111results about How to "Valid description" patented technology

The invention provides a continuous voice recognition method based on a deep long and short term memory recurrent neural network. According to the method, a noisy voice signal and an original pure voice signal are used as training samples, two deep long and short term memory recurrent neural network modules with the same structure are established, the difference between each deep long and short term memory layer of one module and the corresponding deep long and short term memory layer of the other module is obtained through cross entropy calculation, a cross entropy parameter is updated through a linear circulation projection layer, and a deep long and short term memory recurrent neural network acoustic model robust to environmental noise is finally obtained. By the adoption of the method, by establishing the deep long and short term memory recurrent neural network acoustic model, the voice recognition rate of the noisy voice signal is improved, the problem that because the scale of deep neutral network parameters is large, most of calculation work needs to be completed on a GPU is avoided, and the method has the advantages that the calculation complexity is low, and the convergence rate is high. The continuous voice recognition method based on the deep long and short term memory recurrent neural network can be widely applied to the multiple machine learning fields, such as speaker recognition, key word recognition and human-machine interaction, involving voice recognition.

An inter-frame prediction method in a hybrid video coding standard belongs to the field of video coding. In order to effectively process deformation movement existing in a video sequence, the present invention puts forward an inter-frame prediction method in a hybrid video coding standard for further improving coding performance of a video. The inter-frame prediction method comprises the steps of: obtaining motion information of a plurality of adjacent coded blocks around a current coding block; obtaining a reference index of each dividing unit in the current coding block according to obtained reference indexes of the adjacent coded blocked; and processing motion vectors of the adjacent coded blocks according to the obtained reference indexes of the adjacent coded blocks and the obtained reference index of each dividing unit in the current coding block, so as to obtain a motion vector of each dividing unit in the current coding block. According to the inter-frame prediction method of the present invention, motion information of the current coding block is predicted through the motion information of the adjacent coded blocks of the current coding block, so that deformation movement existing in the video sequence can be effectively described, and coding efficiency can be further improved.

The invention belongs to the field of computer vision and pattern recognition and discloses a moving target tracking method based on an improved multi-example learning algorithm. Firstly, a random measurement matrix is designed according to the compression perception theory. Then a multi-example learning algorithm is used to sample an example in a current tracking result small neighborhood to form a positive package, and at the same time, sampling an example is carried out in a large neighborhood ring to obtain a negative package. For each example, the characteristic of a character target is extracted at an image surface, and the random measurement matrix is utilized to carry out dimensionality reduction on the characteristic. According to the extracted example characteristic, online learning weak classifiers are utilized, and weak classifiers with strong discrimination ability are selected from a weak classification pool to form a strong classifier. Finally, when a new target position is tracked, according to a similarity score of the current tracking result and a target template, the online adaptive adjustment of classifier update degree parameters is carried out. According to the method, a problem that a tracking result in the existing algorithm is easily affected by an illumination change, an attitude change, the interference of a complex background, target fast motion and the like is solved.

The invention discloses a three-dimensional laser radar positioning and navigation method for intelligent inspection. The method comprises the steps of S1 acquiring original point cloud data of a to-be-inspected area through a laser radar, and establishing a global point cloud map; S2 converting the global point cloud map into a grid map, and performing global path planning according to preset patrol points to obtain an optimal patrol path; S3 estimating to obtain an initial position, obtaining initial point cloud data of the initial position, and matching the initial point cloud data with theglobal point cloud map to obtain an initial positioning value based on a global map coordinate system; and S4 through matching of the current frame point cloud data obtained in real time and the global point cloud map, obtaining a current inspection pose. As the method is realized by utilizing the three-dimensional laser radar, the functions of three-dimensional perception, autonomous path planning and positioning navigation of the environment can be well realized, and the inspection precision of the inspection robot is improved; GPS positioning navigation does not need to be adopted, and theanti-electromagnetic interference capacity is high.

The invention provides a medical image registration method based on combination of mutual information and gradient information, comprising steps of performing weighting on mutual information similarity and adding difference values of a gradient model into gradient similarity, performing space changing on a floating image to be registered and calculating similarity measure of a reference image and the changed floating image. The invention provides a similarity measure function of weighting mutual information and gradient information. The similarity measure can attain a maximization value through constantly optimizing search and changing space, and the registration is successful. The similarity measure function is employed in the process of the registration, which can effectively solve the problems that the medical image registration algorism is low in accuracy and the robustness is not strong.

The invention relates to an audio / video keyword identification method based on decision-making level fusion. The method mainly includes the following steps that (1) a keyword audio / video is recorded, a keyword and non-keyword voice acoustic feature vector sequence and a visual feature vector sequence are obtained, and accordingly a keyword and non-keyword acoustic template and a visual template are trained; (2) acoustic likelihood and visual likelihood are obtained according to the audio / video in different acoustic noise environments, so that the acoustic mode reliability, visual mode reliability and optimal weight are obtained, and accordingly an artificial neural network can be trained; (3) secondary parallel keyword identification based on the acoustic mode and the visual mode is conducted on the audio / video to be detected according to the acoustic template, the visual template and the artificial neural network. According to the audio / video keyword identification method based on decision-making level fusion, the acoustic function and the visual function are fused at a decision-making level, the secondary parallel keyword identification based on the dual modes is conducted on the audio / video to be detected, the contribution of visual information in the acoustic noise environment is fully utilized, and therefore identification performance is improved.

The invention relates to a micro-CT technology-based reservoir core three-dimensional entity model reconstruction method. The method includes the following steps that: a micro-CT technology is utilized to scan a reservoir core, so that a sectional scanning image of the reservoir core can be obtained, so that a CT image of a real reservoir core can be obtained, and a watershed algorithm is adopted to perform image segmentation on the CT image, so that the three-dimensional data of the image can be obtained; a Marching Cubes algorithm is adopted to generate a three-dimensional surface model of the reservoir core; and with the three-dimensional surface model of the reservoir core adopted as constraints, a constrained Delaunay tetrahedralization algorithm is utilized to generate the three-dimensional entity model of the reservoir core. According to the method of the invention, the three-dimensional entity model of the real reservoir core is established based on the sectional scanning image of the reservoir core, so that the obtained three-dimensional entity model of the reservoir core is more approximate to the structure of a real reservoir core. The method of the invention has the advantages of high accuracy and high efficiency. With the method adopted, defects of poor accuracy and low efficiency in the reconstruction of the three-dimensional entity model of a reservoir core in the prior art can be eliminated, and an effective guarantee can be provided for reservoir core simulation characteristic research.

The invention discloses a texture image classification method based on a BoF (bag of feature) and multi-feature fusion. The method comprises: performing local image selection on a texture image to form a fragment set; extracting GGCM (gray level-gradient co-occurrence matrix) features and SIFT (scale invariant feature transform) local features of all fragments, and performing importance weighted fusion on different features; generating a feature word for fusion feature clusters and performing preference and weighting on the words by using DWDPA (dynamic weighted discrimination power analysis), and assigning a fusion feature vector by using the preference and weighted word to form a training set fusion feature word bag model; computing the fusion feature vector of the texture image to be tested and acquiring a corresponding fusion feature word bag; and training the feature word bag model by using a SVM (support vector machine) as a classifier. The method effectively overcomes a defect that the GGCM is low in accuracy for large texture classification, compensates a weakness of information loss of the BoF feature space, and is more accurate and good in robustness.

An infrared and visible light video fusion method based on moving target detection belongs to the technical field of video fusion. A moving target is detected through low-rank sparse representation, the target is extracted using the low-rank and sparse time slices along X-T and Y-T planes, the integrity of the target detected is ensured by spatial information, and the robustness and accuracy of detection are improved. Then, the structure information of an image can be better described through non-subsample shearlet transform with a good fusion effect. The method has the advantages of high computational efficiency, unlimited filtering direction, translational invariance, and the like, and improves the fusion effect. In addition, the low-frequency sub-band coefficients are fused according toa sparse representation rule, the high-frequency sub-band coefficients are fused according to a weighting rule taking a target saliency map as a weighting factor, and a better fusion effect is achieved. Therefore, the infrared and visible light video fusion method based on moving target detection presented by the invention is more robust, and has a better fusion effect.

The invention provides a domain driving design-based MVVM design model (DDMVVM). An overall structure of the domain driving design-based MVVM design model is divided into four layers: a view layer, a view model layer, a domain model layer and an infrastructure layer. A domain driving design-based MVVM model method specifically comprises the following steps: designing at two aspects: strategic design and tactics design; developing a frame on the basis of a domain driving design-based MVVM model; realizing each layer of the domain driving design-based MVVM design model under a wechat mini-application development platform. The domain driving design-based MVVM design model is suitable for the field of software engineering and software architecture in a graph user interface program; the domain driving design-based MVVM design model can effectively guide and normalize the software developer to emphasize on the field of system core business; the complexity of the software system can be effectively reduced; the quality attributes, such as maintainability and expandability, of the software system can be effectively promoted.

The invention provides a method for judging typesetting directions of text regions, belonging to the OCR field. The method is characterized by carrying out statistic analysis according to the obtained projective histogram, finding out respective most representative characteristic data triples and judging the typesetting directions of the text regions with less than three character lines accordingto the length-width ratios of the external rectangles of the text regions; judging the typesetting directions of the text regions with not less than three character lines according to the number and statistical positions of abnormal projecting cylinders; judging the text typesetting directions which can not be judged by the above methods according to the first moment between normal projecting cylinders; judging the typesetting directions which can not be judged by the above methods according to the indent of the text characters; and giving up judging in the regions of which the typesetting directions still can not be judged at least. The method can accurately judge whether the normal text regions are horizontal or vertical and judge whether the text regions with low inclination angles or slight geometric distortion are horizontal or vertical and has good judgment effect, high speed and good application value.

The present invention discloses a cloud particle detection method and a cloud particle detector. According to the cloud particle detection method, the backscattering signals of the cloud particles are received to detect the depolarization condition of the cloud particles on laser so as to obtain the phase state of the cloud particles; and the forward scattering signals of the cloud particles are received to detect the equivalent optical particle size of the cloud particles. According to the present invention, with the cloud particle detection method and the cloud particle detector, the small ice crystal (less than 50 [mu]m) in the mixing phase state cloud can be effectively detected.

The invention discloses a recording terminal clustering method based on voice time-frequency transform features and integral linear programming. The method comprises the following steps: a, carrying out preprocessing on read recording; b, extracting GTCC features from each frame of voices; c, constructing a depth self-coding network with a bottleneck layer, and extracting bottleneck features; d, constructing gaussian supervector features; e, carrying out spectrogram feature extraction on the voice obtained after preprocessing in step a; f, splicing the spectrogram features in step e to the gaussian supervector in step d, thus obtaining voice time-frequency transform features describing the features of the recording terminal finally; and g, carrying out clustering on the voice time-frequency transform features of all the recording samples by adopting the integral linear programming algorithm, wherein the clustering result is taken as the final classification basis. The feature extraction and clustering of the method provided by the invention are unsupervised, and compared with the existing supervising method, the method provided by the invention has better universality.

The invention provides a target identification method based on quality evaluation. The target identification method based on quality evaluation includes the steps: constructing a target identificationmodel which includes a quality evaluation network, a feature extraction network, and a feature aggregation network, wherein the target identification model is used for extracting the target feature from a video so as to characterize the overall structural information and local information of the target; training the target identification model, and adjusting the parameters of the quality evaluation network and the feature extraction network during the training process so as to enable the target identification model to output the target feature according with the preset demand; and performingtarget identification on the video through the trained target identification model. Therefore, the target identification method based on quality evaluation solves the target identification problem caused by changeable appearance and irregular image quality in a video sequence, and adds the interframe correlation information in quality evaluation so as to obtain more effective target information toenable characterization of the target to be more accurate, thus improving the identification accuracy.

The invention provides a space cost distance calculating method based on the grid spatial data. The method inputs data representing space information through a ruled data format, the data describes the cost information of space access. Then on the input basis, the necessary smallest space cost distance value and path from an arbitrary position to another space position in the space is calculated. Through extending one grid value to nine values, on the basis of not changing the space position corresponding to the central grid, the cost relationship to the surrounding grids is intuitively described. The method can describe the passing distance differences between different directions in the hypsography areas; by using the calculating method, the conditions of mountainous areas, semi-closed roads, and one-way roads can be described and processed, therefore, the grid cost distance calculated result is closer to the fact, which meets the need of practical applications.

An opportunistic spectrum access method based on Markov chain and CSMA includes that the access predication statistical module method of the Markov chain is combined with the CSMA access protocol to adopt different access modes according to different induction areas, before the user access is recognized, the spectrum detection is carried out according to spectrum induction method, and the first and second dual thresholds Lambda 1 and Lambda 2 of the interference temperature of the master user, the judgment area whether the master user exists is divided according to the two thresholds, when the interference temperature value is less than the Lambda 1, the user is recognized and accessed directly; when the interference temperature value is more than the Lambda 2, the CSMA access mode is adopted for back-off access; when the interference temperature is between the Lambda 1 and Lambda 2, a Markov chain access predication statistical module with the master user as the priority is built in queuing and non-queuing modes, and the Markov chain module is adopted to describe the opportunistic spectrum access. During access process, the first and second interference temperature thresholds are regulated dynamically to realize the maximization of the whole network throughput.

The invention discloses a moving object segmenting method based on local space-time manifold learning. The method comprises the following steps: 1. segmentation of a space-time cube of an input video and extraction of an illuminant invariance feature; 2. establishment of a background model based on the local space-time manifold learning; and 3. segmentation of a moving object based on the local space-time manifold and on-line updating maintenance of a model. According to the moving object segmenting method based on the local space-time manifold learning, local space-time variation can be effectively described, the scale adaptability of an SIFT (scale invariant feature transform) feature point set in an extended image in the input video is treated, the problem that a dynamic background and illumination change can not be effectively eliminated when the moving object is segmented is solved, and the reliable and effective moving object can be provided for an intelligent monitoring platform.

The invention discloses an anti-occlusion particle filtering target tracking method based on the fusion of color features and local binary pattern features. The method comprises the following steps: carrying out target initialization: carrying out feature extraction on the color integral histogram and the local binary pattern integral histogram of an area of interest; calculating the feature determinacy coefficient of the color features and the local binary pattern features; according to different current target states, selecting different tracking methods; if the target state is normal, carrying out target tracking by a particle filtering method which fuses the color features and the local binary pattern features, carrying out target tracking by a blocking particle filtering method which fuses the color features and the local binary pattern features if the target state is partial occlusion, and predicting a target position by a least square method if the target state is serious occlusion; updating the current target state; when the target is under the normal state, updating the target; resampling particles; and carrying out particle propagation. By use of the anti-occlusion particle filtering target tracking method, the stability and the robustness of the target tracking under an occlusion situation can be improved.

The invention discloses a human body posture feature extracting method based on 3D joint point coordinates. Two features are extracted in total, one is the global feature based on the body part location (BPL), the other is the local feature based on the local joint structure (LJS), and the final feature is the feature formed after the two features are fused. The experiment proves that the extracted features can effectively describe human body postures, and the human body postures can be well recognized when the features are applied to a support vector machine classification model.

The invention provides a supervision cross-modal Hash search method based on a nonparametric Bayesian model, used for solving the technical problem that in the existing cross-modal Hash search method, the search accuracy is low. The method comprises the implementation steps of operating to acquire normalized training data and test data; classifying the normalized training data; operating to acquire three training data parameters of the normalized training data; operating to acquire the probability that normalized image training data and normalized text training data belong to the same class; operating to acquire the posteriori probability of the training data; operating to acquire a unified Hash code of the normalized image training data and the normalized text training data; operating to acquire the Hash code of the test data; operating to acquire a Hamming distance matrix of the Hash code of the test data and the unified Hash code of the normalized image training data and the normalized text training data; and operating to acquire a search result of the test data. The method provided by the invention is high in search accuracy, and can be applied to mutual search service for images and texts of mobile terminal equipment and the Internet of things.

The invention discloses an LBP (Local Binary Pattern) image and block encoding-based iris feature extracting method comprising the following steps of: firstly, adopting an LBP operator for a normalized iris image to obtain an LBP iris image; then, extracting an iris feature from the LBP image by using an iris statistical information based block encoding method; and finally, taking a Hamming distance as a classifier to obtain a recognition result. Parameters of the LBP operator and block encoding in the feature extracting method are obtained through training. Compared with the traditional feature extracting method, the LBP image and block encoding-based iris feature extracting method has higher recognition rate and strong robustness for illumination change.

The invention provides a network attack path prediction method based on an attacker characteristic index, and the method comprises the steps: firstly, providing a quantitative index of a network attack path from the perspective of an attacker in combination with an attack graph and a hidden Markov model, wherein the attack cost, the attack income and the attack profit reflect different optimal attack path selections of attackers with different intentions; secondly, an attack path in the attack graph being quantified and analyzed based on a quantitative index, and a network attack and defense scene being described more effectively; and finally, respectively adding the attack cost, the attack income and the attack profit of all vulnerabilities on each attack path to obtain the total attack cost, the total attack profit and the total attack profit of the whole attack path, and comparing the index values of the attack paths to obtain the attack cost, the total attack profit and the total attack profit of the whole attack path. Therefore, one or more attack paths which may be attacked by an attacker with relatively high risk can be found more accurately, a network administrator can be helped to know the network security condition more comprehensively, and the security of a network system can be ensured more efficiently.

The invention relates to an image processing method and device and an object detection method and system. The image processing method comprises a color classification step, a mask generation step and a feature extraction step, wherein the color classification step is used for classifying multiple pixels in an image region into at least one dominant color category; the mask generation step is used for generating a chromatic aberration mask for each dominant color category, and the chromatic aberration masks reflect the relations between the multiple pixels and the corresponding dominant color categories; the feature extraction step is used for extracting a feature vector from each corresponding chromatic aberration mask.

The invention discloses a visualization analysis method of a document hot point. The visualization analysis method mainly comprises the steps of 1) researching data acquisition and processing, particularly determining the field needed to be researched, crawling field periodical data, extracting a field needed to be researched, and processing an acquired keyword field; 2) researching subject detection; and 3) visually researching a subject and a subject network. Change is performed from a data source, a data acquisition method is changed, the field periodical is selected, data crawling is performed according to a set condition, a large of representative data relevant to research can be obtained, visualization is performed by a method combining a strategic map and the subject network, a result is analyzed from multiple angles, the accuracy of the analysis can be further improved, and effective analysis and explanation are expected to be performed on research trend in future.

The invention relates to a cross-scattering model-based polarization radar remote sensing image urban target decomposition method. The method includes the following steps that: a PolarSAR image single-view polarization scattering matrix is converted into a polarization covariance matrix or a polarization coherence matrix; a PolarSAR image pixel polarization azimuth is calculated on the basis of the polarization covariance matrix or the polarization coherence matrix; the cross-scattering model of a building is constructed on the basis of the polarization azimuth and a dihedral angle reflector;a polarization target decomposition solving equation is constructed according to a surface scattering model, an even-order scattering model, a volume scattering model, a spiral scattering model and the cross-scattering model, and the scattering coefficients of each model are solved; and the scattering weighting coefficients of each model are corrected, and surface scattering power, even-order scattering power, volume scattering power, spiral scattering power, and cross-scattering power are calculated on the basis of the scattering weighting coefficients of each model, and target decompositionis completed. With the cross-scattering model-based polarized radar remote sensing image urban target decomposition method of the present invention adopted, the HV scattering of the building can be effectively described.

The invention discloses a heterogeneous network community structure and a community discovery method based on the structure. A zero model of a heterogeneous network is provided, and a basic randomized network model is provided for research of a heterogeneous network structure in a complicated network; on this basis, modularity applicable to a function heterogeneous network of the heterogeneous network is provided, and a new method for evaluation of heterogeneous network communities is provided; on the basis of the heterogeneous network modularity, the heterogeneous network community structure is defined, and the heterogeneous network community discovery method based on maximization of the heterogeneous network modularity is provided on the basis of the community structure. The community structure in the heterogeneous network is discovered effectively. According to the structure and the method, through adoption of homogeneous and heterogeneous adjacent matrixes, a connection relationship of the heterogeneous network is effectively described.

The invention discloses a method for reducing the reconstruction distortion degree of a viewpoint in light field drawing, and the method comprises the following steps: 1) constructing a 3D scene witha shielding phenomenon; 2) carrying out quantitative description of shielding; 3) obtaining the camera position with the minimum shielding degree; and 4) reconstructing a new viewpoint at any positionand in any direction. According to the method, data acquisition and multi-view drawing in an actual complex scene can be realized; compensation of edge information loss in computational imaging and virtual reality caused by a shielding phenomenon can be solved, the method can even be applied to quantitative description of a three-dimensional scene shielding phenomenon of a 3D video technology, and the interactivity and the three-dimensional experience of the 3D video technology can be improved.

The invention provides a remote control robot for underwater precise measurement. Motion of four or more degrees of freedom of the underwater robot is achieved, the underwater robot has the capabilityof depth keeping hover, direct movement, rotation and lateral movement, and can maintain good communication with land control equipment, electronic components are all subjected to watertight pressure-resisting protection, and the remote control robot is suitable for passing through an underwater narrow area to conduct underwater detection and underwater measurement operation; the precise measurement underwater robot specially plans a path to improve the cruise and measurement efficiency; a multi-beam depth finder and a binocular precision measuring instrument are utilized to synergistically carry out underwater measurement, high-precision and high-density measurement of underwater detection targets is achieved through a multi-beam depth measuring technology, binocular precision measurement helps the underwater robot to execute certain complex underwater operation tasks, and two measurement methods synergistically complete high-quality underwater precision measurement; and the relativeerror of underwater measurement is small, and the operating requirements of the underwater robot for underwater target grabbing and the like are met.

The invention provides an image denoising method based on super pixel clustering and sparse representation to solve a technical problem of a low denoised image peak signal to noise ratio and detain information loss of an existing image donoising method. The method comprises the steps of (1) inputting an image to be denoised, (2) carrying out super pixel segmentation and super pixel clustering on the image and obtaining multiple clusters of similar super pixels, (3) carrying out image block extraction and dictionary training on each cluster of similar super pixels, (4) calculating the sparse coefficient of each image block under a corresponding dictionary, (5) searching the similar image block of each image block and calculating the sparse coefficient weighted sum of similar image blocks, (6) restraining the sparse decomposition process of each image block by using the sparse coefficient weighted sum of the similar image blocks, and obtaining a new sparse coefficient, (7) judging whether a current number of iterations is larger than a maximum number of iterations Lambda or not, executing a step (8), otherwise adding 1 to the number of iterations, and executing the step (5), and (8) reconstructing the image to be denoised , and obtaining a denoised image.

The invention discloses a method for predicting interaction between sRNA (small RNA) and target mRNA (messenger RNA) of sRNA in bacteria. The method includes the steps: 1) data collection and sorting;2) feature extraction, conversion of a data set into a matrix; 3) F-score feature optimization; and 4) training and construction of an SVM model, prediction and obtaining of a prediction result. Themethod for predicting interaction between sRNA and target mRNA of sRNA in bacteria can effectively characterize RNA sequence information and improve the prediction accuracy of sRNA-target mRNA interaction. At the same time, the method for predicting interaction between sRNA and target mRNA of sRNA in bacteria has the advantages of low cost, low time consumption and high prediction speed.