42results about How to "Improve feature extraction accuracy" patented technology

The invention discloses a signal feature extraction method used for a distributed optical fiber vibration sensing system. The method mainly comprises the steps of in an improved ensemble empirical mode decomposition (MEEMD) processing process, reading original data, and performing vibration signal locating and phase demodulation; introducing two groups of white noises with a mean value of zero toperform EMD; performing permutation entropy calculation for a first IMF component; comparing an entropy value with a set threshold, and if the entropy value is higher than the set threshold, repeatingthe steps until the entropy value is lower than the threshold; performing EMD on residual data to obtain residual IMF components of a vibration signal; and performing Hilbert analysis on the IMF components to obtain an eigenvector of vibration signal mode identification. By applying the method provided by the invention, the problems of mode mixing, false components and the like in a conventionaldecomposition method can be solved; the processing process is simplified; the reconstruction precision is improved; the data processing time is shortened; and the method is of important significance for improving the mode identification timeliness and accuracy of the distributed optical fiber vibration sensing system.

The invention relates to an image processing method and device, electronic equipment and a storage medium. The method comprises the steps: obtaining an iris image group which comprises at least two iris images to be compared; detecting an iris position in the iris image and a segmentation result of an iris region in the iris image; performing multi-scale feature extraction and multi-scale featurefusion processing on an image region corresponding to the iris position to obtain an iris feature map corresponding to the iris image; and executing comparison processing by using the segmentation results and the iris feature maps corresponding to the at least two iris images, and determining whether the at least two iris images correspond to the same object based on a comparison result of the comparison processing. According to the embodiment of the invention, accurate comparison of iris images can be realized.

The invention discloses an EKF-SLAM (Extended Kalman Filter-Simultaneous Localization And Mapping) algorithm based on a straight line feature map, which is characterized in that straight line features in an environment are extracted by using laser radar, and a feature map is established according to the straight line features. The EKF-SLAM algorithm taking a straight line as the feature map is creatively designed. In an experimental environment with obvious straight line features, straight lines features of the environment are extracted very easily. The complexity of the algorithm is low, the operation overhead of a processor is reduced, the real-time performance of the system is improved, and the environment processing means of the SLAM algorithm is enriched.

The invention discloses a foggy day vehicle detection method based on deep learning. The method comprises the following steps: carrying out image preprocessing on an acquired foggy traffic vehicle picture; carrying out feature extraction on the preprocessed foggy traffic vehicle picture by adopting a deep residual network model; obtaining a plurality of feature maps with different sizes, carryingout multi-scale detection on a plurality of feature maps with different sizes to obtain a multi-scale detection feature map and improve feature extraction accuracy; and finally, training the deep residual error network model by adopting a transfer learning method according to the obtained multi-scale detection feature map to obtain a vehicle detection network model in foggy days. A network structure is simplified by adopting a transfer learning method; the method is advantaged in that not only can detection speed be improved, but also target detection precision is improved, clustering is carried out by utilizing the K-means clustering method, the size of the initial prior frame required by the network is acquired, deepening of the shallow network and simplification of the integral framework are carried out, detection speed is improved, the loss function and the predicted output tensor are simplified, and positioning efficiency is improved.

The invention belongs to the field of behavior recognition, particularly relates to a behavior feature extraction method, system based on space-time frequency domain hybrid learning, and a device, andaims to solve the problem of low skeleton behavior feature extraction precision. The method comprises the steps of obtaining a video behavior sequence based on a skeleton, and extracting a time-spacedomain behavior feature map through converting a network; inputting the time-space domain behavior feature map into a frequency domain attention network, performing frequency selection, inverting toa time-space domain, and adding the obtained behavior feature map to the time-space domain behavior feature map; synchronously performing local and non-local reasoning, and performing high-level localreasoning; and globally pooling the time-space domain behavior feature map obtained through reasoning to obtain the behavior feature vector of the video behavior sequence. The method can be applied to behavior classification, behavior detection and the like. According to the method, an effective frequency mode is adaptively selected in a frequency domain, a network with local affinity fields andnon-local affinity fields is adopted in a time-space domain for space-time reasoning, local details and non-local semantic information can be synchronously mined, and therefore the behavior recognition precision is effectively improved.

The invention belongs to the pantograph operating state monitoring field. In view of problems in the prior art, the present invention provides a line-scan camera-based pantograph image extraction method and device. Whether the working state of a pantograph is abnormal is monitored according to an extracted pantograph image, so that maintenance can be guided. According to the line-scan camera-basedpantograph image extraction method and device, feature detection and preliminary pantograph positioning are performed on an image to be detected according to an established pantograph classificationmodel; if a pantograph exists in the image to be detected, the pantograph is positioned again, a pantograph slide plate and a contact line position in the image to be detected can be judged; whether atrue pantograph exists in the image to be detected is detected according to the relative positions of the pantograph slide plate and the contact line; and if no true pantographs exist in the image tobe detected, preliminary pantograph positioning is performed on a next image to be detected.

The invention discloses a local feature extraction method based on deep learning, and the method comprises the following steps: carrying out the network training, training a pre-constructed network on an image data set MS-COCO, segmenting the data set into a training set and a verification set which respectively comprise 82783 and 40504 images, carrying out the image matching, in an experiment, using a standard local feature pipeline to evaluate the performance of the local feature extraction method, wherein the standard local feature pipeline is used for extracting and matching features from any given pair of images in an experiment, then carrying out repeated score (Repeatability) calculation, then carrying out matching score (M-Score) calculation, and finally performing homography estimation effect evaluation. The detection step is postponed to description, so that more stable key points are obtained, compared with a traditional non-machine learning mode, the method has a more flexible feature searching process, and feature extraction precision is improved while a large number of key points are obtained.

The embodiment of the invention relates to an MCASPP neural network fundus image optic cup and optic disc segmentation model based on an Attention mechanism. The model comprises a model body, the system comprises a feature extraction module, an attention mapping module, a multi-scale cavity convolution module and an output module. Extracting a first image feature in the input image through a feature extraction module; the attention mapping module is used for extracting a second image feature of the input image; and obtaining a first feature according to the advanced feature, the low-level feature and the second image feature in the first image feature, through the multi-scale cavity convolution module, the advanced features are subjected to multiple convolution to obtain the second features, and the output module obtains the prediction probability graph according to the first features and the second features, so that the feature extraction precision of the image segmentation network can be improved, and the technical problem of relatively low optic disc precision of optic cups segmented by a full convolution network in related technologies is avoided.

The invention aims to provide a pipe abnormality detecting method based on sample generation and interval Markov features and relates to the pipeline abnormality detection field. The pipe abnormalitydetecting method based on sample generation and interval Markov features comprises the following steps that firstly, extracting and filtering processing are conducted on historical data samples; secondly, regularization and down-sampling treatment are conducted on the historical data samples; thirdly, an ELM model is established and trained; fourthly, based on the ELM model, sample generation is conducted, and an expanded input sample set is obtained; fifthly, for each sample in the expanded input sample set, when t is larger than Q, the state of each moment is replaced by the average state inthe first Q time intervals, and the interval Markov features are extracted; sixthly, based on an SVM model or an RF model, pipeline abnormality is identified. By the adoption of the pipe abnormalitydetecting method based on sample generation and the interval Markov features, the technical problem that in the prior art, under complicated work conditions, weak pipeline leakage signals and work condition adjustment signals are difficult to identify is solved, and the precision of pipeline abnormality detection can be improved.

The invention provides an on-load tap-changer mechanical fault diagnosis method. The method comprises the steps of collecting vibration signals of an on-load tap-changer device at at least two positions in an operation state through an acceleration sensor, and carrying out the preprocessing of the vibration signals; carrying out GA-VMD decomposition on the at least two preprocessed channel vibration signals, and extracting an energy entropy and a fuzzy entropy to construct a feature vector to serve as observation value matrix input of a coupling hidden Markov model; inputting the feature vector obtained in the step 2 into a coupling hidden Markov model for training, and establishing a coupling hidden Markov model operation state model; and inputting the test set into the operation state model obtained in the step 3, and judging the state and the fault type of the on-load tap-changer according to the maximum likelihood probability. According to the on-load tap-changer mechanical fault diagnosis method provided by the invention, the diagnosis result is high in accuracy, the structure is simple, and the operability is high.

The invention discloses a material cage stacking method and device based on depth cameras, an electronic device and a system. The method comprises the following steps that 1, a first depth map is collected through a left depth camera; 2, a first feature at the left side of an upper material cage and a second feature at the left side of a lower material cage are extracted from the first depth map, and a position p1 and a position p2 of the first feature and the second feature relative to the left depth camera are calculated; 3, a second depth map is collected through a right depth camera; 4, a third feature at the right side of the upper material cage and a fourth feature at the right side of the lower material cage are extracted from the second depth map, and a position p3 and a position p4 of the third feature and the fourth feature relative to a right depth camera are calculated correspondingly; 5, the deviation distance and the deviation angle of the upper material cage relative to the lower material cage are calculated according to the position p1, the position P2, the position P3 and the position p4 and the pose conversion relation of the right camera relative to the left camera; and 6, a forklift AGV is controlled, so that the deviation distance and the deviation angle are close to 0, and stacking of the upper material cage and the lower material cage is completed.

The invention discloses an infant auxiliary nursing method, and relates to the field of infant nursing. The infant auxiliary nursing method comprises the steps of receiving training voices by a service center, wherein the training voices are composed of infant voices with state labels; based on training voices, training a depth learning network and obtaining an infant voice recognition network model; inputting a to-be-recognized infant voice into the infant voice recognition network model and obtaining a corresponding state label; sending the state label to a client and outputting the state label conveniently by the client. According to the method, the requirements of infants can be accurately recognized, and the nursing burdens of the nursing personnel of the infants can be relieved. Theinvention further discloses a service center, an infant auxiliary nursing device, an infant auxiliary nursing system and a readable storage medium.

The invention discloses an indoor article searching and positioning method for visually impaired people. The method comprises the following steps: S1, enabling the visually impaired people to input the name of a target article through a voice module and then collect images indoors through a binocular camera; s2, designing an adaptive sigmoid transfer algorithm (ASTF) based on a neural network, andcombining the ASTF with a Laplace operator to enhance the brightness of the acquired image and reduce the distortion degree; s3, designing a variable-scale convolutional neural network to convolve the images obtained in the step S2 to the same size; s4, designing a convolutional neural network fused with a multi-level attention mechanism to extract feature information of the image obtained in S3,and matching the feature information with feature data of a target object in a database; s5, if matching succeeds, obtaining the position of the target object, and outputing position information of the target object through a voice module; and if the matching is not successful, outputting the information that no information exists through the voice module. According to the invention, visually impaired people can be effectively helped to accurately search articles in a weak light environment.

The invention discloses a forensic identification report examination method and system, and relates to the field of semantic extraction and analysis. The method comprises the steps of: 1, after multidimensional features of a case forensic identification report in a database are extracted through NLP, using an HMM model of a multi-order Markov hypothesis based on word frequency distribution and synonym analysis for carrying out multidimensional feature segmentation and fitting, and obtaining and storing a feature model; 2, after a to-be-examined report is input, selecting an examination task option based on an interaction unit, and calling the feature model for format examination and damage part examination; and 3, generating and displaying a legal medical expert identification and examination report according to examination results of format examination and injury part examination. According to the method, the HMM model of the multi-order Markov hypothesis constructed by utilizing the characteristics associated with common term synonyms and contexts in forensic medicine is constructed, and the feature model is established, so that legal persons are helped to quickly and accurately examine and identify flaws and errors existing in reports.

The invention relates to an SLS-based metal surface crack detection system, which comprises a transmitting module, a receiving module, a three-dimensional mobile platform and a control module, and ischaracterized in that a detected workpiece is placed on the three-dimensional mobile platform; the transmitting module comprises a high-frequency optical fiber pulse ultrasonic laser, a scanning galvanometer mechanism and a gathering mirror, and the high-frequency optical fiber pulse ultrasonic laser excites a beam of pulse ultrasonic laser to be reflected by the scanning galvanometer mechanism and then excited on a measured workpiece through the gathering mirror; the receiving module comprises a plurality of piezoelectric sensors, a charge amplifier and a high-speed data acquisition card; thepiezoelectric sensors are arranged around the surface of a measured workpiece; signals acquired by the piezoelectric sensors are connected with the high-speed data acquisition card through the chargeamplifier to complete real-time acquisition of the signals. Compared with the prior art, the method has the advantage that the metal surface cracks can be accurately and efficiently detected withoutdamage.

The invention discloses an upper limb motion recognition method and system based on surface electromyogram signals. The method comprises the following steps of obtaining electromyogram signals of muscles of multiple positions of an upper limb; extracting an active segment signal in the electromyographic signal based on a framing energy method; performing feature extraction and feature fusion on the active segment signal according to the optimal feature fusion mode; and performing action recognition by adopting an SVM classifier according to the fused features. According to the method, by searching for the optimal feature combination mode, it is guaranteed that the needed features can be rapidly and accurately obtained and fused, and the efficiency and accuracy of action recognition are improved.

The embodiment of the invention discloses a recommendation method and device based on context and user long and short term preference adaptive learning, and belongs to the technical field of data analysis mining, and the method comprises the steps: carrying out the adaptive learning of a weight multi-type context fusion candidate article description matrix through a CNN network, and obtaining the features of candidate articles; according to the historical interaction article sequence of the target user, combining two factors of weight multi-type context and user long-term and short-term preferences, performing adaptive learning based on a GRU network, and obtaining preference features of the target user; and performing score prediction based on the user long-term and short-term preference weights of adaptive learning, determining a score prediction result, and performing article recommendation based on the score prediction result. According to the method and device, a more excellent recommendation effect can be obtained, and the predicted user score is more accurate, so that the recommendation accuracy is improved, and the user satisfaction is improved.

The invention discloses a GS-SSD traffic large-scale scene vehicle target rapid detection method comprising the following steps: 1, training and screening an optimal Gabor convolution kernel group, placing the optimal Gabor convolution kernel group in a shallow VGG16 network, using SIP to map advanced semantic information of a feature map deep layer to a feature map shallow layer, and finally building a novel GS-SSD network model in a Caffe environment; step 2, acquiring a training and testing image data set required for establishing a novel GS-SSD network model; 3, training and testing the built novel GS-SSD network model in the Caffe environment; and step 4, carrying out frame sampling on a video collected by a traffic intersection camera, inputting the obtained image into the tested novel GS-SSD network, carrying out target detection on the obtained image by using the tested novel GS-SSD network model, and finally outputting a detection result. The method is high in precision and detection speed.

The invention relates to the technical field of hyperspectral image feature extraction, in particular to a feature extraction method suitable for a hyperspectral image, which comprises the following steps: S1, preprocessing an acquired hyperspectral image to obtain a data set; S2, constructing a feature extraction model based on a least square support vector machine principle, and performing feature extraction on the data set; and S3, respectively obtaining spectral characteristics and spatial information in the data set. According to the feature extraction method suitable for the hyperspectral image, the hyperspectral image is preprocessed, the feature data to be extracted is preliminarily determined, secondary extraction is performed by using the constructed feature extraction model, andthree times of judgment and verification are performed in combination with the Bayesian strategy so as to ensure that the extracted features are obvious and unique identification points of the hyperspectral image; therefore, the accuracy of feature extraction is improved, the accuracy of spectral image recognition and analysis is also improved, and positive significance is brought to further research of spectral images.

The invention provides a decorrelation multi-frequency empirical mode decomposition method, which can be used for solving the mode aliasing phenomenon existing in the EMD method and causing the problem of low feature extraction accuracy. Firstly, the masking signals of multiple frequencies are added to the initial signal, and the signal components with different frequency ratios are preliminarilydecomposed to obtain multiple IMF components. Secondly, the correlation coefficients between adjacent IMFs are calculated and decoupled to further separate the aliased IMF and obtain the optimal IMF.Finally, the optimal IMF is subtracted from the original signal and the above steps are repeated until the residual component is constant or monotonic. Because the IMFs are independent of each other and do not interfere with each other, the phenomenon of modal aliasing is significantly reduced, and the feature extraction accuracy is effectively improved. The innovation of this method is that it combines the masking signal processing method with the correlation processing method, which makes it possible to decompose mixed signals with different frequency ratios adaptively, suppress modal aliasing phenomena and improve feature extraction accuracy.

The invention relates to a multiscale analysis method of factory building vibration, characterized by comprising the steps of 1), acquiring horizontal vibrational speed signals of a middle hood +Y ofa hydropower station building under different conditions; 2), using adaptive iterative filtering to decompose the horizontal vibrational speed signals of the middle hood +Y of the hydropower station building under different conditions so as to obtain a plurality of stable components corresponding to the horizontal vibration speed signal of the middle hood +Y of the hydropower station building under each condition; 3), using the multiple stable components corresponding to the horizontal vibration speed signal of the middle hood +Y of the hydropower station building under each condition to construct a standard characteristic vector of building vibration based on a percentile method, and using the standard characteristic vector to recognize heath state of building structure so as to finish multiscale analysis of building vibration. The method of the invention is widely applicable to the field of structural operation maintenance for hydropower station buildings.

The invention provides a feature extraction model training method and device, equipment, a storage medium and a computer program product. The method comprises the steps of performing feature extraction on a plurality of training samples carrying sample tags through a feature extraction model of at least one historical version obtained through training to obtain a plurality of sample features corresponding to each historical version; determining a sample similarity between any two training samples based on a plurality of sample features corresponding to each historical version; based on the determined sample similarity, at least one training sample pair is selected from the multiple training samples, and the training sample pair comprises a first training sample and a second training sample; training a target feature extraction model based on the selected at least one training sample pair to obtain a trained feature extraction model; according to the invention, the training efficiency of the feature extraction model can be improved, and the feature extraction precision of the feature extraction model obtained through training can be improved.

The invention discloses a photovoltaic power generation power prediction method based on deep learning. The photovoltaic power generation power prediction method comprises the following steps: A, acquiring photovoltaic power generation data and sending the photovoltaic power generation data to a memory for storage; B, performing feature extraction on the stored photovoltaic power generation data;C, encrypting the data subjected to feature extraction; D, using the encrypted data as input of a BP neural network, wherein the output of the BP neural network is to-be-predicted photovoltaic power generation power; And E, performing deep training on the BP neural network to obtain the photovoltaic power generation prediction power. The prediction method is high in precision and prediction rate,and the adopted data preprocessing method can realize data sorting, noise reduction and data filtering, so that the subsequent data processing efficiency is improved; According to the adopted featureextraction method, the first keyword and the second keyword are searched, so that the extraction difficulty can be reduced, and the feature extraction precision is improved; The adopted encryption method can perform multiple encryption on the photovoltaic data, so that the security and confidentiality of the data are improved.

The invention discloses an image feature extraction method, a computer readable storage medium and a computer terminal. The method comprises the following steps: acquiring a target image; and enabling the target image to be subjected to feature extraction through a feature extraction model, and obtaining target features of the target image, wherein the target features and the target image have rotation equivariability. The feature extraction model comprises a feature rotation equivariability layer and at least one feature fusion layer which are connected in sequence, and the resolutions of output results of the at least one feature fusion layer are different. Each feature fusion layer comprises a rotation equivariant sub-layer and a plurality of fusion sub-layers, and the structures of the plurality of fusion sub-layers are the same. The technical problem that the accuracy of feature extraction of the target image corresponding to the target with variable angles is low in the prior art is solved.

The invention discloses a repositioning method of a robot. The method comprises the following steps: acquiring a feature map comprising feature points, feature descriptors and historical poses of each historical frame image of the robot; when the robot positioning fails, acquiring a current frame image of an environment where the robot is located; inputting the current frame image into the convolutional neural network model for feature extraction to obtain feature points and feature descriptors of the current frame image; determining feature points and feature descriptors of a target historical frame image matched with the current frame image according to feature points and feature descriptors of each historical frame image and the current frame image; performing feature point matching according to the feature points and the feature descriptors of the target historical frame image and the current frame image to obtain a feature point corresponding relation; determining the actual pose of the robot according to the corresponding relation of the feature points and the historical pose of the robot in the target historical frame image. According to the method, the repositioning precision of the robot in an environment with illumination variation can be improved.

The invention relates to a multi-scale analysis method for powerhouse vibration, which is characterized in that it comprises the following steps: Step 1): Obtain the horizontal vibration velocity signals of the wind cover + Y in the middle layer of the powerhouse of a hydropower station under different working conditions; Step 2): adopt The self-adaptive iterative filtering method decomposes the horizontal vibration velocity signal of the wind cover + Y in the middle layer of the hydropower plant under each working condition, and obtains several stationary components corresponding to the horizontal vibration velocity signal of the wind cover + Y in the middle layer under each working condition ; Step 3): based on the percentile method, using several stationary components corresponding to the horizontal vibration velocity signal of the wind cover +Y of the middle layer of the hydropower plant building under each working condition, constructing the standard eigenvector of the plant vibration, which is used to identify the structure of the plant The multi-scale analysis of powerhouse vibration is completed, and the invention can be widely used in the field of operation and maintenance of powerhouse structures of hydropower stations.

The invention relates to the technical field of deep learning and artificial intelligence, and discloses a multi-scene-based image search test method, which comprises the following steps of: selecting data of the same category from a database to obtain an initial data group, and selecting a retrieval data group from the initial data group; loading data of the initial data set and the retrieval data set into a memory through a LOADDATA method in a test script; extracting features in the data of the retrieval data set through an improved RepVGG algorithm; comparing the features to obtain a test result; and outputting the test result to generate a test report. According to the method, the retesting efficiency is improved, the practicability of an original algorithm model in a specific scene is higher, and the feature extraction precision is improved. The invention further discloses a multi-scene-based image searching and testing device, a product and a storage medium.

Examples of the present invention involve a McASPP neural network eye -looking cup visual disk segmentation model based on the Attention mechanism.Among them, this model includes: feature extraction module, attention mapping module, multi -scale empty convolutional module, and output module. The first image feature of the input image is extracted through the feature extraction module.The second image features, and the first feature of the high -level characteristics, low -level characteristics, and the second image characteristics in the first image characteristics. Through multi -scale empty convolutional modules, the second features are obtained by multiple convolution of high -level features. The output module is output module.According to the first and second features, the predictive probability chart can improve the characteristics of the characteristics of the image segmentation network and avoid technical problems with lower accuracy of the visual cup with full convolutional network segmentation in related technologies.