50results about How to "Achieve high-precision recognition" patented technology

The invention discloses a terahertz polarization real-time imaging method, which comprises the following steps of: (1) putting samples to be tested in a terahertz polarization real-time imaging device; (2) continuously changing the optical path difference between pump light and detection light, and measuring a polarized component Ex of a terahertz pulse passing through the samples; (3) changing the polarization state of the detection light, continuously changing the optical path difference between the pump light and the detection light again, and measuring another polarized component Ey of the terahertz pulse; and (4) processing two groups of terahertz polarized components of the samples, and forming relative-intensity images. The terahertz polarization real-time imaging method can rapidly and accurately measure different polarized components of terahertz light waves, highly precisely identifies target objects, and has wide application range.

The invention relates to an early-stage automatic winter wheat drawing method based on Sentinel-2 image data. The method includes following steps of obtaining Sentinel-2 image data of early winter wheat in the study area; performing atmospheric correction and cloud mask processing on the data; calculating and generating NDVI vegetation index; combining the prior knowledge of winter wheat phenologyto generate phenological characteristics of multiple winter wheat targeting; establishing the classification rules of winter wheat by using the generated phenological characteristics. The distribution results of winter wheat in the study area and the corresponding area monitoring information are obtained by using the classification rules. According to the method, winter wheat phenological characteristics are used for constructing winter wheat-targeted phenological characteristics, high-precision winter wheat recognition independent of ground data can be achieved, and a refined winter wheat drawing result with the spatial resolution of 10 meters can be obtained 5 months before winter wheat is harvested.

The invention relates to a 3D point cloud recognition method based on B-spline surface similarity detection, which is characterized in that the method comprises the following steps of1 modeling and fitting the three-dimensional point cloud by using B-spline surface equation; 2 sampling that fitting B-spline surface by adopt different parameter sampling methods; 3 calculating a similar point by defining a geometric local characteristic of a description point, and carry out point-to-point matching; 4 calculating the equidistant distance between the features of the calculated point pairs; 5 classifying that equidistant transformation types of two point in a point pair; 6 comparing equidistant point pairs after equidistant classification; 7 applying that spectral clustering algorithm to obtainthe local correspond similar parts among the objects, and completing the detection of the similar area; 8 judging whether the proportion of the similarity point cloud and the sampling point cloud islarger than the set threshold value to complete the identification of the whole point cloud. The method of the invention realizes efficient and accurate object recognition.

The invention provides a target detection and tracking method based on a plurality of irregular ROI. The method comprises the following steps: 1) selecting a plurality of irregular regions of interest(ROI) in collected original images; 2) carrying out set union operation on the plurality of regions of interest (ROI) to obtain a fused region to be detected; 3) carrying out intersection operation on the region to be detected and a blank image to obtain a mask image; 4) carrying out intersection operation on the original images and the mask image to obtain a synthesized image; and 5) carrying out identification, target detection and tracking on the region to be detected of the synthesized image. By selecting different shapes of ROI regions of different original images in a customized mannerand carrying out processing on the regions, the new region is allowed to be close to the shape of detection portion to the maximum degree, a lot of redundant areas are removed and image processing efficiency is improved; and by carrying out fusion on a plurality of regions in different sizes and shapes, high-precision identification, detection and tracking are realized.

The invention discloses a face attribute analysis method based on transfer learning, belonging to the technical field of calculation and calculation, in particular to the technical field of computer vision for recognizing the face attribute. The invention combines training sample sets on a multi-attribute prediction network to predict feature attributes, migrating the convergent multi-attribute prediction network to the master attribute prediction network, the master attribute prediction network is trained and the parameters are fine-tuned until the loss function of the master attribute prediction network converges, The main attributes include but are not limited to the face attributes based on logical regression and the main attributes based on linear regression, which not only prevents the local minima, but also avoids the precision degradation caused by the excessive complexity of tasks, and is more accurate and flexible in practical application.

The invention discloses a concrete crack width measuring method based on U-net CNN image recognition and pixel calibration. The method comprises the following steps: using a single lens reflex and a laser range finder to obtain a plurality of pictures of concrete structure surface cracks with known shooting distances and focal lengths; training a U-net CNN network to realize automatic identification of cracks and extract crack forms; marking the crack width direction, and extracting the number of pixels in the crack width direction; carrying out an indoor target experiment, constructing a nonlinear model of the shooting distance, the focal length and the pixel actual size, and calibrating the pixel actual size; and calculating the crack width in combination with the pixel size in the crackwidth direction and the calibrated pixel actual size. Compared with the prior art, the method is novel, intelligent, efficient and accurate in width parameter identification.

The invention discloses an animal long-time posture recognition system based on an infrared image and a wearable optical fiber. The system comprises an attention mechanism module M1, a local posture recognition model training module M2, a local posture recognition prediction module M3, and a global posture restoration module M4. According to the invention, high-precision identification of the keynode of the animal behavioral infrared low-resolution video wearing the optical fiber equipment is realized, the time length of manual verification is greatly reduced, the influence of animal curling,rotation, shooting illumination fluctuation and the like is solved, and the system can be widely applied to the field of brain imaging.

The invention discloses a point cloud registration method based on a convolution restricted Boltzmann machine. The multi-angle point cloud image of the object can be recognized, and the pose of the object is obtained. The method comprises the following steps: firstly, obtaining each basic angle of a to-be-identified object for scanning to obtain point cloud image data of the to-be-identified object, and preprocessing the point cloud image data; then, carrying out convolution calculation on the model by using a convolutional neural network to obtain a relatively robust node; constructing a restricted Boltzmann machine training model, determining each parameter of the model, inputting a result obtained by the convolutional network into the model for training, and obtaining optimized bias vectors a and b of the restricted Boltzmann machine and a weight matrix W; and carrying out reverse calculation according to the parameters to obtain a visible layer, wherein the feature vector with thehighest score is the feature vector of the object most possibly corresponding to the point cloud image, so that the object is identified.

The invention relates to a spectrum matching method, and is suitable for the field of all kinds of spectral analyses. The spectrum matching method comprises the fields of infrared / near infrared spectrum analyses, hyperspectral remote sensing, micro-imaging spectral analyses and the like and the steps that firstly a tested matter spectrum is obtained, and then the tested matter spectrum is matched with a known matter spectrum of a matter materiel spectral library. The specific steps comprise that firstly a feature-enhanced factor matrix is obtained, therefore a spectral feature-enhanced space is obtained, then spectral data are projected to the feature-enhanced space, and feature-enhanced spectrum matching is conducted. The feature-enhanced space is obtained by using a principal component analysis and space projection, substantial features of the spectrum is highlighted, noise is suppressed, the accuracy and adaptability of the spectrum matching are improved, and therefore high-accuracy recognition of matter kinds is achieved.

The embodiment of the invention provides an in-situ wetting angle measuring device and a wetting angle determining method based on deep learning. The in-situ wetting angle measuring device comprises oil-water pretreatment equipment and wetting angle measuring equipment, and the method is a wetting angle determining method based on deep learning. Particles with different particle sizes and properties are placed in the oil-water pretreatment equipment to simulate various porous media and used for simulating full contact reaction of crude oil, water and rock, so that crude oil components are adsorbed on the surface of the rock and the mass transfer effect between oil and water is fully generated to obtain an oil reservoir in-situ fluid, and the actual condition of an oil reservoir is simulated to the maximum extent. The wetting angle measurement result is more accurate; furthermore, the convolutional deep learning network is introduced into wetting liquid form capture and simulation, high-precision recognition of the wetting angle is achieved, the problem that the calculation error of the wetting angle is large due to human factors and formula applicability differences is avoided, and the measurement precision of the wetting angle is improved.

The invention relates to a construction method of a user position identification model and a user position identification method. The construction method of the user position identification model comprises the following steps: sample user characteristic data, sample user real position data corresponding to the sample user characteristic data, test user characteristic data and test user real position data corresponding to the test user characteristic data; training a preset neural network model according to the sample user feature data and the sample user real position data to obtain a trainedneural network model; inputting the test user feature data into a trained neural network model to obtain a user position recognition result output by the trained neural network model; and screening the trained neural network model according to the user position identification result and the real position data of the test user, and taking the screened trained neural network model as a user positionidentification model. High-precision identification of the user position can be realized through the user position identification model.

The invention relates to an age identification method and device, electronic equipment and a storage medium. The method comprises the following steps of performing initial age identification on a target object in a first target image to obtain a first identification age of the target object in the first target image, acquiring height information of a target object in the first target image, and determining a second identification age of a target object in the first target image according to a matching condition between the first identification age and the height information in combination with the first target image. According to the embodiment of the invention, the age identification accuracy of specific objects such as children or teenagers can be improved, and high-precision age identification is realized.

The invention relates to a method for generating a target detection football candidate point of a Nao robot based on a Heatmap. The method comprises the steps of selecting a convolutional neural network as a target detection model; simulating competition environments, collecting a plurality of groups of pictures to make a data set for training and testing, generating a Heatmap, processing to obtain a Heatmap visualization result, reconstructing the convolutional neural network to accelerate the network calculation speed, setting a proper threshold value, taking points greater than the set threshold value in the Heatmap as candidate points of a ball, and finally sending the candidate points into a classifier to obtain a final accurate identification result. According to the method, the adaptive capacity of the Nao robot vision system to the light environment of the competition field is enhanced; high-precision recognition of the football can be achieved in different light environments;feature extraction is completed through few convolution layers; the recognition real-time performance is guaranteed; and meanwhile the football recognition accuracy is greatly improved through the method that football candidate points are generated and then enter the classifier to be recognized.

The invention discloses a portable extra-large-breadth high-precision scanning device. The characteristics lie in that the portable extra-large-breadth high-precision scanning device comprises a scanner master machine, a scanner slave machine, roadways and a stepper motor, the scanner master machine and the stepper motor are connected and arranged on the roadways, and the scanner master machine transmits acquired image information to the scanner slave machine which is wirelessly connected with the scanner master machine. The device disclosed by the invention selects a portable handheld scanner structure, adopts a block image acquisition mode, combines an image splicing technology, adopts a constant current source to drive a sensor five-color light source and adopts the stepper motor to drive transverse constant-speed scanning, thereby realizing portable large-breadth high-precision image acquisition.

The embodiment of the invention provides a method for constructing a finger vein recognition system and the finger vein recognition system, and the method comprises the steps: A1, constructing an initial finger vein recognition system which comprises a preprocessing module, a feature extraction module employing a non-finger vein data set for pre-training, and a domain algorithm module; a2, obtaining finger vein grey-scale maps corresponding to a plurality of fingers, and training a preprocessing module to preprocess the finger vein grey-scale maps so as to extract finger vein vein maps of thefingers; a3, training the feature extraction module to extract a finger vein feature vector of the finger according to the finger vein map transfer learning, wherein training data in a triple form generated based on the finger vein map is adopted as the training data; and a4, training the domain algorithm module to establish different recognition areas for different fingers according to the fingervein feature vectors, and analyzing the similarity between the target finger and the finger corresponding to each recognition area according to the finger vein feature vectors.

The invention discloses an automatic tobacco marketing RPA method and system. The automatic tobacco marketing RPA system comprises an image acquisition device, an image processing subsystem, a task processing subsystem and an evaluation management subsystem. The image acquisition device acquires image data related to the retail terminal; the image processing subsystem analyzes and processes the image data and identifies image content; the task processing subsystem identifies a task type according to the image content and the acquisition information corresponding to the image data and performstask allocation; and the evaluation management subsystem generates corresponding evaluation data according to task execution results under different task types and predetermined rules of marketing activities. By the adoption of the automatic tobacco marketing RPA method and system, automatic management of tobacco marketing can be achieved, and intelligent quantitative evaluation of marketing is achieved, and the reliability of activity evaluation is guaranteed.

The invention discloses a remote sensing detection system for detecting field forest and grassland fire. The remote sensing detection system comprises a transmitting system, a receiving system, a spectral analysis system and a data acquisition processor; the transmitting system comprises a laser; the receiving system comprises a telescope; the spectral analysis system comprises a collimating mirror and a dichroic mirror which are sequentially arranged on a light path of a light beam emitted by the telescope; a first narrow-band filter and a first converging mirror are sequentially arranged on a transmission light path of the dichroic mirror, the first converging mirror converges the light to a first detector; a second narrow-band filter and a second converging mirror are sequentially arranged on a reflected light path of the dichroic mirror, the second converging mirror converges the light to a second detector; and the first photoelectric detector and the second photoelectric detector are both in signal connection with the data acquisition processor. The invention also discloses a remote sensing detection method for detecting field forest and grassland fire. The method has the characteristics of high detection sensitivity and large detection range.

The invention discloses a coal-rock interface recognition system. The coal-rock interface recognition system comprises a detection module, a simulation test bench module, a visual image module and a data image module, and the detection module is used for recognizing a coal-rock mining interface on the basis of machine vision; the simulation test bench module comprises a physical simulation unit and a working face laboratory unit; the visual image module is used for extracting image information in a stratum medium; and the data image module comprises an analysis unit. According to the coal-rock interface recognition system, a coal-rock visual image information fuzzy set, a feature database and a rule base can be established, boundary conditions of high-resolution imaging of a complex geometrical shape heteroplasmon are determined, and coal-rock three-dimensional self-correlation back projection imaging is formed; and an image recovery unit is used for forming a surface layer visual detection three-dimensional geological distinguishing and pre-judging model of a coal face, texture features are constructed by using different scale decomposition coefficients, comparative analysis is carried out on coal-rock images, and high-precision recognition of a coal-rock interface is realized.

The invention provides a visual and auditory aesthetics evaluation method and system based on electroencephalogram signals, and relates to the technical field of brain cognitive neural functions. The visual and auditory aesthetics evaluation method comprises the following steps: collecting original electroencephalogram signals of a subject during aesthetic activities, and obtaining EEG original data of the subject; preprocessing the obtained EEG original data to obtain pure EEG signals with the same duration under different conditions; calculating energy density spectrums of five different frequency bands of delta, theta, alpha, beta and gamma through short-time Fourier transform, and obtaining energy density spectrum feature matrixes of the different frequency bands; and screening an optimal feature by using a recursive feature based on a support vector machine to perform classifier training, performing classification identification on aesthetic subjective judgment of the participant, obtaining a subjective judgment result of the participant, and feeding back the subjective judgment result to the participant through sound and picture reminding. According to the method, high-precision identification of aesthetic judgment of participants can be realized, objective evaluation of visual and auditory aesthetic can be realized, and thoughts are developed and a technical foundation is laid for fairness of aesthetic evaluation.

The invention discloses a space-spectrum information combined satelliteborne hyperspectral image segmentation and clustering method, which comprises the following steps of: preprocessing an acquired satelliteborne hyperspectral image, and removing a wave band with a serious radiation quality problem and a wave band with a serious atmospheric absorption influence to obtain a reflectivity image; judging the complexity level of the image, and determining segmentation parameters according to the complexity level; selecting part of wave bands or all wave bands in the reflectivity image to carry out principal component transformation, retaining the first N principal components after transformation, and carrying out normalization processing; performing spatial dimension filtering processing on the image; principal component transformation is carried out on the reflectivity image after spatial dimension filtering, and the first N principal components after transformation are reserved and normalized; obtaining an initial pattern spot segmentation result and a pattern spot adjacency relation graph based on seed point region growth and image initial segmentation of SAD; splitting and combining the pattern spots; and calculating spectrum and texture features of the pattern spots, constructing a pattern spot spatial spectrum feature set, and clustering the pattern spots by using a K-means method to obtain an image clustering result image.

The invention discloses an intelligent processing feature recognition method based on point cloud semantic segmentation, and the method comprises the steps: converting a machining feature intelligentrecognition problem of a part CAD model into a semantic segmentation problem of a three-dimensional point cloud model, building a three-dimensional part CAD model library, and carrying out the preprocessing of the three-dimensional part CAD model library, and obtaining a training data set; furthermore, inputting the data set into an improved PointNet semantic segmentation network, enabling the network to take the PointNet semantic segmentation network as a basic framework, and improving the segmentation precision by introducing a residual block structure of a ResNet network; and finally, inputting the various machining features into a detection module, verifying the machining feature types through abnormal point detection and a DBSCAN clustering algorithm, and determining the number of themachining features. According to the intelligent process feature identification method, cross-part type and cross-feature type machining feature extraction is achieved, the problem that special-shaped machining features, free-form surfaces and composite machining features are difficult to recognize is solved, and system integration of full-life-cycle information of enterprise reuse machining features and manufacturing industry computer-aided software is facilitated.

The invention discloses a vehicle recognition method, a vehicle recognition model training method and related devices. The method comprises the steps: obtaining a picture with a vehicle; inputting thepictures into a trained vehicle recognition model so as to classify the vehicles; identifying the types of the vehicles, and determining the confidence coefficient of classifying the vehicles into dangerous vehicles; if the confidence coefficient of classifying the vehicle into the dangerous vehicle is smaller than a preset threshold value, dangerous goods label detection is conducted on the picture, and when it is detected that the picture comprises the dangerous goods label, it is determined that the vehicle is the dangerous vehicle. Through the above mode, high-precision recognition of thedangerous goods vehicle can be more effectively realized through a mode of cascading and matching the recognition detection of the vehicle type with the detection of the dangerous goods label, and the accuracy of the recognition of the dangerous goods vehicle is effectively improved.

The invention discloses a quantitative analysis method of metal phase volume fraction gradient distribution of a cutting surface layer of a duplex titanium alloy. The method comprises: processing a titanium alloy cut sample into a test piece through wire-electrode cutting and inset, polishing and corroding the test piece to obtain a surface to be detected, scanning the surface to be detected through a scanning electron microscope to obtain an electron microscope image, converting the electron microscope image into a gray image through gray scale processing, trimming the electron microscope image to obtain an effective electron microscope image, dividing the effective electron microscope image into a phase region alpha and a phase region beta according to gray scales, carrying out binarization processing on the effective electron microscope image, carrying out dissociation layering along the cutting depth direction, counting the number of phase alpha pixels and phase beta pixels in eachlayered image, wherein the percentage of the phase alpha pixels and phase beta pixels in total pixels in each layered image is a volume fraction of the phase alpha and the phase beta in layered images, and acquiring the characteristic of the metal phase volume fraction gradient distribution of the side profile of a cut test sample according to the volume fractions of the phase alpha and phase beta in each layered image.

The invention discloses a human body behavior recognition system based on a graph convolutional neural network, and relates to the technical field of human body behavior recognition. The method comprises the following steps: 1, constructing an undirected space-time skeleton diagram for human joint data collected by a depth sensor, and taking the undirected space-time skeleton diagram as an input signal of space-time diagram convolution; 2, inputting the constructed skeleton graph into a space-time graph convolutional network for action feature extraction to realize human body action recognition; 3, embedding the gating unit recurrent neural network into space-time diagram convolution to optimize the network, and better realizing synchronous extraction of space domain features and time domain features; 4, realizing man-machine interaction in a virtual environment by utilizing 3D modeling software. According to the method, human skeleton action information collected in an NTU RGB + D data set is utilized, a space-time skeleton graph network structure is constructed for time sequence representation of human skeleton joint point positions and a space cooperation relation, and end-to-end human skeleton action recognition based on a space-time diagram convolutional neural network is achieved.

The invention relates to an intelligent security check method based on suspected dangerous goods evaluation, and belongs to the technical field of security check, and the method comprises the following steps: obtaining the volume data of a packaged object through security check CT equipment, carrying out the feature value extraction of a segmented independent object, and obtaining a dangerous goods similarity rate vector of the object; and acquiring a dangerous goods classification identification probability vector of the goods by using a dangerous goods identification module, further performing dangerous goods comprehensive evaluation on the goods, and judging whether to give an alarm based on an evaluation result. According to the method, suspected dangerous objects are comprehensively recognized, missing detection events can be reduced, the security check efficiency can be greatly improved while the security check quality is ensured, and an effective guarantee is provided for application scenes needing high-precision security check.

The invention relates to a chaotic microwave breast tumor imaging classification device and method based on deep learning. According to the method, a plurality of lasers with different wavelengths areinjected into a multi-longitudinal-mode laser to generate a plurality of quasi-orthogonal chaotic ultra-wideband signals, a neural network for a proposed double-classification model is built througha deep learning algorithm, and an image classification effect model is realized. In supervised learning, a data set with existing attributes and marks is provided for a neural network, a training model capable of classifying images is established by initializing weights, performing gradient descent, checking correctness of gradient descent and the like, the input images are judged, and tumor recognition and benign and bad classification are performed by using a deep learning double-classifier model. The chaotic microwave photon radar with excellent characteristics is combined with deep learning, high-precision identification of breast tumors is realized, and the problems of complexity and misjudgment of intuitive tumor judgment of doctors are solved.

An SSVEP multi-scale noise transfer and characteristic frequency detection method based on FHN-CCA fusion comprises the steps that firstly, multi-channel SSVEP data collection is carried out, then an optimal channel X is determined through a pre-experiment and serves as an output signal of a formal experiment, then signal preprocessing is carried out, useless data in the first 0.1 s are cut off, and useless noise is filtered out through a Butterworth filter; then FHN model parameter initialization and FHN model processing are carried out, preprocessed signals and noise are sent to an FHN model for stochastic resonance processing, target frequency is identified based on a CCA method of spatial filtering and template matching, and finally frequency matching detection is carried out; according to the method, high-precision identification of the characteristic frequency is realized, and the identification accuracy and the information transmission rate of the SSVEP target frequency are greatly increased.

The invention discloses a steel rail crawling online monitoring terminal, system and method. The monitoring terminal comprises a camera assembly, an image processing unit, a signal acquisition unit, a communication unit, a temperature measurement sensor and a light source. The camera assembly shoots an image containing a measurement identifier, and the image processing unit calculates position information of the measurement identifier according to the image shot by the camera assembly and then sends the position information to the signal acquisition unit. The light source is used for supplementing light for image shooting of the measurement identifier through triggering of the signal acquisition unit under a weak light condition. And the signal acquisition unit simultaneously acquires a steel rail temperature signal measured by the temperature measurement sensor, and sends the steel rail temperature signal and the position information of the measurement identifier to the outside through the communication unit. The technical problems that an existing line steel rail state monitoring system is installed and measured in a contact mode, potential safety hazards to a line are large, and maintenance is difficult can be solved.