CN109784142B - Hyperspectral target detection method based on conditional random projection - Google Patents
Hyperspectral target detection method based on conditional random projection Download PDFInfo
- Publication number
- CN109784142B CN109784142B CN201811422749.XA CN201811422749A CN109784142B CN 109784142 B CN109784142 B CN 109784142B CN 201811422749 A CN201811422749 A CN 201811422749A CN 109784142 B CN109784142 B CN 109784142B
- Authority
- CN
- China
- Prior art keywords
- vector
- conditional random
- feature
- sample
- projection
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
Images
Landscapes
- Image Analysis (AREA)
- Investigating Or Analysing Materials By Optical Means (AREA)
Abstract
The invention provides a hyperspectral target detection method based on conditional random projection characteristics, which can realize accurate and rapid detection of targets and can improve the sensitivity to the number of training samples. A supervised feature selection and feature extraction method based on conditional random projection is provided in the aspect of representation of hyperspectral image spectral data. A projection parameter matrix related to the data and the label is obtained, and conditional random characteristics are obtained through the projection parameter matrix, so that the accuracy of target detection is improved; meanwhile, screening estimation and sampling are added in the whole operation obtaining process, the time for the operation of screening estimation and sampling is short, the training time is short, the operation of screening estimation and sampling is completed in an off-line stage, and the result obtained by training is directly used in actual operation, so that the method has the advantages of high detection accuracy and high test operation efficiency.
Description
Technical Field
The invention belongs to the technical field of remote sensing image processing, and relates to a hyperspectral remote sensing target rapid detection method based on a conditional random projection feature extraction strategy.
Background
With the rapid development of the hyperspectral remote sensing technology, the application of hyperspectrum in military and civil life is more and more extensive, so that the requirement of hyperspectral data processing aiming at specific application environments is higher and higher. The hyperspectral remote sensing image typical ground object target detection is an important research direction, and has significance in urban configuration scientificity, agricultural planting distribution planning and military sensitive target classification extraction.
Currently, in the field of hyperspectral surface object typical target detection, detection problems are mainly converted into data classification problems, and the following two main classification methods are relied on:
the method comprises the following steps: the higher high-resolution spectral information of the hyperspectral remote sensing image means that the spectral data contain more detailed information which can be used as a basis for determining the ground object type. The method is realized by adopting a threshold segmentation idea on a target surface material characteristic spectrum section (such as the characteristic of steel on the surface of an airplane or a ship in a near-infrared spectrum section is obvious) of hyperspectral observation data, and comprises self-adaptive threshold segmentation methods such as Otsu and the like. However, due to the influence of factors such as an imaging angle and an atmospheric environment, a spectrum of a complex ground object background may be similar to a target spectrum (for example, ground objects such as target shadows and building roof asphalt have similar spectral properties to a target in a near-infrared spectral band, namely, a foreign object co-spectrum), so that the problem of over-high false alarm is easily caused in the threshold segmentation-based method; furthermore, using only data of a specific spectral band is also a serious waste of hyperspectral observation data.
Secondly, a supervision method is adopted: the method is a classification method commonly used in the field of hyperspectral remote sensing image processing. In order to solve the problems of foreign body same spectrum, same object different spectrum and the like, the idea of supervised sample learning and training is adopted, the function mapping relation between target spectrum information and a target label is directly fitted, and a target identification task is completed. Typical supervised models are neural networks, support vector machines, etc.
The hyperspectral ground object target detection method based on the deep learning model which is widely applied at present belongs to one of supervised methods, and has been greatly improved in the aspect of target detection performance. However, the existing target detection method based on the deep learning model has inherent defects, which are mainly expressed in the following aspects:
firstly, as for the deep learning model, the training time is very long and tedious, massive training samples need to be consumed, large-scale parallel computing resources are required to be relied on, and the black box attribute is still obvious as not fully proven in theory. Specifically, the robustness of the depth model object detection is improved at the cost of improving the model Capacity (Capacity). The model structure is complex and huge, the parameters are numerous, the training and actual operation calculation amount of the model is huge, and the model depends on cluster parallel computing platforms such as FPGA, GPU and the like. On the basis, the feature learning model with the characteristics of high capacity and most parameters needs to be trained based on a large number of samples so as to prevent the over-fitting phenomenon.
Secondly, the problem of target detection is limited by directly applying a deep learning technology. Deep learning techniques essentially solve the data classification problem. Each category in the classification task is a closed set, and cost functions of the models treat various types of data without distinction, so that the data of different categories have Cluster structures (Cluster) in the constructed feature space and the constructed output space; in the pattern recognition task, the positive sample set is a 'closed set', the negative sample set is an 'open set', and the cost function of the method is to reduce the characteristic distance of the positive sample (target) data and increase the characteristic distance between the positive sample and the negative sample. In the hyperspectral target spectrum detection task, target spectrum data are clear, but the range of negative samples is too wide, and the feature extraction and matching process of the 'positive/negative samples' is not reasonable without the difference of a deep learning model.
Therefore, how to combine the essential difference and the connection between the data classification and the target detection task, give full play to the advantages of the data classification and the target detection task, how to quickly and effectively realize the detection of the target under the condition of ensuring the detection accuracy because the number of training samples is not sufficient, and is a key problem in the field of hyperspectral image processing at present.
Disclosure of Invention
In view of this, the invention provides a hyperspectral target detection method based on conditional random projection features, which can realize accurate and rapid detection of targets and can improve the sensitivity to the number of training samples.
In order to achieve the above object, the present invention provides a hyperspectral target detection method based on conditional random projection, which comprises the following steps:
step 1, collecting randomly generated spectrum vectors to form a training set and normalizing the training set to obtain sample spectrum data and corresponding labels thereof; the sample spectral data comprises positive sample data and negative sample data;
step 2, projecting the sample spectrum data by using a randomly generated projection parameter matrix to obtain a random feature vector;
step 3, screening the feature dimension of the random feature vector according to the label information of the spectral data, so that the geometric position relationship of the screened random feature vector in the feature space is consistent with the geometric position relationship of the corresponding label in the output space;
step 4, collecting column vectors corresponding to the screened random feature vectors in a projection parameter matrix, and estimating probability distribution of the column vectors;
step 5, sampling the column vectors according to the probability distribution obtained in the step 4, taking the sampling result as a transformation parameter, and projecting the sample spectrum data and the spectrum vectors to be detected respectively by using the transformation parameter to obtain respective conditional random characteristics of the sample spectrum data and the spectrum vectors to be detected;
step 6, based on the conditional random features of the sample data, constructing a probability density distribution function of the conditional random features of the spectral vector to be detected; based on the probability density distribution function, the probability of the conditional random features of the spectral vector to be detected is obtained, the spectral vector to be detected, which is higher than the probability of the conditional random features of the spectral vector to be detected by a set threshold value, is judged as a target, and hyperspectral target detection is realized.
In step 6, the probability density distribution function of the conditional random features of the spectral vector to be measured is constructed only based on the conditional random features of the positive sample data.
Wherein, the generalized Fourier transform method based on Walsh-Hadamard transform in the step 2 obtains random feature vectors.
In step 3, feature dimension screening is realized by maximizing an objective function, where the objective function is:
s.t.q＞0,qT1＝1
wherein, superscript T represents transposition; the vector q is used for measuring the strength of the correlation between the random feature vector and the label information; the vector y is a label vector corresponding to the input sample, and the vector y-Is the label vector corresponding to the negative sample, and the random feature vector phi is the feature mapping vector of the sample data
Mapping vectors for features of negative sample data; (| | q-q)ref||2P) is a regular term, where the vector qrefIs an a priori estimate of a vector q, each column vector being 1/k, k being the length of the vector q, p being the vector qrefAn upper error bound with vector q; λ is the weight of the regularization term;
the characteristic dimension screening mode is as follows: and (3) maximizing the objective function by adjusting the vector q, wherein when the objective function is maximized, the random characteristic dimension corresponding to the non-zero term column vector in the vector q is the characteristic dimension needing to be reserved.
The projection parameters corresponding to the feature dimensions reserved after screening are subjected to Gaussian mixture distribution with the composition of 2, and the mean value, covariance and component probability of the Gaussian mixture distribution are estimated by using an expectation maximization algorithm in an iterative approximation mode, so that the conditional probability distribution of the projection parameters is obtained.
Wherein, the step 5 adopts a Markov method to sample the random variable.
In step 6, a gaussian kernel function K is used for windowing, and a probability density distribution function of the conditional random features of the spectral vector to be measured is constructed based on the conditional random features:
wherein x is the conditional random feature of the spectral vector to be measured, xiIs a conditional random feature of the ith positive sample spectral vector, i ═ 1,2,3 … Np,NpIs the total number of positive samples.
In step 6, the set threshold is a minimum value of the probability density values obtained by substituting each positive sample in the training samples as the spectral vector to be measured into the probability density distribution function.
Has the advantages that:
1. the method provides a supervised feature selection and feature extraction method based on conditional random projection in the aspect of representation of hyperspectral image spectral data. A projection parameter matrix related to the data and the label is obtained, and conditional random characteristics are obtained through the projection parameter matrix, so that the accuracy of target detection is improved; meanwhile, screening estimation and sampling are added in the whole operation obtaining process, the time for the operation of screening estimation and sampling is short, the training time is short, the operation of screening estimation and sampling is completed in an off-line stage, and the result obtained by training is directly used in actual operation, so that the method has the advantages of high detection accuracy and high test operation efficiency.
2. The method fully considers the difference between the target detection problem and the data classification problem, utilizes the data difference between the positive sample and the negative sample to construct an optimization model different from the deep learning technology, can theoretically construct the characteristic space of the positive sample, directly judges the similarity between the sample to be detected and the positive sample through the characteristic space of the positive sample, judges whether the target to be detected is the target or not, improves the sensitivity to the number of training samples, and the identification mode is more in line with the learning and cognition habits of a human intelligent system on target data.
Drawings
FIG. 1 is a flow chart of the present invention.
FIG. 2 is a schematic diagram of a positive sample generation characterization and feature matching method according to the present invention.
Detailed Description
The invention is described in detail below by way of example with reference to the accompanying drawings. The flowchart of this embodiment is shown in fig. 1, and includes the following steps:
step 1, collecting spectral vectors of an image to be processed to form a training set and normalizing the training set to obtain sample spectral data and a corresponding label; the sample spectral data comprises positive sample data and negative sample data;
the specific process is as follows: for the image to be processed acquired by the same hyperspectral detector, extracting the spectral vector corresponding to each pixel point under the spatial dimension, dividing the spectral vector into spectral data containing a target (positive sample) and a background (negative sample), and marking the spectral data. The marked spectrum vector is divided into a training set and a test set according to a required proportion, and a schematic diagram of a positive sample generating type feature description and feature matching method is shown in fig. 2.
Normalizing each dimension of the spectrum vector: because the measurement units of the indexes in the index system are different, in order to enable all the indexes to participate in evaluation calculation, all the indexes need to be subjected to standardization processing, and the numerical values of the indexes are mapped to a certain numerical value interval through function transformation. The dimensions of the spectral vector (the D dimensions) are scaled to fall within a small specific interval: the mean value of each dimension is 0 and the variance is 1; the normalization is as follows:
xn＝(xn-μn)/sn
wherein xnIs a vector, mu, formed by a certain spectral dimension of the N spectral vectorsnIs the corresponding mean value, snThe standard deviation is that N is 1,2,3 … … D, N is the total number of pixel points, and N pixel points correspond to N spectral vectors, and the specific implementation function can be realized by the zscore function in Matlab.
Traversing each dimension of the spectral vector corresponding to the image to be processed to obtain a sample matrix X formed by the spectral feature vectors after N X D normalization, wherein the sample matrix X is sample spectral data;
step 2, carrying out random projection on the sample matrix X by using the set projection parameter matrix omega to obtain a random eigenvector of the sample spectrum data; the probability distribution obeyed by the projection parameter matrix is irrelevant to the sample spectrum data and the labels corresponding to the sample spectrum data, and the important difference exists between the probability distribution obeyed by the projection parameter matrix and the labels corresponding to the sample spectrum data and the fine iterative tuning of all projection parameters according to the sample labels in the deep learning technology. The whole random projection process can be represented by the following formula:
φ(X,ω)
specifically, firstly, a D x k dimensional projection parameter matrix omega is randomly generated, k is a positive integer, the value is the length of a random feature vector to be obtained, each column vector in the projection parameter matrix omega is independently and identically distributed and follows standard normal distribution, the specific realization can depend on a randn function in matlab, and the process has a very fast algorithm, so that the process can be efficiently and quickly completed. Multiplying the sample matrix X with a projection parameter matrix omega to obtain X omega, and obtaining N k-dimensional characteristic vectors; and (3) each column vector of the N k-dimensional feature vectors is substituted into a nonlinear activation function phi (such as sigmoid, tanh and the like) to obtain a random feature vector.
When the multiplication (X omega) of the sample matrix X and the projection parameter matrix omega is realized, data can be directly mapped into a high-dimensional space, or the inner product operation in the high-dimensional space is directly converted into the original low-dimensional space for calculation by adopting a kernel function method, the mapped result does not need to be explicitly written, and even the situation of infinite dimensionality can be processed; the optimal scheme is to adopt a fast operation method (such as a generalized Fourier transform method based on Walsh-Hadamard transform), and has extremely high operation efficiency and storage advantages even if high-dimensional mapping is carried out (k is far more than D).
Step 3, screening the feature dimension of the random feature vector according to the label information of the spectral data, so that the geometric position relationship of the screened random feature vector in the feature space is consistent with the geometric position relationship of the corresponding label in the output space;
in feature dimension screening, random features of positive sample data have a clustering structure in a feature space, and the distance between positive and negative samples is increased, so that if negative samples are clustered in the feature space, the optimization of feature mapping of the positive samples is influenced, and therefore, the negative samples are not clustered in the feature space. Specifically, feature dimension screening is achieved by maximizing an objective function, which is:
s.t.q＞0,qT1＝1
wherein, superscript T represents transposition; the vector q is used for measuring the strength of the correlation between the random feature vector and the label information; the vector y is a label vector corresponding to the input sample, and the vector y-Is the label vector corresponding to the negative sample, and the random feature vector phi is the feature mapping vector of the sample data
Mapping vectors for features of negative sample data; (| | q-q)ref||2P) is a regular term, where the vector qrefIs an a priori estimate of a vector q, each column vector being 1/k, k being the length of the vector q, p being the vector qrefAn upper error bound with vector q; λ is the weight of the regularization term;
the characteristic dimension screening mode is as follows: and (3) maximizing the objective function by adjusting the vector q, wherein when the objective function is maximized, the random characteristic dimension corresponding to the non-zero term column vector in the vector q is the dimension needing to be reserved.
The process of adjusting the vector q is a classical linear programming process and can be quickly solved by a standard CVX convex optimization tool box.
Step 4, collecting column vectors corresponding to the screened random feature vectors in the projection parameter matrix and estimating the probability distribution of the column vectors as the probability distribution of the projection parameters is irrelevant to both the data and the labels;
specifically, the projection parameters corresponding to the feature dimensions retained after the screening are a plurality of column vectors, which are denoted as ω'. These high-dimensional vectors can be assumed to follow a mixture gaussian distribution with a composition of 2. While the probability modelImplicit variables exist inside the model, resulting in an unresolved solution for the maximum likelihood estimation of the distribution parameters. Therefore, the mean of the Gaussian mixture distribution (component 1/2 mean vector μ) is estimated here by means of iterative approximation using the expectation-maximization (EM) algorithm1、μ2) Covariance (component 1/2 covariance matrix Σ1，∑2) With component probability pi1、π2From this, a conditional probability distribution (gaussian mixture distribution) of the projection parameters is derived:
p(ω′)＝π1p(ω′；μ1,∑1)+π2p(ω′；μ2,∑2)
step 5, sampling the column vectors according to the probability distribution obtained in the step 4, taking the sampling result as a transformation parameter, and projecting the spectrum vectors to be detected and the sample spectrum data respectively by using the transformation parameter to obtain respective conditional random characteristics;
this procedure is equivalent to the fact that, given a known probability distribution function P (ω '), the random variable ω' is sampled so as to satisfy the probability distribution P (ω '), so that a markov method (MCMC sampling) can be used, i.e. if a markov chain with a probability transition matrix P can be constructed, the distribution of which is stationary is the known probability distribution P (ω'). Arbitrary initial state s0Performing state transition on the Markov chain to obtain a series of random sequences s0，s1，sn+1…. If the Markov chain converges from the nth step, then from snThe initially obtained states all satisfy the probability distribution p (ω), i.e. the random variables that need to be sampled that satisfy a given probability distribution p (ω). Based on the conditional probability distribution p (ω ') of the projection parameters obtained in step 4, the algorithm can be used to sample the random variable ω' to obtain the required transformation parameters. The transformation parameters can be obtained by the method, and the operation in the step 2 is repeated to project the spectral vector to be measured and the sample spectral data, so that the conditional random characteristics of the spectral vector to be measured and the sample spectral data are obtained.
Step 6, based on the conditional random features of the sample data, constructing a probability density distribution function of the conditional random features of the spectral vector to be detected; based on the probability density distribution function, the probability of the conditional random features of the spectral vector to be detected is obtained, the spectral vector to be detected, which is higher than the probability of the conditional random features of the spectral vector to be detected by a set threshold value, is judged as a target, and hyperspectral target detection is realized.
The threshold value may be set as a minimum value of probability density values obtained by substituting each positive sample in the training samples as the spectral vector to be measured into the probability density distribution function.
Since the random features of the positive samples have a cluster structure in the feature space in steps 3 to 5, the higher the cluster density is, the more "positive sample commonality" is reflected, and the greater the probability of belonging to the positive sample is. Feature matching in deep learning models is essentially a "method of elimination" based on discriminant ideas (inter-class differences): a "positive sample" is confirmed by excluding the possibility that the input feature is a "negative sample" and vice versa. In the target recognition task, although the "positive sample" may constitute a "closed space", the space constituted by the "negative sample" is open. Feature matching in this case should focus on the similarity between the input features and the "positive samples". Therefore, only the conditional random features of the sample data are subjected to probability density estimation to form a probability density distribution function of the sample data. And performing conditional random projection on the spectral vector to be identified to obtain conditional random features, calculating the probability that the conditional random features belong to the positive sample, and judging that the spectral vector to be detected is the target when the probability value is higher than a set threshold value, thereby completing target detection.
The probability density distribution function of the conditional random features of the spectral vector to be measured can be obtained by windowing the Gaussian kernel function K:
wherein, x is the condition random feature of the spectral vector to be measured, xiIs the bar of the ith positive sample spectral vectorRandom features, i ═ 1,2,3 … Np,NpIs the total number of positive samples.
The above description is only a preferred embodiment of the present invention, and is not intended to limit the scope of the present invention. Any modification, equivalent replacement, or improvement made within the spirit and principle of the present invention should be included in the protection scope of the present invention.
Claims (8)
1. A hyperspectral target detection method based on conditional random projection is characterized by comprising the following steps:
step 1, collecting randomly generated spectrum vectors to form a training set and normalizing the training set to obtain sample spectrum data and corresponding labels thereof; the sample spectral data comprises positive sample data and negative sample data;
step 2, projecting the sample spectrum data by using a randomly generated projection parameter matrix to obtain a random feature vector;
step 3, screening the feature dimension of the random feature vector according to the label information of the spectral data, so that the geometric position relationship of the screened random feature vector in the feature space is consistent with the geometric position relationship of the corresponding label in the output space;
step 4, collecting column vectors corresponding to the screened random feature vectors in a projection parameter matrix, and estimating probability distribution of the column vectors;
step 5, sampling the column vectors according to the probability distribution obtained in the step 4, taking the sampling result as a transformation parameter, and projecting the sample spectrum data and the spectrum vectors to be detected respectively by using the transformation parameter to obtain respective conditional random characteristics of the sample spectrum data and the spectrum vectors to be detected;
step 6, based on the conditional random features of the sample data, constructing a probability density distribution function of the conditional random features of the spectral vector to be detected; based on the probability density distribution function, the probability of the conditional random features of the spectral vector to be detected is obtained, the spectral vector to be detected, which is higher than the probability of the conditional random features of the spectral vector to be detected by a set threshold value, is judged as a target, and hyperspectral target detection is realized.
2. The hyperspectral target detection method based on conditional random projection as claimed in claim 1, wherein in the step 6, a probability density distribution function of the conditional random features of the spectral vector to be measured is constructed only based on the conditional random features of the positive sample data.
3. The hyperspectral target detection method based on conditional random projection as claimed in claim 1 or 2, wherein the generalized Fourier transform method based on Walsh-Hadamard transform in step 2 obtains random feature vectors.
4. The hyperspectral target detection method based on conditional random projection as claimed in claim 1 or 2, wherein in the step 3, the feature dimension screening is realized by maximizing an objective function, wherein the objective function is as follows:
s.t.q＞0,qT1＝1
wherein, superscript T represents transposition; the vector q is used for measuring the strength of the correlation between the random feature vector and the label information; the vector y is a label vector corresponding to the input sample, and the vector y-For the label vector corresponding to the negative examples, vector phiTMapping a transpose of a vector for features of sample data, the vector
Transpose of feature mapping vector for negative sample data; (| | q-q)ref||2P) is a regular term, where the vector qrefIs an a priori estimate of a vector q, each column vector being 1/k, k being the length of the vector q, p being the vector qrefAn upper error bound with vector q; λ is the weight of the regularization term;
the characteristic dimension screening mode is as follows: and (3) maximizing the objective function by adjusting the vector q, wherein when the objective function is maximized, the random characteristic dimension corresponding to the non-zero term column vector in the vector q is the characteristic dimension needing to be reserved.
5. The hyperspectral target detection method based on conditional random projection as claimed in claim 1 or 2, wherein the projection parameters corresponding to the feature dimensions retained after screening are subjected to Gaussian mixture distribution with composition of 2, and the mean, covariance and component probability of the Gaussian mixture distribution are estimated by using an expectation maximization algorithm in an iterative approximation manner, thereby obtaining the conditional probability distribution of the projection parameters.
6. The hyperspectral target detection method based on conditional random projection as claimed in claim 1 or 2, wherein the markov method is used to sample the random variable in step 5.
7. The hyperspectral target detection method based on conditional random projection as claimed in claim 2, wherein in the step 6, a gaussian kernel function K is used for windowing, and a probability density distribution function of conditional random features of a spectral vector to be detected is constructed based on the conditional random features:
wherein x is the conditional random feature of the spectral vector to be measured, xiIs a conditional random feature of the ith positive sample spectral vector, i ═ 1,2,3 … Np,NpIs the total number of positive samples.
8. The method as claimed in claim 1, wherein in step 6, the threshold is set as a minimum value of probability density values obtained by substituting each positive sample in the training samples as the spectral vector to be measured into the probability density distribution function.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811422749.XA CN109784142B (en) | 2018-11-27 | 2018-11-27 | Hyperspectral target detection method based on conditional random projection |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN201811422749.XA CN109784142B (en) | 2018-11-27 | 2018-11-27 | Hyperspectral target detection method based on conditional random projection |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN109784142A CN109784142A (en) | 2019-05-21 |
| CN109784142B true CN109784142B (en) | 2020-12-04 |
Family
ID=66496440
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN201811422749.XA Active CN109784142B (en) | 2018-11-27 | 2018-11-27 | Hyperspectral target detection method based on conditional random projection |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN109784142B (en) |
Families Citing this family (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN110197481B (en) * | 2019-05-30 | 2021-04-09 | 苏州大学 | Graphene fingerprint peak analysis method based on big data analysis |
| CN112907589A (en) * | 2021-04-02 | 2021-06-04 | 联通(上海)产业互联网有限公司 | Deep learning algorithm for detecting abnormality and segmenting abnormal region in image |
| CN113111587A (en) * | 2021-04-20 | 2021-07-13 | 北京理工雷科电子信息技术有限公司 | Reusable and extensible machine learning method based on plug-in model |
Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102314685A (en) * | 2011-07-23 | 2012-01-11 | 北京航空航天大学 | Hyperspectral image sparse unmixing method based on random projection |
| CN104268856A (en) * | 2014-09-15 | 2015-01-07 | 西安电子科技大学 | Method for extracting pixel purity index based on end member of image processor |
| CN104484681A (en) * | 2014-10-24 | 2015-04-01 | 西安电子科技大学 | Hyperspectral remote sensing image classification method based on space information and ensemble learning |
| CN105784113A (en) * | 2016-03-08 | 2016-07-20 | 深圳市未来媒体技术研究院 | Imaging system and imaging method for differentiating fluorescent radiation spectral image from reflective spectral image |
| CN108197650A (en) * | 2017-12-30 | 2018-06-22 | 南京理工大学 | The high spectrum image extreme learning machine clustering method that local similarity is kept |
-
2018
- 2018-11-27 CN CN201811422749.XA patent/CN109784142B/en active Active
Patent Citations (5)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN102314685A (en) * | 2011-07-23 | 2012-01-11 | 北京航空航天大学 | Hyperspectral image sparse unmixing method based on random projection |
| CN104268856A (en) * | 2014-09-15 | 2015-01-07 | 西安电子科技大学 | Method for extracting pixel purity index based on end member of image processor |
| CN104484681A (en) * | 2014-10-24 | 2015-04-01 | 西安电子科技大学 | Hyperspectral remote sensing image classification method based on space information and ensemble learning |
| CN105784113A (en) * | 2016-03-08 | 2016-07-20 | 深圳市未来媒体技术研究院 | Imaging system and imaging method for differentiating fluorescent radiation spectral image from reflective spectral image |
| CN108197650A (en) * | 2017-12-30 | 2018-06-22 | 南京理工大学 | The high spectrum image extreme learning machine clustering method that local similarity is kept |
Non-Patent Citations (4)
| Title |
|---|
| An Insight into Extreme Learning Machines: Random Neurons,Random Features and Kernels;Guang-Bin Huang;《Cognitive Computation》;20140403;第376-390页 * |
| Extreme learning machine and adaptive sparse representation for image classification;JiuwenCao等;《ELSEVIER》;20160623;第91-102页 * |
| Nonlinear Dimensionality Reduction via the ENH-LTSA Method for Hyperspectral Image Classification;Weiwei Sun等;《IEEE JOURNAL OF SELECTED TOPICS IN APPLIED EARTH OBSERVATIONS AND REMOTE SENSING》;20140228;第7卷(第2期);第375-388页 * |
| 基于条件随机场的高光谱遥感影像农作物精细分类;余铭等;《中国农业信息》;20180630;第30卷(第3期);第74-82页 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN109784142A (en) | 2019-05-21 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Wang et al. | Remote sensing image classification based on the optimal support vector machine and modified binary coded ant colony optimization algorithm | |
| CN108319964B (en) | Fire image recognition method based on mixed features and manifold learning | |
| Kim et al. | Color–texture segmentation using unsupervised graph cuts | |
| US20180144182A1 (en) | Analyzing digital holographic microscopy data for hematology applications | |
| CN101140624A (en) | Image matching method | |
| CN109784142B (en) | Hyperspectral target detection method based on conditional random projection | |
| Bhattacharya et al. | Outlier detection using neighborhood rank difference | |
| CN110298235B (en) | Hyperspectral anomaly detection method and system based on manifold constraint self-coding network | |
| CN108492298B (en) | Multispectral image change detection method based on generation countermeasure network | |
| Yousefi et al. | Comparison of different algorithms for land use mapping in dry climate using satellite images: a case study of the Central regions of Iran | |
| CN103955709B (en) | Weighted synthetic kernel and triple markov field (TMF) based polarimetric synthetic aperture radar (SAR) image classification method | |
| Tao et al. | Density-sensitive fuzzy kernel maximum entropy clustering algorithm | |
| CN113723492A (en) | Hyperspectral image semi-supervised classification method and device for improving active deep learning | |
| CN113221065A (en) | Data density estimation and regression method, corresponding device, electronic device, and medium | |
| Zou et al. | Survey on clustering-based image segmentation techniques | |
| Berge et al. | Structured Gaussian components for hyperspectral image classification | |
| CN110097067B (en) | Weak supervision fine-grained image classification method based on layer-feed feature transformation | |
| CN113111969B (en) | Hyperspectral image classification method based on mixed measurement | |
| Veracini et al. | Fully unsupervised learning of Gaussian mixtures for anomaly detection in hyperspectral imagery | |
| Zhao et al. | Spectral–spatial classification of hyperspectral images using trilateral filter and stacked sparse autoencoder | |
| Han et al. | Automatic binary and ternary change detection in SAR images based on evolutionary multiobjective optimization | |
| Yang et al. | Supervised land-cover classification of TerraSAR-X imagery over urban areas using extremely randomized clustering forests | |
| CN112966710B (en) | FY-3D infrared hyperspectral cloud detection method based on linear discriminant analysis | |
| Ding et al. | Ensemble classification of hyperspectral images by integrating spectral and texture features | |
| Yang et al. | Progressive neighbors pursuit for radar images classification |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |