125results about How to "Improve classification ability" patented technology

The invention discloses an image classification method based on a semi-supervised self-paced learning cross-task deep network. The method includes the steps of randomly selecting a small amount of labeling samples from the whole image data set, reserving the labels, and remaining all the samples as unlabelled samples having the real labels to be unknown in the whole process, wherein the weight ofthe labeled samples is constant to be one in the training process, the weight of the unlabelled samples is initialized to be zero, and only the labeled samples are used as a training set in the initial process; S2, training a cross-task deep network by the training set; S3, according to the trained cross-task deep network, predicting the pseudo labels of all the unlabelled samples, and giving a corresponding weight of each unlabelled sample; S4, according to a self-paced learning normal form, selecting an unlabelled sample with a high confidence degree, and adding to the training set; and S5,repeating the steps S2-S4 until the cross-task deep network performance is saturated or reaches a preset cycle number. According to the method, the human design feature is not needed to be input, andthe classification can be realized by directly inputting the original image.

The invention relates to a recognition method of digital music emotion, belonging to the field of computer pattern recognition; the recognition method solves the problem that the existing recognition method of digital music emotion can not recognize sampling-based digital music format, the sorting technology based on a multi-class support vector machine is adopted, acoustic characteristic parameters and music theory characteristic parameters are combined, so as to carry out emotion recognition of digital music; the recognition method comprises the following steps: (1) pretreatment; (2) characteristic extraction; (3) training the multi-class support vector machine; (4) recognition. The music emotion is classified into happiness, impassion, sadness and relaxation, the emotion recognition is carried out based on a sampling-based digital music format file, the common acoustic characteristics in the speech recognition field are not only extracted, and a series of music theory characteristics are extracted according to the theory characteristics of music; meanwhile, the sorting method based on the support vector machine is adopted, the leaning speed is rapid, the sorting precision ratio is high and the recognition efficiency is improved.

The invention discloses a wavelet transformation and deep learning-based rolling bearing weak fault diagnosis method. The method comprises the following steps: a rolling bearing vibration signal is obtained, and the acquired vibration signal is subjected to continuous wavelet transformation to obtain a time-frequency diagram; autocorrelation operation is performed on wavelet coefficients corresponding to each frequency on the time-frequency diagram to filter out noise interference and extract periodic fault components; the Hilbert transformation is used to perform envelope demodulation to obtain a fault characteristic frequency; a processed time-frequency diagram is input as a feature diagram, and categories of early faults are determined by training a deep learning classification model. Theoretical and experimental results show that fault categories can be determined accurately at an early stage when weak fault of a rolling bearing occur based on an improved wavelet time-frequency diagram used as the classification model for input training, right determination results can be given when the method is applied to different bearings, high training speed can be realized, and high actual application value can be achieved.

The invention discloses a face recognition method of a deep convolutional neural network, which reduces the time complexity, and enables a weight in the network to still have a high classification capacity under the condition of reducing the number of training samples. The face recognition method comprises a training stage and a classification stage. The training stage comprises the steps of (1) randomly generating a weight wj between an input unit and a hidden unit and an offset bj of the hidden unit, wherein j equals to 1,...,L and represents the number of the weight and the offset, and the total number is L; (2) inputting a training image Y and a label thereof, by using a forward conduction formula hw, b(x)=f(W<T>x), wherein hw, b(x) is an output value, x is input, and an output value hw, b(x<(i)>) of each layer is calculated; (3) calculating the offset of the last layer according to a label value and an output value of the last layer; (4) calculating the offset of each layer according to the offset of the last layer, and acquiring the gradient direction; and (5) updating the weight. The classification stage comprises the steps of (a) keeping all parameters in the network to be unchanged, and recording a category vector outputted by the network of each training sample; (b) calculating a residual error delta, wherein delta=||hw, b(x<(i)>)-y<(i)>||<2>; and (c) classifying a tested image according to the minimum residual error.

The invention discloses a plastic pipe surface defect detection method, system and equipment and a computer readable storage medium. The method comprises the steps that surface defect features corresponding to the surface image of a plastic pipe to be detected are acquired so as to obtain the defect features to be detected; and the defect features to be detected are inputted to a pre-created trained model so as to obtain the surface defect types outputted by the trained model, wherein the trained model is the model obtained by training the model to be trained created on the basis of a BP neural network algorithm by using training samples in advance, and the training samples include the historical surface defect features and the corresponding surface defect types. The highly nonlinear mapping relation of the input and the output of the BP neural network and the great fault tolerance capability, the classification capability and the self-learning capability are utilized so that the detection accuracy of the plastic pipe surface defects can be enhanced, a lot of labor resources can be saved and the detection efficiency can be greatly enhanced.

The invention discloses a night pedestrian detection method based on statistical features of infrared pedestrian brightness. According to the method, gray average information of parts of one pedestrian and a negative sample in a sample database is firstly subjected to statistical processing, a mapping interval boundary is determined by the aid of the information, and a brightness histogram feature for distinguishing vote interval division is constructed; then, a histogram feature in the gradient direction is calculated, and the two features are combined to form a final feature descriptor; secondarily, model training is performed by the aid of Adaboost in combination of a decision tree, and pedestrian determination and positioning are performed with a sliding window scanning method; finally, when a classifier obtains lower degree of confidence through classification judgment on a certain detection frame, a brightness interval template is adopted for secondary judgment, so that night pedestrian detection is realized. Pedestrians in the night environment are effectively detected, and the method has the characteristics of high detection rate and adaptability.

The invention discloses a radar scanning-mode sea clutter local multi-fractal target detector, belonging to the field of processing radar signals. The working of the radar in a residing mode is supposed in the existing sea clutter local multi-fractal target detection method, thus having larger dimensionality calculation and computation workload and being lack of instantaneity. The invention can be used for carrying out multi-fractal calculation on the basis that the radar works in a scanning mode, and a primary scanning process is not needed to acquire sufficient sampling points; the calculation of each scanning period dimensionality is updated by an iterative approach, thus ensuring the instantaneity; and the fuzzy characteristics of a local multi-fractal spectrum is extracted to form detection statistic. The detector has favorable capabilities of classifying the sea clutter and the target and detecting weak targets in strong sea clutter, is suitable for tracking target signals into fractal unstable clutter, and has generalization and application value.

The invention discloses a method for structuring a text classifier by carrying out characteristic expansion treatment on text information objects in a training set and a text classifying device applied to the method. Through carrying out the characteristic expansion treatment on texts to classify the texts and carrying out the characteristic expansion treatment on training texts to structure the classifier, the performance of the classifier is improved and harmful information in short text information can be intercepted and filtered. The invention has favorable identification ability and classification ability on short texts and is particularly suitable for the processing of the texts in instant messaging systems such as QQ, MSN and cellphone short messages and the texts in network comments.

The invention provides a micro-expression recognition method based on a difference slice energy diagram and Gabor transformation. The micro-expression recognition method comprises the following steps: constructing a micro-expression sequence difference slice energy diagram, calculating micro-expression frames containing change regions in a micro-expression periodic sequence, graying the micro-expression sequence to extract micro-expression difference slices, and overlapping micro-expression difference slice frame sequences to obtain a micro-expression difference energy diagram; extracting the characteristics of the difference slice energy diagram, constructing a Gabor filter kernel function, performing Gabor characteristic extraction on the difference slice energy diagram, sampling the extracted characteristics, writing the characteristics of each sample in a column vector form, maximizing the quotient of inter-class divergence and intra-class divergence through linear discriminant analysis with supervisory information, and further extracting the characteristics of each sample; performing classification and identification, training a model according to training data, and predicting and classifying test samples according to the model. The method provided by the invention has general applicability. Compared with an existing method, the method provided by the invention has a higher identification rate.

The invention discloses a semi-supervised learning-based pedestrian detection method. The method includes the following steps that: the training samples of a source image set and the categories of the training samples are obtained, pedestrian labeling is performed on a part of images in a target scene image set, and training samples and sample features corresponding to target scene images are obtained; a decision-making forest is generated through training based on the training samples of the source image set, training samples of which the categories are known in the target scene image set are adopted to screen decision-making trees in the decision-making forest, and after the decision-making trees are reorganized, a new decision-making forest can be generated; the new decision-making forest is adopted to score training samples of which the categories are unknown in the target scene image set, and training samples with high confidence are labeled as pedestrian training samples; the training samples of which the categories are known in the target scene image set and the pedestrian training samples are adopted to train a neural network; and test samples are inputted into the new decision-making forest, test samples with high confidence are made to pass through the neural network, so that a pedestrian detection result is obtained. The semi-supervised learning-based pedestrian detection method is advantageous in high pedestrian detection accuracy.

The invention discloses a method for quickly recognizing a front face through video monitoring, which is used for recognizing the most front image frame in one segment of video monitoring. The method comprises the following steps of firstly, extracting front face images from a standard video base and monitoring videos to use as a training positive sample, extracting side face images as a training negative sample, then extracting features of an integral channel, and utilizing an Adaboost algorithm to train a strong classifier from the features. When the front face is recognized, in the inputted monitoring videos, a face is detected by the existing matured face detection algorithm, a detection window is recognized, the images are scored, and the image with highest score is selected, namely the most front face frame. The method has the advantages that the recognizing speed and recognizing precision of the face gesture are improved, the front face is accurately recognized, and the calculation workload of face gesture recognizing is reduced.

The invention is applicable to the field of computer technology, and provides a collaborative recommendation method and device and a terminal device. The method comprises the following steps of: acquiring a predetermined amount of item data set information, wherein the item data set information comprises user scoring information of items in the item data set; based on the project data set information, carrying out an iterative search according to a crow search algorithm to find the position of the cluster center of the optimal fuzzy C-means cluster; taking the position of the cluster center ofthe optimal fuzzy C-means cluster as the position of the initial cluster center; based on the optimized position of the initial cluster center, calculating the similarity between users according to the fuzzy C-means clustering method and classifying the users according to the similarity to construct a recommendation model; based on the recommendation model, predicting the items that are not graded by the target user, obtaining the prediction score by similarity measurement, and recommending the first N items to the target user based on the ranking information of the prediction score.

The invention discloses a pedestrian detection method based on incremental learning. The pedestrian detection method based on the incremental learning can be used in the field of pedestrian detection in the background of big data. The pedestrian detection method based on the incremental learning comprises the following steps: firstly, an initial classifier is obtained by utilizing an image sample to intensively train a classifier; secondly, a novel training set is utilized to carry out incremental learning of the classifier, and a renewed classifier is obtained; thirdly, an image to be detected is detected by utilizing the renewed classifier to obtain a pedestrian detection result. The pedestrian detection method based on the incremental learning can occupy little internal storage in short time, and utilizes a plurality of continuously collected training sets in the background of big data to effectively carry out study and renew the classifier. Compared with a common method, the pedestrian detection method based on the incremental learning is high in speed, small in occupied internal storage, and convenient to use, and a classifying effect quite similar to the common method can be learnt.

The invention discloses a multi-source self-adaptive fault-tolerant federated filtering integrated navigation system and a navigation method. The system comprises a strapdown inertial navigation system, a satellite navigation system, a Doppler velocity measurement system, a celestial navigation system, and a main filter and three sub-filters that are in information connection with the systems respectively; the three sub-filters are all in information connection with the strapdown inertial navigation system, and are connected to the main filter through a fault detection and isolation module, the strapdown inertial navigation system is connected to the fault detection and isolation module through a state transmitter, and an output result of the fault detection and isolation module passes through an information sharing factor calculation module, then the calculation result is input into the main filter; and the main filter outputs the fused information and synchronizes a fusion result tothe three sub-filters and the state transmitter. According to the system and the method disclosed by the invention, the state of each sub-filter can be tracked more accurately, and a more precise fusion result can be obtained.

The invention provides a quality control chart pattern recognition method based on improved genetic algorithm optimization. The quality control chart pattern recognition method comprises the followingsteps: simulating various pattern characteristics of a control chart by using a Monte Carlo method; generating data of a corresponding mode through the parameter values; adopting the PCA principal component analysis method to carry out dimension reduction and denoising on the original data, main features of the data are extracted, shortening the training time of the model and improvintg the recognition accuracy; establishing a probabilistic neural network model, and carrying out pattern classification recognition by utilizing the characteristics of simple structure and convenient training; optimizing a main parameter smoothing factor of the probabilistic neural network by virtue of an improved single-objective optimization genetic algorithm; searching possible abnormal reasons from different aspects according to the identification result.The method solves the problems that all abnormal conditions cannot be monitored and recognized when an existing enterprise carries out quality control, effective abnormal information is difficult to find from a control chart, and appropriate measures cannot be taken to correct the abnormal conditions in the production process.

The invention relates to a multi-label text data feature selection method and device, and belongs to the technical field of text data processing. The multi-label text data feature selection method comprises the steps of firstly considering second-order correlation between marks in a text data set, grouping the marks to enable the marks to be better suitable for a multi-mark data set, determining afinal score of each feature according to scores calculated by the features in each mark group, and selecting a set number of features with higher scores from the final scores to form a feature set; and then, based on the obtained feature set, determining the neighborhood granularity of each sample according to the classification interval of each sample in the text data set for the marks to obtaina multi-mark neighborhood decision-making system, calculating importance by utilizing the dependency degree of the improved neighborhood rough set, and screening the obtained feature set, thereby realizing feature selection of the multi-mark text data. Compared with an original neighborhood rough set feature selection method for all attributes, the multi-label text data feature selection method is lower in time complexity, and the optimal feature subset is more accurate.

The invention belongs to the technical field of image classification, and particularly relates to a zero-sample image classification method of an adversarial network based on meta-learning. The methoduses a meta-learning training mode in a zero-sample classification task, and simulates the learning task of zero-sample image classification in a training stage by inputting visual features and semantic features into a network in sequence. According to the method, the generation process of the visual features is completed, the alignment relation of different classifiers is guaranteed, meanwhile,knowledge obtained by tasks of each epinode is fully utilized, semantic classifiers are better trained under the supervision of the visual classifiers, and therefore the visual features and the semantic features which are closer to real distribution are synthesized, and the visual features and the semantic features which are closer to real distribution are obtained. A zero-sample image classification technology suitable for a real situation is designed. According to the method, the generalized zero-sample image classification capability can be more outstanding, the generalization capability ofthe model is improved, and the problem of domain offset generally existing in zero-sample learning is relieved.

The invention relates to a target tracking method based on a TLD algorithm framework. The method comprises steps that (1) a user selects a tracking target in an initial frame, and initializes the algorithm; (2) a local tracker in a tracker module predicts the position in the next frame by an optical flow method and a cascaded predictor; (3) a detector detects the target of the current frame through the sliding window scanning and the cascaded classifier, and sends a classification result to a learning module, wherein the tracker and the detector work in parallel; (4) an integrator determines afinal target position by determining and integrating a target result obtained by the tracker and the detector; (5) the learning module corrects the results of the tracker and the detector through structural constraints and updates a target sample; (6) whether the video frame is finished is determined, and if not, the steps (2)-(4) are repeated until the end of the tracking. The method can improvethe robustness of the tracking target and ensures the real-time performance in the tracking process.

The invention provides a heart sound classification method and system based on CNN combined with improved frequency wavelet slice transformation, and the method comprises the steps: carrying out the preprocessing of an obtained heart sound signal, finding the position of each cardiac cycle through a hidden half Markov model, and carrying out the interception of each cardiac cycle signal; converting each intercepted one-dimensional cardiac cycle signal into a two-dimensional time-frequency image by using improved frequency slice wavelet transform; calculating sample entropies of the acquired heart sound signals respectively; comparing the two-dimensional time-frequency image with a preset sample entropy threshold, when the sample entropy of the heart sound signal is greater than the presetsample entropy threshold, performing network training and classification according to the two-dimensional time-frequency image by using a first convolutional neural network, otherwise, performing network training and classification according to the two-dimensional time-frequency image by using a second convolutional neural network; firstly, signals with different interference degrees are distinguished by using sample entropy, and then classification is performed by using different convolutional neural network models for different signals, so that the accuracy of heart sound signal classification is greatly improved.

The invention is suitable for the mode identification technology field and provides a supervised learning optimization method under a tensor mode and a system thereof. The method comprises the following steps of receiving an input training tensor data set; introducing a within-class scattering matrix into a target function so that the target function maximizes a between-class distance and simultaneously minimizes a within-class distance; constructing an optimization framework of a target function of an OPSTM subproblem; constructing an optimization framework of a target function of an OPSTM problem; solving a modified dual problem and outputting a Lagrangian optimal combination and an offset scalar b; calculating a projection tensor w*; calculating an optimal projection tensor w; according to the w and the b, constructing a decision function; carrying out rank decomposition on tensor data to be predicted and then inputting into the decision function so as to carry out prediction. In the invention, problems of a dimension curse, over learning, a small sample and the like generated when a vector mode algorithm is used to process the tensor data are overcome and a time-consuming alternative projection iteration process in an existing tensor mode algorithm is effectively avoided.

The invention provides a fault diagnosis method and apparatus for a main sintering exhaust fan. The fault diagnosis method comprises: collecting a vibration signal of a main sintering exhaust fan; and carrying out wavelet packet decomposition on the vibration signal, extracting a test sample and acquiring fault information of the main sintering exhaust fan by using an extreme learning machine model. With the wavelet packet analysis method, effective extraction and decomposition can be carried out on high-frequency and low-frequency components; while information integrity is ensured, characteristics of non stability and short duration of the signal can be reflected with deepening of a fault degree; the severe working environment of the main sintering exhaust fan and the harmonic characteristic of the vibration signal can be adapted effectively. Meanwhile, because of the powerful classification capability of the extreme learning machine, adaptability to various fault types of the main sintering exhaust fan is improved. Therefore, with the fault diagnosis method, a technical problem of poor accuracy of fault diagnosis in the prior art is solved.

The invention discloses an air pollution monitoring system based on NB-IoT technology and edge computing. The air pollution monitoring system comprises a sensing layer, a transmission and processing layer and a platform layer. The sensing layer takes an STM32 processor as a main control chip, collects atmospheric pollutant information by using a laser dust sensor and a gas sensor, and sends monitoring data to an edge computing server through an NB-IoT network base station to be cooperatively processed with a cloud server, so that real-time analysis and processing of atmospheric pollutant data are realized; and the processed data are uploaded to an upper computer, so that the position of a pollution source on a map can be positioned on line, and an early warning function can be realized through a short message mode or a mode of checking pollution indexes in real time through a Web page. According to the invention, a mode of cooperative processing of a plurality of edge computing servers and the cloud server is implemented through NB-IoT wireless communication network transmission, the load of a cloud platform is reduced, the data transmission efficiency and the accuracy of atmospheric pollution data monitoring are improved, acquisition, transmission, analysis processing, early warning and positioning of atmospheric pollution monitoring are realized in an all-around manner, and it is convenient for a user to know the air pollution condition in time.