74results about How to "Retain structural information" patented technology

The invention discloses an image saliency detection method based on region description and priori knowledge. The method comprises the following steps that: (1) an image to be detected is subjected to pre-segmentation, superpixels are generated, and a pre-segmentation image is obtained; (2) a fusion feature covariance matrix of the superpixels is generated; (3) feature different region descriptors and space distribution region descriptors of each superpixel are calculated; (4) the initial saliency value of each pixel point of the image to be detected is calculated; (5) a priori saliency region and a background region of the image are obtained; (6) the saliency weight of each pixel point of the image to be detected is calculated; and (7) the final saliency value of each pixel point is calculated. The image saliency detection method has the advantages that the saliency region can be uniformly highlighted in an obtained final saliency map; the background noise interference is inhibited; a good saliency detection effect can be achieved in ordinary images, and the processing on the saliency detection of complicated images can also be realized; and the processing on subsequent image key region extraction and the like can also be favorably carried out.

The invention discloses a network graph embedding method based on edges. The method comprises steps that an Edge2vec algorithm model is constructed; a random gradient descent method is utilized to train the Edge2vec algorithm model; the Edge2vec algorithm model is utilized to realize network graph embedding. The method is advantaged in that the edges of the network graph are directly mapped to thelow dimension vector space through a neural network model based on an automatic depth encoder, the local neighbor degree information and the global neighbor degree information between the edges can be reserved, compared with the prior art, the structural information of the edges of the network graph can be effectively reserved, better performance can be realized for network graph analysis tasks of the edges, and the method can be applied to the network graph analysis tasks of the edges.

The invention discloses an automatic generation method of a handwritten Chinese character library, which comprises the following steps of: predicting a Chinese character skeleton flow field by adopting a two-stage convolutional neural network G through a writing track synthesis stage and a font style rendering stage on the basis of a deep stacking network and structure information guidance; firstly, the writing style is learned from a small amount of handwritten Chinese characters written by a user, so that the writing track of unwritten Chinese characters is synthesized; The handwriting styleof the target is rendered; and generating a complete GB2312 Chinese handwriting word stock file with the writing style of the user. The method can ensure the structural accuracy and style consistencyof the generated fonts at the same time, is simple, efficient, low in cost and high in quality, and can meet the actual application requirements of ordinary people for quickly making personalized handwriting fonts.

The invention discloses a hyperspectral data classification method based on space-spectrum combination information, and the method provided by the invention employs a convolution neural network and asuperpixel dividing method, and solves a utilization problem of the space information of a current hyperspectral image. The method comprises the steps: 1, building a convolution neural network model,carrying out the feature extraction, and obtaining an extracted feature vector; 2, carrying out the superpixel dividing of the hyperspectral image through an M-SLIC algorithm, and obtaining a label image after superpixel dividing; 3, carrying out the clustering of a hyperspectral feature image, generating a new feature vector through combining a BoVM model, and completing the classification process. The method achieves the extraction of high-dimensional nonlinear features through the convolution neural network, multiple convolution layers and a downloading layer, reduces the impact on the spectrum information from the photographing condition difference through adding spatial information, achieves the clustering of the feature spectrum image, replaces the feature spectrum obtained through primary feature extraction via the convolution neural network by the secondary features obtained through the BoVM model, further reduces the classification errors, and is higher in theoretical and engineering practice significance.

The invention realizes a tree point cloud three-dimensional reconstruction method based on a local structure and direction perception, and belongs to the technical field of space information. Considering the continuous characteristic of the branch direction and the point cloud density in the local structure, the method deduces a connecting relation of branches at a data missing position, an optimization equation and the iteration process of point cloud repair are designed, points in point cloud data are driven to contract and diffuse in the same time along the direction of a skeleton line by employing skeleton point cloud, the optimized point cloud does not produce extra noises, the contracted skeleton is employed to inherit the space association between adjacent iterations instead of depending on the optimized point cloud, and structural information of the original point cloud is fully retained. The radius of each node of the skeleton is calculated by employing a plant growth model, and the skeleton is expanded to a three-dimensional tree model according to the radius of each node. Compared with the conventional tree three-dimensional reconstruction method, point cloud of a deletion region can be more accurately repaired, and a better modeling result is obtained.

The invention discloses a fuzzy clustering-based brain MR image segmentation method. The invention mainly aims to solve the problem of the incapability of a traditional FCM (Fuzzy C-Means) algorithm and an improved method thereof to simultaneously eliminate noises and bias fields during a brain MR image segmentation process. According to the method of the invention, local spatial information, local grayscale information and non-local information in an image are fully utilized so as to construct a multi-local information fuzzy factor and a non-local weight, and the details of the image are preserved as much as possible while the anti-noise performance of the algorithm is improved; a bias field model is established to remove gray nonuniformity in the brain MR image; the multi-local information fuzzy factor and the non-local weight are embedded into an FCM method with the bias field model, so that the segmentation of the brain MR image under the noise and bias field condition can be realized. With the method of the invention adopted, noises in the brain MR image can be effectively suppressed, and the influence of the bias field on the segmentation of the brain MR image can be effectively eliminated. The method has better segmentation performance.

The present invention relates to a multi-layer automatic coding method and system based on deep learning. Combining the principle of deep learning and tensor algorithm, the original data is expressed in the form of tensor, which can fully excavate the original data without destroying the structure of the original data. The original information, and through multi-layer learning, obtain more essential abstract features, so as to overcome the limitations of vector expression, retain the structural information of the original data to a large extent, and obtain more robust feature extraction and pattern learning , which is conducive to the reflection of the essence of the original data and the subsequent pattern classification.

This invention discloses an image identification method based on the Gabor phase mode including: picture selection: exchanging Gabor to the being compared images to get Gabor character images for them, picking up global and local Gabor phase modes from each of the character image and evaluating the character modes to connect the evaluation results in series to high dimension character vectors to be compared to get the similarity degree among the vectors to identify images.

The invention discloses a high-dimensional multimedia data classifying method based on maximum margin tensor study. The method includes the following steps that (1) a training data set of multimedia data is built; (2) the training data set is modeled and analyzed to obtain a classifying model; (3) according to a user inquiry data set and the classifying model, the inquiry data set is classified. According to high-dimensional performance and structure performance of the multimedia, the multimedia data is expressed through tensor, and high-dimensional multimedia data is classified through a maximum margin classifier method. Classifying is finished while the multimedia data is subjected to decomposition analysis, structural information in the multimedia data is reserved, dimensionality curse caused by high-dimensional data generated through a traditional splicing method is avoided, and the method is more accurate than a traditional multimedia data classifying method and facilitates calculation.

The invention discloses an SAR (Synthetic Aperture Radar) image target identification method based on multilinear principal component analysis and tensor analysis. The method comprises the following steps that: constructing a four-order tensor training sample; utilizing the multilinear principal component analysis to obtain a multilinear projection matrix; constructing a core tensor; carrying out linear discriminant analysis on the core tensor to obtain the weight vectors of one group of linear discrimination functions; and carrying out classified identification on test samples. By use of the method, the four-order tensor sample is constructed for the SAR image, the multilinear principal component analysis is adopted to extract features, image structure information is effectively kept, and a correct recognition rate of a target is improved.

The embodiment of the invention provides a method and a device for classifying texts, a storage medium and a processor, and belongs to the technical field of computers. The method comprises the stepsof performing word segmentation on a to-be-classified text; determining a word vector corresponding to each word obtained by performing word segmentation on the to-be-classified text based on a word vector model, and forming a matrix by the word vectors corresponding to the words belonging to one sentence; processing each matrix based on the sentence vector model to obtain a sentence vector corresponding to each matrix; and processing each sentence vector based on the sentence classification model to obtain a category score vector corresponding to each sentence vector, and determining the typeof the sentence vector corresponding to the category score vector according to each category score vector to realize classification of the to-be-classified text. Therefore, the defects that the constructed word frequency or feature matrix is very sparse and the relationship between words is ignored when the short texts are classified are overcome, and the effect of classifying the texts is improved.

The invention relates to the field of computer graphics, and discloses an image same name point matching method and device. At least two images are acquired; feature extraction of the images is performed so as to acquire a stable first feature point set; feature point matching is performed between two different first feature point sets so as to generate a second feature point set, and the corresponding pixel point set of the second feature point set in the corresponding images is acquired by using a correlation inverse algorithm; similar area expanding is performed in the pixel point set according to the image scale and the image features; the area ratio of the similar areas in the two images is calculated; the Euclidean distance or the correlation coefficient of the point sets in the similar areas in the two images is calculated so as to calculate the matching degree of the similar areas in the two images; and a same name point set is acquired when the matching degree is greater than a preset threshold so that same name point matching identification of the photographed images can be performed without acquiring the technical parameters of image photographing.

The invention discloses a texture fusion method for RGB-D camera real-time three-dimensional reconstruction. The method includes: processing the RGB-D data stream to obtain the definition of a color image, selecting a key frame to extract a foreground, performing filtering and denoising on a depth image, calculating a normal vector of a depth image point cloud, and establishing a reconstructed data stream; quantitatively establishing an adaptive weight field of the color image in a mode of combining a probability method and a heuristic method, and taking the adaptive weight field as confidencecoefficient distribution of a real-time frame for describing color data; and by comparing the confidence coefficient weight in the adaptive weight field of the real-time frame with the latest confidence coefficient weight of the reference point cloud, selecting an operation from three operations of replacement, fusion and reservation to update a texture result, thereby realizing texture fusion applied to three-dimensional reconstruction. According to the method, high-quality data can be extracted, blurring of texture fusion can be effectively reduced, a clear texture reconstruction result isrealized, and the method is embedded into an RGB-D reconstruction framework with relatively low calculation cost, so that the texture reconstruction precision is remarkably improved.

The invention provides a method for constructing a kerogen average molecular structure model. The method comprises the steps of: calculating a hydrogen index HI of each sample and a kerogen average molecular structure parameter of regional shale; extracting the kerogen sample in a shale powder sample for XPS energy spectrum analysis and FTIR spectral analysis, determining the elemental compositionin the kerogen sample, the atomic ratio of carbon atoms to heteroatoms and the existence form of heteroatoms; fitting a peak area according to the characteristic peaks of each group in a FTIR spectrum, and calculating the chemical structure parameters of kerogen; determining the total carbon number of the kerogen average molecular structure, and constructing the initial kerogen average molecularstructure; maintaining the main carbon skeleton structure and the heteroatoms functional group unchanged, continuously adjusting the carbon number of each aliphatic chain and the position of the heteroatoms functional group in the initial kerogen average molecular structure, until the adjusted model structure parameters are consistent with the XPS and FTIR experimental results, obtaining the kerogen average molecular structure model of the regional shale.

The invention relates to a digital image blind forensics technology based on PatchMatch, comprising the following steps: 1) extracting BRISK key points from an image to be tested and generating feature descriptors; 2) calculating that offset between each BRISK feature descriptor according to the PatchMatch algorithm and conduct conduction and random search to find a stable matching point; 3) performing Median filtering with a circular window with radius; 4) filter and taking a circular neighborhood with a radiu of, The fitting error of the least mean square linear model was calculated Settingthe fitting error thresholdof the least mean square linear model calculated from 4); 6) Because similar background regionsare usually close to each other, deleting region pairs with a distance of lessthan pixels; 7) Because similar details (non-copy paste areas) are usually small, deleting area pairs with less thanpixels are deleted; 8) carrying out The geometrical expansion operation with the circular structure with radius to realize the accurate positioning of the copying and pasting area in the drawing. The invention obtains good tamper detection effect and is suitable for copying and pasting tamper detection.

The invention relates to a trap printing method and system based on a boundary and belongs to the technical field of trap printing. The trap printing method comprises the following steps of: firstly, extracting the boundary of an image to be subjected to the trap printing; determining the distance and direction from each pixel point within a pre-set width range T1 at two sides of the boundary to the boundary which is closest to the pixel point; and then, determining trap printing attributes of the boundary and color channels at two sides of the boundary; and then, determining the gradient amplitude and direction of the pixel point needing to be subjected to the trap printing; determining weight coefficients of the adjacent pixel points in a trap printing template according to the gradient amplitude and direction of a central pixel point; and finally, carrying out the trap printing on the pixel point needing to be subjected to the trap printing through the trap printing template. According to the invention, structural information of an original image can be maintained as much as possible and the trap printing quality of the image is improved.

The invention discloses an image denoising method based on a structural similarity and total variation hybrid model. The method comprises (1), designing a functional E (u); (2), introducing a new auxiliary variable to the E (u) and turning the original model into two simple sub-models by using an alternate iterative method; (3), performing numerical solution on the two sub-models respectively by using a gradient descent method and a chambolle projection method to obtain a discrete mathematical model; (4), inputting a noisy image f; (5), performing iteration denoising on the f by using the discrete mathematical model; and (6), stopping until the iteration reaches set end conditions and outputting the denoised image. By means of the image denoising method, structural information of images can be well maintained while denoising is performed effectively, visual effects of the images are improved, and the method is applicable to denoising of natural images.

The invention relates to a mobile storage device with an authentication function. The mobile storage device comprises a mobile storage device body and an iris recognizer connected with the mobile storage device body through electrical signals. The iris recognizer comprises a sampling module, a preprocessing module, a characteristic encoding module and a code matching module, wherein the characteristic encoding module is used for extracting and encoding characteristics of an iris image and comprises a first-time LBP operator processing submodule, a second-time LBP operator processing submodule, a third-time LBP operator processing submodule and a fourth-time LBP operator processing submodule. Relevance between the center point and other neighbourhoods around the center point is improved, image texture different in scale and frequency can be met, after multi-time processing of the LBP operator processing submodules, the encoding length is reduced continuously without affecting the relevance between the center point and other neighbourhoods, storage space is saved, the calculation amount is reduced, recognition speed is increased, the recognition accuracy is improved, and robustness is high.

The invention discloses a terahertz time-domain spectrum article classification method based on a neural network. The method is realized through the following steps: (1) measuring the terahertz spectrum data of articles to be classified; (2) calculating the optical constant of the articles to be classified; (3) extracting a Pauli decomposition eigenvalue; (4) constructing a convolutional neural network; (5) constructing the characteristic matrix of a training sample and the characteristic matrix of a test sample; (6) training the convolutional neural network; (7) obtaining the class label of each data point in the test sample; and (8) outputting a classification result according to the difference of class labels. According to the terahertz time-domain spectrum article classification methodprovided by the invention, a terahertz time-domain spectrum of the articles is measured and the convolutional neural network is applied to classify the articles, so that the method has the advantagesof broad application scenes, non-contact, non-damage and high classification accuracy.

The invention discloses an ATM (automatic teller machine) input device with identification functions implemented by the aid of irises. The ATM input device comprises an ATM input device body and an iris identifier. The iris identifier is in electric signal connection with the ATM input device body and comprises a sampling module (1), a preprocessing module (2), a feature encoding module (3) and an encoding matching module (4), and the feature encoding module is used for extracting and encoding features of iris images and comprises a first LBP (local binary pattern) operator processing sub-module, a second LBP operator processing sub-module, a third LBP operator processing sub-module and a fourth LBP operator processing sub-module. The ATM input device has the advantages that the relevance of central points and other surrounding neighborhoods can be improved, requirements on image textures with different scales and frequencies can be met, the encoding lengths can be continuously reduced after the iris images are repeatedly processed by the LBP operator processing sub-modules without influence on the relevance of the central points and the surrounding neighborhoods, accordingly, storage spaces can be saved, the computational complexity can be reduced, the identification speeds can be increased, the identification accuracy can be improved, and the ATM input device is high in robustness.

The invention discloses a target tracking method based on fast tensor singular value decomposition feature dimensionality reduction. The target tracking method comprises: extracting multiple featuresfrom each frame of video data, and constructing a tensor structure; performing singular value decomposition on the constructed tensor; and training related filters by using the features after dimension reduction, and tracking the target. According to the method, the number of features can be effectively reduced, the tracking speed is increased, and compared with a traditional vector-based principal component analysis feature dimensionality reduction mode and other modes, the structure information of the features is better reserved; the tensor singular value decomposition has invariance to therotation of the feature to enhance the robustness of the tracker to the target rotation.

The invention relates to a hardware Trojan horse detection method and system based on a bidirectional graph convolutional neural network. The method comprises the following steps of firstly, preprocessing a netlist file, creating a corresponding directed graph representation, encoding gate device information as a feature representation X, and constructing circuit directed graph data, respectively creating a forward circuit diagram for describing a circuit signal propagation structure and a reverse circuit diagram for describing a circuit signal dispersion structure, respectively constructing corresponding graph neural network feature extractors to extract structural features, and combining the structural features into final gate device features, constructing a multi-layer perceptron classification model, forming a hardware Trojan horse gate classification model by the multi-layer perceptron classification model and a graph neural network feature extractor, and learning model parameters by using a weighted cross entropy loss function to obtain a trained hardware Trojan horse gate classification model, and converting a to-be-detected netlist into a directed graph, inputting the directed graph into the trained hardware Trojan horse gate classification model for detection, and outputting a suspicious door device list. According to the method, the exit-level hardware Trojan horse can be effectively detected.

The invention provides an automatic control device based on iris recognition. The automatic control device comprises an automatic control device and an iris recognizer which is in electrical signal connection with the automatic control device. The iris recognizer comprises a sampling module (1), a preprocessing module (2), a feature coding module (3) and a coding matching module (4). The feature coding module (3) is used for extracting and coding features of an iris image, and comprises a first LBP operator processing sub-module, a second LBP operator processing sub-module, a third LBP operator processing sub-module and a fourth LBP operator processing sub-module. According to the invention, the correlation between a center point and other surrounding areas is increased; image texture of different scales and frequencies can be realized; after repeated processing of the LBP operator processing sub-modules, coding length is constantly reduced without affecting the correlation between the center point and the surrounding areas; storage space is saved; calculation amount is reduced; the recognition speed is improved; the recognition accuracy is enhanced; and high robustness is acquired.

The invention discloses a method for converting a low-resolution image into a high-resolution image. The method comprises the steps of carrying out concentrated sample matching according to a given input low-resolution image Il<t> and according to similarity measure to obtain an image block set and taking out the best image block with the use of the prior knowledge of texture to generate a texture gradient function Ut (p). According to the method, Il<t> is interpolated to generate an image Is in the direction of a retained image, an edge gradient function Ue (p) is generated with the use of a learned edge function ft (vl), and Ut (p) and Ue (p) are weighted with the use of a weighting function W(p) to generate a gradient function U, thereby obtaining a high-resolution image Ih<t> amplified by S times. The method for converting a low-resolution image into a high-resolution image has the characteristic that a reconstructed high-resolution image is high in quality and is more real.

The invention discloses a digital restoration method for flaking diseases of ancient wall paintings based on compressed sensing, and an intelligent terminal system. The method employs a wall painting image preprocessing module, a wall painting flaking disease marking module and a wall painting image restoration module. The method comprises the steps: enabling the wall painting image preprocessing module to carry out the image denoising and HSV color space transformation, enabling the wall painting flaking disease marking module to achieve the V component extraction of a wall painting image, drawing a brightness contour map through a contour function, carrying out the threshold segmentation of the brightness contour map, and obtaining wall painting segmented images; obtaining a closed interval of a disease region through the morphological corrosion operation, carrying out the overlapping of the segmented images and a damaged wall painting image, and marking the flaking disease region of the wall painting; enabling the wall painting image restoration module to carry out the restoration of the flaking disease region of the wall painting through an image restoration algorithm based on compressed sensing of a PMLE mechanism and an EM mechanism. The method can independently complete the marking of the flaking disease region in the ancient wall painting and the intelligent restoration, and provides support for the digital virtual display of the wall painting.

The invention relates to a navigator started based on iris recognition, wherein the navigator includes a navigation device and an iris recognition device in electrical single connection with the navigation device; the iris recognition device comprises: (1) a sampling module; (2) a preprocessing module; (3) a feature encoding module used for extraction and encoding of characteristics of an iris image and including a first LBP operator processing sub module, a second LBP operator processing sub module, a third LBP operator processing sub module and a fourth LBP operator processing sub module; and (4) an encoding matching module. The relationship of a center point with other surrounding neighborhood regions is increased, and image textures with different scales and frequencies can be met; after processing with several LBP operator processing sub modules, under a condition without affecting the relationship of the center point with the surrounding neighborhood regions, the encoding length is reduced continuously, storage space is saved, the calculated quantity is reduced, the recognition speed is improved, the recognition accuracy rate is enhanced, and relatively high robustness is obtained.

The invention puts forward an electronic book content representation method based on a local reconfiguration model. The method comprises the following steps that: A: tree structure expression: for each electronic book, dividing the electronic book into a plurality of pages, dividing each page into a plurality of paragraphs, and organizing each electronic book into a three-layer tree structure of "electronic book-page-paragraph"; B: node feature expression: constructing a vocabulary, calculating a word distribution vector, and using principal component analysis to carry out dimensionality reduction and compression on the word distribution vector of each level of nodes; C: local reconstruction model establishment: using the information of a child node to reconstruct the parent node information, i.e., establishing the local reconstruction model, solving the local reconstruction model, and obtaining a reconstruction coefficient vector; D: the uniform vector expression of the tree structure: according to the reconstruction coefficient vector obtained in C, carrying out information fusion on the node and the child node thereof, and updating the feature vector expression of the node; andE: carrying out electronic book retrieval and recommendation based on contents.

The invention discloses a motion random exposure based super-resolution imaging system and method, belonging to the technical field of image acquisition. The imaging system comprises a black box, a motion route control cableway, a low-resolution camera, an added shutter, a binary random number generator and an image reconstruction processor, wherein the low-resolution camera is fixed on the support of the motion route control cableway and moves in the black box along a given route at the fixed speed, meanwhile, the added shutter in front of a camera lens flickers under the control of the pulse signal sequences generated by the binary random number generator to carry out motion random exposure and transmits the low-resolution images obtained after exposure to the image reconstruction processor, a high-resolution image reconstruction model is established according to the sparsity of the high-resolution images, and the high-resolution images are reconstructed by a nonlinear optimization method. The invention has the advantages of simple structure, easy implementation and low cost and is used for acquisition and reconstruction of the super-resolution images.

The invention discloses a multi-path interference correction method and system for a TOF (Time of Flight) module and electronic equipment. The multi-path interference correction method for the TOF module comprises the following steps: synthesizing a TOF cross-correlation graph with multi-path interference through an optical signal propagation simulation model with multi-path interference so as to obtain a synthesized TOF cross-correlation graph set with multi-path interference; constructing TOF depth maps which are in one-to-one correspondence with the TOF cross-correlation maps with the multi-path interference and do not have the multi-path interference, so as to obtain a real TOF depth map set without the multi-path interference; training a multi-path interference correction model based on the synthesized TOF cross-correlation graph set with the multi-path interference and the real TOF depth graph set without the multi-path interference to obtain a trained multi-path interference correction model; and correcting TOF data collected by a TOF module through the trained multipath interference correction model to obtain a TOF corrected depth map.

The invention discloses an elevator car identified through iris. The elevator car comprises an elevator car and an iris identifier in electric signal connection with the elevator car; and the iris identifier includes a sampling module (1), a preprocessing module (2), a characteristic encoding module (3) for extracting and encoding characteristics of iris images and including a first LBP operator processing submodule, a second LBP operator processing submodule, a third LBP operator processing submodule and a fourth LBP operator processing submodule, and a code matching module (4). The adopted technical scheme by the elevator car adds relevance between a center point and other surrounding adjacent areas, can satisfy image textures with different sizes and frequencies, continuously reduces the encoding length without influencing relevance between the center point and the surrounding adjacent areas after multiple times of processing by the LBP operator processing submodules, saves the storage space, reduces the calculated amount, accelerates the identifying speed, enhances the identifying accuracy, and is higher in robustness.