87results about How to "Improve feature extraction efficiency" patented technology

The invention discloses a collaborative filtering recommendation method and a collaborative filtering recommendation system of a convolution neural network integrating an attention model, which relates to the technical field of data mining recommendation, improves feature extraction efficiency and scoring prediction accuracy, reduces operation and maintenance cost, simplifies cost management mode,and is convenient for joint operation and large-scale promotion and application. The invention relates to a collaborative filtering recommendation method of a convolution neural network integrating an attention model, comprising the following steps: step S1, splicing a user feature vector and an item feature vector into a new vector; S2, sending the new vector as an input vector into the multi-layer perceptron to learn and predict the score; The attention model is fused into the object potential vector to obtain the convolution neural network of the object feature vector or the hidden layer of the multi-layer perceptron.

The invention discloses a Dalvik instruction abstraction-based Android malicious code detection method. The method comprises the following steps of 1) detecting a malicious code and training a classification model: extracting a Dalvik operation code from a smali file, performing abstraction simplification to obtain an instruction symbol, performing statistics and normalization processing on N-Gram sequence characteristics of the abstract Dalvik instruction symbol, and finally establishing a malicious code detection model and a malicious family classification model by adopting a machine learning-based classification algorithm; and 2) preprocessing a to-be-detected APK file firstly, extracting Dalvik instruction characteristics, performing abstraction simplification and N-Gram serialization processing, and preliminarily judging whether the code is a malicious code or not through detection of the malicious code detection model; and if not, directly giving a detection result, or if yes, obtaining a malicious code family type further through the malicious family classification model. The method is high in speed and relatively high in validity.

The invention discloses a method of video monitor object automatic detection and a system thereof. Combined with object motion information and form information in a video, based on a Gentle AdaBoost algorithm and an expanded Harr characteristic, a classifier is trained and object detection in the video is carried out automatically, a return value is determined after employing a training sample window to pass through each layer of a cascade classifier, when the return value is a positive number, an object is searched in real time in video monitoring and is highlighted, and a problem of low inquire efficiency of mass video data in the prior art is solved. The method and the system have the characteristics of simple design, fast detection speed, high precision and strong robustness, efficiency of extracting a characteristic in the video is raised, and the method and the system can be widely used for pedestrian detection and tracking in the video retrieval field.

The invention provides a point cloud posture standardization-based method for extracting the linear characteristic of point cloud, which comprises six steps, is used for extracting the linear characteristic in disordered and three-dimensional point cloud, can conveniently measure the relative posture of a target, and belongs to the technical field of the three-dimensional measurement and the machine vision. The method comprises the following steps of: firstly, building a KD-TREE structure of point cloud, so that the searching speed of the adjacent point set of the point cloud can be improved; secondly, building the adjacent point set of each point according to the density of the whole point cloud, obtaining the main direction of the point set, and building a Householder transformation matrix to adjust the posture of the point cloud; thirdly, carrying out surface fitting on the adjacent point set to obtain two main curvature of the point based on a curved surface equation, and selecting the main curvature with the higher absolute value as the curvature estimation of the point; and finally, obtaining the curvature estimation value of all point cloud, and taking the point which is larger than a given threshold value as a linear characteristic point, so that the linear characteristic can be extracted.

The invention discloses a HOG (Histograms of Oriented Gradients) type quick feature extracting method, and belongs to the technical field of computer die recognition. The method comprises the steps of 1, preprocessing an image; 2, calculating gradient of each pixel of the image, and dividing the image into cells and blocks; 3, calculating the HOG of each cell; 4, normalizing the HOG of each cell in each block; 5, collecting the HOG in each block, and combining into the HOG feature of the image; when calculating the HOG of each cell, each cell is divided into four sub-cells, the sub-cell adjacent to any cell M is selected from other cells in M blocks, and then all pixels in the cell M and all pixels in the selected sub-cells are subjected to tri-interpolation so as to obtain the HOG of the cell M. With the adoption of the method, the feature extracting speed can be greatly increased whole ensuring the effectiveness of the features.

The invention discloses a novel efficient power quality disturbance image feature extraction and recognition method. The method comprises the following steps: converting an electric energy quality signal into a gray level image, enhancing disturbance characteristics by using three methods of gamma correction, edge detection and peak-valley detection to obtain a binary image, and extracting nine characteristics of area, Euler number, angle second moment, contrast ratio, correlation, mean value, variance, inverse difference moment and entropy to construct an original characteristic set; carryingout sorting on the basis of the feature Gini importance degree, and determining the feature with the maximum influence on classification; and comprehensively considering the classification precisionand efficiency, determining the number of trees in the random forest, and constructing a random forest classifier by using the optimal feature subset to identify the power quality disturbance signal.According to the invention, 8 types of common power quality disturbance signals of voltage sag, voltage sag, voltage interruption, flickering, transient oscillation, harmonic waves, voltage cutting marks and voltage peaks under different noise environments can be identified efficiently and accurately, and the feature extraction efficiency of the disturbance signals is improved.

The invention discloses a method for extracting space conical reentry target micro-motion features based on empirical mode decomposition, which mainly solves the problem of easiness in failure of feature extracting in the prior art. The method adopts the scheme with the following steps of 1, according to a narrow-band linear frequency modulating signal model, calculating a transmitting signal sequence in a pulse repeating cycle; 2, according to a transmitting signal and a received target echo signal, establishing a pulse compression signal matrix, and establishing a Doppler echo signal of a conical reentry target according to the matrix; 3, according the calculated Doppler echo signal of the conical reentry target, utilizing the empirical mode decomposition to obtain a plurality of feature mode functions; 4, according to the obtained feature mode functions, reestablishing a Doppler signal of a conical reentry target scattering center; 5, according to the reestablished scattering center signal, establishing a time-frequency map of the conical reentry target scattering center; 6, extracting the micro-motion features of the target from the time-frequency map. The method has the advantage that while the micro-motion features are accurately extracted, the feature extracting efficiency is improved, so that the method can be used for identifying the target.

The invention relates to a method for extracting characteristics of a three-dimensional model based on a section ray, which aims at the problems of a high calculation complexity, a larger distortion, not having a common research method and being difficult to obtain higher research precision in the existing method for extracting the characteristics of a mode based on contents. The method takes distance sets from specific points on a principal axis section of the model to the model surface as a characteristic description of the model, thus a problem of the model matching is transformed into the problem of comparing a group of distance sets. The method is divided into two sections of characteristic extraction and characteristic matching; when extracting the characteristics, the method can simply and rapidly obtain the model characteristics which have small reconfiguration distortion and can correctly describe a stereochemical structure and structural characteristics of the three-dimensional model. The method has the advantages of simplification, wide applicability, good research quality and high efficiency, therefore, the method has higher practical applicability and wider application range.

The invention discloses an image super-resolution reconstruction method capable of gradually fusing features and an electronic device, and the image super-resolution reconstruction method comprises the steps: obtaining an original image and a reconstruction network, carrying out the feature extraction of the original image through a first feature extraction module, and the original feature map passing through the plurality of second feature extraction modules and the plurality of third feature extraction modules in sequence, and using the image reconstruction module for performing super-resolution reconstruction on the intermediate feature map to obtain a target image with higher resolution. According to the method, features useful for image super-resolution reconstruction are screened out through multiple times of feature fusion and multiple times of step-by-step screening, the proportion of useful feature information finally input into the image reconstruction module is larger, the model feature extraction effect is better, and useful information loss and useless information redundancy are reduced. And meanwhile, the occupied computer memory in the model training and running process is smaller.

The invention discloses a vehicle detection method based on improved YOLOv3. According to the method, a feature weight optimization structure SEnet is introduced into a residual structure of a featureextraction network Darknet53 in YOLOv3 to optimize features extracted by the YOLOv3; residual error module reduction is performed on the feature extraction network after the SEnet structure is introduced; and the heights and the widths of vehicle labels in a vehicle data set are clustered by adopting a k-means method so as to optimize the original YOLOv3 target detection priori box. During batchtesting, the improved YOLOv3 has a vehicle detection map of 96.49% and a detection speed of 45 graphs / second, so that the requirements of real-time performance and accuracy of traffic detection are met, and a high application value is achieved in traffic detection.

The invention discloses a method for identifying cells in an embryo light microscope image. The method comprises the following steps: preprocessing the embryo light microscope image; carrying out labeling processing on the preprocessed embryo light microscope picture; inputting the marked embryo light microscope picture into a FasterRCNN recognition model trained in advance to generate a cell prediction result, wherein the FasterRCNN recognition model comprises a feature extraction network, an RPN network, a Roi Align network, a classification regression network and a C-NMS network; and carrying out cell identification according to the cell prediction result. The invention further discloses a system for identifying the cells in the embryo light microscope image, computer equipment and a computer readable storage medium. By adopting the method and the system, accurate extraction of the cells in the embryo light microscope picture is realized through deep optimization of the Faster RCNNnetwork, meanwhile, a brand-new CNMS network is constructed, the detection score is flexibly adjusted through detection of the overlapping proportion and the area proportion of the detected objects, and the omission ratio is remarkably reduced.

The invention discloses a laser point cloud super-resolution reconstruction method based on a self-attention generative adversarial network, and the method comprises the steps of carrying out the feature extraction of a laser point cloud image in a generator network, and obtaining the laser point cloud features; carrying out feature expansion on the laser point cloud features, and then carrying out coordinate reconstruction to obtain dense point cloud data; identifying the dense point cloud data to determine a corresponding confidence coefficient; pre-judging the corresponding dense point cloud data according to the confidence coefficient of the dense point cloud data, if the confidence coefficient value is close to 1, predicting that the input is possibly from target distribution with high confidence coefficient by the discriminator, otherwise, performing feature integration on the dense point cloud data by a generator to obtain an output feature; and training the adversarial networkthrough the output features to obtain final dense point cloud data. According to the invention, feature information sharing among different feature extraction units can be realized, the size of the model is reduced while the reconstruction precision is improved, and lightweight of the network model is facilitated.

The invention relates to a radar clutter identification method based on a convolutional neural network. The method comprises the steps of obtaining clutter data; Dividing the clutter data into training data and test data; Constructing a convolutional neural network; Using the training data to train the convolutional neural network; Testing the trained convolutional neural network by using the testdata; And when the recognition accuracy of the convolutional neural network is greater than a preset accuracy, obtaining an optimal convolutional neural network. According to the method provided by the invention, the clutter data is processed and input into the convolutional neural network, and implicit feature extraction is carried out on the clutter data by utilizing the convolutional neural network, so that the feature extraction efficiency is improved; By reducing the number of clutter data, the sampling cost is reduced, and meanwhile, the recognition efficiency is improved; The convolutional neural network is used to classify and identify the clutter data, the identification accuracy rate reaches 99.51%, and the identification accuracy rate is greatly improved.

The invention provides a video target tracking method and device. The method comprises the following steps: respectively inputting a tracking target image and a search area image into a first feature extraction module and a second feature extraction module, and carrying out image feature extraction; inputting the tracking target features and the search area features into a feature fusion module based on an interactive attention mechanism for feature fusion; and inputting the fusion features into a classification and regression module, and outputting the image category in the bounding box and the position and size information of the bounding box. According to the invention, image feature extraction is carried out by adopting an attention mechanism, and the region-of-interest features can be obtained to improve the feature extraction efficiency; and an interactive attention mechanism is adopted to carry out feature fusion, so that target features and search area features are fully interacted, and the problem that targets are lost due to illumination, deformation, shielding and the like in the prior art is solved.

The invention discloses a stroke identification method and system based on an FMCW radar system. The method comprises the following steps: obtaining intermediate frequency signal data of at least oneto-be-identified stroke contained in handwritten Chinese characters based on the FMCW radar system; preprocessing the intermediate frequency signal data of each to-be-identified stroke to obtain a feature map set of each to-be-identified stroke; obtaining a trained Chinese character basic stroke recognition model, wherein the Chinese character basic stroke recognition model is a convolutional neural network model taking a feature map set as an input parameter and taking a basic stroke category as an output parameter; and inputting the feature map set of each to-be-identified stroke into the Chinese character basic stroke identification model, and obtaining a basic stroke category which is output by the Chinese character basic stroke identification model and is matched with each to-be-identified stroke. According to the method, the data volume used for representing the gesture motion trend is reduced, the feature extraction efficiency is improved, and the basic stroke category can be accurately judged.

The invention discloses an Alzheimer's disease feature extraction method and an Alzheimer's disease feature extraction system based on collective correlation coefficients. The Alzheimer's disease feature extraction method comprises the steps of: acquiring magnetic resonance imaging data of the Alzheimer's disease; and adopting a genetic algorithm based on the collective correlation coefficients toperform feature optimization on the acquired magnetic resonance imaging data, so as to obtain key features of the Alzheimer's disease, wherein the genetic algorithm based on the collective correlation coefficients regards the collective correlation coefficients as heuristic knowledge and regards an optimal classification effect as a target to extract the key features. The Alzheimer's disease feature extraction method and the Alzheimer's disease feature extraction system adopt the genetic algorithm based on the collective correlation coefficients to perform feature optimization on the acquiredmagnetic resonance imaging data, combine the collective correlation coefficients with the genetic algorithm to optimize the traditional feature extraction process, improve the feature optimization efficiency of the genetic algorithm by regarding the collective correlation coefficients as heuristic knowledge, regard the optimal classification effect as the target, effectively improve the feature extraction efficiency on the premise of ensuring the classification effect, and can be widely applied to the field of data mining.

The invention discloses a target region extraction method for a multi-modal medical image based on a convolutional neural network. The method comprises the following steps: 1) constructing a mask region convolutional neural network for target region extraction in a multi-modal medical image; 2) training the constructed mask region convolutional neural network; and 3) inputting a to-be-processed multi-modal medical image into the trained mask region convolutional neural network to perform target region extraction. According to the invention, the automatic and accurate segmentation of a target area in the multi-modal medical image can be realized, the subjective difference problem and the time-consuming and labor-consuming defects of the manual segmentation of the target area can be overcome, and the accuracy of the extraction of the target area in the multi-modal medical image can be improved; according to the invention, the feature image extraction of the multi-modal medical image canbe realized through a plurality of parallel SE-Resnet, and the feature extraction efficiency of the medical image and the information fusion efficiency of the multi-modal medical image can be improvedby integrating an extrusion excitation block into a feature extraction network.

The invention discloses a video motion characteristic extraction method based on a fuzzy concept lattice, which is mainly used for solving the problems of background interference and motion ghost existing in the conventional method. The method comprises the following implementation methods of: firstly, partitioning a video shot, generating a motion characteristic association rule of all lenses byusing the fuzzy concept lattice and extracting an interesting lens according to the association rule; secondly, generating a motion characteristic association rule of all target frames in the interesting lens by using the fuzzy concept lattice and extracting an interesting target frame according to the association rule; and lastly, extracting the motion characteristic of the interesting target frame according to fuzzy concept lattices of all image blocks in the interesting target frame. According to the method, the video motion characteristic can be extracted quickly; and the method can be applied to occasions needing to process mass video data such as target tracking, video monitoring and the like.

The invention relates to a face recognition method based on MBLBP and DCT-MB2DPCA, belongs to the technical field of computer vision process, and solves the problem that the recognition rate is low due to a single feature extract method in an existing face recognition algorithm. Based on a multi-scale block local binary system mode and discrete cosine transform bi-directional module two-dimensionmain component analysis, the face recognition method is achieved by the following steps. A face image is converted from a spatial domain to a frequency domain through DCT, and then the face image is reconstructed through IDCT; feature extract is carried out on the converted face image by using an MBLBP operator; a characteristic matrix is acquired through BM2DPCA; a nearest neighbor classifier isused to recognize a test sample. The face recognition method is applicable to two-dimension face recognition in the fields of safety system, identification, personal equipment login and the like.

The invention relates to a feature extraction method and device based on reinforcement learning, and a computer device. The method comprises the steps of obtaining a feature extraction code of a learning object; wherein the feature extraction code is determined according to manual writing; acquiring state features of the learning object according to the feature extraction code; training a deep network structure based on reinforcement learning by adopting the state features; obtaining an optimal network structure and an optimal weight parameter of the trained deep network structure; generatingan optimal feature extraction strategy according to the optimal network structure and the optimal weight parameter; wherein the optimal feature extraction strategy is used for extracting portrait features of the insurance service user so as to analyze insurance demands of the insurance service user according to the portrait features. By adopting the method, the feature extraction codes are set tobe applied to model training, so that the feature extraction efficiency can be improved, namely, a modeling effect is used as a learning reward to stimulate a computer to continuously optimize a learning strategy so as to learn a new feature extraction mode.

The invention discloses a four-value weight and multiple classification-based human face feature extraction method. The method comprises the steps of constructing a human face training sample database; establishing a convolutional neural network; adjusting a caffe framework; preprocessing a human face image sample, inputting the human face image sample to the convolutional neural network for performing training, until the network is completely converged, and storing a generated human face identification model; and preprocessing a to-be-extracted human face image, producing a mean value file, inputting the mean value file to the human face identification model to obtain a feature graph, rotating the feature graph for multiple different angles to extract features respectively, performing addition fusion on different angle features of the same image, and finally obtaining a main human face feature. The method has the beneficial effects that the problems of huge memory consumption and insufficient storage space of network training is radically solved; and the feature with a strongest expression capability is obtained in a multi-feature extraction fusion mode, so that the human face identification accuracy is remarkably improved.

The invention discloses an HTTP (Hyper Text Transport Protocol) traffic feature recognition and extraction method based on machine learning. The method comprises the following steps: step 1, carryingout HTTP traffic recognition and acquisition; step 2, carrying out feature detection and generating rules and step 3, extracting HTTP traffic characteristics. Compared with feature extraction based onregular expressions on the existing market, the method has the advantages that the feature accuracy is improved, the probability of mistakenly extracting dirty data by the regular expressions is reduced, and compared with a feature marking method based on manpower, the labor cost input and the response feedback time to novel features are reduced. Meanwhile, in the patent, feature / rule generationand feature extraction are separated, a unique extraction engine can be designed, and the feature extraction efficiency is improved.

The invention provides an image dimension reduction method of an extreme learning machine based on graph embedding, and belongs to the field of machine learning and data mining. The method comprises the following steps: firstly, selecting an original image sample set, and constructing a sample relation matrix by utilizing inter-sample distances and label information of original image samples; secondly, according to the constructed sample relation matrix, firstly, carrying out random mapping on an input vectorized image sample, and then, learning a feature extraction matrix by minimizing a weighted sample reconstruction error; and finally, performing data dimension reduction on vectorized image data by using the learned feature extraction matrix. The method is short in training time and efficient in data compression, and the compression quality and dimension reduction stability of the data are effectively improved.

The invention discloses an image sample library feature representing method based on grayscale distribution statistical information. The image sample library feature representing method based on the grayscale distribution statistical information comprises the steps of selecting a certain number of a position dot pair collection according to an image size and a feature of a certain type of samples, then confirming mutual relations among position spot pairs to gray level average value in a field of all sample calculation position points of the type according to two gray level average values of the position spot pairs in the samples, confirming reliability and relevancy of mutual relations of the position dot pairs between the same type of samples and the type of samples and other types of samples, and finally selecting parts of position dot pairs which are high in reliability and small in relevancy and mutual relations of the position dot pairs from the initial position dot pair collection, wherein the position dot pairs and the mutual relations of the position dot pairs are used as character representations of the types of the samples. The image sample library feature representing method based on the grayscale distribution statistical information is especially suitable for an image sample library such as a feature extraction and representation of auto logos and road signs, wherein the image sample library is low in resolution, and an image structure feature of the image sample library is obvious,.

The invention provides a malicious software family classifier generation method and device based on weak coupling SGAN, and a readable storage medium, which are used for adapting to family classification model training of malicious software with a part of family-label-free malicious software, and finally determining that to-be-detected software belongs to a certain type of malicious software family. According to the method, a function of extracting original graphic features of malicious software is realized through a binary file of the malicious software in combination with an improved malicious software image scaling algorithm, an original malicious software family classifier is trained by utilizing 1D-CNN of a VGG model and the malicious software with family tags, then a weakly coupled semi-supervised generative adversarial network model is adopted, a malware family classifier, a research and judgment device in a semi-supervised generative adversarial network and a generator are trained by utilizing unlabeled malware, and finally the malware family classifier with a wider application range is obtained. The method has a good effect on classification of malicious software with unknown family tags or inaccurate family tags.

The invention provides an industrial audio fault monitoring system and method based on a deep neural network. The feature extraction efficiency is improved by selecting and constructing an industrial audio feature set; a deep learning model is introduced as a classifier, so that the accuracy of fault analysis in the field of industrial audio analysis is improved; the deep learning classification model is trained through the normal audio and the abnormal audio generated during the operation of the industrial equipment, so that the burden of manual decision making is reduced, the accuracy of judging the industrial audio fault probability is improved, and the functions of monitoring the audio fault of the industrial equipment in real time and performing fault early warning in a complex environment are realized. The system has the functions of online real-time monitoring, early warning and the like, has the advantages of low deployment cost, high function integration degree and high fault recognition rate, and has the capability of wide popularization.

The invention discloses a voice feature matching method based on a convolutional neural network. The method comprises the steps of S1, carrying out preprocessing, extracting Mel spectrograms in audiosignals, cutting the Mel spectrograms into image segments in a time domain, carrying out Fourier transform in the image segments to obtain spectrum signals, and extracting feature vectors; S2, arranging the feature vectors of audio samples according to time sequences, carrying out pooling to form voice record files, and converting the voice record files into binary feature sequences; and S3, carrying out voice feature matching, comparing voice query files with the voice record files, and searching the voice record files which have the same content as the voice query files. According to the method, a voice recognition accuracy rate is improved, complexity of a voice recognition system is reduced, and software robustness is improved.

The invention provides a high and low frequency interleaved edge feature enhancement method suitable for pedestrian target detection and a method for constructing an enhancement network, and belongs to the technical field of target detection. The method is characterized by comprising the following steps: S1, selecting a convolution module to perform dimension transformation, adjusting the scale ofa feature map, and extracting high and low frequency feature components according to a frequency distribution coefficient; S2, fusing the output high-frequency component with the low-frequency component through a pooling and convolution module; S3, fusing the output low-frequency component with the high-frequency component through a convolution and up-sampling module; and S4, returning the outputhigh-frequency and low-frequency fusion components to the original feature scale through deconvolution, and outputting feature fusion information under the combined action. The method has the advantages that the method can serve as an independent unit to be embedded into a deep neural network pedestrian target detection system, edge contour feature information of pedestrian targets can be remarkably enhanced, and detection precision is improved.

The embodiment of the invention discloses a feature extraction method and device, computer equipment and a storage medium, and belongs to the technical field of computers. The method comprises the following steps: obtaining a data configuration file, obtaining a feature extraction framework, calling the feature extraction framework based on the data configuration file, and executing the followingsteps: determining an operator matched with each association relationship type according to the association relationship type and the matching relationship between at least two first data tables, andcalling the operator, and performing feature extraction on the at least two first data tables to obtain feature information of the plurality of objects. According to the method provided by the embodiment of the invention, a universal feature extraction framework is provided, feature information of a plurality of objects contained in the data configuration file is automatically extracted through the feature extraction framework, a developer does not need to develop feature extraction codes for a network model, the time consumed for developing the feature extraction codes is reduced, therefore,the feature extraction efficiency is improved, and the data calculation of the data configuration file is realized.

The invention relates to an operation code frequency-based malicious code visual analysis method. The method comprises the following steps of extracting operation code character sequences of maliciouscodes, and converting the operation code character sequences into constant sequences; performing normalization operation on the obtained constant sequences; converting the normalized constant sequences into RGB values to obtain RGB color sequences; rearranging the RGB color sequences according to a specified sequence and then filling pictures; and according to the obtained pictures, performing visual analysis. The method can improve the analysis efficiency and is suitable for similarity comparison of malicious samples.