62 results about "Ensembles of classifiers" patented technology

Techniques are disclosed for multi-modal identification that utilize a classifier combination framework. One embodiment of the present invention provides a multi-modal identification system that includes a collection of classifiers that classify feature streams derived from audio and / or video sources. A classifier combination scheme is used to combine the classifier outputs having varying degrees of confidence, but in a robust way by using a confidence-based weighting scheme that operates on a “per-class” basis, rather than (or in addition to) the traditional “per-classifier” basis. The system can be distributed across several machines running independent feature classifiers on the subscription basis.

An adaptation method includes using a first classifier trained on projected representations of labeled objects from a first domain to predict pseudo-labels for unlabeled objects in a second domain, based on their projected representations. A classifier ensemble is iteratively learned. The ensemble includes a weighted combination of the first classifier and a second classifier. This includes training the second classifier on the original representations of the unlabeled objects for which a confidence for respective pseudo-labels exceeds a threshold. A classifier ensemble is constructed as a weighted combination of the first classifier and the second classifier. Pseudo-labels are predicted for the remaining original representations of the unlabeled objects with the classifier ensemble and weights of the first and second classifiers in the classifier ensemble are adjusted. As the iterations proceed, the unlabeled objects progressively receive pseudo-labels which can be used for retraining the second classifier.

An image focus assessment method is provided that works reliably for images of a variety of relatively dissimilar workpieces or workpiece features. The focus assessment method is based on analysis of a single image (without the benefit of comparison to other images). The robustness of the focus assessment method is enhanced by the use of at least one classifier based on a plurality of focus classification features. In one application, a primary advantage of assessing focus from a single image is that an overall workpiece inspection time may be reduced by avoiding running an autofocus routine if an image is already in focus. In various embodiments, the focus assessment method may include an ensemble of classifiers. The ensemble of classifiers can be trained on different training data (sub)sets or different parameter (sub)sets, and their classification outcomes combined by a voting operation or the like, in order to enhance the overall accuracy and robustness of the focus assessment method.

Time lapse, characteristic of aging, is a complex process that affects the reliability and security of biometric face recognition systems. Systems and methods use deep learning, in general, and convolutional neural networks (CNN), in particular, for automatic rather than hand-crafted feature extraction for robust face recognition across time lapse. A CNN architecture using the VGG-Face deep (neural network) learning produces highly discriminative and interoperable features that are robust to aging variations even across a mix of biometric datasets. The features extracted show high inter-class and low intra-class variability leading to low generalization errors on aging datasets using ensembles of subspace discriminant classifiers.

A method and system for fusing a collection of classifiers used for an automated insurance underwriting system and / or its quality assurance is described. Specifically, the outputs of a collection of classifiers are fused. The fusion of the data will typically result in some amount of consensus and some amount of conflict among the classifiers. The consensus will be measured and used to estimate a degree of confidence in the fused decisions. Based on the decision and degree of confidence of the fusion and the decision and degree of confidence of the production decision engine, a comparison module may then be used to identify cases for audit, cases for augmenting the training / test sets for re-tuning production decision engine, cases for review, or may simply trigger a record of its occurrence for tracking purposes. The fusion can compensate for the potential correlation among the classifiers. The reliability of each classifier can be represented by a static or dynamic discounting factor, which will reflect the expected accuracy of the classifier. A static discounting factor is used to represent a prior expectation about the classifier's reliability, e.g., it might be based on the average past accuracy of the model, while a dynamic discounting is used to represent a conditional assessment of the classifier's reliability, e.g., whenever a classifier bases its output on an insufficient number of points it is not reliable.

Various embodiments relate to estimating the location of social network users. In one embodiment, a plurality of social media messages generated by a given user is received. A plurality of location features is extracted from the social media messages. Each of the location features is processed with at least one classifier from an ensemble of classifiers. A location classification is generated by each of the classifiers for each of the social media messages. Each classification comprises a location and a weight associated with that location. Once of the locations is selected from the location classifications as the location of the given user based on a combination of the weights of the location classifications.

The invention discloses a method of intelligently classifying texts, which comprises the following steps: a great number of training text sets are prepared; each text is segmented; feature items are extracted and stored; weight computation is carried out on each feature item; the text is converted into a text vector and is stored in a classifier and finally a feature item set and a classifier set are formed. The texts to be classified are segmented; features in the feature item set are defaulted to be matched with the text to be classified; and the weight computation of the feature items is carried out so that the features which are not matched with the text to be classified are filtered and the features which are matched with the text to be classified are left as the features of the text. The feature items are converted into the text vector. The text vector is compared with the vector in the classifier through a similarity algorithm and the classification of the text is determined according to the similar text vector. The texts can be classified and extracted more accurately through the effective method disclosed by the invention and the speed of the method is faster than the speed of the prior art.

A method and system for fusing a collection of classifiers used for an automated insurance underwriting system and / or its quality assurance is described. Specifically, the outputs of a collection of classifiers are fused. The fusion of the data will typically result in some amount of consensus and some amount of conflict among the classifiers. The consensus will be measured and used to estimate a degree of confidence in the fused decisions. Based on the decision and degree of confidence of the fusion and the decision and degree of confidence of the production decision engine, a comparison module may then be used to identify cases for audit, cases for augmenting the training / test sets for re-tuning production decision engine, cases for review, or may simply trigger a record of its occurrence for tracking purposes. The fusion can compensate for the potential correlation among the classifiers. The reliability of each classifier can be represented by a static or dynamic discounting factor, which will reflect the expected accuracy of the classifier. A static discounting factor is used to represent a prior expectation about the classifier's reliability, e.g., it might be based on the average past accuracy of the model, while a dynamic discounting is used to represent a conditional assessment of the classifier's reliability, e.g., whenever a classifier bases its output on an insufficient number of points it is not reliable.

The invention discloses a multilevel semantic feature-based face feature extraction method and recognition method. The method includes the following steps that: 1) organ areas of each image in a facial image set A are divided; 2) bottom-level features of each organ are extracted and clustered; two clusters are extracted from clustering results and are adopted as positive and negative samples, and the positive and negative samples are trained in a paired combination manner such that a classifier set can be obtained, and the results of discrimination which is performed on the bottom-level features by the classifier set are united so as to obtain the middle-level features of the organ; the images in the A are the classified according to tags; any two classifications are selected from classification results of the tags and are adopted as positive and negative samples, and the positive and negative samples are trained in a paired combination manner such that a classifier set can be obtained, and the results of classification and discrimination which are performed on all the middle-level features in the A by the classifier set are united so as to obtain high-level features of the tags; the bottom-level features, the middle-level features and the high-level features are adopted to construct face features of the images; face features Vq are generated for any image q to be searched; and the face features Vq are matched with the face features in the A, and query results are returned. With the multilevel semantic feature-based face feature recognition method and recognition method adopted, recognition accuracy and stability can be improved.

The invention discloses a local region matching-based face search method. The method includes the following steps that: 1) faces of each image in a face image set A are aligned with a face of a standard format, and areas of various organs are divided; 2) bottom-level feature vectors of each organ are extracted from and are clustered; 3) any two classifications are selected from clustering results of each organ and are adopted as positive and negative samples, and a support vector machine classifier is trained; training is performed in a paired combination manner, such that a classifier set of the organs can be obtained, and the results of discrimination of the bottom-level feature vectors which is performed by each classifier in the classifier set are united so as to form new feature vectors, namely, middle-level feature vectors of the organs; 4) the ratio of the distance of each key point on each face contour to left and right eyes to the distance between the two eyes is calculated and is adopted as the middle-level feature vector of the corresponding face contour; the above middle-level feature vectors are combined such that Vr can be obtained; and 5) a middle-level feature vector Vq is generated for a face image q to be searched; and the Vq is matched with the Vr in the A, and query results are returned. With the local region matching-based face search method of the invention adopted, a search effect of similar faces can be improved.

Techniques are disclosed for multi-modal identification that utilize a classifier combination framework. One embodiment of the present invention provides a multi-modal identification system that includes a collection of classifiers that classify feature streams derived from audio and / or video sources. A classifier combination scheme is used to combine the classifier outputs having varying degrees of confidence, but in a robust way by using a confidence-based weighting scheme that operates on a “per-class” basis, rather than (or in addition to) the traditional “per-classifier” basis. The system can be distributed across several machines running independent feature classifiers on the subscription basis.

The invention discloses an integrated learning anti-fraud test method, and aims at overcoming deficiencies of risk control and anti-fraud technologies in the industry of internet finance in the priorart, and providing an application integrated learning technology to carry out risk control test on users and transaction behaviors. The method comprises the following steps of: extracting a training sample set and extracting features of user information in the training sample set; training a base classifier, and training a feature view by adoption of a multi-term classification algorithm so as toobtain a base classifier set; processing the base classifier set by utilizing an integrated learning method so as to obtain an integrated classification model; and classifying test samples by using the integrated classification model so as to obtain a test result, and carrying out integration by using another machine learning algorithm so as to obtain a final result. The invention furthermore discloses an integrated learning anti-fraud test system on the basis of the method. According to the method and system, fraud users in internet transactions can be effectively tested, and relatively goodgeneralization and higher stability are provided at the same time.

The invention discloses a self-paced reinforcement image classification method and system. The method comprises the following steps: S10, inputting image data for classification and type labels, and performing feature extraction on the data; S20, based on reinforcement learning and self-paced learning frameworks, establishing a mathematic model; S30, updating parameters of the model and a weak classifier set of the model in an iteration mode until convergence is realized; and S40, predicting types of newly input test images. The method is characterized in that intrinsic consistency and complementarity of a reinforcement learning method and a elf-paced learning method are fully utilized, a distinguishing capability of a classification model and the reliability of image samples participating in learning are simultaneously highlighted in a learning process, and at the same time, effective learning and robust learning are realized. Compared to a conventional image classification method, the method and system have higher classification accuracy and higher robustness for label noise.

Online learning of an ensemble of classifiers or regressors is performed to predict a non-stationary time-varying parameter over a time series of episodes. For an episode, an ensemble action is selected from a set of ensemble actions based on ensemble action quality values (Q values) at an error state for the episode. The selected ensemble action is executed to update the ensemble. A cost of executing the selected ensemble action is computed or retrieved, and a reward is computed indicating how well the updated ensemble predicts the non-stationary time-varying parameter over the episode. The Q value for the selected ensemble action at the error state for the episode is updated based on both the reward and the cost of executing the selected ensemble action. The cost may be based on the ensemble action alone, or on both the ensemble action and the error state for the episode.

An autofocus method includes acquiring an image of a scene that includes one or more out of focus faces and / or partial faces. The method includes detecting one or more of the out of focus faces and / or partial faces within the digital image by applying one or more sets of classifiers trained on faces that are out of focus. One or more sizes of the one of more respective out of focus faces and / or partial faces is / are determined within the digital image. One or more respective depths is / are determined to the one or more out of focus faces and / or partial faces based on the one or more sizes of the one of more faces and / or partial faces within the digital image. One or more respective focus positions of the lens is / are adjusted to focus approximately at the determined one or more respective depths.

The invention discloses a shielded human face detection method. The method comprises the following steps of establishing an image pyramid of each frame of picture; based on a human face detection dictionary obtained by training by adopting a Boosting method previously, judging whether a human face exists or not and the human face is shielded or not; performing training by using multiple human faceand non human face pictures; according to a human face local position adopted during feature extraction by each weak classifier, classifying weak classifiers; based on a weak classifier set {hi(j)} obtained after category classification, performing human face detection and abnormal human face judgment; and combining human face detection sets to obtain a human face or abnormal human face detectionresult. The method has the characteristics that whether the human face is shielded or not can be correctly judged for the situations of wearing mouth masks with different colors, wearing eye patcheswith different colors, wearing sunglasses with dark color, and the like.

The present invention includes 1, determining two kinds of sample, according to inputting sample data, extracting high dimensional selected features; 2, determining one weak sorter assembly relative to single features training, each weak sorter in said assembly corresponding one feature; 3, respectively setting initial weight(ed) value for said two kinds training sample configuring; 4, according to current weight(ed) value inputting training sample to proceed training; 5, according to training result, selecting minimal error ratio preset number weak sorter corresponded feature as current turn feature selection result; 6, updating all training sample weight(ed) value, re-executing 4-6 steps until ending preset turn. The present invention solves high store content and calculated amount problem in current selected method.

A target tracking method with self-restoration capacity based on a multi-stage detector includes the steps of selecting a plurality of detectors of different types to be connected in series in combination with the concept of cascading Adaboost multi-stage weak classifiers into a strong classifier, selecting a significance detector for first-stage detection, selecting a classifier assembly detecting module for second-stage detection, substituting the classifier assembly detecting module into a random tree to calculate the posterior probability of a positive sample in a probable area detected byfirst-stage detection, selecting a related filtering detector for third-stage detection, calculating the relevance between a sample with the posterior probability larger than a certain threshold andthe positive sample initialized or obtained in the last frame so as to reduce accumulated errors caused by long-term tracking, determining the position with the maximum relevance as the target area inthe current frame through the multi-stage detector, sampling the determined position, supplementing the positive and negative sample number of concentrated removed samples to ensure the reliability and number consistency of the samples, activating a redetection mechanism if the maximum value of the relevance is larger than a certain threshold, and detecting the area near the position again to search for a target.

Various embodiments of the invention relate to estimating the location of social network users. In one embodiment, a plurality of social media messages generated by a given user is received. A plurality of location features is extracted from the social media messages. Each of the location features is processed with at least one classifier from an ensemble of classifiers. A location classification is generated by each of the classifiers for each of the social media messages. Each classification comprises a location and a weight associated with that location. Once of the locations is selected from the location classifications as the location of the given user based on a combination of the weights of the location classifications.

The embodiment of the invention provides an image recognition method and device, computer equipment and a storage medium, and belongs to the field of artificial intelligence. The method comprises thesteps that: an object recognition result set associated with an image according to a classifier set is obtained, wherein the classifier set comprises at least two classifiers, wherein the at least twoclassifiers are models in different parameter training stages, the at least two classifiers are used for identifying target objects contained in the image, and each classifier corresponds to one identification result; recognition confidence coefficients corresponding to the at least two classifiers respectively are obtained; and a target identification result corresponding to the image is determined according to the identification confidence coefficients and the identification results contained in the object identification result set. With the method provided by the embodiments of the invention adopted, the accuracy of image recognition can be improved.

The invention discloses a power communication network fault positioning method, and the method comprises the steps: firstly carrying out the preprocessing of historical warning data, obtaining a plurality of important warning attributes, and distributing influence factors; secondly building base classifiers serving as sub-prediction models, respectively predicting the historical warning data, and enabling prediction accuracies to serve as base classifier weights; thirdly enabling the plurality of base classifiers to divided into base classifier sets, obtaining the mean weight and mean influence factor of each group through estimation, obtaining the comprehensive weight of each group, selecting the maximum comprehensive weight, wherein a fault type corresponding to the maximum comprehensive weight is a final prediction result; building a combined prediction model at this moment; predicting existing warning data through employing the combined prediction model, and obtaining a final fault positioning prediction result. The method solves problems of low accuracy and speed in the technology of fault positioning, employs the built combined prediction model for fault positioning, remarkably improves the accuracy of fault positioning, and greatly shortens the fault positioning time.

The invention provides a method for identifying the authenticity of a banknote based on banknote anti-counterfeiting point multiple-feature fusion. The method comprises the following steps: the same weak classifier is trained by an AdaBoost algorithm for different training sample sets; the weights of the training samples and the weak classifier are adjusted according to wrongly-classified training samples in the training sample sets and the situation of training sample wrong classification in the weak classifier; the weak classifier is retrained after the training sample sets are redistributed; the above steps are repeated, and T weak classifiers and the weight corresponding to each weak classifier are obtained after the weak classifier is trained T times; the T weak classifiers are integrated into a final strong classifier; and the trained strong classifier tests a test sample set, statistical analysis is made on the test error rate, and the steps are repeated until the steps are repeated a set maximum number of times T and a classification model based on fusion of anti-counterfeiting points of the sample sets is obtained.

The present invention relates to the field of data processing. A defect prediction method and device, the method comprising: from a pre-stored product failure record selecting a training attribute set based on a target attribute, and combining the target attribute with the training attribute set to form a training set (101); the target attribute is the defect attribute of a historical faulty product; generating a classifier collection according to the training set, the classifier collection comprising at least two tree classifiers (102); and utilizing the classifier collection as a prediction model to predict the defects of a faulty product (103). The method is used in the defect prediction process of a faulty product to realize accurate and quick locating of a faulty product.

The invention discloses a classification model training and using method and device, equipment and a medium, and relates to the technical field of artificial intelligence. The training method comprises the steps of utilizing a training sample set, adopting an ensemble learning algorithm to train a classification model to obtain at least one classifier set, wherein the classifier set comprises at least two base classifiers; testing the classifier set by adopting a test sample set to determine the classification performance of each base classifier, and determining the classification weight of each base classifier according to the classification performance, wherein the classification weight is used for determining the weight of the classification result of each base classifier when the classifier set is used for sample classification. According to the embodiment of the invention, the diversity of the base classifiers in ensemble learning can be ensured, and the classifiers have the common classification performance for minority classes and majority classes, so that the classification model is suitable for the condition of unbalanced samples.

The invention provides a target classifying method based on a local and depth feature assembling. The method mainly includes the steps of depth convolution feature extracting, local feature extracting, encoding and assembling through classifiers, wherein in the process, Fisher vectors of descriptors and scale invariant feature transformation (SIFT) descriptors are extracted in a final complete connection layer of a depth network, a support vector machine (SVM) is trained for each feature with the Fisher vectors as the encoding strategies, the assembling through the classifiers is trained and tested, an input data set is optimized and classified, and finally voting is conducted and a final decision is made. By means of an intermediate layer of the depth network, the classification capacityof the features obtained from the complete connection layer can be enhanced, one independent classifier is trained for each feature, and therefore good classification performance is achieved; meanwhile, computation cost is low, and various applications of the target classification technology can be easily obtained.

The invention discloses a method and a system for identifying a set-top box type. The method for identifying the set-top box type comprises the following steps: identifying an EPG interface in at least one frame of channel change image output by a set-top box to be detected in a channel change state according to a pre-generated first classifier set to obtain a first identification result set; and determining the type of the set-top box to be detected based on the first identification result set. The method and the system provided by the technical scheme of the invention can be used for automatically identifying the type of the set-top box.