US-12620194-B2 - Hyperspectral image band selection method and system based on latent feature fusion

US12620194B2US 12620194 B2US12620194 B2US 12620194B2US-12620194-B2

Abstract

A hyperspectral image band selection method based on latent feature fusion comprises: S 11 , inputting a hyperspectral image cube and segmenting the inputted hyperspectral image cube into several regions by superpixel segmentation; S 12 , learning low-dimensional latent features corresponding to the several regions from the several regions respectively to obtain a latent feature matrix of all the regions; S 13 , calculating an average Laplacian matrix and an average latent feature matrix of the hyperspectral image cube; S 14 , fusing the latent feature matrix, the average Laplacian matrix, and the average latent feature matrix of all the regions to obtain a low-dimensional self-representation matrix of the hyperspectral image cube; and S 15 , clustering the low-dimensional self-representation matrix by a k-means algorithm to obtain an optimal band subset of the hyperspectral image cube.

Inventors

Xinzhong ZHU
Huiying XU
Chang Tang
Jianmin Zhao

Assignees

ZHEJIANG NORMAL UNIVERSITY

Dates

Publication Date: 20260505
Application Date: 20220317
Priority Date: 20210425

Claims (10)

1 . A hyperspectral image band selection method based on latent feature fusion, comprising: S1, inputting a hyperspectral image cube and segmenting the hyperspectral image cube into N regions by superpixel segmentation to obtain segmented regions; S2, learning low-dimensional latent features corresponding to the N regions from the N regions respectively to obtain a latent feature matrix of all the regions; S3, calculating an average Laplacian matrix and an average latent feature matrix of the hyperspectral image cube; S4, fusing the latent feature matrix, the average Laplacian matrix, and the average latent feature matrix of all the regions to obtain a low-dimensional self-representation matrix of the hyperspectral image cube; and S5, clustering the low-dimensional self-representation matrix by a k-means algorithm to obtain an optimal band subset of the hyperspectral image cube.
2 . The hyperspectral image band selection method according to claim 1 , wherein in S1, the hyperspectral image cube into the N regions by superpixel segmentation is segmented by adopting an ERS entropy rate superpixel segmentation algorithm.
3 . The hyperspectral image band selection method according to claim 2 , wherein in S1, a number of the segmented regions is represented as: N = 5 ⁢ 0 ⁢ 0 × N z P + 4 ⁢ 0 × N z N b × res wherein N represents an optimal segmentation number for the hyperspectral image cube; N z represents a number of non-zero regions at an edge of a detected hyperspectral image; P represents pixels contained in each band; N b represents a fixed constant; res represents a spatial resolution of the detected hyperspectral image.
4 . The hyperspectral image band selection method according to claim 1 , wherein in S2, the low-dimensional latent features corresponding to the N regions are learned from the N regions respectively to obtain the latent feature matrix of all the segmented regions, expressed as: max Y ( i ) Tr ⁡ ( Y ( i ) T ⁢ E ( i ) ⁢ Y ( i ) ) ⁢ s . t . Y ( i ) T ⁢ Y ( i ) = I wherein Y (i) represents a low-dimensional latent feature matrix corresponding to an i-th segmented region; E (i) represents a Laplacian matrix corresponding to the i-th segmented region; I represents an identity matrix; the Laplacian matrix corresponding to each of the segmented regions is specifically represented as follows: E ( i ) = D - 1 2 ⁢ WD - 1 2 wherein W represents a similarity matrix between samples in each of the segmented regions; D represents a diagonal matrix, represented as: D jj = ∑ W : , j wherein j represents a j-th sample in the segmented regions; D jj represents elements in a j-th row and j-th column of the diagonal matrix D ; ∑ W : , j represents a sum of all elements in a j-th column of the similarity matrix W.
5 . The hyperspectral image band selection method according to claim 4 , wherein in S4, the latent feature matrix, the average Laplacian matrix, and the average latent feature matrix of all the regions are fused to obtain the low-dimensional self-representation matrix of the hyperspectral image cube, expressed as: max F , R , γ Tr ⁡ ( F T ⁢ ∑ i = 1 N ⁢ γ i ⁢ Y ( i ) ⁢ R ( i ) ) + λ ⁢ Tr ⁡ ( F T ⁢ F _ ) + β ⁢ Tr ⁡ ( F T ⁢ LF ) s . t . F T ⁢ F = I d , R ( i ) T ⁢ R ( i ) = I d , ∑ i = 1 N ⁢ γ i 2 = 1 , γ i ≥ 0 wherein F represents a low-dimensional self-representation matrix of a fused hyperspectral image; γ i represents a contribution rate of each of the segmented regions; Y (i) represents a low-dimensional latent feature matrix corresponding to the i-th segmented region; R (i) represents a rotation matrix corresponding to the i-th segmented region; F represents an average latent feature matrix; L represents an average Laplacian matrix; λ and β both represent equilibrium parameters; T represents a transposition of a matrix; I d represents an identity matrix with a size of d*d.
6 . A hyperspectral image band selection system based on latent feature fusion, comprising a processor comprising: a hyperspectral image segmentation module, used for inputting a hyperspectral image cube and segmenting the hyperspectral image cube into N regions by superpixel segmentation to obtain segmented regions; a latent feature learning module, used for learning low-dimensional latent features corresponding to the N regions from the N regions respectively to obtain a latent feature matrix of all the regions; a calculation module, used for calculating an average Laplacian matrix and an average latent feature matrix of the hyperspectral image cube; a latent feature fusion module, used for fusing the latent feature matrix, the average Laplacian matrix, and the average latent feature matrix of all the regions to obtain a low-dimensional self-representation matrix of the hyperspectral image cube; and a hyperspectral band selecting module, used for clustering the low-dimensional self-representation matrix by a k-means algorithm to obtain an optimal band subset of the hyperspectral image cube.
7 . The hyperspectral image band selection system according to claim 6 , wherein in the hyperspectral image segmentation module, the inputted hyperspectral image cube into the N regions by superpixel segmentation is segmented by adopting an ERS entropy rate superpixel segmentation algorithm.
8 . The hyperspectral image band selection system according to claim 7 , wherein in the hyperspectral image segmentation module, a number of the segmented regions is represented as: N = 5 ⁢ 0 ⁢ 0 × N z P + 4 ⁢ 0 × N z N b × res wherein N represents an optimal segmentation number for the hyperspectral image cube; N z represents a number of non-zero regions at an edge of a detected hyperspectral image; P represents pixels contained in each band; N b represents a fixed constant; res represents a spatial resolution of the detected hyperspectral image.
9 . The hyperspectral image band selection system according to claim 6 , wherein in the latent feature learning module, the low-dimensional latent features corresponding to the N regions are learned from the N regions respectively to obtain the latent feature matrix of all the segmented regions, expressed as: max Y ( i ) Tr ⁡ ( Y ( i ) T ⁢ E ( i ) ⁢ Y ( i ) ) ⁢ s . t . Y ( i ) T ⁢ Y ( i ) = I wherein Y (i) represents a low-dimensional latent feature matrix corresponding to an i-th segmented region; E (i) represents a Laplacian matrix corresponding to the i-th segmented region; I represents an identity matrix; the Laplacian matrix corresponding to each of the segmented regions is specifically represented as follows: E ( i ) = D - 1 2 ⁢ WD - 1 2 wherein W represents a similarity matrix between samples in each of the segmented regions; D represents a diagonal matrix, represented as: D jj = ∑ W : , j wherein j represents a j-th sample in the segmented regions; D jj represents elements in a j-th row and j-th column of the diagonal matrix D ; ∑ W : , j represents a sum of all elements in a j-th column of the similarity matrix W.
10 . The hyperspectral image band selection system according to claim 9 , wherein in the latent feature fusion module, the latent feature matrix, the average Laplacian matrix, and the average latent feature matrix of all the regions are fused to obtain the low-dimensional self-representation matrix of the hyperspectral image cube, expressed as: max F , R , γ Tr ⁡ ( F T ⁢ ∑ i = 1 N ⁢ γ i ⁢ Y ( i ) ⁢ R ( i ) ) + λ ⁢ Tr ⁡ ( F T ⁢ F _ ) + β ⁢ Tr ⁡ ( F T ⁢ LF ) s . t . F T ⁢ F = I d , R ( i ) T ⁢ R ( i ) = I d , ∑ i = 1 N ⁢ γ i 2 = 1 , γ i ≥ 0 wherein F represents a low-dimensional self-representation matrix of a fused hyperspectral image; γ i represents a contribution rate of each of the segmented regions; Y (i) represents a low-dimensional latent feature matrix corresponding to the i-th segmented region; R (i) represents a rotation matrix corresponding to the i-th segmented region; F represents an average latent feature matrix; L represents an average Laplacian matrix; λ and β both represent equilibrium parameters; T represents a transposition of a matrix; I d represents an identity matrix with a size of d*d.

Description

CROSS REFERENCE TO THE RELATED APPLICATIONS This application is the national phase entry of International Application No. PCT/CN2022/081429, filed on Mar. 17, 2022, which is based upon and claims priority to Chinese Patent Application No. 202110447625.2, filed on Apr. 25, 2021, the entire contents of which are incorporated herein by reference. TECHNICAL FIELD The present application relates to the field of hyperspectral remote-sensing image band selecting technology, and in particular to a hyperspectral image band selection method and system based on latent feature fusion. BACKGROUND Hyperspectral sensors capture spectral and spatial information of target scenes by collecting dozens or even hundreds of continuous hyperspectral bands. Compared with RGB images, hyperspectral images have richer information and higher resolution. With the continuous development and maturation of hyperspectral imaging techniques and image classification techniques, hyperspectral images are widely used in various fields, such as salient object detection, medical image processing, mineral exploration, and the like. However, it is important to perform dimensionality reduction processing on the hyperspectral images to solve the problem of dimensional disasters. Hyperspectral dimensionality reduction can be roughly divided into two categories, namely feature extraction and feature selection, wherein in the hyperspectral field, the feature selection is also known as band selection. Usually, band selection is performed in the original feature space, i.e. only some representative bands are selected from the entire hyperspectral images to form a feature band subset without changing the original data information, thus the physical meaning of the original data can be reserved, and band redundancy in the original hyperspectral data is reduced. According to the separability of category labels, the band selection can be further divided into two types: supervised band selection and unsupervised band selection. The unsupervised band selection only requires selecting a subset of feature bands based on the importance of the bands. Some measurement indicators are provided for the importance of a certain band, such as information divergence, minimum noise value, Euclidean distance, and the like. Clustering-based methods have attracted much attention in the unsupervised band selection in recent years. In most clustering-based band selection methods, firstly, each band is stretched into a single feature vector, and then hyperspectral bands are selected based on the corresponding objective function. However, actually, for a certain band, different regions often correspond to different objects, and their spectral characteristics are different. In addition, high-dimensional pixel features of each band contain a large amount of redundant information, which limits the performance of hyperspectral band selection. Therefore, it is not appropriate to treat each band directly as a single feature vector. SUMMARY In response to the defects in the prior art, the present application aims to provide a hyperspectral image band selection method and system based on latent feature fusion. In order to achieve the above objective, the present application adopts the following technical solutions: A hyperspectral image band selection method based on latent feature fusion comprises: S1, inputting a hyperspectral image cube and segmenting the inputted hyperspectral image cube into several regions by superpixel segmentation;S2, learning low-dimensional latent features corresponding to the several regions from the several regions respectively to obtain a latent feature matrix of all the regions;S3, calculating an average Laplacian matrix and an average latent feature matrix of the hyperspectral image cube;S4, fusing the latent feature matrix, the average Laplacian matrix, and the average latent feature matrix of all the regions to obtain a low-dimensional self-representation matrix of the hyperspectral image cube; andS5, clustering the obtained low-dimensional self-representation matrix by a k-means algorithm to obtain an optimal band subset of the hyperspectral image cube. Further, in S1, segmenting the inputted hyperspectral image cube into several regions by superpixel segmentation is completed by adopting an ERS entropy rate superpixel segmentation algorithm. Further, in S1, the inputted hyperspectral image cube is segmented into several regions, with a number of the segmented regions represented as: N=5⁢0⁢0×NzP+4⁢0×NzNb×res wherein N represents an optimal segmentation number for each hyperspectral image cube; Nz represents a number of non-zero regions at an edge of a detected hyperspectral image; P represents pixels contained in each band; Nb represents a fixed constant; res represents a spatial resolution of each hyperspectral image. Further, in S2, the low-dimensional latent features corresponding to the several regions are learned from the several regions respectively