51 results about "Nearest neighbour classifiers" patented technology

The invention provides a method for extracting, matching and recognizing characteristics of finger veins, which comprises the following steps: acquiring a finger vein image by an infrared image acquiring device, preprocessing the image, extracting the characteristics, recognizing and analyzing, wherein the preprocessing step comprises the concrete steps of graying a color image, extracting a finger area, adopting direction filtering and enhancing, extracting a finger vein grain according to a finger outline mark and binarizing, de-noising by an area eliminating method, and standardizing the image size to a uniform image size; the characteristic extracting method comprises the steps of carrying out sub-block partitioning on the finger vein grain image, and extracting the characteristics ofeach sub-block image by a bidirectional two-dimensional principal component analysis method of bidirectional characteristic value weighting partitioning; and the recognizing and analyzing step recognizes the characteristics of each sub-block as a whole by a nearest neighbor sorter. The invention can reduce the calculated amount of a high dimensional image matrix in recognition of the finger veins, can obviously improve the recognition speed of the finger vein, and the recognition rate is stable and high.

The invention discloses a human face identification method based on manifold learning, and belongs to the technical field of image processing. The method solves the problem of excessive resource consumption of the traditional method for directly processing high-dimension images. The method is combined with two kinds of methods including the nearest characteristic sub space classifier method and the local linear embedding method for realizing the dimension reducing processing on human face images, then, the nearest classifier is adopted for identifying the data subjected to dimension reduction, firstly, the human face image high-dimension data is firstly built, and the human face image samples are stretched into one-dimension vectors in lines; then, the built human face image high-dimension data is subjected to dimension reduction processing, and the low-dimension expression of all obtained human face images is obtained; and finally, the data is embedded into the space at the low dimension. Through the training on the images, the images to be tested are collected in real time, the human face identification is carried out, the method is more reasonable than a local linear embedding method based on Euclidean distance, the identification accuracy is higher, the method has lower operation complexity than a method of directly adopting high-dimension data for identification, and the method is simpler and more convenient.

The invention provides a method for extracting finger vein features for matching recognition, which comprises the following steps: (1) collecting finger vein images with an image acquisition device; (2) pre-treating the collected finger vein images, comprising making color images gray, extracting finger areas, eliminating salt-and-pepper noise and gaussian noise respectively with a combined filter, splitting images and performing binaryzation for split images through a local dynamic threshold algorithm, de-noising in an area elimination method, extracting finger vein context images according to finger contour markers, and finally standardizing the sizes of the images to form unified images; (3) extracting finger vein features through two-dimensional linear discriminant analysis algorithms weighted both in the line direction and in the row direction; and (4) carrying out matching and recognition with a nearest classifier. The invention significantly improves the speed of finger vein recognition and achieves stable and high recognition rate.

The invention discloses a three-dimensional face recognition device and method based on the three-dimensional point cloud. The three-dimensional face recognition device based on the three-dimensional point cloud comprises an data input unit, a characteristic region detecting unit, a noise filtering unit, a characteristic point selecting unit, a data registration and calculation unit and a nearest neighbor classifier calculation unit, wherein the data input unit is used for obtaining three-dimensional point cloud face data, the characteristic region detecting unit is used for positioning three-dimensional point cloud characteristic regions obtained by the data input unit, the noise filtering unit is used for conducting smoothness on the positioning data so as to filter noise, the characteristic point selecting unit is used for selecting global face characteristic points from the pre-processed face data, the data registration and calculation unit is used for conducting registration on the selecting global face characteristic point selected by the characteristic point selecting unit, and the nearest neighbor classifier calculation unit is used for classifying the data registered through the data registration and calculation unit. The three-dimensional face recognition method based on the three-dimensional point cloud comprises the steps of imputing, characteristic region positioning, rough registration, noise filtering, characteristic point selecting, re-registration and recognition. According to the three-dimensional face recognition device and method based on the three-dimensional point cloud, recognition is conducted after noise filtering and registration are conducted on the three-dimensional face data, and thus the recognition precision is high.

The invention belongs to the field of proteomics classification, relating to a classification method of differential proteomics. The method comprises: selecting characteristics by univariate statistics analysis, sequential feature selection and a genetic algorithm; extracting characteristics by main ingredient analysis and a partial least squares method; connecting and integrating with linear discriminant analysis, a k-nearest neighbor classifier, a support vector machine, a decision tree, a naive Bayes classifier and an artificial neural network in series to obtain serial integrated classifiers which are connected in parallel and combined; endowing each base classifier with a weighting coefficient according to classification accuracy rate; and taking a fuzzy attribute value as a classification result output mode to obtain the classification judgment result of a target sample. The method has higher classification accuracy rate and robustness on classifying and judging abnormal and normal samples in the differential proteomics, and is suitable for classifying and analyzing the data of differential proteomics.

The invention provides a gait recognition method based on multi-linear mean component analysis. The gait recognition method comprises a training stage and a recognition stage. The training stage refers to performing dimension reduction processing on a half-cycle gait sequence image subjected to linear interpolation through the following algorithms, and training to obtain a projection conversion matrix of the algorithms: projecting the half-cycle gait sequence image into a low-dimension multi-linear subspace through general tensor discriminant analysis; performing further feature extraction by adopting the multi-linear mean component analysis; enabling a training tensor sample to be finally projected into a low-dimension vector space by adopting linear discriminant analysis. According to the recognition stage, conversion matrixes obtained through training the to-be-recognized half-cycle gait sequence image sample with the general tensor discriminant analysis and multi-linear mean component analysis algorithms are subjected to twice projection in a tensor space, the conversion matrix obtained through training the to-be-recognized half-cycle gait sequence image sample with the linear discriminant analysis algorithm is projected in the vector space, and a nearest neighbor classifier is adopted in the vector space during recognition. The gait recognition method can be used for improving the accuracy of gait recognition and has high robustness under different environments.

A leaf identifying method based on manifold learning belongs to the technical field of biological feature identification. The method includes: (1) the collection and the pretreatment on the data of the leaf; (2) the characteristic extraction of the leaf; (3) the training of a sorter and the testing of the leaf data. At first, the leaf data is collected by a device; then de-noising, smoothing, dividing, normalizing, graying and vectorization treatments are carried out; then a semi-monitoring manifold learning arithmetic is used for extracting the linear characteristics from the leaf data; finally, a nearest neighboring sorter is adopted for sorting. A plurality of pre-treatments are carried out on the collected data, thus effectively restraining the effect of noise. The semi-monitoring manifold learning arithmetic can effectively detect the inside structure of the distribution of the leafa data and the monitoring information is led in for improving the separability of the data. The semi-monitoring manifold learning arithmetic is a linear characteristic extraction method which extensively reduces the calculation complexity of the arithmetic.

The invention discloses a high-texture image classification method and a system based on visual saliency detection, wherein, the method comprises the following steps of: segmenting a saliency region through an image pixel saliency value detection method of color contrast; the texture features and gradient features of salient regions are extracted by the complete local binary mode operator and thedirectional gradient histogram algorithm, and the image details are represented by the effective tandem fusion strategy. And classifying the extracted fusion vector by the nearest neighbor classifierto obtain the recognition rate. This method utilizes the visual attention model fitting human visual mechanism to locate the salient target region in the image, in order to get rid of the redundant background information, the high-texture image is segmented, and then the feature of the salient region is extracted and the image classification is completed, so that the different textures in the image (field of view) are automatically separated, which makes the subsequent scene description or target recognition possible.

The invention discloses a face recognition method based on the multi-resolution multi-threshold local binary pattern, and pertains to the technical field of pattern recognition. Through the local binary pattern, spatial structure information of positions of face facial organs is hard to acquire, and at the same time, the LBP operator adopts size relationship between local region neighbour pixels and the central pixel to perform encoding, so the noise effect exists; in order to solve the problems, the method adopts different thresholds to perform LBP image encoding, and finally, face local and overall information is acquired through different division methods, so the MRTLBP characteristic information extraction is more discriminative. The characteristics are used as the face discrimination characteristics for classification and recognition. Classification and recognition are performed by adopting nearest neighbor classification devices. The experimental analysis shows that the method of the invention has good face characteristic expressing capability and has high robustness for illumination, expressions and poses.

The invention discloses a target tracking method based on fuzzy learning. The method mainly comprises the steps that firstly, a sparse representation algorithm is used to track a target in the aspect of a tracker; secondly, in the aspect of a detector, cascade classification detectors (a variance classifier, a collection classifier and a nearest classifier) are used to detect the target position; and finally, in the aspect of learning update, a fuzzy learner is used to integrate the output results of the tracker and the detector, and the final target position is acquired according to the membership of four constraints of time continuity, the spatial uniqueness , the similarity and the target size consistency. According to the target tracking method based on fuzzy learning, the real-time performance is ensured, and at the same time the adaptability to the target illumination change is great; and the discriminant ability of a learner is improved; the tracking accuracy and robustness of the algorithm are improved; and the method has important theoretical and practical significances for the research and practical application development of target tracking.

The invention discloses a human body movement recognition method based on surveillance isometric mapping, and belongs to the field of pattern recognition and computer vision. The human body movement recognition method comprises the following steps; S1, performing foreground extraction through codebook method for the video to acquire a binarized human body foreground image; S2, performing morphology processing and normalization for the human body foreground image to acquire a human body silhouette image; S3, performing periodization analysis for the human body silhouette image sequence, wherein each video is represented by a series of key frames comprising a complete movement period; S4, performing vectorization for the key frames of the human body silhouette image, and performing characteristic dimension reduction through surveillance isometric mapping; S5, recognizing the characteristic after dimension reduction through the nearest categorizer according to Hausdorff distance. The human body movement recognition method breaks through the limitation of the conventional algorithm, and reduces complexity of the algorithm while increasing the categorizing accuracy, thereby being more feasible in practical engineering application.

The invention discloses a handwriting number identification method, which comprises the following steps that an intra-class abutting figure and an inter-class abutting figure are constructed on training samples according to the distance and the label types; a target dimension number and a projection conversion matrix are determined according to a matrix formed by the intra-class abutting figure, the inter-class abutting figure and the training samples; the training samples are converted in to a discrimination sub space according to the projection conversion matrix; the projection conversion matrix is utilized for mapping samples to be tested into the discrimination sub space to obtain test samples; a neighbor classifier is utilized for classifying the test samples. The method has the advantages that each sample is respectively subjected to neighbor finding in identical-class samples and neighbor finding in different-class samples, so the problem of poor classification performance caused by sample distribution unbalance is solved.

The invention relates to a ground moving vehicle target classification and recognition method and system based on a high-resolution distance image, and the method comprises the steps: carrying out the averaging and energy normalization of multiple collected continuous original HRRP echoes of a ground moving vehicle target; calculating the HRRP echo power spectrum features according to the HRRP echoes after averaging and energy normalization; calculating the distances between the HRRP echo power spectrum features and various types of preset power spectrum feature templates, and obtaining a plurality of distances; comparing the distances, and determining the class of the ground moving vehicle target according to a comparison result. The method calculates the distances between the HRRP echo power spectrum features and various types of preset power spectrum feature templates through a nearest neighbor classifier, recognizes the class information of the ground moving vehicle target through comparing the distances between the HRRP echo power spectrum features and various types of preset power spectrum feature templates, achieves the precise classification and recognition of the ground moving vehicles, and provides support for the classification and recognition of military moving vehicles on the ground.

The invention discloses a bi-visual-angle face identification method based on multi-local correlation characteristic learning. Multiple local information in correlation characteristic learning is used to better know real non-linear structures among data, so that the face identification accuracy is improved. The realization process includes the steps of determining a plurality of local blocks of each training sample, constructing a uniform optimization framework of related characteristic learning and multi-local fusion, solving the correlation projection direction and multi-local fusion coefficient through alternative iteration, conducting feature extraction and characteristic fusion for training and testing samples, and using the nearest neighbor classifier for identification. Compared with the prior art, the bi-visual-angle face identification method is more effective and robust.

The invention provides a low-resolution face recognition method coupling gait characteristics. The method includes that all low-resolution faces and gait samples are converted to a vector space, core transformation characteristics of faces and the gait characteristics are calculated by adopting inner product and revised, then low-resolution face images and gait samples indicating the same individual form sample pairs, sample pairs under different collections are close enough, essence geometrical characteristics of a data local structure are kept, a goal optimization model is built and transformed into marks to resolve generalized characteristics to obtain two different transformation matrixes, and the low-resolution face images and gait images obtain new characteristics. During testing, a testing sample vectors first and is mapped to a high-dimensional space from the vector space through the core transformation, and an affiliation category is predicted by adopting the nearest neighbor sorter. The method does not need to estimate high-resolution face images and can conduct identity recognition directly by using a gait coupling machine study method on the low-resolution face images.

The invention relates to a method of detecting life state of a spatial rolling bearing based on vibration-sensitive time-frequency characteristics and belongs to the technical field of reliability estimation and life prediction for rolling bearings. The method includes: establishing a life state characteristic set of the spatial rolling bearing via vibration frequency-domain characteristic parameters to represent the life state; removing non-sensitive characteristics in the life state characteristic set by means of life sensitivity index computation, and establishing a life state sensitive characteristic subset to enhance representation of the life state; subjecting the life state sensitive characteristic subset to dimensionality reduction and redundancy removal by a linear local tangent space alignment algorithm so as to obtain a low-dimensional life state characteristic set with good class characteristics; using a nearest neighbor classifier to recognize different life states of the spatial rolling bearing. The method allows the life state of the spatial rolling bearing to be accurately recognized and is better in application effect.

The invention discloses a two-dimension linearity discrimination analysis face identification method, comprising steps of calculating an affine matrix element according to a chosen training sample matrix, calculating an intra-class divergence matrix and an extra-class divergence matrix according to the affine matrix element, calculating the characteristic value and the characteristic vector of a matrix of the product between the inverse of the intra-class divergence matrix and the extra-class divergence matrix according to the intra-class divergence matrix and the extra-class divergence matrix, solving a projection matrix, using the projection matrix to project the training sample matrix to the projection space to obtain the projected matrix, adopting a neighbor classifier to perform classification processing on the projected matrix and the test sample, and calculating the recognition rate. The invention fuses the advantages of two-dimension linearity discrimination analysis and the two-dimension local projection maintaining method, which not only performs effective dimension reduction on the original data, but also maintain the local characteristics of the data. Furthermore, the invention avoids the fact that the two-dimension structure of the sample data is damaged when the sample data is stretched into one-dimension data, and has high recognition rate and high efficiency.

The invention discloses a large space depth metric learning based gait recognition method. The large space depth metric learning based gait recognition method is characterized by comprising a trainingstage and a recognition stage; the training stage adopts a gait energy image to describe a gait sequence, a convolutional neural network is trained through a large space depth metric learning model,the convolutional neural network is taken as a feature extracting function in the recognition stage, the optimized gait feature is extracted from the gait energy image, and the identity of a gait person is recognized through nearest neighbor classifier. The gait recognition extracts the discriminative gait feature through the large space depth metric learning model, and the gait feature is matchedand recognized through similarity calculation. The large space depth metric learning based gait recognition method can be applied to various fields like intelligent video monitoring and intelligent vision robots, and has a broad application prospect.

The present invention provides a sea surface ship intelligent tracking system and method based on machine vision. The system comprises an image collection module, a Mean-shift tracking module, a Kalman filter module, a multi-stage parallel detection module and online learning module. The sea surface ship intelligent tracking system and method based on the machine vision employs a Haar+AdaBoost combination algorithm to detect the sea surface ships, and the detection result is taken as the starting frame of the tracker so as to initialize the tracker and the detector and replace the manual delineation of a target area. The Kalman filter is configured for the detection and the tracking of the ship so as to shorten the detection range, reduce the calculation amount and improve the timeliness. Based on the Mean-shift tracking module, the stable feature, anti-shielding, simple calculation and insensitivity of target deformation, rotation and background movement; the multi-stage parallel detection module is employed to combine the set classifier based on the random forest, the Haar classifier and the nearest neighbor classifier so as to solve the difficulty that the ship tracking is failed caused by shielding and deformation and improve the tracking timeliness, the accuracy and the robustness.

The present invention relates to a rolling bearing service life state same-scale characterization and recognition method under different rotating speeds. The method is characterized in that a dimensionless parameter which is low in rotating speed sensitivity and is irrelevant with the vibration energy and a multi-domain information entropy are utilized to reflect the service life state of a rolling bearing, and the dimensionless parameter and the multi-domain information entropy are fused to construct a rolling bearing service life state characteristic set, thereby realizing the same-scale quantitative characterization of the service life state under different rotating speeds; a service life sensitive index algorithm is used to filter the characteristics poor in service life sensibility, the characteristics good in service life sensibility are selected to construct a service life state sensitive characteristic set having stronger characterization capability, and an orthogonal neighborhood preserving embedding algorithm is used for the nonlinear dimensionality reduction and is used to remove the redundant information to thereby obtain a low-dimension service life state sensitive characteristic set of good classification characteristic; and then a weighted nearest neighbor classifier of good robustness is applied to realize the classification recognition of different service life states, thereby finally realizing the rolling bearing service life state same-scale characterization and recognition method under different rotating speeds. The method enables the rolling bearing service life state to be recognized accurately under different rotating speeds, and has a better application effect.

The invention discloses a foundation cloud picture classification method based on a completion local three value model. The method comprises the following steps that the local information of each training sample is decomposed into local difference value vectors and center pixels; each local difference value vector is decomposed into the products of sign vectors and amplitude vectors; the three-value mode coding is adopted for the sign vectors, the amplitude vectors and the center pixels, and in addition, the rotating unchanged consistency characteristics are respectively calculated; the rotating unchanged consistency characteristics are merged to obtain the final characteristic expression of the training samples; the final characteristic expression of the foundation cloud picture is calculated; and on the basis of the final characteristic expression of the foundation cloud picture and the training samples, the nearest adjacent classifier is adopted to obtain the classification results of the tested foundation cloud picture. The foundation cloud picture classification method has the advantages that the local information of images is considered in three aspects of sign, amplitude and center pixels, the local three-value mode is adopted for coding, and the final coding is carried out to obtain the final characteristic expression of the images, so better noise robustness and classification accuracy can be obtained.

The invention discloses a cross-domain image classification method based on coupling knowledge migration, and the method comprises the steps: searching a common low-dimensional subspace of a source domain and a target domain based on a maximum mean value difference criterion, and eliminating the difference between the data edge distribution and class condition distribution of the source domain andthe target domain; constructing respective adjacent graphs according to the label information of the source domain data and the pseudo label information of the target domain data, keeping the structural consistency of the data from the original space to the low-dimensional subspace, dynamically adjusting the structures of the adjacent graphs, and promoting the positive migration of knowledge in the domain; training a nearest neighbor classifier by utilizing the source domain data with the label information in the low-dimensional subspace, and carrying out continuous iterative optimization onthe pseudo label information of the target domain data to obtain final label information of the target domain data, thereby completing cross-domain image classification; in addition, according to themethod, different confidence coefficients are given to the pseudo tags of the target domain image by designing a sample reweighting strategy, so that the negative migration of knowledge in the domainis effectively reduced, and the cross-domain image classification precision is improved.

The invention discloses a single-sample face recognition method based on Gabor feature extraction and spatial transformation, and mainly solves the problem that a traditional face recognition method can not be applied since a within class scatter matrix is zero under a condition that only single training sample image is in the presence. The method comprises the following steps that: adopting a Gabor wavelet to extract a spatial feature vector from an original single sample image; then, carrying out fusion on the extracted spatial feature vector with an original spectral feature vector, utilizing a feature spatial transformation method to carry out low-dimension feature space transformation on a fused feature matrix, and transforming the fused feature matrix to a low-dimension subspace; and finally, utilizing a nearest neighbor classifier to finish recognition. By use of the method, single-sample face recognition can be accurately finished, recognition accuracy is improved, and calculation cost is lowered. Compared with the prior art, the face recognition method which is put forward by the invention exhibits better effectiveness and robustness.

The invention relates to an illumination face recognition method based on a completed local convex-and-concave pattern, and belongs to the pattern recognition field. The method includes the steps of dividing an image; carrying out bilinear interpolation for each piece of the image; encoding the symbol characteristic and the amplitude characteristic of a local difference of each pixel point in each piece of the image to obtain a symbol characteristic matrix and an amplitude characteristic matrix of each piece of the image; encoding pixel points of each piece of the image to obtain a central pixel characteristic matrix of each piece of the image, extracting the histogram characteristics of the three characteristic matrixes to obtain three characteristic vectors, and successively connecting the three characteristic vectors to obtain histogram characteristic vectors of all pieces of the image; and finally connecting the histogram characteristic vectors of all pieces of the image to obtain a histogram characteristic vector of the original image, sending the characteristic vector to the nearest neighboring classifier to be classified, and verifying the identity of an original face image. The method is an image texture description method based on second-order differential, and is capable of effectively identifying human faces in an illumination environment.