555 results about "Incremental learning" patented technology

Described is an active learning system for fingerprinting an object identified in an image frame. The active learning system comprises a flow-based object segmentation module for segmenting a potential object candidate from a video sequence, a fixed-basis function decomposition module using Haar wavelets to extract a relevant feature set from the potential object candidate, a static classifier for initial classification of the potential object candidate, an incremental learning module for predicting a general class of the potential object candidate, an oriented localized filter module to extract features from the potential object candidate, and a learning-feature graph-fingerprinting module configured to receive the features and build a fingerprint of the object for tracking the object.

The invention discloses a voiceprint identification method. The voiceprint identification method comprises the following steps of: 1, preprocessing segmented speech data of each speaker in a training speech set to form a group of sample sets corresponding to each speaker; 2, extracting Mel-frequency cepstrum coefficients from each sample in all sample sets; 3, selecting a sample set one by one and randomly selecting the Mel-frequency cepstrum coefficients of part samples of the sample set, and training a Gaussian mixture model for the sample set; 4, performing incremental learning on the samples which are not selected and trained in the step 3 and the Gaussian mixture model of the sample set corresponding to the sample set one by one to obtain all optimized Gaussian mixture models, and optimizing a model library by utilizing all optimized Gaussian mixture models; and 5, inputting and identifying test voice data, identifying the Gaussian mixture model of the sample set corresponding to the test voice data by utilizing the optimized model library in the step 4, and adding the test voice data to the sample set corresponding to the speaker.

The invention discloses a method for identifying a disturbance event in a distributed type optical fiber pipeline security early-warning system. When the disturbance event exists, wavelet de-noising is conducted on two routes of sampling signals. The characteristic values of one sampling signal where wavelet de-noising is conducted are extracted, wherein the characteristic values include the vibration fragment length, the time domain energy, the k-order original point distance, the k-order center distance, the skewness, the kurtosis and low frequency wavelet coefficient energy Ej, obtained through wavelet decomposition, of all layers, and j ranges from 1 to 7. The thirteen extracted characteristic values are sent to a decision tree classification device, and the type of the disturbance event is obtained through the decision tree classification device. Man-machine interaction incremental learning is achieved by changing the type of the disturbance event stored in a database under the condition that a new type of the disturbance event appears or the type, obtained through the decision tree classification device, of the disturbance event is wrong, and online training is conducted on the decision tree classification device according to the modified type of the disturbance event. By means of the method, the type of the disturbance event can be accurately obtained.

The invention, which belongs to the technical field of network intrusion detection, discloses an intrusion detection method and intrusion detection system based on sustainable ensemble learning. A multi-class regression model is constructed by using a class probability output and a classification confidence product of an individual learner as training data, so that the decision-making process of the ensemble learning has high adaptability to the attack type to improve the detection accuracy. At the model updating stage, parameters and decision results of historical models are added into the training process of a new model, thereby completing incremental learning of the model. According to the invention, on the basis of the ensemble learning fusion plan of the multi-regression model, the decision-making weights of the individual learner during the detection processes for different attack types are allocated in a fine granularity manner; and the parameters and results of the historical models are used for training the new model, so that the stability of the model is improved and the sustainability of the learning process is ensured. Besides, the experiment result is compared with theexisting MV and WMV plans, the accuracy, stability and sustainability of the intrusion detection method and intrusion detection system are verified.

A sensor enhanced communication device (200) is provided with a wake mode, a standby mode and sleep mode. The sleep mode is a periodic occurrence within the standby mode which places a cluster of sensors and transducers (202) into a state of arousal in which the sensitivity of the transducers and sensors is increased while the sampling rate is decreased. Incremental learning can occur during the sleep mode as well as basic memory transfers. Since the cluster does not have to re-acquire information upon entering wake mode, the overall power efficiency is improved.

The invention discloses a phishing website detection method based on uniform resource locator (URL) classification. Firstly, a modeling is performed for URL characteristics, a method specific to a domain name imitation phenomenon of the characteristics is provided for calculating the similarity between a suspicious domain name and a protected domain name by a dynamic programming thought, in order to collect phishing URL high frequency suspicious character features, a suspicious character extraction algorithm based on a generalized suffix tree is provided, then on the basis of characteristic modeling, a support vector machine (SVM) algorithm is utilized to perform classified training for experimental training set, a SVM classification model is obtained after the training, the SVM classification model is used for classifying the URL to be detected, and a server for detecting a phishing website updates the current SVM classification model according to a specific online incremental learning strategy.

The invention discloses a dynamic hand gesture recognition method based on self incremental learning of the hidden Markov model. The method includes the following steps that firstly, a hand gesture is detected and tracked; secondly, feature extraction and vector quantization are carried out; thirdly, model training and hand gesture recognition are performed; fourthly, incremental learning is performed. According to the dynamic hand gesture recognition method based on self incremental learning of the hidden Markov model, dynamic hand gesture operation by a hand gesture operator in front of a camera can be accurately recognized, the recognized hand gesture data can be applied to incremental learning of an original model to adjust model parameters. Thus, the original model can dynamically adapt to novel variation generated in future hand gesture data and high adaptability to adjustment and alternation of the hand gesture data can be achieved. Thus, the model can be adjusted continuously along with the hand gesture data and better robustness on the unknown hand gesture recognition in the future is achieved.

The invention discloses a method for recognizing and retrieving video faces based on on-line study and Bayesian inference. The method comprises the following steps: step one: establishing an initialization model of a face recognition model, (i.e. the face recognition model adopts a GMM face recognition model); step two: establishing a face category model, (i.e. the model renewal of the initialization face model is performed by adopting an incremental learning manner); step three: recognizing and retrieving video faces. The test sequence and the category model are assigned, the sequence recognition information of the accumulation video in the Bayesian inference process is utilized, the probability density function of the identity is propagated according to information of a time axis, and the method provides recognition results of the video faces for users based on the MAP rules to obtain recognition scores. The invention establishes a model training frame based on non-supervised learning completely, according to spatial distribution of the training sequence, the initialization model is evolved for the category model in different modes, and the distribution of spatial data is better fitted through adjusting Gaussian mixture number of the face category model.

The invention relates to an intelligent lung cancer early cell pathological image recognition processing method, which comprises image pretreatment, image segmentation, laminate cell separation and reconstruction, cell character extraction and cell classification. The invention has the advantages that the image segmentation based on reinforcement learning uses incremental learning and continuous interaction with environment to search for optimized segmentation threshold value to obtain the segmentation effect which average value is 91%, the laminate cell separation and reconstruction uses B spline and modified deBoor-Cox method to simulate true cell edge better, the classifier uses general vote method to avoid low classifying accuracy of single classifier and improve total classifying accuracy, the application of two-stage classifier can reduce the possibility of false positive and false negative. Tests prove that the classifying accuracy of cancer or no cancer can average reach 93.8%, the classifying accuracy of squamous cell carcinoma, adenocarcinoma and small cell carcinoma average reaches 75%, and the false positive and negative average reach 4-6%.

Within the frameworks of hierarchical neural feed-forward architectures for performing real-world 3D invariant object recognition a technique is proposed that shares components like weight-sharing (2), and pooling stages (3, 5) with earlier approaches, but focuses on new methods for determining optimal feature-detecting units in intermediate stages (4) of the hierarchical network. A new approach for training the hierarchical network is proposed which uses statistical means for (incrementally) learning new feature detection stages and significantly reduces the training effort for complex pattern recognition tasks, compared to the prior art. The incremental learning is based on detecting increasingly statistically independent features in higher stages of the processing hierarchy. Since this learning is unsupervised, no teacher signal is necessary and the recognition architecture can be pre-structured for a certain recognition scenario. Only a final classification step must be trained with supervised learning, which reduces significantly the effort for adaptation to a recognition task.

The invention discloses an incremental equipment fault diagnosis method based on knowledge distillation and hidden layer sharing, and relates to the field of bearing equipment fault diagnosis. According to the method, a knowledge distillation and hidden layer sharing technology is utilized, so that a shallow layer equipment fault diagnosis model is guaranteed to have relatively good data characteristic extraction capability, and the fault classification performance of the shallow layer equipment fault diagnosis model is improved. For continuous increase of industrial data and update of an edgeequipment fault diagnosis model, methods of effective sample identification, data set reconstruction, pre-training model fine adjustment and the like are used for realizing incremental learning of the model. The requirements on the network bandwidth and the network delay in a massive real-time industrial equipment data transmission process are met; the accuracy of a shallow layer equipment faultdiagnosis method is improved; and the incremental learning is supported. Through a simulation experiment of bearing operation state data, under the condition that calculation resources are limited, the edge cloud collaborative data transmission efficiency is improved and the fault prediction classification accuracy is realized; and the incremental data learning and processing are supported.

A bio-inspired actionable intelligence method and system is disclosed. The actionable intelligence method comprises recognizing entities in an imagery signal, detecting and classifying anomalous entities, and learning new hierarchal relationships between different classes of entities. A knowledge database is updated after each new learning experience to aid in future searches and classification. The method can accommodate incremental learning via Adaptive Resonance Theory (ART).

The invention discloses a DoS / DDoS attack detecting and filtering method based on light-weight intrusion detection. The problem that intrusion detection accuracy still needs to be improved in the prior art is solved through the method. The method includes the steps of 1, flow early warning based on a time window; 2, abnormal flow characteristic processing; 3, rapid attack detection based on rule matching; 4, excavation and detection of attacks of unknown types; 5, attack filtering based on an IP list. Compared with the prior art, through the combination of the light-weight intrusion detection technology with characteristic selection, the problem that an original DoS / DDoS attack detection method based on classification detection is poor in instantaneity is basically solved; through the combination of online incremental learning and characteristic selection, the problems existing in detection of attacks of unknown types are solved, wherein pre-built attack modes of the attacks are not matched.

The invention discloses a combustion optimization control method for a boiler. The combustion optimization control method is characterized by comprising the following steps: sampling a combustion nonlinear system of the boiler to obtain input / output data at the current moment; training the real-time sampled input / output data by an online incremental learning fuzzy neural network, building an online incremental learning predicting model of the combustion nonlinear system of the boiler; performing a nonlinear prediction control algorithm on the online incremental learning predicting model for realizing the optimization and the control of the combustion process of the boiler. According to the combustion optimization control method for the power station boiler of the online incremental learning fuzzy neural network, the nonlinear optimization problem in the predication control algorithm is solved by utilizing a particle swarm optimization algorithm through the online identification of the boiler combustion optimization model; the real-time optimization and control of the boiler combustion process are realized.

The invention provides an audio content classification system, which comprises a training end and a test end, wherein the training end extracts characteristics of audio test samples through an audio characteristics extracting module, and trains classifier parameters through a classifier training module; and the test end comprises the audio characteristics extracting module shared by the training end, a classifier decision module, a transient characteristics extracting module, a transient characteristics smoothing module and an incremental learning module, wherein the audio characteristics extracting module is used for extracting audio characteristics of input signals; the classifier decision module takes output audio characteristics of the audio characteristics extracting module as input to classify the classifier parameters obtained by training a first frame through a training part; simultaneously, the transient characteristics extracting module extracts transient characteristics of the input signals, and outputs the transient characteristics of the input signals to the transient characteristics smoothing module; the transient characteristics smoothing module corrects and outputs an output result of the classifier decision module; and simultaneously, an incremental learning module utilizes classified class information and characteristic information of audio frames as a group of incremental learning samples to update the classifier parameters.

The invention discloses an incremental learning-based traffic exception detection method and device, and a storage medium. The method comprises the steps of acquiring traffic data of a user terminal;detecting exception of the traffic data by using a pre-created traffic detection classifier in an exception detection device; and when abnormal data is detected, obtaining training sample data based on the abnormal data, and performing online training on the traffic detection classifier by using the training sample data. The generalization capability of the classifier is improved by diversifying the training samples.

The invention provides a crop disease identification method based on incremental learning. When new data arrive, continuous learning is carried out based on an original learning result, and the capability of progressive learning is achieved, which means that new knowledge can be obtained from new samples obtained by batch and the performance is gradually improved under a condition that original knowledge is effectively kept. Firstly, a crop disease sample database is collected, and simulation incremental learning of disease images in the sample database is carried out using a negative correlation integrated neural network as main technical means, so that an initial parameter of a negative correlation learning system is determined, an integrated neural network classifier based on negative correlation learning is initialized based on the initial parameter, and the classifier is trained using a sample in an initial stage; in an incremental learning stage, when an expert adds a new sample in the sample database, the integrated neural network classifier based on negative correlation learning only is updated by only training the newly-added sample data, so that the object of incremental learning is achieved; and finally, a diagnosis result of a disease picture and control measures are fed back to a user, so that the pest and disease can be accurately identified and diagnosed, and the object of comprehensive crop control is achieved.

The invention relates to an incremental learning-fused support vector machine multi-class classification method, and aims to reduce sample training time and improve classification precision and anti-interference performance of a classifier. The technical scheme comprises the following steps of: 1, extracting partial samples from total samples at random to serve as a training sample set D, and using the other part of samples as a testing sample set T; 2, pre-extracting support vectors from the training sample set D; 3, performing support vector machine training on a pre-extracted training sample set PTS by using a cyclic iterative method so as to obtain a multi-class classification model M-SVM; 4, performing binary tree processing on the multi-class classification model M-SVM to obtain a support vector machine multi-class classification model BTMSVM0; 5, performing incremental learning training on the multi-class classification model BTMSVM0 to obtain a model BTMSVM1; and 6, inputting the testing sample set T in the step 1 into the multi-class classification model BTMSVM1 for classification. The incremental learning-fused support vector machine multi-class classification method is used for performing high-efficiency multi-class classification on massive information through incremental learning.

The invention discloses a method, a device and a system for automatic recommendation of network content. The method for automatic recommendation of network content includes receiving network text data and extracting keywords from the network text data, training a keyword classifier by an incremental learning process according to the keywords, receiving the network text data of specific users, classifying the specific users according to classification results of the keywords extracted from the network text data of the specific users by the aid of the keyword classifier, and recommending network content, interested by the specific users, to the specific users according to the classification of the network content predefined on the basis of the classification of the specific users. The method, the device and the system for automatic recommendation of the network content have the advantages that recommended site categories can be changed according to changes of personal interests of the users, and a great quantity of manual data annotation during training of the classifier is omitted.

The invention discloses a microwave drying PID (proportion integration differentiation) control method based on an increment improved BP (back propagation) neural network, selecting a three-layer BP neural network as a protomodel and combining incremental learning, an L-M (Levenberg-Marquardt) optimization algorithm, a BP neural network and PID control to realize online adjustment of PID control parameters. The method comprises the following steps: firstly carrying out offline system identification on a neural network by adopting the actual production data in a microwave drying process as training data, comparing the output value of the network with the measured value until the mean square error of network training meets the requirement, determining weight and threshold for each layer of the network, and carrying out online dynamic adjustment on the PID control parameters by taking parameters of a controlled object which are measured in the actual production process as the input of the neural network, wherein the output of the neural network is namely parameters of a PID controller Kp, Ki and Kd.

The invention discloses a SAR target recognition method based on a sparse least squares support vector machine, which belongs to the technical field of image processing and mainly solves the problem that the existing method need a long time for SAR target recognition. The realization process comprises the following steps of: firstly respectively implementing feature extraction to the selected target images with known classification information and images to be recognized to obtain training samples and test samples; and then applying iterative training to the training samples by using the combination of incremental learning method and reversal learning method to select a sparse support vector set and obtain a Lagrange multiplier and deflection corresponding to the support vectors in the set; and finally using a classification decision function to recognize the test samples according to the obtained support vector set, the Lagrange multiplier and deflection corresponding to the support vectors. The invention has the advantage of shortening recognition time under the condition of equivalent recognition precision and can be used for detection and recognition of SAR target.

The invention discloses a structure adaptive CNN (Convolutional Neural Network)-based face recognition method, and belongs to the field of identity recognition. According to the method, the advantages of a conventional CNN directly extracting characteristics from a two-dimensional image for recognition are maintained, and in addition, a network structure is adaptively constructed to overcome the shortcoming of excessive dependence of the conventional CNN on human experiences. According to the structure adaptive CNN-based face recognition method, the network is extended according to network requirements, so that the controllability and adjustability of the network structure are achieved, ineffective training is also avoided, difficulty in training for face recognition is lowered, and in addition, an optimal face recognition network structure is obtained; by adaptive network extension advantages, a newly added face sample can be relearned on the basis of maintaining early recognition results, retraining overhead is reduced, and incremental learning is implemented; the automatic intelligent face recognition method is low in training difficulty and high in accuracy under a big data condition.

The invention discloses a data driven incremental integration based screw type fault diagnosis model. A method comprises the following steps that data points are collected and divided into normal samples and fault samples; random sampling is conducted, and unbalance samples of different slope rates are obtained and divided into four groups; relative balance samples are obtained through a resampling method based on dividing neighbors; the relative balance samples are input into DAE to extract fault features, when new data exists, feature patterns are incrementally combined, and then the samplesare input into SVM for fault diagnosis; cases which have information content and are rich in representativeness are selected, and dynamic comprehensive evaluation is conducted on effective features and the cases; an effective case set and the new data are combined, and the incremental learning process is conducted again. The model obtains balance data beneficial to accurate fault type identification on the condition that sample noise and distribution features are fully considered, by conducting dynamic evaluation and incremental combination through selection features and the cases, effectiveinformation is reserved and passed on, and then rapid and efficient incremental learning and classification diagnosis of equipment faults are achieved.

The invention discloses a Chinese short-text sentiment classification method based on text characteristic insertion. Characteristic extraction is carried out in a text characteristic insertion manner; firstly, words are represented by vectors having a relatively short length through word insertion; on one hand, the characteristic dimension is reduced; on the other hand, the similar relationship between words can be described better; then, word vectors are weighted according to positions of different words in a text; therefore, text characteristic insertion is carried out; text characteristics having relatively low dimension can be obtained; on this basis, Chinese short-text sentiment classification is carried out, so that the Chinese short-text sentiment classification precision is increased; furthermore, because network languages are rapid to change and the randomness of Chinese short-texts is high; the word vectors are continuously updated in an incremental learning manner; therefore, the word vectors can keep up with change of a text to be trained; and thus, the text sentiment classification precision is improved.

The invention discloses a face tracking method which is used for tracking the position of a face in a continuous video image sequence. The method comprises the following steps: determining the position of the face in a first frame of image; determining a feature subspace with the first frame of image through the maximum class interval standard; reading the images in the video sequence in turn from a second frame of image; concerning each frame of image being read, determining the position of the face in the current frame of image through the position of the face and the feature subspace of the previous frame of image; and updating the feature subspace at intervals of preset number of frames. The invention adopts the maximum class interval standard, thereby strengthening the robustness of face feature extraction. Simultaneously the invention also adopts an increment learning method for on-line updating of the feature subspace, thereby not only improving tracking effect but also meeting real-time requirements.

An image increment learning method based on a variational auto-encoder comprises the following steps: 1) constructing a main network taking a ResNet50 network layer structure as a prototype, and introducing an MMD distance metric; 2) setting an optimizer and a learning rate, and adopting a knowledge distillation and weight punishment strategy; 3) training the ResNet 50 model by adopting a restricted sample increment method in the training mode; and 4) reloading the optimal model trained recently, and repeatedly using the restrictive sample increment method for training until all increment dataare trained, so that the generation of the disturbance resistance is not limited by the influence of a plurality of environmental factors in practice, and the practical value is higher.

The invention discloses an incremental learning image classification training method under the big data scene. The method comprises steps that step 1, an initial image classifier is trained through utilizing original image data; step 2, if a new class image appears in an application process, incremental training of an initial model is carried out, and the image classifier after update is acquired; and step 3, to-be-classified images are identified through utilizing the trained image classifier, and test image classes are acquired. The method is advantaged in that the convolutional neural network is utilized to carry out image characteristic extraction, manual characteristic definition can be avoided, and the high identification degree is realized; the convolutional neural network is utilized to carry out image identification, memory occupation is small, the calculation speed is fast, incremental leading of the new class image can be carried out, storage of the original training data is not needed, the training time and the storage time can be substantially saved, and the method is especially suitable for big data image classification occasions.

The invention provides a content sensing method based on network stream behaviors. The content sensing method comprises the following steps: acquiring network flow in an outer network environment andextracting an observational characteristic as a training sample; training a model by utilizing the training sample; inputting a network stream with an unknown type into the model and identifying the content; carrying out incremental learning by utilizing recognized network stream flow and historical model parameters; updating the model parameters to ensure the continuity of model classes. According to the content sensing method provided by the invention, a dynamic modeling capability of a hidden Markovmoder model and a powerful nonlinear expression capability of a deep neural network are utilized; an experiment result shows the feasibility of the method and compares performance advantages of an existing technical scheme.