WO2019090509A1

WO2019090509A1 - Hyperspectral image classification method and system

Info

Publication number: WO2019090509A1
Application number: PCT/CN2017/109914
Authority: WO
Inventors: 李岩山; 王贤辰; 谢维信
Original assignee: 深圳大学
Priority date: 2017-11-08
Filing date: 2017-11-08
Publication date: 2019-05-16

Abstract

The present invention is applicable to image classification. Provided is a hyperspectral image classification method, comprising: dividing hyperspectral images into a training set and a test set and extracting local feature points; calculating the local feature points of the training set by means of a K-means algorithm to form a dictionary; forming the nearest neighbor words in the dictionary for local feature points to be classified of the test set by using a KNN algorithm, and searching for the nearest neighbor feature points for feature points to be classified of the images in the test set; searching for the neighbor word with the shortest spectral dimension distance from among the nearest neighbor words; introducing the three constraints of the neighbor feature points, the neighbor words and the spectral dimension distance; solving the constraint least absolute fitting problem to obtain a coding coefficient; and pooling the coding coefficient by means of a max-pooling algorithm and classifying the test set by taking the obtained coding coefficient as a feature descriptor of the hyperspectral images. The problem of uncertainty when a mapping relationship between a hyperspectral image feature point and a dictionary word is established is solved, and the discernment of similar images is improved.

Description

Method and system for classifying hyperspectral image

Technical field

The invention belongs to the technical field of image processing, and in particular relates to a method and a system for classifying hyperspectral images.

Background technique

Compared with grayscale and RGB color maps, hyperspectral images contain spatial and spectral information, large amounts of data, and have important applications in detecting camouflage and concealed military targets, as well as in civilian search and rescue targets. With the development of hyperspectral technology, the current hyperspectral image exhibits high spatial resolution. The object target has rich texture and structural information on the hyperspectral image, and the spectral information it contains is also very complicated and rich.

As a major breakthrough in the field of remote sensing, hyperspectral imagery uses many narrow electromagnetic wave bands to obtain useful information from objects of interest. While retaining high spatial resolution, its spectral resolution is greatly improved, reaching nanometers. An order of magnitude that can be used to detect and identify undetectable feature categories in traditional panchromatic and multispectral remote sensing. Compared with traditional multi-spectral remote sensing images, hyperspectral remote sensing images have the characteristics of large amount of information and high spectral resolution, which greatly enhances the ability to describe and distinguish the types of features. It makes the original in multi-spectral remote sensing. Ground objects that cannot be effectively detected are detected. However, due to the high data dimension of hyperspectral images, conventional image classification methods have great limitations in processing hyperspectral images. How to quickly and accurately mine the required information from a large number of hyperspectral data to achieve high The classification of precision is still an urgent problem to be solved.

Traditional hyperspectral image classification mainly uses pixel-level classification methods. Traditional remote sensing monitoring methods are mainly composed of medium and low resolution. In hyperspectral images, the size of an object on an image is the same size as a pixel. The pixel-level classification method is suitable for such hyperspectral image classification. However, with the development of hyperspectral sensor technology, the resolution of space and spectrum in hyperspectral images has been greatly improved. Traditional pixel fraction at this time Class methods have not been well adapted to hyperspectral classification.

The word bag model is widely used due to its simplified representation of hyperspectral images and efficient coding of image features and visual words. In general, the classification method of the word bag model is shown in Figure 1. It is mainly divided into the following steps: 1 Extracting feature points from the image and describing them (such as SIFT features, Scale-invariant feature transform, scale-invariant feature transform characteristics) 2; K-means and other methods are used to train the generated feature points into a visual dictionary; 3 through the image feature coding method, the image features to be classified are mapped to visual words in the visual dictionary; 4 through the pooling algorithm to form image descriptors 5 categorize the image descriptors by a classification algorithm such as a support vector machine. Hyperspectral feature coding is the process of quantifying the feature points of the image to be classified into visual words. The coding error is the main factor affecting the correct rate of image classification.

The traditional hyperspectral remote sensing image has low resolution, and the optical image has low resolution and low spectral resolution. Most optical image resolutions are based on pixel-level spectral curves to analyze data. Most of the cases require pixel de-mixing, while low spectral resolution yields less information, which is prone to "foreign matter" and "same spectrum". The phenomenon of foreign matter, and the uncertainty of the mapping relationship between hyperspectral image feature points and hyperspectral words, the recognition of similar images is poor. The large-scale high-resolution hyperspectral is a high-dimensional fine big data feature space. With the development of remote sensing technology and spectroscopic imaging systems, new requirements have been put forward for target detection.

Summary of the invention

The technical problem to be solved by the present invention is to provide a classification method and system for hyperspectral images, aiming at solving the uncertainty existing in the prior art in establishing a mapping relationship between hyperspectral image feature points and hyperspectral words, and the similar image recognition power difference The problem.

The present invention is achieved in such a manner that a method for classifying hyperspectral images includes:

The hyperspectral image is divided into a training set and a test set, and local feature points of the training set and local feature points of the test set to be classified are respectively extracted from the training set and the test set, according to local feature points of the training set and The local feature points to be classified of the test set constitute a training set feature point set and a test set to be classified feature point set;

Calculating local feature points of the training set by a K-means algorithm to form a dictionary;

Using a KNN algorithm to form a nearest neighbor word for the local feature points of the test set to be classified in the dictionary;

Using the KNN algorithm to find the nearest neighbor feature points for the feature points to be classified of the test set;

Finding a neighbor word with the shortest spectral distance in the nearest neighbor word;

The neighboring feature points, the neighboring words and the spectral dimension distance triple constraint are introduced. By solving the constrained least squares fitting problem, the local feature points of the test set to be classified and the coding coefficients of the dictionary words are obtained.

The coding coefficients are pooled by a maximum pooling algorithm, and the encoded coefficients obtained by the pooling are used as feature descriptors of the hyperspectral image, and the test set of the hyperspectral image is classified according to the feature descriptor.

Further, the test set to Y represents a group of local feature point set to be classified, when B is a dictionary, a Y y _i to the i-th local feature point to be classified, y _i expressed in [mu] _i of Corresponding to the band information, the coding coefficient is represented by Z, then:

Where d _i is the Euclidean distance between the local feature point y _i to be classified and the dictionary word, d _ij is the Euclidean distance between the local feature point to be classified and the neighbor feature point, h _i is the spectral dimension between y _i and the dictionary word The Euclidean distance, λ ₁ , λ ₂ and λ ₃ are penalty factors.

Further, the problem of solving the constrained minimum multiplication by solving includes:

By solving

Obtaining a local feature point to be classified and a coding coefficient Z of the dictionary word in the test set to be classified in the test set.

The invention also provides a classification system for hyperspectral images, comprising:

a feature point extracting unit, configured to divide the hyperspectral image into a training set and a test set, respectively extracting a local feature point of the training set and a local feature point of the test set to be classified from the training set and the test set, according to the The local feature points of the training set and the local feature points of the test set to be classified constitute a training set A set of feature points to be classified and a set of test points to be classified;

a feature point calculation unit, configured to calculate a local feature point of the training set by a K-means algorithm to form a dictionary, and adopt a KNN algorithm to form a nearest neighbor of the local feature points of the test set to be classified in the dictionary a word, using a KNN algorithm, searching for a nearest neighbor feature point for the feature point to be classified of the test set, and finding a neighbor word with the shortest spectral distance in the nearest neighbor word;

The image classification unit is used to introduce the neighboring feature points, the neighboring words and the spectral dimension distance triple constraint. By solving the constrained least squares fitting problem, the local feature points and dictionary words to be classified in the test set to be classified are obtained. a coding coefficient, the coding coefficient is pooled by a maximum pooling algorithm, and the encoded coefficient obtained by the pooling is used as a feature descriptor of the hyperspectral image, and the test set of the hyperspectral image is determined according to the feature descriptor sort.

Further, the step of the image classification unit solving the constraint minimum multiplication fitting problem comprises:

By solving

Compared with the prior art, the invention has the beneficial effects of introducing the local constraint of the hyperspectral image band information and the local correlation information of the neighbor feature points as the image feature points on the basis of the local constraints of the hyperspectral word points on the hyperspectral features. The classification criterion solves the uncertainty existing in the mapping relationship between hyperspectral image feature points and dictionary words, and improves the recognition ability of similar images.

DRAWINGS

1 is a flow chart of a method for classifying a word bag model provided by the prior art;

2 is a flowchart of a method for classifying hyperspectral images according to an embodiment of the present invention;

FIG. 3 is a schematic structural diagram of a classification system for hyperspectral images according to an embodiment of the present invention.

Detailed ways

The present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It is understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

FIG. 2 is a diagram showing a method for classifying hyperspectral images according to an embodiment of the present invention, including:

S201. The hyperspectral image is divided into a training set and a test set, and local feature points of the training set and local feature points of the test set to be classified are respectively extracted from the training set and the test set, according to local features of the training set. a point and a local feature point of the test set to be classified constitute a training set feature point set and a test set to be classified feature point set;

S202: Calculate local feature points of the training set by using a K-means algorithm to form a dictionary;

S203, using a KNN algorithm, in the dictionary, forming a nearest neighbor word for the local feature points of the test set to be classified;

S204: Using a KNN algorithm, searching for a nearest neighbor feature point for the feature point to be classified of the test set;

S205. Find a neighbor word with the shortest spectral distance in the nearest neighbor word;

S206, introducing a neighboring feature point, a neighboring word and a spectral dimension distance triple constraint, and solving a constrained least squares fitting problem, obtaining a local feature point of the test feature to be classified and a coding coefficient of the dictionary word in the test set to be classified. In this step, by solving

S207. The coding coefficient is pooled by a maximum pooling algorithm, and the encoded coefficient obtained by the pooling is used as a feature descriptor of the hyperspectral image, and the test set of the hyperspectral image is performed according to the feature descriptor. classification.

The classification method provided by the embodiment of the present invention proposes a triple three-dimensional locality constrained linear coding algorithm, which introduces hyperspectral image band information and local related information of neighboring feature points on the basis of local constraints of hyperspectral words on hyperspectral feature points. The local constraint is used as the distinguishing item of image feature point classification, which solves the uncertainty of the mapping relationship between hyperspectral image feature points and hyperspectral words, and improves the recognition ability of similar images. The specific triple constraint refers to the local constraint of hyperspectral words, neighbor feature points and hyperspectral image band information on hyperspectral feature points. Three-dimensional refers to the introduction of spectral dimensions on the basis of hyperspectral two-dimensional space. The three-dimensional three-dimensional local constraint linear coding algorithm model is as follows:

Let Y = [y ₁ , y ₂ , ..., y _N ] ∈ R ^{D × N} represent a set of feature points of the test set of the hyperspectral image, B = [b ₁ , b ₂ , ..., b _M ] ∈ R ^D×M denotes a dictionary, y _i denotes the i-th local feature point to be classified in Y, and μ _i denotes corresponding band information indicating y _i , where [μ ₁ , μ ₂ , . . . , μ _i ]∈R ^D×N The coding coefficient Z=[z ₁ ,z ₂ ,...,z _M ]∈R ^D×M of the feature point y _i and the dictionary B is obtained by the triple three-dimensional locality constraint linear coding algorithm, and the calculation formula is as follows:

Where d _i is the Euclidean distance between the feature point y _i and the dictionary word, the feature point d _ij is the Euclidean distance between the feature point to be classified and the neighbor feature point, and h _i is the spectral dimension between the feature point y _i and the dictionary word European distance.

λ ₁ , λ ₁ and λ ₃ are penalty factors, and λ ₁ , λ ₁ and λ _{3 are set} according to the data set size of the hyperspectral image and the image feature similarity. Constraint 1 ^T z _i =1 also satisfies the requirement of translation invariance of LLC (Locality-constrained Linear Coding) coding coefficients. Item 1 in the above formula

For signal fidelity, ensure that the classification signal energy is not lost; item 2

Is the coefficient z _i constrained by the Euclidean distance of the nearest neighbor of the feature point, ensuring that the feature point is mapped to the nearest neighbor word;

The coefficient z _i is constrained by the neighboring feature points to eliminate the ambiguity and uncertainty in the classification of hyperspectral image caused by external factors such as light changes and shooting angles;

It is to use the coefficient z _{i to} be constrained by the band of the neighbor word of the feature point, and to ensure that the feature point is mapped to the word with the closest band.

FIG. 3 shows a classification system for hyperspectral images provided by an embodiment of the present invention, including:

The feature point extraction unit 301 is configured to divide the hyperspectral image into a training set and a test set, and extract local feature points of the training set and local feature points of the test set from the training set and the test set, respectively, according to the The local feature points of the training set and the local feature points of the test set to be composed constitute a training set feature point set and a test set to be classified feature point set;

The feature point calculation unit 302 is configured to calculate a local feature point of the training set by using a K-means algorithm to form a dictionary, and adopt a KNN algorithm in which a local feature point to be classified of the test set is formed in the dictionary. a neighboring word, using a KNN algorithm, searching for a nearest neighbor feature point for the feature point to be classified of the test set, and finding a neighbor word with the shortest spectral distance in the nearest neighbor word;

The image classification unit 303 is configured to introduce a neighboring feature point, a neighboring word, and a spectral dimension distance triple constraint. By solving the constraint minimum multiplication fitting problem, the local feature point and the dictionary word to be classified in the test feature to be classified in the test set are obtained. a coding coefficient, the coding coefficient is pooled by a maximum pooling algorithm, and the encoded coefficient obtained by the pooling is used as a feature descriptor of the hyperspectral image, and the test set of the hyperspectral image is determined according to the feature descriptor sort.

Further, the step of the image classification unit 303 solving the constraint minimum multiplication fitting problem includes:

By solving

The above is only the preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions and improvements made within the spirit and principles of the present invention are intended to be included within the scope of the present invention.

Claims

A method for classifying hyperspectral images, comprising:

The hyperspectral image is divided into a training set and a test set, and local feature points of the training set and local feature points of the test set to be classified are respectively extracted from the training set and the test set, according to local feature points of the training set and The local feature points to be classified of the test set constitute a training set feature point set and a test set to be classified feature point set;

Calculating local feature points of the training set by a K-means algorithm to form a dictionary;

Using a KNN algorithm to form a nearest neighbor word for the local feature points of the test set to be classified in the dictionary;

Using the KNN algorithm to find the nearest neighbor feature points for the feature points to be classified of the test set;

Finding a neighbor word with the shortest spectral distance in the nearest neighbor word;

The neighboring feature points, the neighboring words and the spectral dimension distance triple constraint are introduced. By solving the constrained least squares fitting problem, the local feature points of the test set to be classified and the coding coefficients of the dictionary words are obtained.

The coding coefficients are pooled by a maximum pooling algorithm, and the encoded coefficients obtained by the pooling are used as feature descriptors of the hyperspectral image, and the test set of the hyperspectral image is classified according to the feature descriptor.
The classification method according to claim 1, wherein a set of local feature points to be classified of the test set is represented by Y, the dictionary is represented by B, and the i-th to be classified in Y is represented by y i For local feature points, the corresponding band information of y i is represented by μ i , and the coding coefficient is represented by Z, then:

Where d i is the Euclidean distance between the local feature point y i to be classified and the dictionary word, d ij is the Euclidean distance between the local feature point to be classified and the neighbor feature point, h i is the spectral dimension between y i and the dictionary word The Euclidean distance, λ 1 , λ 2 and λ 3 are penalty factors.
The classification method according to claim 2, wherein said solving the constraint minimum multiplication The fitting problems include:

By solving
Obtaining a local feature point to be classified and a coding coefficient Z of the dictionary word in the test set to be classified in the test set.
A classification system for hyperspectral images, comprising:

a feature point extracting unit, configured to divide the hyperspectral image into a training set and a test set, respectively extracting a local feature point of the training set and a local feature point of the test set to be classified from the training set and the test set, according to the The local feature points of the training set and the local feature points of the test set to be classified constitute a training set feature point set and a test set to be classified feature point set;

a feature point calculation unit, configured to calculate a local feature point of the training set by a K-means algorithm to form a dictionary, and adopt a KNN algorithm to form a nearest neighbor of the local feature points of the test set to be classified in the dictionary a word, using a KNN algorithm, searching for a nearest neighbor feature point for the feature point to be classified of the test set, and finding a neighbor word with the shortest spectral distance in the nearest neighbor word;

The image classification unit is configured to introduce a neighboring feature point, a neighboring word and a spectral dimension distance triple constraint, and obtain a coding of the local feature point to be classified and the dictionary word in the feature set to be classified in the test set by solving the constraint minimum multiplication fitting problem. a coefficient, the coding coefficient is pooled by a maximum pooling algorithm, and the encoded coefficient obtained by the pooling is used as a feature descriptor of the hyperspectral image, and the test set of the hyperspectral image is performed according to the feature descriptor classification.
The classification system according to claim 4, wherein a set of local feature points to be classified of the test set is represented by Y, the dictionary is represented by B, and the i-th to be classified in Y is represented by y i For local feature points, the corresponding band information of y i is represented by μ i , and the coding coefficient is represented by Z, then:

Where d i is the Euclidean distance between the local feature point y i to be classified and the dictionary word, d ij is the Euclidean distance between the local feature point to be classified and the neighbor feature point, h i is the spectral dimension between y i and the dictionary word The Euclidean distance, λ 1 , λ 2 and λ 3 are penalty factors.
The classification system according to claim 5, wherein said image classification unit solves about The steps of the beam minimum multiplication fit problem include:

By solving
Obtaining a local feature point to be classified and a coding coefficient Z of the dictionary word in the test set to be classified in the test set.