78 results about "Multiclass classification" patented technology

a method for improving stacking schema for classification tasks, according to which predictive models are built, based on stacked-generalization meta-classifiers. Classifications are combined to build a new scheme from at least two layers and multiclass classification problems are converted into binary classification problems. One-against-all class binarization and regression learners are used for each class model and ensemble classifiers are improved using stacking. Accuracy differences, accuracy ratio, and runtime classification in multiclass datasets are also improved and the class of a value is then predicted.

The invention discloses a brand new inquiry-related multi-ranking-model integration method. The brand new inquiry-related multi-ranking-model integration method comprises the following steps of: establishing sub-ranking models for each inquiry and inquiry-related document; performing vectoring representation on the sub-ranking models so as to convert a plurality of inquiry-related ranking models into characteristic data and integrate the plurality of ranking models; establishing a new loss function as an optimization target at an inquiry level and a sample level by using a ranking support vector machine as the sub-ranking model; adjusting weight among losses generated by different inquiries by using the loss function; and providing a multiple inquiry-related ranking support vector machine fusion algorithm. Compared with the traditional model, the inquiry-related ranking models can achieve better properties when the inquiry-related multi-ranking-model integration algorithm provided by the invention is applied to actual tasks. The multi-model fusion algorithm provided by the invention can be applied to ranking learning, can also be applied to multi-element classification, sequence labeling and the like, and has a wide application prospect in the fields of information retrieval, network search and the like.

The method includes steps: (1) building ASM model, and initializing ASM model so as to obtain initial position of model; (2) generating multiple classes of training samples corresponding to each feature point on human face in order to train classifier in multiple classes of vector machine; (3) using training sample corresponding to each feature point trains the train classifier in multiple classes of vector machine; (4) using initial position of model as initial position of ASM searching, and using the said classifier to locate each feature point. The method is related to detecting eye, training classifier, converting classifier in two classes to classifier in multiple classes, and ASM location for feature point. Further, the invention is applicable to recognition of human face, recognition of sexuality etc. possessing high precision.

The present invention provides a machine learning based vehicle logo identification method. The method comprises the following steps: 1) performing vehicle logo location: searching out a vehicle license plate according to a vehicle license plate location method, and on this basis, determining a region where a vehicle logo is possibly located, i.e. an ROI region as a vehicle logo location image; 2) extracting an SIFT feature set of the vehicle logo location image; 3) by using a clustering method, generating K keywords and a weight of each keyword from the SIFT feature set; 4) obtaining the K keywords and the weight of each keyword from the SIFT feature set as an input of a classifier; and 50 performing classification by using a well-trained multi-class SVM classifier. The present invention provides a machine learning based vehicle logo identification method that reduces time costs, is convenient for use and is high in accuracy.

The invention discloses a method for automatically classifying three-dimensional models based on a support vector machine. The method comprises the steps as follows: carrying out dimension reduction and characteristic decomposition on a geodesic distance matrix; creating binary classifiers for the characteristics of the three-dimensional models by using the support vector machine; and combining every two of the binary classifiers in a 'one-to-one' way by using the support vector machine to form a polynary classifier. According to the method, the three-dimensional models can be automatically classified, so that the robustness of the characteristic extraction process of the models is higher, the computing speed is higher, the characteristic extraction time is greatly shortened, and fewer three-dimension model training sample conditions can be well corresponded. The method has higher generalization performance and good expansion capability and non-linearity performance.

The invention provides a breast ultrasonic image multi-classification system and method based on cross correlation characteristics. The system comprises an ultrasonic image preprocessing unit; an interest area extraction unit used for extracting an image of an interest area; an internal cross correlation density feature extraction unit used for extracting an internal cross correlation density characteristic value of an interest image; a conventional feature extraction unit used for extracting a plurality of conventional characteristic values of the interest area image; and a multi-classification unit used for training classifiers, inputting an internal cross correlation density characteristic value vector and conventional characteristic vectors to three trained classifiers for classification, and regarding the category with the most prediction as a final classification result. According to the classification method, the internal cross correlation density characteristic based on the interest area is additionally provided, the auxiliary diagnosis effect of a breast ultrasonic computer can be effectively improved, the classification category of galactocele which is the benign lesion is added, and the requirement of a breast ultrasonic computer auxiliary system by doctors is further met.

The invention provides a turbulence target detection method based on BP neural network multi-class classification. The method comprises the step of according to a turbulence echo mode, under differentsignal to noise ratio conditions, generating echo amplitude sequences, and training sets and test sets of a BP neural network formed by corresponding strength grades so as to allow the trained BP neural network to divide the turbulence intensity into multiple grades, thereby achieving turbulence detection. According to the invention, an experience formulas and parameterization models are not required, by use of the multi-class classification function of the neural network, the echo amplitude sequences of a weather target radar are used as input data of the training set of the neural network;the turbulence intensity grades are used as output data of the training set; the relation between the radar echoes and turbulence intensity can be effectively determined only through learning of massecho data; and an objective that turbulence detection is achieved by use of the neural network to classify the turbulence intensities can be achieved.

The invention discloses a method and a system for detecting local impedance change of objects based on controlled knocking and a classifier. Classification is carried out by that an object has different local impedances and different response signals after being knocked and by the combination with a pattern recognition method. The method and the system not only can be used for detecting bolt screwing moment, but also can be used for detecting the damage status of composite material. The device adopted by the method is simpler, the method is simple and practical, the detection result is accurate, and multiclass classification can be realized.

The invention relates to a cyberspace threat knowledge extraction method and device. The cyberspace threat knowledge extraction method comprises the steps of extracting ontology concepts of a to-be-extracted text and relationships among the ontology concepts according to a pre-trained multi-label classifier; extracting examples corresponding to the ontology concepts in the text to be extracted according to a pre-trained multi-class classifier of each ontology concept; and according to the entities corresponding to the ontology concepts and the relationship between the ontology concepts, determining the relationship between the instances. By adopting the cyberspace threat knowledge extraction method, the extraction problem of cyberspace threat knowledge in the unstructured text can be wellsolved by combining the hidden entity and the ontology relationship.

A computer device may include a memory storing instructions and processor configured to execute the instructions to obtain a text file generated based on a communication session; select topics for classifying communication sessions; and apply one or more trained machine learning models to the obtained text file to determine one or more most likely topics for the communication session. The processor may be further configured to apply regular expression binary classifiers, each associated with a different topic, to the obtained text file to determine a likelihood the communication session is associated with a particular topic; select a topic for the communication session based on the determined one or more most likely topics and the determined likelihood the communication session is associated with the particular topic; and classify the communication session based on the selected topic.

The invention relates to a motion analysis method, which comprises the steps of obtaining a first motion set sequence of a trainer according to a human body posture estimation algorithm OpenPose and amotion video of the trainer; obtaining a one-dimensional action feature vector and a two-dimensional action feature point set of the action video according to the first action set sequence; obtainingaction classification of the action video according to the one-dimensional action feature vector and a multi-class classifier; According to the standard two-dimensional action feature point set corresponding to the action classification and the two-dimensional action feature point set, obtaining an action score of the trainee. According to the method, large-scale computing resources are not needed, and the accuracy of fitness action recognition and evaluation is improved. The invention further relates to a motion analysis device, a storage medium and equipment.

The invention relates to a text classification method based on representation enhancement and fusion. The method comprises the steps that a text classification model based on representation enhancement and fusion is constructed, and the processing steps of an input text in the text classification model based on representation enhancement and fusion are as follows: discrete characters of the inputtext are converted into continuous feature vectors in a data representation layer to obtain multiple representation vectors; adding disturbance into the representation vector in the representation enhancement layer to obtain a representation enhancement vector; further extracting and abstracting the characterization enhancement vector in the characterization abstraction layer to obtain an abstractcharacterization vector; classifying the abstract representation vectors in a classification layer to obtain output text tags; and synthesizing each output text label in the fusion layer to obtain afinal text label. The method can effectively solve the problems that in existing text multi-class classification, distribution of sample data among classes is unbalanced, and correct classification isdifficult when the number of samples in a small number of classes is insufficient.

The invention discloses a protein structure class prediction method based on weighted composition position vectors and a support vector machine. Firstly, protein structure class normal data sets are selected or established, weighting factors are introduced based on composition position vectors, and to-be-predicted protein sequences are represented through the weighted composition position vectors; the weighted composition position vectors are combined with the support vector machine, and the protein structure class prediction method is established through a direct multiclass classification strategy. According to the method, information of amino acid compositions is contained, position information of each amino acid residue in the protein sequences is also contained, the one-to-one correspondence function relation is formed between the position information and the protein sequences, the weighting factors are introduced to the composition position vector representation method, and prediction accuracy can be obviously improved by means of adjusting the weighting factors. The method is simple, rapid and sensitive, and is expected to be applied to other protein prediction fields.

The invention discloses a data classification method, which comprises the following steps: determining the probability that a prediction data set respectively belongs to each category; Constructing probability intervals belonging to each category of the prediction data set according to the probability; randomly generating a random number between [0, 1] is randomly generated, and determining a category of the prediction data set according to a probability interval in which the random number is located. A data classification apparatus is also disclosed. This scheme is mainly applied to the multi-class classification problem in machine learning field. This method not only enriches the internal logical structure of multi-class logistic regression algorithm, but also enhances the performance of multi-class logistic regression algorithm by judging the class and fully considering the influence of training sample error.

A system and method for semantic segmentation with a soft cross-entropy loss is provided. The system inputs a first color image to an input layer of a semantic segmentation network for a multi-class classification task. The semantic segmentation network generates, at an auxiliary stride, a first feature map as an output of an auxiliary layer of the semantic segmentation network based on the input first color image. The system extracts the generated first feature map from the auxiliary layer and computes a probability map as a set of soft labels over a set of classes of the multi-class classification task, based on the extracted first feature map. The system further computes an auxiliary cross-entropy loss between the computed probability map and a ground truth probability map for the auxiliary stride and trains the semantic segmentation network for the multi-class classification task based on the computed auxiliary cross-entropy loss.

The invention relates to the field of industrial control system intrusion detection, in particular to an intrusion detection method and system for an industrial control system. The method comprises the following steps: performing frame-by-frame acquisition processing on acquired original Modbus TCP data, and converting values of different functional characteristics and attack category labels into a form capable of being programmed and identified; performing principal component analysis dimension reduction processing on the attack classification feature set data, removing redundant data, and generating a multi-class classification data set; combining the attack data of different categories with normal data to generate a second-category classification data set; training to generate a second-class classification detector and a multi-class classification detector; connecting the plurality of second-class classifiers through an OR gate to obtain a distributed second-class classification detector intrusion detection system, and deploying the multi-class classifiers to form a multi-class classification detector intrusion detection system; and optimizing the algorithm model, inputting Modbus TCP data into an optimized intrusion detection system for classification processing, and obtaining an intrusion detection result. The accuracy and efficiency of intrusion detection in the industrial control system are improved.