CN108540802B - Local constraint linear coding method and system for hyperspectral image - Google Patents

Abstract

The invention is applicable to the technical field of image processing, and provides a locally constrained linear encoding method for hyperspectral images, including: extracting features to be encoded in hyperspectral images to obtain a feature set; dictionary; obtain the band information of the feature set and the band information of the words in the dictionary respectively; take the band information of the feature point and the band information of the word as the discriminant constraint item, calculate the Euclidean between the feature point of the hyperspectral image and the word and obtain the feature encoding coefficient; encode the feature of the hyperspectral image according to the feature encoding coefficient. In the embodiment of the present invention, while using the local features of hyperspectral images and the local linear constraints of visual words, the band information of hyperspectral images is introduced as a discriminant constraint for the mapping of feature points to dictionary words, thereby reducing the number of feature points. The ambiguity of the mapping with dictionary words enhances the representation of hyperspectral images.

Description

A Locally Constrained Linear Coding Method and System for Hyperspectral Images

technical field

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

Background technique

In terms of feature coding for common image classification, the existing coding algorithms mainly include hard allocation, sparse coding, local feature aggregation descriptor VLAD (vector of locally aggregated descriptors), Fisher vector and local constrained linear coding algorithm LLC (Locality- constrained linear coding), etc.

Compared with traditional grayscale images, hyperspectral images contain rich spatial, radiometric and spectral triple information. Extending the traditional application from ordinary image LLC to 3-dimensional LLC and applying it to hyperspectral image classification can achieve good results. The obtained feature points may show strong similarity, while the feature points extracted from images of the same category show differences, which leads to the blurring of the mapping of feature points to dictionary words in the LLC encoding process. "meaning" or "multi-word synonym", the representation ability of hyperspectral images is weak.

SUMMARY OF THE INVENTION

The technical problem to be solved by the present invention is to provide a local constrained linear encoding method and system for hyperspectral images, which aims to solve the problem that the mapping of feature points to dictionary words in the existing encoding method is ambiguous, and the representation ability of hyperspectral images is weak. The problem.

The present invention is implemented in this way, a method for locally constrained linear encoding of hyperspectral images, comprising:

Extract the features to be encoded of the hyperspectral image to obtain a feature set;

determining the dictionary of the hyperspectral image according to the feature points in the feature set;

respectively acquiring the band information of the feature set and the band information of the words in the dictionary;

Using the band information of the feature point and the band information of the word as a discriminant constraint item, calculate the Euclidean distance between the feature point of the hyperspectral image and the word, and obtain the feature encoding coefficient;

The feature of the hyperspectral image is encoded according to the feature encoding coefficient.

Further, the feature set is represented by X=[x ₁ ,x ₂ ,...,x _N ]∈R ^D×N , and the dictionary is represented by B=[b ₁ ,b ₂ ...b _M ]∈ R ^D×M , the dictionary B includes M words, the band information of the feature set is represented by Ω, Ω=[μ ₁ , μ ₂ ,...,μ _K ]∈R ^D×K , the The band information of the word is represented by U, U=[u ₁ , u ₂ ,...,u _L ]∈R ^D×L , _xi represents the ith feature point in the feature set X, and the feature encoding coefficient is represented by _zi means;

The feature coding coefficient is calculated by the following formula with _zi :

in:

dif(μ_i,U)＝[dif(μ_i,u₁),dif(μ_i,u₂),...,dif(μ_i,u_N)]^T，⊙表示各元素点乘，d_ij表示特征点x_i与近邻特征点x_j之间的欧式距离，H为距离阈值，λ₁和λ₂为惩罚因子，λ₁＝λ₂＝0.005，dif(μ_i,u_j)表示所述高光谱图像的特征点的波段μ_i与该特征点映射的单词的波段u_j的差值。

dif(μ _i ,U)=[dif(μ _i ,u ₁ ),dif(μ _i ,u ₂ ),...,dif(μ _i ,u _N )] ^T , ⊙ denotes the dot product of each element, d _ij represents the Euclidean distance between the feature point x _i and the neighboring feature point x _j , H is the distance threshold, λ ₁ and λ ₂ are penalty factors, λ ₁ =λ ₂ =0.005, dif(μ _i , u _j ) represents the The difference between the band μ _i of the feature point of the hyperspectral image and the band u _j of the word mapped by the feature point.

The embodiment of the present invention also provides a locally constrained linear encoding system for hyperspectral images, including:

The feature extraction unit is used to extract the feature to be encoded of the hyperspectral image to obtain a feature set;

a dictionary determining unit, configured to determine the dictionary of the hyperspectral image according to the feature points in the feature set;

a band acquisition unit, configured to obtain the band information of the feature set and the band information of the words in the dictionary respectively;

A coefficient calculation unit, for using the band information of the feature point and the band information of the word as a discriminant constraint item, calculating the Euclidean distance between the feature point of the hyperspectral image and the word, to obtain a feature encoding coefficient;

A feature encoding unit, configured to encode the feature of the hyperspectral image according to the feature encoding coefficient.

Further, the feature set is represented by X=[x ₁ ,x ₂ ,...,x _N ]∈R ^D×N , and the dictionary is represented by B=[b ₁ ,b ₂ ...b _M ]∈ R ^D×M , the dictionary B includes M words, the band information of the feature set is represented by Ω, Ω=[μ ₁ , μ ₂ ,...,μ _K ]∈R ^D×K , the The band information of the word is represented by U, U=[u ₁ , u ₂ ,...,u _L ]∈R ^D×L , _xi represents the ith feature point in the feature set X, and the feature encoding coefficient is represented by _zi means;

The coefficient calculation unit calculates and obtains the characteristic coding coefficient z _i through the following formula:

in:

dif(μ_i,U)＝[dif(μ_i,u₁),dif(μ_i,u₂),...,dif(μ_i,u_N)]^T，⊙表示各元素点乘，d_ij表示特征点x_i与近邻特征点x_j之间的欧式距离，H为距离阈值，λ₁和λ₂为惩罚因子，λ₁＝λ₂＝0.005，dif(μ_i,u_j)表示所述高光谱图像的特征点的波段μ_i与该特征点映射的单词的波段u_j的差值。

dif(μ _i ,U)=[dif(μ _i ,u ₁ ),dif(μ _i ,u ₂ ),...,dif(μ _i ,u _N )] ^T , ⊙ denotes the dot product of each element, d _ij represents the Euclidean distance between the feature point x _i and the neighboring feature point x _j , H is the distance threshold, λ ₁ and λ ₂ are penalty factors, λ ₁ =λ ₂ =0.005, dif(μ _i , u _j ) represents the The difference between the band μ _i of the feature point of the hyperspectral image and the band u _j of the word mapped by the feature point.

Compared with the prior art, the present invention has the beneficial effects that: in the embodiment of the present invention, the feature to be encoded of the hyperspectral image is extracted to obtain a feature set including feature points, and the dictionary of the hyperspectral image is determined according to the feature points in the feature set, Obtain the band information of the feature set and the band information of the words in the dictionary, and use this as a discriminant constraint, calculate the Euclidean distance between the feature points of the hyperspectral image and the word, and obtain the feature coding coefficient, which is encoded according to the feature. The coefficients encode the characteristics of the hyperspectral image. In the embodiment of the present invention, while using the local features of hyperspectral images and the local linear constraints of visual words, the band information of hyperspectral images is introduced, and it is used as a discriminant constraint for the mapping of feature points to dictionary words, thereby reducing the number of feature points. The ambiguity of the mapping with dictionary words enhances the representation of hyperspectral images.

Description of drawings

1 is a flowchart of a method for locally constrained linear encoding of hyperspectral images provided by an embodiment of the present invention;

FIG. 2 is a schematic structural diagram of a locally constrained linear encoding system for hyperspectral images provided by an embodiment of the present invention.

Detailed ways

In order to make the objectives, technical solutions and advantages of the present invention clearer, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention, but not to limit the present invention.

1 shows a locally constrained linear encoding method for hyperspectral images provided by an embodiment of the present invention, including:

S101, extracting the feature to be encoded of the hyperspectral image to obtain a feature set;

S102, according to the feature points in the feature set, determine the dictionary of the hyperspectral image;

S103, obtain the band information of the feature set and the band information of the words in the dictionary respectively;

S104, using the band information of the feature point and the band information of the word as a discriminant constraint, calculate the Euclidean distance between the feature point of the hyperspectral image and the word, and obtain a feature encoding coefficient;

S105: Encode the feature of the hyperspectral image according to the feature encoding coefficient.

The basic idea of LLC in the prior art is to use the linear combination of the nearest k dictionary words to express the feature X to be encoded. Suppose X=[x ₁ ,x ₂ ,...,x _N ]∈R ^D×N is a set of feature sets obtained by extracting hyperspectral images, D represents a D-dimensional descriptor, B=[b ₁ ,b ₂ , ...,b _M ]∈R ^D×M is a dictionary containing M words, and the feature encoding coefficient z _i is calculated from the Euclidean distance between the feature point x _i of the hyperspectral image and the word b _j , as follows:

Among them, ⊙ represents the dot product of each element, λ is the penalty factor, d _i ∈ R ^M represents the local adapter, which can assign different degrees of freedom to each basis vector, and the assigned degrees of freedom are related to the basis vector and the input feature point x _i . The similarity is proportional, as shown in formula (2):

d _i =exp(dist( _xi ,B)/σ) (2)

in:

dist( _xi ,B)=[dist( _xi ,b ₁ ),dist( _xi ,b ₂ ),...,dist( _xi ,b _M )] ^T (3)

where dist( _xi , b _i ) represents the Euclidean distance between _xi and b _i , and σ is used to adjust the weight decay rate of the local adapter. Usually, dist(x _i ,B) is subtracted from max(dist(x _i ,B)) to normalize d _i to the (0,1] range, the constraint 1 ^T _zi =1 satisfies the translation of LLC code The requirement of invariance. The LLC code in Eq. (1) does not have sparseness in terms of l ⁰ norm, but since its solution has only a few valid values, from this point of view, the LLC code in Eq. (1) has sparseness In practice, those smaller coefficients are usually thresholded to zero.

The difference between hyperspectral image and ordinary two-dimensional image is that it contains not only two-dimensional image information but also spectral information. In view of the above-mentioned mapping of image features to dictionary words, the ambiguity of "polysemy of one word" or "polysemy of multiple words" may occur. In order to classify the hyperspectral image more accurately, the embodiment of the present invention introduces the band information of the feature points of the hyperspectral image as a discriminant constraint item, and proposes a new locally constrained linear encoding of the hyperspectral image as shown in FIG. 1 . method.

Suppose X=[x ₁ ,x ₂ ,...,x _N ]∈R ^D×N is a set of feature sets obtained by extracting hyperspectral images, B=[b ₁ ,b ₂ ...b _M ]∈R ^D×M is a dictionary, there are M words in dictionary B, Ω is the band information of the feature point set X of the hyperspectral image to be classified, Ω=[μ ₁ , μ ₂ ,...,μ _K ]∈R ^D×K , U represents the band information of the word in the dictionary B of the hyperspectral image, U=[u ₁ , u ₂ ,...,u _L ]∈R ^D×L , R represents the dimension, and D represents the D-dimensional descriptor. After improving the formula (1) in the traditional LLC, the following formula is obtained:

in,

in,

dif(μ _i ,U)=[dif(μ _i ,u ₁ ),dif(μ _i ,u ₂ ),...,dif(μ _i ,u _N )] ^T (6)

Z∈R ^D×N is the feature point of the hyperspectral image to be classified and the coding coefficient of the word in the dictionary solved by Equation (4). In formula (4), d _ij represents the Euclidean distance between the feature point x _i and the neighboring feature point x _j , and H is the distance threshold. λ ₁ and λ ₂ are penalty factors, set λ ₁ =λ ₂ =0.005. In formula (5), dif(μ _i , u _j ) represents the difference between the band μ _i of the feature point of the hyperspectral image to be classified and the band u _j of the word mapped by the feature point. The influence weight of feature points is larger than that of words with large band difference. Constraint 1 ^T _zi =1 also satisfies the translation invariance requirement of LLC coding. The first item in formula (4) is the signal fidelity, which ensures that the classification signal energy is not lost, and the second item is that the coefficient _zi is constrained by the Euclidean distance of the nearest neighbor words of the feature point, which ensures that the feature point is mapped to the nearest neighbor word, The third item is to use the coefficient _zi to be constrained by the band of the neighboring words of the feature point to ensure that the feature point is mapped to the word with the closest band.

The locality-constrained linear coding method of a hyperspectral image provided by the embodiment of the present invention is improved based on a locality-constrained linear coding (LLC) algorithm of a common two-dimensional image. Compared with ordinary color images, hyperspectral images have increased band information and are three-dimensional data structures. The ordinary LLC algorithm can only work on ordinary color images, and cannot be directly used for encoding local features of hyperspectral images.

The locally constrained linear coding method provided by the embodiment of the present invention utilizes the local features of hyperspectral images and the local linear constraints of visual words, and at the same time, introduces the band information of hyperspectral images, and uses it as the judgment of the mapping of feature points to dictionary words Constraints, thereby reducing the ambiguity of the mapping between feature points and dictionary words, and enhancing the representation of hyperspectral images.

In practical applications, the main problem of hyperspectral images is the large number of bands and the large amount of data, which brings great difficulties to the classification and identification of hyperspectral images. The information redundancy is high, the space required for data storage is large, and the processing time is long. Due to the large number of hyperspectral image bands, the phenomenon of dimensionality disaster is prone to occur, that is, the classification accuracy decreases. Therefore, the dimensionality reduction process that reduces the amount of data and saves resources is very important. Needed. Therefore, the embodiment of the present invention combines the feature encoding algorithm of the two-dimensional image with the hyperspectral image according to the characteristics of the hyperspectral image, thereby improving the classification accuracy of the hyperspectral image. mainly include:

1) A new locally constrained linear coding LLC method for hyperspectral images is proposed. Based on the local constraints of the dictionary words of hyperspectral images on feature points, the band information of hyperspectral images is introduced as a constraint discriminant item to solve the problem of high spectral density. The uncertainty in the mapping relationship between spectral image feature points and dictionary words improves the representation ability of hyperspectral images.

2) Classify hyperspectral images by using the improved locally constrained linear coding LLC method of the new hyperspectral image proposed above. It is proved by experiments that the coding algorithm proposed in this embodiment can better classify hyperspectral images.

The embodiments of the present invention can be implemented in the following applications: 1. Hyperspectral image classification and recognition; 2. Hyperspectral image target detection; 3. Substance sorting.

The embodiment of the present invention also provides a locally constrained linear encoding system for hyperspectral images as shown in FIG. 2 , including:

The feature extraction unit 201 is used to extract the feature to be encoded of the hyperspectral image to obtain a feature set;

a dictionary determination unit 202, configured to determine the dictionary of the hyperspectral image according to the feature points in the feature set;

Band acquisition unit 203, for respectively acquiring the band information of the feature set and the band information of the words in the dictionary;

The coefficient calculation unit 204 is used to calculate the Euclidean distance between the feature point and the word of the hyperspectral image with the band information of the feature point and the band information of the word as a discriminant constraint to obtain a feature encoding coefficient;

A feature encoding unit 205, configured to encode the feature of the hyperspectral image according to the feature encoding coefficient.

Further, the feature set is represented by X=[x ₁ ,x ₂ ,...,x _N ]∈R ^D×N , and the dictionary is represented by B=[b ₁ ,b ₂ ...b _M ]∈ R ^D×M , the dictionary B includes M words, the band information of the feature set is represented by Ω, Ω=[μ ₁ , μ ₂ ,...,μ _K ]∈R ^D×K , the The band information of the word is represented by U, U=[u ₁ , u ₂ ,...,u _L ]∈R ^D×L , _xi represents the ith feature point in the feature set X, and the feature encoding coefficient is represented by _zi means;

The coefficient calculation unit 204 calculates and obtains the characteristic coding coefficient z _i by the following formula:

in:

dif(μ_i,U)＝[dif(μ_i,u₁),dif(μ_i,u₂),...,dif(μ_i,u_N)]^T，⊙表示各元素点乘，d_ij表示特征点x_i与近邻特征点x_j之间的欧式距离，H为距离阈值，λ₁和λ₂为惩罚因子，λ₁＝λ₂＝0.005，dif(μ_i,u_j)表示所述高光谱图像的特征点的波段μ_i与该特征点映射的单词的波段u_j的差值。

dif(μ _i ,U)=[dif(μ _i ,u ₁ ),dif(μ _i ,u ₂ ),...,dif(μ _i ,u _N )] ^T , ⊙ denotes the dot product of each element, d _ij represents the Euclidean distance between the feature point x _i and the neighboring feature point x _j , H is the distance threshold, λ ₁ and λ ₂ are penalty factors, λ ₁ =λ ₂ =0.005, dif(μ _i , u _j ) represents the The difference between the band μ _i of the feature point of the hyperspectral image and the band u _j of the word mapped by the feature point.

An embodiment of the present invention also provides a terminal, including a memory, a processor, and a computer program stored in the memory and running on the processor, characterized in that, when the processor executes the computer program, the high-level shown in FIG. 1 is implemented. Individual steps of a locally constrained linear encoding method for spectral images.

An embodiment of the present invention further provides a readable storage medium on which a computer program is stored, characterized in that, when the computer program is executed by a processor, the locally constrained linear encoding of the hyperspectral image as shown in FIG. 1 is implemented. the various steps of the method.

In addition, each functional module in each embodiment of the present invention may be integrated into one processing module, or each module may exist physically alone, or two or more modules may be integrated into one module. The above-mentioned integrated modules can be implemented in the form of hardware, and can also be implemented in the form of software function modules.

If the integrated modules are implemented in the form of software functional modules and sold or used as independent products, they may be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention is essentially or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product, and the computer software product is stored in a storage medium , including several instructions for causing a computer device (which may be a personal computer, a server, or a network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of the present invention. The aforementioned storage medium includes: U disk, removable hard disk, Read-Only Memory (ROM, Read-Only Memory), Random Access Memory (RAM, Random Access Memory), magnetic disk or optical disk and other media that can store program codes.

The above descriptions are only preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent replacements and improvements made within the spirit and principles of the present invention shall be included in the protection of the present invention. within the range.

Claims (2)

1. A local constraint linear coding method of a hyperspectral image is characterized by comprising the following steps:

extracting the features to be coded of the hyperspectral image to obtain a feature set;

determining a dictionary of the hyperspectral image according to the feature points in the feature set;

respectively acquiring the wave band information of the feature set and the wave band information of words in the dictionary;

calculating the Euclidean distance between the feature points of the hyperspectral image and the words by taking the wave band information of the feature points and the wave band information of the words as discrimination constraint items to obtain a feature coding coefficient;

the set of features is expressed as X ═ X₁，x₂，...，x_N]∈R^D×NSaid dictionary is represented by B ═ B₁，b₂...b_M]∈R^D×NThat is, the dictionary B includes M words, and the band information of the feature set is represented by Ω, and Ω ═ u₁，u₂，...，u_K]∈R^D×NThe band information of the word is represented by U, and U is [ U ═ U₁，u₂，...，u_L]∈R^D×L，x_iRepresenting the i-th feature point in a feature set X, said feature encoding coefficients in z_iRepresents;

characteristic coding coefficient in z_iCalculated by the following formula:

wherein:

an element indicates that the element is dot-by-dot, d_ijRepresents a characteristic point x_iAnd neighbor feature point x_jOf between, is₁And λ₂As a penalty factor, λ₁＝λ₂＝0.005,dif(u_i,u_j) A band u representing a feature point of the hyperspectral image_iThe band u of the word mapped with the feature point_jA difference of (d);

and coding the characteristics of the hyperspectral image according to the characteristic coding coefficient.

2. A locally constrained linear coding system for hyperspectral images, comprising:

the characteristic extraction unit is used for extracting the characteristics to be coded of the hyperspectral image to obtain a characteristic set;

a dictionary determining unit, configured to determine a dictionary of the hyperspectral image according to the feature points in the feature set;

the band acquisition unit is used for respectively acquiring band information of the feature set and band information of words in the dictionary;

the coefficient calculation unit is used for calculating the Euclidean distance between the feature points of the hyperspectral image and the words by taking the wave band information of the feature points and the wave band information of the words as discrimination constraint items to obtain feature coding coefficients;

the set of features is expressed as X ═ X₁，x₂，...，x_N]∈R^D×NSaid dictionary is represented by B ═ B₁，b₂...b_M]∈R^D×NThat is, the dictionary B includes M words, and the band information of the feature set is represented by Ω, and Ω ═ u₁，u₂，...，u_K]∈R^D×NThe band information of the word is represented by U, and U is [ U ═ U₁，u₂，...，u_L]∈R^D×L，x_iRepresenting the ith feature point in a feature set X, said features being codedCode coefficient in z_iRepresents;

the coefficient calculation unit calculates a characteristic coding coefficient z by the following formula_i：

Wherein:

an element indicates that the element is dot-by-dot, d_ijRepresents a characteristic point x_iAnd neighbor feature point x_jOf between, is₁And λ₂As a penalty factor, λ₁＝λ₂＝0.005,dif(u_i,u_j) A band u representing a feature point of the hyperspectral image_iThe band u of the word mapped with the feature point_jA difference of (d);

and the characteristic coding unit is used for coding the characteristics of the hyperspectral image according to the characteristic coding coefficient.

Images