A classifier training method and apparatus for training, a linear classifier trained by the method, and its use, are disclosed. In training the linear classifier, signatures for a set of training samples, such as images, in the form of multi-dimension vectors in a first multi-dimensional space, are converted to a second multi-dimension space, of the same or higher dimensionality than the first multi-dimension space, by applying a set of embedding functions, one for each dimension of the vector space. A linear classifier is trained in the second multi-dimension space. The linear classifier can approximate the accuracy of a non-linear classifier in the original space when predicting labels for new samples, but with lower computation cost in the learning phase.

Mobile device (102) comprising a number of sensing entities (230) for obtaining data indicative of the context of the mobile device and / or user thereof, a feature determination logic (230) for determining a plurality of representative feature values utilizing the data, and a context recognition logic (228) including an adaptive linear classifier (234), configured to map, during a classification action, the plurality of feature values to a context class, wherein the classifier is further configured to adapt (236) the classification logic thereof on the basis of the feature values and feedback information by the user of the mobile device. A method to be performed by the mobile device is presented.

An image classification method is provided. The method includes: inputting a to-be-classified image into a plurality of neural network models; obtaining data output by multiple non-input layers specified by each neural network model to generate a plurality of image features corresponding to the plurality of neural network models; respectively inputting the plurality of corresponding image features into linear classifiers, each of the linear classifiers being trained by one of the plurality of neural network models for determining whether an image belongs to a preset class; obtaining, using each neural network model, a corresponding probability that the to-be-classified image comprises an object image of the preset class; and determining, according to each obtained probability, whether the to-be-classified image includes the object image of the preset class.

The invention provides a pulmonary nodule benignity and malignancy predicting method based on convolutional neural networks. The pulmonary nodule benignity and malignancy predicting method based on the convolutional neural networks comprises obtaining multi-scale pulmonary nodule image blocks from pulmonary nodule coordinates; according to the multi-scale pulmonary nodule image blocks, structuring the convolutional neural network corresponding to every multi-scale pulmonary nodule image block; according to a loss function, training the convolutional neural networks through the multi-scale pulmonary nodule image blocks; extracting low-dimensional features of pulmonary nodules through the trained convolutional neural networks; training a non-linear classifier through the low-dimensional features and predicting unknown pulmonary nodule image blocks. The pulmonary nodule benignity and malignancy predicting method based on the convolutional neural networks can accurately predict the benignity and malignancy of the unknown pulmonary nodule image blocks.

The invention discloses a face recognition method based on dictionary learning models. The method comprises the following steps of: mapping trained and tested face images to a low-dimension space to acquire a training signal set matrix; establishing the dictionary learning models which comprise an irrelevant dictionary learning (IDL) model and an unconstrained irrelevant dictionary learning (U-IDL) model; inputting the training signal set matrix into the IDL and U-IDL models, and solving the models to acquire an irrelevant dictionary and a linear classifier; acquiring a corresponding sparse vector of each picture belonging to a test sample based on the dictionary acquired in the last step by using a sparse expression algorithm; and inputting the sparse vectors into the linear classifier to acquire category labels of test sample pictures, wherein the result expressed by the category labels is used as the face recognition result. The invention provides the new models and the new method for dictionary learning problems in sparse expression, and the models and the method can be applied to mode identification and image classification problem under common conditions; and particularly, aiming at face recognition application, the dictionary learning method can achieve relatively high face recognition accuracy.

The invention disclosers a human body behavior recognizing method based on an RGB-D video and belongs to the technical field of computer vision behavior recognition. The method includes extracting the dense Moving Pose feature, the SHOPC feature and the HOG3D feature from the RGB-D video acquired from an RGB-D camera according to the principle that human body behaviors of different classes in the RGB-D video have different moving information, geographic information and texture information, adopting an edge-limited multi-core learning method to conduct feature fusion on the three types of features, and finally adopting an Exemplars-SVM linear classifier is adopted to judge human body behavior action. Compared with the prior art, the three types of features extracted have the advantages of illumination invariance, scale invariance and view angle invariance, obvious robustness is achieved for the appearance difference and the behavior action process difference of action executers , and the human body behavior action recognition accuracy is improved to some extent.

The invention relates to a microblog-based neologism emotional tendency judgment method, belonging to the field of natural language processing. The microblog-based neologism emotional tendency judgment method disclosed by the invention comprises the following steps: dividing words of microblog corpuses through a Chinese word division tool, blocking the corpuses, the words in which are divided, by taking stop words in a word division result as a division point, pairwise combining adjacent word strings in each block, calculating the combined word string frequency, and taking the word strings, the frequencies of which are higher than a threshold value, as neologism candidate strings; filtering the neologism candidate strings according to a word formation rule of Chinese linguistics and an adjacent change number rule so as to obtain neologisms; calculating the similarity between co-occurrence words and hownet emotional words by utilizing an emotional dictionary of a hownet; calculating the relevancy between the neologisms and the co-occurrence words; constructing an image model; and obtaining the emotional polarity distribution of the neologisms by utilizing a label propagation algorithm, and obtaining the emotional tendency of the neologisms by constructing a linear classifier. By means of judgement of the emotional tendency of the neologisms, a blogger can express views better; and furthermore, the emotional tendency of the blogger can be accurately known by users.

The invention provides a deep learning algorithm-based classification method of bacterial pneumonia and viral pneumonia in children. According to the method, a source data set is manually labeled; onthe basis of the combination of a full convolutional network semantic segmentation algorithm and a convolutional neural network algorithm, the full convolutional network semantic segmentation algorithm is adopted to perform lung region foreground segmentation on an image so as to obtain a region of interest, the extracted region of interest is inputted to a convolutional neural network model so asto train a classifier, and therefore, the category of an unknown chest X-ray image can be predicted, and the high-dimensional features of the region of interest are extracted; and a traditional imageprocessing method is adopted to extract the low-dimensional features of the region of interest; and the high-dimensional features and the low-dimensional features are used to train a non-linear classifier; and the category of the unknown X-ray image is predicted, and the type of the pneumonia of a patient can be judged. Since a main component analysis algorithm is used to perform dimensionality reduction on the features, and therefore, the amount of calculation can be reduced; and the features which have been subjected to mixed dimensionality reduction are inputted into the nonlinear classifier, and the category of the unknown X-ray image can be predicted.

The invention relates to detection device and method for multi-class targets. The detection device comprises an input unit, a joint classifier and a discrimination unit, wherein the input unit is configured to input data to be detected; and the joint classifier comprises multiple strong classifiers which can treat multiple classes of target data, wherein each strong classifier is formed by summarizing a group of weak classifiers, and each weak classifier carries out weak classifying on the data to be detected by using one feature; the discrimination unit is configured to discriminate which class of target data the data to be detected belongs to based on the classifying results of the multiple strong classifiers, the joint classifier contains a shared feature list, and each feature in the list is respectively shared by one or more weak classifiers belonging to different strong classifiers; and the weak classifiers which have the same feature but belong to different strong classifiers have mutually different parameter values.

A method for computer aided detection of anatomical abnormalities in medical images includes providing a plurality of abnormality candidates and features of said abnormality candidates, and classifying said abnormality candidates as true positives or false positives using a hierarchical cascade of linear classifiers of the form sign(wTx+b), wherein x is a feature vector, w is a weighting vector and b is a model parameter, wherein different weights are used to penalize false negatives and false positives, and wherein more complex features are used for each successive stage of said cascade of classifiers.

The invention relates to improved methods and computer-based systems and software products useful for deriving and optimizing linear classifiers based on an adjusted sparse linear programming methodology (A-SPLP). This methodology is based on minimizing an objective function, wherein the objective function includes a loss term representing the performance of the objective function on a training dataset comprising at least two separate, adjustable weighting constants associated with classification errors for data points in-class and not-in-class, respectively.

The invention discloses a hyperspectral image classification method based on linear regression Fisher discrimination dictionary learning (LRFDDL). The method includes the following steps that the textural features and the spectral features of each hyperspectral image are fused and normalized; a dictionary and a linear classifier are obtained by learning through an LRFDDL algorithm; model prediction is conducted. According to the hyperspectral image classification method based on LRFDDL, the LRFDDL algorithm is applied to hyperspectral image classification, so that the classification accuracy of the hyperspectral images is improved; in addition, the textural features of the hyperspectral images are added into the classification, and thus the classification effect of the hyperspectral images is further improved.

The invention provides a method for dynamically generating a tree-shaped combination classifier, which is used to test walkers. The method comprises the following steps that: all samples are read in and have characteristics picked up, and characteristic vectors are generated; the tree-shaped combination classifier is initialized, which makes the structure as a tree with only one root node; whether extensible leaf nodes are existed in the tree is judged; one extensible leaf node is chosen to be a father node of a training single classifier, and a training sample is chosen for the training single classifier; a single classifier is obtained by the AdaBoos algorithm; whether the classifier obtained by training meets the fissionable requirement is judged, if the classifier fails to meet the fissionable requirement, the classifier is added in the tree; the sample used to train the single classifier is divided into two parts for retraining, and two single classifiers are obtained and added in the tree; the combination classifier is constructed until the classifier meets the requirement; the tree-shaped combination classifier obtained is utilized to classify testing targets, and testing results are obtained. The method of the invention has the advantages of lowering the rate of false alarm and improving the testing rate.

The invention relates to a human body behavior recognition method based on global characteristics and sparse representation classification. The method comprises the following steps: performing Gaussian kernel convolutional filtering preprocessing on a video frame, and extracting a moving foreground pixel by using a differential method; sampling a pixel value according to a time space dimension ofa parameter, determining a moving area, adjusting the size of the video frame, performing primary dimension reduction, splicing video frames in rows to form a vector group, and acquiring characteristic vectors; splicing the characteristic vectors in rows to form a characteristic matrix, performing secondary dimension reduction, calculating a primary characteristic dictionary of the characteristicmatrix, initializing the dictionary, after dictionary initialization, performing dictionary learning by using a class accordant K-time matrix singular value decomposition method, calculating an inputsignal sparse code according to the dictionary, inputting the code into a classifier, and outputting a behavior type; and counting dictionary learning parameters, and performing behavior recognition in real time. By adopting the method, dictionaries and linear classifiers with both reconstitution functions and classification functions are acquired, human body behavior recognition efficiency is improved, and the method is applicable to scientific fields such as security monitoring, video search based on contents and virtual reality.

The invention relates to the technical field of intelligent information processing and discloses a method for classifying data streams in a dynamic data environment. The method comprises the following steps: partitioning the data streams; establishing different classifiers for different concept drift; storing in a characteristic data pool of the classifiers; when a new data block arrives, judging whether the concept drift occurs or not by Kullback-Leibler (KL) divergence; if the concept drift does not occur, classifying by using the classifier at the last moment; if the concept drift occurs, seeking the proper classifiers from the characteristic data pool of the classifiers by the KL divergence and classifying; and if no coincident classifier exists, training a new classifier, adding the new classifier into the characteristic data pool of the classifiers and deleting the outdated classifiers. By the method, stable and mutational concept drift can be detected simultaneously; when the concept drift occurs, classification is performed by selecting the proper classifier to guarantee the efficiency of a model; and the performance of the model is guaranteed by deleting the outdated classifiers.

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

The invention discloses a multi-classifier-based convolutional neural network quick classification method. According to the method, an activation function and a linear classifier are added after each convolutional layer except the last one. During network training, image features of the convolutional layers are obtained firstly and the classifiers after the convolutional layers are trained by using cross entropy loss functions. After the training, the activation functions are adjusted to enable the classification accuracy to reach the best. During execution of an image classification task, the classifiers of all the layers are activated in sequence in a forward propagation process, the image features after convolution are subjected to calculation analysis through the classifiers, a judgment value is obtained, and if the judgment value meets the activation requirements of the activation functions, classification results of the classifiers are directly output and the classification process is ended. On the contrary, forward propagation activates the next convolutional layer to continue executing the classification task. The method can classify images easy to classify in advance to finish the forward propagation process of a network, so that the network classification speed is increased and the classification time is shortened; and the method has good practical value.

The invention relates to a target tracking method and a system based on an on-line initialization gradient enhancement regression tree. In the system consisting of a video input end, a tracking target output end and an on-line training classifier, the method comprises the steps of 1) selecting a tracking target from a video series and extracting positive and negative samples of a Haar-like feature, 2) randomly establishing the on-line classifier according to the positive and negative samples to obtain a training residual error, 3) conducting training amendment by taking the training residual error as a training sample of the on-line classifier and establishing a target model, and 4) acquiring an image confidence map from a next frame of video image, determining a maximum position of a confidence value in a target window, and accomplishing tracking. According to the method and the system, the tree can be converged to an optimum point quickly to ensure that the random forest detection optimization is accomplished, the classifier is updated through on-line study, and the problems of appearance variation, rapid movement and shielding of a target are solved well.

The invention discloses a three-dimensional flow radiographic method and system based on optical coherence tomography of feature space. OCT (optical coherence tomography) scattering signals of a scattering signal sample in a three-dimensional space are collected through a collector; a two-dimensional feature space is constructed through a theoretically established classifier in combination with local signal-to-noise ratios of the OCT scattering signals and a decorrelation coefficient; dynamic flow signal and stationary tissue classification is achieved. The specific steps include calculating and analyzing the OCT scattering signals through first-order and zero-order auto-covariance to obtain the signal-to-noise ratios of the OCT scattering signals and two decorrelated features; constructing a signal-to-noise ratio reciprocal-decorrelation coefficient two-dimensional feature space; constructing a linear classifier of ID spaces based on the principle of multivariate time series, and removing background of the stationary surrounding tissues. The method and system herein can evidently inhibit the influence of system noise upon flow radiography, contrast of flow images is increased, vessel visibility of deep tissues is particularly increased, and accuracy of blood flow can be improved.

Frequency features to be used for binary classification of data using a linear classifier are selected by determining a set of hypotheses in a d-dimensional space using d-dimensional labeled training data. A mapping function is constructed for each hypothesis. The mapping functions are applied to the training data to generate frequency features, and a subset of the frequency are selecting iteratively. The linear function is then trained using the subset of frequency features and labels of the training data.

The invention discloses a time-space condition information based moving object detection method. The method comprises the following steps: building a target detection time-space domain model through considering the significance of human visual time-space domains; calculating a conditional probability that a detection image belongs to a time-space domain reference background; carrying out nonlinear transformation on the conditional probability through negative logarithm checking so as to extract time-space conditional information; carrying out weighted summation on the conditional information of image in an adjacent domain through considering the local consistency of image characteristics; and as characteristics, carrying out object detection by using a linear classifier. The conditional probability is rapidly calculated by using a color histogram, and an image block replacing a single pixel is adopted for carrying out modeling and detection, thereby reducing the algorithm complexity and the storage space requirements; and through combining with an image block difference pre-detection mechanism, the object detection speed is increased. The method disclosed by the invention is low in algorithm complexity, less in storage space requirements and high in algorithm instantaneity, and can effectively suppress the background disturbance interference and isolate the noise influence; and by using the method, the real-time detection of moving objects on the existing computers is realized, therefore, the method is applicable to embedded intelligent camera platforms.

The invention relates to a binocular-vision-based real-time extraction method and system for three-dimensional hand information. The method comprises: obtaining image information collected by a left camera and a right camera of a binocular camera system in real time; detecting hands in the images collected by the left camera and the right camera in real time; detecting a hand at a first frame by using a histogram of oriented gradient (HOG) and a support vector machine (SVM) linear classifier and detecting hands at follow-up frames by using a target tracking algorithm; extracting information of centers of palms and fingertips of the detected hands in real time and using the information as feature points of the hands; and according to the binocular-vision principle, calculating three-dimension coordinates of the feature points based on the extracted feature points so as to obtain real-time three-dimensional hand information. In addition, the system is composed of a binocular camera system and an information processing device; and the information processing device includes a hand detection module, a feature point extraction module, and a three-dimensional reconstruction module. According to the invention, real-time reliable detection and tracking of hands under a complex background can be realized; and thus reliable reconstruction of the three-dimensional hand information can be realized based on the detection and tracking.

Aiming at the problems of long manual labeling time and large expenses of a mass data hierarchy classifier, the invention provides a semi-supervised mass data hierarchy classification method comprising the following steps of: carrying out feature extraction on webpages in a webpage base; automatically generating the training set of a leaf node on the existing hierarchy classification body base by utilizing a rule set and an expansion rule; clustering the webpage of each existing leaf node; classifying unlabelled sets on the basis of clustering, and adding examples in the unlabelled sets, which are similar to the training set, to the training set of the corresponding leaf node to expand the scale of the training set; training the canonical linear classifier of each node by using a random gradient descent method; if the result of the classifier meets stopping conditions, stopping; otherwise, entering the step F; classifying the unlabelled sets by using the hierarchy classifier established by the steps C and D; adding the high-confidence classified webpages to the training set; and returning to the step C and repeating the steps form C to F.

The invention provides a face detection method based on the HOG characteristics. The face detection method comprises the steps of firstly, training a large number of partial SVM classifiers based on the HOG characteristics, then selecting some partial SVM classifiers which have the good classification effect from the partial SVM classifiers, combining classification results of the selected partial SVM classifiers to a training sample into new characteristic vectors, and finally using SVMs to train the vectors to obtain the final classifier. According to the face detection method based on the HOG characteristics, the HOG partial characteristics are used for being combined with the partial classifiers, and the good robustness is achieved on the conditions such as uneven facial expression illumination and uneven partial illumination.

The invention discloses a lightweight fine-grained image recognition method for cross-layer feature interaction in a weak supervision scene, and the method comprises the steps: constructing a novel residual module through employing multi-layer aggregation grouping convolution to replace conventional convolution, and enabling the novel residual module to be directly embedded into a deep residual network frame, thereby achieving the lightweight of a basic network; then, performing modeling on the interaction between the features by calculating efficient low-rank approximate polynomial kernel pooling, compressing the feature description vector dimension, reducing the storage occupation and calculation cost of a classification full-connection layer, meanwhile, the pooling scheme enables the linear classifier to have the discrimination capability equivalent to that of a high-order polynomial kernel classifier, and the recognition precision is remarkably improved; and finally, using a cross-layer feature interaction network framework to combine the feature diversity, the feature learning and expression ability is enhanced, and the overfitting risk is reduced. The comprehensive performance of the lightweight fine-grained image recognition method based on cross-layer feature interaction in the weak supervision scene in the three aspects of recognition accuracy, calculation complexity and technical feasibility is at the current leading level.