63 results about "Low rank matrix decomposition" patented technology

The invention relates to a low-rank decomposition and space spectrum constraint-based hyperspectral image time domain change feature extraction method. The method mainly comprises the following stepsof: calculating a difference image between two hyperspectral remote sensing images acquired at a same place and at different times; establishing a low-rank matrix decomposition model with a space spectrum constraint by utilizing internal data structural characteristics of the difference image; and solving each component of the model through an alternate iteration manner so as to extract a time domain change feature. According to the method, the low-rank matrix decomposition model with the space spectrum constraint and a solution algorithm of the model are disclosed by sufficiently utilizing internal structures of data, so that the time domain change feature is effectively extracted, multiple forms of noises are removed, real changes are strengthened and false changes caused by noises are suppressed, thereby improving the change detection precision of time domain change hyperspectral remote sensing images.

The invention belongs to the video analysis field and aims at accurately recovering a video sequence background in an online mode. The invention discloses a video sequence background recovery method based on online low-rank background modeling. Based on traditional background modeling, low-rank matrix decomposition of a nuclear norm and motion information estimation of binaryzation are introduced so that a problem which can not be processed in the prior art can be solved. The method comprises the following steps of 1) specifically expressing a background recovery problem in a video sequence as following unrestraint optimization equation solving; 2) constructing a kth frame of motion mapping weight vector wk; 3) using an alternative direction method to solve; 4) solving ck<l+1>; 5) solving ek<l+1>; 6) repeating the above steps 4) and 5) till that algorithm convergence is achieved; 7) using variable substitution to solve; 8) solving Yk<l+1>; 9) solving Lk<l+1>; 10) repeating the above steps 8) and 9) till that the algorithm convergence is achieved; and 11) solving a video background. The method of the invention is mainly used for a video analysis occasion.

The invention discloses a face recognition method based on non-convex low-rank decomposition and superposition linear sparse representation. The method comprises the following steps: 1, according to alow-rank matrix decomposition theory, replacing a nuclear norm with a gamma norm for low-rank matrix decomposition, and introducing a structure incoherent discrimination item to form discriminative non-convex low-rank decomposition; 2, resolving the discriminative non-convex low-rank decomposition, and decomposing the face sample matrix into a low-rank matrix and a sparse matrix; 3, decomposing the low-rank matrix into prototype dictionaries and variation dictionaries through superposition linear representation, and then using the two dictionaries as dictionaries for testing through linear weighting combination; and 4, solving a sparse coefficient of l1 norm by using a homotopy method by using a sparse representation algorithm, carrying out classified recognition on the face images by reconstructing a minimum sparse residual model, and classifying the face samples to be tested into a class with a minimum error, thereby realizing face recognition. According to the method, good robustness and high efficiency can be maintained under the conditions of shielding and noise pollution.

The invention discloses a hyperspectral image denoising method based on non-convex low-rank matrix approximation and total variation regularization. The hyperspectral image denoising method comprisesthe following steps: 1) acquiring hyperspectral image data to be denoised, wherein M, N and K represent the number of width and height spectral bands of a hyperspectral image respectively; 2) constructing a denoising model based on non-convex low-rank matrix decomposition and TV regularization; (3) optimizing the model by adopting an Augmented Lagrange function (ALM) algorithm, and optimizing themodel by adopting an ALM algorithm; 4) optimizing and solving the model by adopting an augmented Lagrange function algorithm; and 5) outputting the hyperspectral image after the mixed noise is removed.

The invention provides a de-raining method of a single image based on sparse and low-rank matrix decomposition. The de-raining method is characterized in that the method comprises following steps: step 1, inputting a rain-containing image, decomposing the rain-containing image into 8*8 image blocks, taking the center of the image blocks as the origin, forming an input matrix by 8 image blocks with 4 of pixel deviation, and performing sparse and low-rank matrix decomposition on the input matrix, wherein a low-rank matrix is regarded as a low-frequency component, and a sparse matrix is regarded as a high-frequency component; step 2, dividing the high-frequency component into a plurality of mutually overlapped high-frequency sub-blocks, obtaining a dictionary by learning through a dictionary learning method, and dividing the dictionary into a rain portion dictionary and a geometric portion dictionary according to HOG characteristics; and step 3, dividing a high-frequency image into a plurality image sub-blocks which are not mutually overlapped after obtaining the rain portion dictionary and the geometric portion dictionary, adding geometric portions in a geometric component and a rain component, and forming a de-rained output image through merging, wherein each image sub-block is represented by the geometric component and the rain component.

The invention provides a multi-dimensional domain joint SAR broadband interference suppression method based on low-rank matrix decomposition. Broadband interference signals exist in a plurality of pulses of current echo data. The short-time Fourier transform matrixes of the pulse echo signals are vectorized respectively; RPCA decomposition is performed to obtain a low-rank matrix and a sparse matrix; each row of the decomposed sparse matrix is rearranged into a short-time Fourier matrix form; the rearranged short-time Fourier matrix is subjected to inverse short-time Fourier transform, the reconstructed interference signal is subtracted from the original echo signal to realize broadband interference suppression, and data subjected to interference suppression is imaged by using an existingimaging algorithm to obtain a high-resolution image. According to the method, the problem of useful signal loss caused by time-frequency filtering is avoided, compared with a traditional method basedon energy characteristic difference, useful signal information can be effectively reserved while interference is suppressed, and the image quality after broadband interference suppression can be improved to a large extent.

The invention discloses a robust image classification method and device based on low-rank two-dimensional local identification image embedding, and the method comprises the steps: constructing an image library unit which is used for obtaining a standard image library, and constructing a new to-be-classified standard image library; the first calculation unit is used for calculating the difference J(P) between the intra-class divergence matrix Sw and the inter-class divergence matrix Sb of the new standard image to be classified; the first image processing unit is used for carrying out low-rankmatrix decomposition on the acquired image X to obtain a low-rank matrix A and a sparse matrix E; the second calculation unit is used for obtaining a final target function according to the combination of results of the first calculation unit and the first image processing unit; the feature matrix calculation unit is used for obtaining a feature matrix Y; and the nearest neighbor classifier unit is used for classifying the images by utilizing a nearest neighbor classifier and outputting a classification result of the images. According to the invention, the technical problems of low classification precision, noise points and singular points in the existing image classification based on the 2DLPP learning model are solved, and the identification precision is improved.

The invention discloses a noise-immunity three-dimensional positioning method for a moving target in a tunnel. The positioning method comprises the steps that (1) beacon nodes, positioning identification cards and data transmission devices are laid according to the environment in the tunnel; (2) sensor nodes perform distance measurement based on an RSSI (Received Signal Strength Indicator) distance measurement method and converge the data to a data center, and an initial partially missing and noise containing Euclidean distance matrix is constructed; (3) mixed Gaussian distribution is introduced to fit unknown noise, a problem of complementing the partially missing Euclidean distance matrix under the complex environment in the tunnel is modeled into a noise-immunity low-rank matrix decomposition model, and the model is solved by adopting the classical expectation maximization algorithm; and (4) the real position of the moving target in the tunnel is calculated by adopting the classicalmulti-dimensional scaling algorithm based on the complemented inter-node real Euclidean distance matrix. The noise-immunity three-dimensional positioning method fully considers influences imposed onsensor node distance measurement by the complex environment in the tunnel, and can perform real-time accurate positioning on the moving target in the tunnel under various types of noise.

The invention relates to an improved convolution matching tracking pipeline video image defogging enhancement method. The fog layer model is obtained by processing the center circle in the video framein the pipeline, and the fog layer is used as the source of convolution dictionary to train the dictionary. The background image and foreground image are obtained by training the low rank matrix decomposition of the first few frames, the residual image is initialized according to the foreground image and the feature response is calculated through the residual image, the foreground reconstructed image is initialized, and the noise energy is calculated. Maximum search for feature response is performed; Reconstruction of foreground image using maximum value and current maximum feature response and updating feature response using maximum value are performed; the residual image is updated, and the residual image energy is calculated. If the residual image energy is less than the noise energy,the final defogging image is calculated. The invention can effectively obtain the fog layer model, and find the fog layer coincident with the video frame through the fog layer model as a convolution dictionary and remove the fog layer. the method can be used in pipeline video image enhancement and other scientific fields.

The invention discloses a defect detection method for a solar cell panel, which is. realized through a measurement system formed by a test rack, a gray area array camera, a telecentric lens, an annular light source and an objective table and includesTest stand, Gray area array camera, A measurement system constructed by a telecentric lens, an annular light source and an objective table is realizedby the following steps of: shooting a picture of a solar cell panel, respectively carrying out expansion processing in horizontal and vertical directions, carrying out edge detection by adopting a canny operator, rotating the image by adopting Hough transformation, and repairing a cavity value of the rotated image; fFiltering and denoising, carrying out brightness equalization processing by adopting gamma correction, and uniformly segmenting the image; converting the segmented image into a matrix D by adopting a robust principal component algorithm, decomposing the matrix D into a low-rank matrix A and a sparse matrix E, and performing recovery processing on the matrix D according to an APG gradient approximation algorithm; and carrying out reverse returning operation on the low-rank matrix A and the sparse matrix E to a single wafer image, carrying out low-rank matrix decomposition technology processing to obtain a defective solar cell panel image, and then detecting a defect position.

The invention provides a damaged image repair method based on multi-view popular low-rank representation. The following operations are carried out on a damaged image: firstly, the multi- type characteristic of the damaged image is extracted, and the multi-view representation of the damaged image is obtained; then the multi-view popular structure is constructed and used, the different local geometric structures of the damaged image are maintained, the damaged image is subjected to low-rank matrix decomposition, and the multi-view popular low-rank representation and damaged matrix representation of the image are obtained; finally the damaged image is repaired based on the integration of multi-view low-rank representations, and a complete original image is obtained. The method has the advantages that the multi-view compensation information and local popular structure characteristic of the damaged are fully considered, the clear and complete original image can be recovered well, the visual experience of a user is improved, and the method can be applied to the image restoration in the fields of biology, medicine, remote sensing and the like.

The invention provides a sign prediction method capable of considering the local labelling characteristics of a node. The method comprises the following steps that: S1: obtaining the form expression of a signed network; S2: analyzing the structural characteristics of the signed network, obtaining a relationship between the rank and structure balance of a network adjacent matrix and a weak structure balance theory, utilizing low-rank matrix decomposition to decompose a low-rank matrix, and meeting a constraint condition that a matrix rank is minimum; S3: according to the conclusion of the structural characteristics of the signed network, converting a sign prediction problem into an optimization problem; and S4: utilizing a stochastic gradient descent method to decompose the low-rank matrix to obtain network considering global characteristics.

The invention provides an endoscopic image highlight removal method based on non-convex low-rank matrix factorization, which comprises the following steps: an absolute highlight detection stage: statistical attributes of an image on saturation and intensity values are calculated, an adaptive threshold is generated, a double-threshold segmentation algorithm is executed, absolute highlight pixels are detected, and an absolute highlight image is generated; a non-convex low-rank matrix decomposition stage: non-convex low-rank matrix decomposition is executed on each image, gradient information isintroduced in the iterative optimization process with the position information of absolute highlight pixels as priori knowledge, and a low-rank image and a sparse image are calculated; an image fusionand reconstruction stage: a finer highlight image is generated, morphological operation and linear filtering operation are executed, a self-attenuation weight image is generated, pixel-level fusion of a low-rank image and an original endoscope image is executed, and an endoscope image without highlight is generated; image blocking and integration: an original image with a large size is divided into a plurality of sub-images by adopting an image blocking technology, and multi-task parallel computing is realized by adopting a multi-thread technology.

The invention discloses a noise-robustness image similarity distinguishing method, and belongs to the field of digital image processing. By the noise-robustness image similarity distinguishing method, images with large noises can be matched with accurate image blocks, the image matching accuracy is improved, and image processing methods such as image stitching and image denoising which are based on block matching can be implemented. A noise-robustness similarity distinguishing function is established by a low-rank matrix decomposition theory which is rapidly developed recently, influences of irregular disturbance such as noises on similarity distinguishing are eliminated, and distinguishing effectiveness under the condition of large noises can be guaranteed.

The invention provides a speaker recognition method based on dictionary learning and low rank matrix decomposition. Thespeaker recognition method comprises the following steps that step 1, speaker audiosare subjected to processing of pre-emphasis, framing, windowing, endpoint detection and the like; step 2, MFCC features corresponding tosentences of speakersare extracted, and a GMM-UBM model is trained; step 3, a global difference spatial matrix T and a global difference spatial factor w are estimated by joint factor analysis (JFA); step 4,i-vector of the sentences of each speaker is obtained; and step 5, the i-vector of the M dimension isextracted from a training set to form a feature matrix, and adistinguish dictionary isgenerated according to the training set and a test set, the obtained dictionary is taken as abackend processing and scoring moduleof thei-vector, anda basis isprovided for the final distinguish. Coding coefficients adapted to the dictionary learning criterion can effectively improve the recognition power, and optimal classification is carried out through structured sparseness.

The invention discloses a bimodal image saliency detection method based on node classification and sparse graph learning, and belongs to the technical field of computer vision. According to the invention, the method comprises the steps: taking a thermal infrared image as an image channel of a color image to carry out superpixel segmentation, extracting color features and multilayer semantic features of each superpixel in a bimodal image, establishing a graph model, and then carrying out low-rank matrix decomposition on the color features of the two modal images, generating a corresponding low-rank matrix and a sparse matrix, and classifying graph model nodes; calculating an initial adjacency matrix according to a node classification distance and an Euclidean distance, and then carrying out sparse graph learning saliency sorting by using the initial adjacency matrix and an indication vector for many times to obtain a saliency graph. Compared with an existing saliency detection method, the method has the advantages that the detection precision is remarkably improved, the image saliency region can be well separated from the background, and the method still has good performance on images shot in a rain and fog environment and an environment with insufficient light.

The embodiment of the invention relates to the technical field of data processing, and discloses a product recommendation method and device, electronic equipment and a computer readable storage medium, and the product recommendation method comprises the steps: constructing a first score recommendation matrix between a user and at least one to-be-recommended financial product according to the related information of the user based on a collaborative filtering method algorithm, wherein the related information comprises personal basic information, scores of financial products and historical purchase information of the financial products; then, based on a low-rank matrix decomposition algorithm, constructing a second score recommendation matrix of the user and each financial product according to a purchase relationship between the user and each financial product; and then, calculating the score of each to-be-recommended financial product according to the first score recommendation matrix and the second score recommendation matrix, and recommending the corresponding to-be-recommended financial product to the user according to the scores. According to the method, the user characteristicsand the product characteristics are fully considered, and the financial product recommendation precision and accuracy are greatly improved.

The invention discloses a method for detecting a hidden object in a millimeter-wave human body image based on block mixture Gaussian low-rank matrix decomposition.The method mainly solves the problemsof low imaging quality caused by weak scatter echo of the hidden object and low detection accuracy of a gray scale value of the hidden object and a human body similarity in the prior art. An implementing scheme of the method comprises the following steps of 1, removingabnormal points in an imaging region background in an original millimeter-wave human body image, and dividing the human body imageinto six parts according to proportions of human body parts; 2, decomposing all regions of thehuman body based on a block mixture Gaussian low-rank matrix decomposition algorithm to obtain a low-rankpart and a sparse part; and 3, binaryzing the sparse part by using a typology method, and removing small noise points to obtain a final detection result graph. The method increases the detection rateof various complicated small targets in the millimeter-wave human body image without a large number of training samples, so that the detected hidden object is more complete; and the method can be used for detecting hidden objects carried by people in public places such as an airport and a bus station.