50 results about "Margin classifier" patented technology

The invention provides a classifier integration method for machine learning. Basic classifiers generated by a RandomForest algorithm are used as the basic classifiers of a new algorithm. An L1_Magic algorithm is used to optimize the weights of the original classifiers, differences of the classifiers are fully utilized to allow the possibility of right classification of each sample in a training set to be consistent. By the voting manner with weights of the new integrated classifier, the accuracy of training set classification is increased. Compared with the RandomForest algorithm, the method has the advantages that the differences of the classifiers are fully utilized to allow the sample data classification accuracy of the integrated classifier to be increased. In addition, the problem of how to use diversity / difference to increase the effect of the integrated classifier in the field of artificial intelligence is solved.

The invention discloses an image classification method and system based on an image salient region. The method includes offline training and online test. The offline training comprises: performing ultra-pixel segmentation on an image to obtain multidimensional segmentation blocks, and calculating the characteristic contrast of the segmentation blocks to obtain a target salient map; performing threshold segmentation on the target salient map to obtain a binary image, performing morphological processing on the binary image, and performing automatic segmentation extraction on the target salient map by employing a segmentation algorithm to obtain the salient region; and inputting the salient region to a convolutional neural network for training to obtain an image classifier based on the image salient region. The online test includes: performing automatic segmentation extraction of the salient region on a test image, inputting a salient region image of the test image to the trained image classifier, and performing image classification to obtain an image class mark. According to the method and system, the segmentation result is guaranteed, the workload of artificial interaction is reduced, and the accuracy of image classification is improved.

A novel statistical learning procedure that can be applied to many machine-learning applications is presented. Although this boosting learning procedure is described with respect to its applicability to face detection, it can be applied to speech recognition, text classification, image retrieval, document routing, online learning and medical diagnosis classification problems.

The invention provides an artificial limb control method which comprises the following steps: acquiring a bioelectric signal; extracting feature information of the bioelectric signal; identifying action types according to the feature information; analyzing the feature information by using a linear distinguish and analytical method and identifying the action types according to pre-stored train results, training an action classifier by a train program module; and controlling an artificial limb to complete corresponding actions according to the action types. By adopting the artificial limb control method, an artificial limb user can naturally and directly select and perform various different physical actions that the user intends to complete just like before amputation. The control technology based on action type identification can realize intuitive control of the myoelectric artificial limb so as to realize natural and intuitive control of the multifunctional artificial limb, thus enhancing use convenience of the artificial limb. In addition, the invention further provides an artificial limb control system.

In one embodiment, a device trains a machine learning-based malware classifier using a first randomly selected subset of samples from a training dataset. The classifier comprises a random decision forest. The device identifies, using at least a portion of the training dataset as input to the malware classifier, a set of misclassified samples from the training dataset that the malware classifier misclassifies. The device retrains the malware classifier using a second randomly selected subset of samples from the training dataset and the identified set of misclassified samples. The device adjusts prediction labels of individual leaves of the random decision forest of the retrained malware classifier based in part on decision changes in the forest that result from assessing the entire training dataset with the classifier. The device sends the malware classifier with the adjusted prediction labels for deployment into a network.

The invention belongs to the technical field of condition monitoring and fault diagnosis of a wind turbine generator system, in particular to a migration diagnosis method of the gearbox fault of a wind turbine generator system. The method comprises the following steps: establishing four neural network structures, namely, a source domain feature extractor, a target domain feature extractor, a domain classifier and a domain discriminator; obtaining predictive label values by forward propagation from annotated source domain data, network training loss functions are calculated according to predictive label and actual label, and source domain feature extractor and domain classifier are pre-trained by back propagation algorithm. The loss functions of the source domain feature, the target domainfeature and the domain discriminator are calculated by forward propagation from the source domain data and the target domain data, and the domain discriminator and the target domain feature extractorare trained by back propagation algorithm respectively. The newly acquired target domain data is input into the target domain feature extractor, the feature is calculated, and the predictive label ofthe new data is obtained by the domain classifier input from the feature.

The invention discloses a garbage type recognition system based on a deep learning method in the field of computer-aided garbage classification. The system comprises a garbage classification data set,a garbage classifier, a garbage classifier training module and a real-time garbage type recognition module. The garbage classifier training module utilizes a garbage classification data set to traina garbage classifier, and the real-time garbage type recognition module collects image information from a camera of the real-time garbage type recognition module and then calls the garbage classifierto recognize the garbage type in an image and outputs a recognition result and a control signal. The system has the advantages that garbage classification does not depend on human garbage classification knowledge any more, and therefore the process of implementing garbage classification in China can be greatly accelerated.

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.

A method for classifier generation includes a step of obtaining data for classification of a multitude of samples, the data for each of the samples consisting of a multitude of physical measurement feature values and a class label. Individual mini-classifiers are generated using sets of features from the samples. The performance of the mini-classifiers is tested, and those that meet a performance threshold are retained. A master classifier is generated by conducting a regularized ensemble training of the retained / filtered set of mini- classifiers to the classification labels for the samples, e.g., by randomly selecting a small fraction of the filtered mini-classifiers (drop out regularization) and conducting logistical training on such selected mini-classifiers. The set of samples are randomly separated into a test set and a training set. The steps of generating the mini-classifiers, filtering and generating a master classifier are repeated for different realizations of the separation of the set of samples into test and training sets, thereby generating a plurality of master classifiers. A final classifier is defined from one or a combination of more than one of the master classifiers.

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 an image classifier establishing method and an image classifier establishing device. The image classifier establishing method comprises steps that a sample image set is acquired, and the sample image set comprises a positive sample comprising a target image and a negative sample without the target image; the deformation processing of the sample images in the sample image set is carried out to acquire the expanded sample image set; according to the expanded sample image set and a deep convolutional neural network model, a classifier aiming at the target image is acquired; the normalization processing of the output of the convolutional layer of the deep convolutional neural network is carried out. By adopting the above mentioned method, the limited sample image sets are identified artificially, and then the limited sample image sets are expanded, and therefore a sample amount is expanded, and the accuracy of the classifier is improved. The identification accuracy of the classifier is further improved by adopting the deep convolutional neural network, and therefore the classifier aiming at the target image and having the higher identification accuracy is provided.

The invention discloses a mid-air gesture recognition method based on an inertial sensor. The method includes the steps that a mid-air gesture signal sequence is extracted aiming at sensing signals acquired by the inertial sensor, then data preprocessing is carried out, then a training sample set, a verification sample set and a test sample are acquired, an LSTM-RNN model is subjected to parameter initialization, the training sample set is used for training the LSTM-RNN model, in the training process, verification samples in the verification sample set are input in the LSTM-RNN model trained in the iteration process, the iteration frequency is controlled according to the recognition error rate of the verification sample set, and a final LSTM-RNN classifier is obtained; finally the test sample is input in the final LSTM-RNN classifier, and a gesture corresponding to the test sample is recognized through the final LSTM-RNN classifier. The method has the advantages of being higher in mid-air gesture recognition precision and accuracy.

Systems and methods for training and using an image classifier are provided. A plurality of points can be generated from respective patches of images annotated as either a first type or a second type. The points corresponding to the images of the second type can be clustered into two clusters. A first cluster of the two clusters can be identified as closer to the points corresponding to the images annotated as the first type. The points in the first cluster can be assigned to a class positive, the points in the second cluster can be assigned to a class negative, and the points corresponding to the images annotated as the first type can be assigned to an anchor class. A plurality of triplets can be generated based on the points in the various classes. Parameters of an image classifier can be adjusted based on a loss function of the triplets.

Image backup using a trained image classifier is disclosed. In various embodiments, an image classifier is applied to a plurality of images to identify one or more images to be included in a save set of images. The save set of images are selectively stored to a second location according to one or more policies.

The invention provides a method and a device for identifying the fatigue state of a driver based on feature selection and integration of face multi-region classifiers, relating to the technical field of the identification of the fatigue state of the driver. According to the method and the device, a relatively stable and reliable identification result is given out, so that a fatigue behavior can be relatively timely detected, and the early warning can be given to the driver earlier. The method comprises a training process and an identifying process. In the training process, features extracted in each region need to be evaluated and selected so as to generate a simplified feature set corresponding to each region, and a group of C4.5 classifiers are trained based on the simplified feature sets; and in the identifying process, the features need to be extracted for classification. A face region is divided in the two processes, and static features and sequence descriptive features relevant with fatigue behaviors in each face region are extracted. According to the method and the device, the identification performance of the fatigue state of the driver is effectively improved, and the relatively high average identification rate of the fatigue state of the driver is achieved by virtue of the simplified feature sets.

The invention relates to a rice detection and classification method based on deep multi-view feature. The method comprises the following steps: 1) division of a data set: dividing a rice drop image data set through a cross-validation method to obtain a training set and a test set; 2) construction of a network model: carrying out feature extraction on an input image through a deep learning method,and meanwhile, continually updating model parameters through stochastic gradient descent; and 3) model analysis and performance evaluation: inputting the test data into a trained rice classifier, andcomparing probability predicted value and classification threshold to obtain a rice detection and classification result. Compared with the prior art, the method is greatly improved in accuracy, and improves defects of rice feature extraction of a conventional method; and the method finishes calculation of husked rice yield of the rice according to the classification result.

The invention relates to the technical field of image information processing, in particular to a fine category classification method based on component polygons. The polygons based on component points are adopted for effectively correcting targets under different postures, fisher encoding features adopted in the classification process have robustness for changes in dimensions and directions of the targets, the adopted greedy algorithm can be used for finding out a set of component point combinations highest in discrimination power, and fine classifiers can be used for distinguishing categories very similar to one another. According to the fine category classification method, the component points of target images are connected, the multiple polygons based on the component points are constructed, and the number of errors caused by correction can be reduced effectively.

The invention discloses a classification method aiming at small sample and high dimensional images. The classification method aiming at the small sample and high dimensional images comprises the following steps: (1) gaining a first classification rule, (2) classifying images on a first level, (3) gaining a second classification rule, (4) classifying the images on a second level, (5) gaining a third classification rule, (6) classifying the images on a third level and gaining a classification result. The classification method aiming at the small sample and high dimensional images is combined with characteristics of industrial manufacture. The first-level image classification has strong manual controllability, and meanwhile combines a manifold dimensionality reduction method and superiorities of a support vector machine, thereby being suitable for the classification of the small sample and high dimensional images. Through combining a direct expression method of image type, the manifold dimensionality reduction method, and a support vector machine classification method with an arborescence topological structure classification method based on position features and barycenter features, a three-level image classification method is established. Due to the fact that the data transmission quantity between the image classifiers of the three levels is small, efficiency can not be affected. The classification method aiming at the small sample and high dimensional images is simple in operation, good in algorithm connection and few in input parameters.

The invention discloses a high-dimensional multimedia data classifying method based on maximum margin tensor study. The method includes the following steps that (1) a training data set of multimedia data is built; (2) the training data set is modeled and analyzed to obtain a classifying model; (3) according to a user inquiry data set and the classifying model, the inquiry data set is classified. According to high-dimensional performance and structure performance of the multimedia, the multimedia data is expressed through tensor, and high-dimensional multimedia data is classified through a maximum margin classifier method. Classifying is finished while the multimedia data is subjected to decomposition analysis, structural information in the multimedia data is reserved, dimensionality curse caused by high-dimensional data generated through a traditional splicing method is avoided, and the method is more accurate than a traditional multimedia data classifying method and facilitates calculation.

The invention provides a mobile application classifying method under imbalanced perception data. The mobile application classifying method comprises the steps that firstly, data subsets which are the same with positive-type samples in number are sampled from a large number of negative-type with-label data through secondary sampling; then, sampling based on similarity is conducted on non-label data by the utilization of the characteristic similarity between the non-label data and the with-label data so as to generate non-label data subsets; a sub-classifier is obtained on each with-label data subset and each non-label data subset through training by the utilization of the semi-supervised studying method; at last, a master classifier is formed by integrating the sub-classifiers. The mobile application classifying method under the imbalanced perception data has the advantages that the mobile application classifying method can be used for deduction of events, activities and backgrounds of current intelligent mobile phone applications so that the designed classifier can be adapted to the screen that the positive-type data in the actual perception data and the negative-type data in the actual perception data are imbalanced in number, and deduction service which is accurate in robustness is provided for mobile phone perception applications.

The invention relates to an image local style migration method based on a decomposition factor. The method comprises the following steps of 1) acquiring a migration task; 2) initializing that style migration network and training with tag data samples; 3) processing the migration task based on the trained style migration network to obtain a composite image, wherein the style migration network comprises two automatic encoders and four tag classifiers, each automatic encoder comprises one encoder and two decoders, the tag classifier is arranged between the encoder and the decoder, and the encoderrealizes the decomposition of a specified factor and a common factor. Compared with the prior art, the method of the invention can better decouple factors, and has the advantages of remarkable reconstruction effect and the like.

Non-negative matrix factorization, NMF, is combined with identification of a maximum margin classifier by minimizing a cost function that contains a generative component and the discriminative component. The relative weighting between the generative component and the discriminative component are adjusting during subsequent iterations such that initially, when confidence is low, the generative model is favored. But as the iterations proceed, confidence increases and the weight of the discriminative component is steadily increased until it is of equal weight as the generative model. Preferably, the cost function to be minimized is:minF,G≥0X-FG2+γ(w2+C∑i=1nL(yi,w·gi+b)).

The invention discloses a visual multi-target tracking method based on multiple single trackers. The visual multi-target tracking method comprises the steps that targets are detected based on a classifier and are tracked through the multiple visual single-target trackers, the targets and the trackers are corrected through different strategies during tracing, and thus the multiple targets are tracked continuously; specifically, firstly, the target classifier is utilized to detect images so as to obtain the targets, then one visual single-target tracker is assigned to each target, and the multiple visual single-target trackers are utilized to jointly complete the task of multi-target tracking; and in order to deal with the cumulative error of the trackers, targets escaping from the view during tracking and new targets, the update cycle for the multiple single-target trackers and the multi-tracker consistency judgment cycle are introduced, thus adaptive management such as update and consistency judgment is conducted cyclically on the multiple single-target trackers, and the multiple targets are effectively tracked. According to the visual multi-target tracking method, the tracking efficiency is greatly improved, and the requirement for multi-target real-time tracking is basically met.

A method (and system) of handling out-of-grammar utterances includes building a statistical language model for a dialog state using, generating sentences and semantic interpretations for the sentences using finite state grammar, building a statistical action classifier, receiving user input, carrying out recognition with the finite state grammar, carrying out recognition with the statistical language model, using the statistical action classifier to find semantic interpretations, comparing an output from the finite state grammar and an output from the statistical language model, deciding which output of the output from the finite state grammar and the output from the statistical language model to keep as a final recognition output, selecting the final recognition output, and outputting the final recognition result, wherein the statistical action classifier, the finite state grammar and the statistical language model are used in conjunction to carry out speech recognition and interpretation.

The invention discloses a video continuous sign language recognition method and system based on a grammar classifier. The method comprises the following steps: segmenting an acquired original sign language video into a plurality of video segments, performing time-space domain feature extraction on each video segment based on a residual connected three-dimensional convolutional neural network, andperforming context learning on the extracted time-space domain features by using a bidirectional long-short-term memory network to obtain features of the sign language video; performing global poolingon the features of the video by adopting a maximum pooling layer to obtain a feature vector of the original sign language video; based on the feature vector, giving a confidence coefficient score corresponding to each word in the sentence by adopting a word classifier module, and giving a confidence coefficient score of each multi-tuple in the sentence by adopting a tuple classifier module; and determining a sign language recognition result based on the confidence score corresponding to each word in the sentence given by the word classifier module and the confidence score of each multi-tuplein the sentence given by the tuple classifier module. The sign language recognition performance can be improved.

The invention discloses an ore mud pie target detection method based on a weak supervision YOLO model, and the method comprises the steps: collecting an ore mud pie image on a conveying belt in real time, and inputting the collected ore mud pie image into a trained WS-YOLO model, so as to obtain a mud pie target in the ore mud pie image. The WS-YOLO model comprises a DarkNet53 network, an FPN network, a first full connection layer and a second full connection layer which are connected in sequence. The target classifier and the target position regression model are connected with the second fullconnection layer, and the active learning module is connected with the target classifier and the target position regression model, the active learning module comprises a US strategy sub-module, an expert labeling sub-module and a sample pool which are connected in sequence, and the output of the sample pool is connected to the input of the DarkNet53 network. According to the method, the problemsof large workload, high cost, long period and the like caused by the fact that a large number of samples need to be accurately labeled in an existing mud pie target detection method can be solved, andthe transplantability of the model among different mines is improved.