A Hyperspectral Remote Sensing Image Denoising Method Combining Depth Prior and Low-Rank Tensor Decomposition

Abstract

The present invention discloses a hyperspectral remote sensing image denoising method that couples deep prior and low-rank tensor decomposition, including: using a first deep denoising prior sub-network to optimize an initial noisy hyperspectral image; using a low-rank tensor to decompose the optimized hyperspectral image to generate spectral orthogonal bases; estimating a spatial reduction factor through a second deep denoising prior sub-network; and obtaining a denoising result of the hyperspectral image by using a tensor train decomposition method based on the spectral orthogonal bases and the spatial reduction factor. This method takes into account the spatial and spectral neighborhood information of hyperspectral images, couples a deep learning network and low-rank tensor decomposition, removes noise in hyperspectral images, and improves the data quality of hyperspectral images, which has important practical significance.

Description

Technical Field

[0001] The present invention belongs to the field of remote sensing image processing, and relates to a hyperspectral remote sensing image denoising method that couples deep prior and low-rank tensor decomposition. Background Art

[0002] Hyperspectral remote sensing, as a newly developed remote sensing technology, captures hyperspectral images of multiple consecutive and narrow bands through imaging hyperspectral sensors. It contains rich spatial and spectral information of all these bands beyond traditional grayscale or RGB images, and can better distinguish subtle physical differences between different surface materials through these rich spectral features. However, in the actual process of optical remote sensing, due to reasons such as atmospheric interference, secondary incidence, sensor jitter, and signal response, almost all collected hyperspectral images are inevitably contaminated by various noises to varying degrees. These noises seriously reduce the image quality and limit the performance of subsequent processing. Therefore, without changing the existing hardware observation conditions, designing corresponding algorithms to denoise hyperspectral images has important research significance for improving the utilization rate of remote sensing data and promoting the implementation of earth observation activities. Currently, the denoising methods commonly used for hyperspectral images are mainly divided into model-driven methods and data-driven methods. Both types of methods have their own advantages and limitations. Model-based denoising methods usually have the disadvantages of high iterative optimization time consumption and unstable sensitivity of set parameters; data-driven denoising methods usually have poor performance due to overfitting effects. Summary of the Invention

[0003] According to the problems existing in the prior art, the present invention discloses a hyperspectral remote sensing image denoising method that couples deep prior and low-rank tensor decomposition. This method combines deep denoising prior and low-rank tensor decomposition, taking into account both the inherent low-rank characteristics of hyperspectral images in model-driven methods and making full use of the powerful feature extraction ability of deep learning in data-driven methods. It has achieved high reconstruction accuracy and execution efficiency in the removal of mixed noises in hyperspectral images, significantly improved the quality of hyperspectral data, and has important practical significance for the subsequent use of hyperspectral data. The method specifically includes the following steps:

[0004] Use the first deep denoising prior sub-network to optimize the initial noisy hyperspectral image;

[0005] Use a low-rank tensor to decompose the optimized hyperspectral image to generate spectral orthogonal bases;

[0006] Estimate the spatial reduction factor through the second deep denoising prior sub-network;

[0007] Based on the spectral orthogonal bases and the spatial reduction factor, use the tensor train decomposition method to obtain the hyperspectral image denoising result.

[0008] Furthermore, when optimizing the initial noisy hyperspectral image:

[0009] Extract the spatial data of the current band and the spectral data of the upper and lower adjacent bands corresponding to the current band of the noisy hyperspectral data band by band, and use the deep denoising prior sub-network one to perform convolutional filtering on the spatial data and spectral data to obtain the preliminary denoising result of the optimized hyperspectral image:

[0010]

[0011] Among them, Γ(g) represents the per-band traversal process in the sub-depth denoising prior sub-network one, is the preliminary denoising result, Y i is the i-th band in the noisy hyperspectral image, is Y i corresponding adjacent spectral gradient.

[0012] Furthermore, generate the spectral orthogonal basis in the following way:

[0013] Based on the low-rank prior of the third-order tensor, the preliminary denoising result of the hyperspectral image is obtained by singular value decomposition to get the estimated spectral orthogonal basis:

[0014]

[0015] Among them, A is the spectral orthogonal product, represents the 3rd-order unfolding form of the tensor , and SVD r (g) represents the truncation result of the rank parameter r in the singular value decomposition mode.

[0016] Furthermore, when estimating the spatial reduction factor through the deep denoising prior sub-network two:

[0017] Reconstruct the third-order tensor unfolding form and spectral orthogonal basis of the preliminary denoised hyperspectral image to obtain the initialization factor. Based on the deep residual learning strategy, use the deep denoising prior sub-network two to perform feature extraction on the initialization factor band by band to eliminate the noise in the initialization factor and obtain the spatial reduction factor. The specific process is as follows:

[0018]

[0019] Among them, the reshape h,w,r (g) function represents the dimension reshaping operation, is the initialization factor. The reshaped initialization factor still contains noise, and use the feature extraction ability of the deep denoising prior sub-network two to estimate the spatial reduction factor:

[0020]

[0021] Among them, the N(g) function represents normalizing each band of , and the IN(g) function is the inverse operation of the normalization of N(g).

[0022] Furthermore, when obtaining the denoising result of the hyperspectral image:

[0023] Based on the low-rank tensor decomposition strategy, the global spectral low-rank property of the hyperspectral image is represented as Tucker decomposition, and the final denoising result of the hyperspectral image is obtained according to the obtained spectral orthogonal basis and spatial reduction factor:

[0024]

[0025] Among them, ×3 represents the modulo 3 tensor matrix product.

[0026] Due to the adoption of the above technical solution, a denoising method for hyperspectral remote sensing images coupling deep prior and low-rank tensor decomposition is provided by the present invention. This method uses Tucker tensor decomposition to describe the global spectral low-rank constraint, uses the coupling of deep denoising prior and low-rank tensor decomposition for the restoration of hyperspectral images, optimizes the spectral orthogonal basis and spatial reduction factor respectively through two deep denoising networks, and realizes the removal of noise interference information in hyperspectral images without complex and time-consuming iterative optimization. In short, the method proposed by the present invention can be effectively used for denoising hyperspectral remote sensing images, improving the interpretation level and application range of subsequent hyperspectral data. BRIEF DESCRIPTION OF THE DRAWINGS

[0027] In order to more clearly illustrate the technical solutions in the embodiments of the present application or the prior art, the following will briefly introduce the drawings required for the description of the embodiments or the prior art. Obviously, the drawings in the following description are only some embodiments recorded in the present application. For those of ordinary skill in the art, other drawings can be obtained based on these drawings without creative efforts.

[0028] Figure 1 Is the flowchart of the method of the present invention DETAILED DESCRIPTION OF THE EMBODIMENTS

[0029] To make the technical solutions and advantages of the present invention clearer, the following will clearly and completely describe the technical solutions in the embodiments of the present invention with reference to the drawings in the embodiments of the present invention:

[0030] As Figure 1 shown, a denoising method for hyperspectral remote sensing images coupling deep prior and low-rank tensor decomposition specifically includes the following steps:

[0031] S1: Considering the correlation between adjacent bands of hyperspectral images, input the noisy hyperspectral data, the spatial data of the current band, and the spectral data of the upper and lower neighboring bands corresponding to the current band band by band. Use the deep denoising prior sub-network 1 to perform convolution filtering on the spatial data and spectral data with a window size of 2D 3×3 for each band, and extract spectral-spatial features through the convolutional layer and ReLU layer to obtain the preliminary denoising result of the optimized hyperspectral image:

[0032]

[0033] Among them, Γ(g) represents the per-band traversal process in sub-network 1. is the preliminary denoising result. Y i is the i-th band in the noisy hyperspectral image, is Y i corresponding adjacent spectral gradient.

[0034] S2: Due to the complexity of the mixed noise distribution, the hyperspectral image still contains residual noise after being optimized by the deep prior denoising sub-network 1, especially in the spectral dimension. To overcome this problem, a low-rank tensor decomposition strategy is introduced to better utilize the low-rank prior of the third-order tensor. Based on the low-rank prior of the third-order tensor, the preliminary denoising result of the hyperspectral image is decomposed by singular value decomposition to obtain the estimated spectral orthogonal basis:

[0035]

[0036] Among them, A is the spectral orthogonal product. represents the 3rd-order unfolding form of the tensor . SVD r (g) represents the truncation result of the rank parameter r in the singular value decomposition mode.

[0037] S3: After obtaining the spectral orthogonal basis in the above process, reconstruct the 3rd-order tensor unfolding form of the preliminary denoising hyperspectral image and the spectral orthogonal basis to obtain the initialization factor. Based on the deep residual learning strategy, use the deep prior denoising sub-network 2 to perform feature extraction and expression on the initialization factor band by band to eliminate the noise in the initialization factor and obtain the spatial reduction factor. The specific process is as follows:

[0038]

[0039] Among them, the reshape h,w,r (g) function represents the dimension reshaping operation. is the initialization factor. The reshaped initialization factor still contains noise. Use the strong feature extraction ability of the deep prior denoising sub-network 2 to estimate the spatial reduction factor:

[0040]

[0041] Among them, the N(g) function represents normalizing each band of . The IN(g) function is the inverse operation of the normalization of N(g).

[0042] S4: Based on the low-rank tensor decomposition strategy, represent the global spectral low-rank property of the hyperspectral image as Tucker decomposition, and obtain the final denoising result of the hyperspectral image according to the obtained spectral orthogonal basis and spatial reduction factor:

[0043]

[0044] Among them, ×3 represents the modulo-3 tensor matrix product.

[0045] This method takes into account the spatial and spectral neighborhood information of the hyperspectral image, couples the deep learning network and low-rank tensor decomposition, removes the noise in the hyperspectral image, and improves the data quality of the hyperspectral image, which has important practical significance.

[0046] The above is only a preferred specific implementation manner of the present invention, but the protection scope of the present invention is not limited thereto. Any person skilled in the art within the technical scope disclosed by the present invention, according to the technical solution and inventive concept of the present invention, makes equivalent substitutions or changes, and should be covered by the protection scope of the present invention.

Claims

1. A hyperspectral remote sensing image denoising method that couples depth prior and low-rank tensor decomposition, characterized in that Including: Using a deep denoising prior sub-network one to optimize the initial noisy hyperspectral image; Using a low-rank tensor to decompose the optimized hyperspectral image to generate spectral orthogonal bases; Estimating the spatial reduction factor through a deep denoising prior sub-network two, specifically: reconstructing the three-way tensor unfolding form of the preliminarily denoised hyperspectral image and the spectral orthogonal bases to obtain an initial factor, and using the deep denoising prior sub-network two to perform feature extraction on the initial factor band by band based on the deep residual learning strategy to eliminate the noise in the initial factor and obtain the spatial reduction factor; Based on the spectral orthogonal bases and the spatial reduction factor, using the tensor train decomposition method to obtain the denoising result of the hyperspectral image.

2. The hyperspectral remote sensing image denoising method based on coupling depth prior and low-rank tensor decomposition according to claim 1, characterized in that: When optimizing the initial noisy hyperspectral image: Extracting the spatial data of the current band of the noisy hyperspectral data and the spectral data of the corresponding upper and lower neighboring bands of the current band band by band, and using the deep denoising prior sub-network one to perform convolutional filtering on the spatial data and the spectral data to obtain the preliminary denoising result of the optimized hyperspectral image; Among them, Γ(·) represents the per-band traversal process in the sub-depth denoising prior sub-network one, is the preliminary denoising result, Y i is the i-th band in the noisy hyperspectral image, is the adjacent spectral gradient corresponding to Y i ​ 3. The hyperspectral remote sensing image denoising method based on coupling depth prior and low-rank tensor decomposition according to claim 2, wherein: Generating spectral orthogonal bases in the following way: The preliminary denoising result of the hyperspectral image based on the low-rank prior of the third-order tensor The estimated spectral orthogonal basis is obtained by singular value decomposition: where A is the spectral orthogonal product, denotes the 3-mode unfolding form of the tensor , and SVD r (·) represents the truncation result of the rank parameter r in the singular value decomposition mode.

4. The hyperspectral remote sensing image denoising method based on coupling depth prior and low-rank tensor decomposition according to claim 2, wherein: When estimating the spatial reduction factor through the deep denoising prior sub-network two, the following algorithm is specifically adopted: Among them, reshape h,w,r (·) function represents the dimension reshaping operation, is the initialization factor. The generated initialization factor after reshaping still contains noise. The feature extraction ability of the deep denoising prior sub-network two is used to estimate the spatial reduction factor: Among them, the N(·) function represents normalizing each band of , and the IN(·) function is the inverse operation of the normalization of N(·).

5. The hyperspectral remote sensing image denoising method based on coupling depth prior and low-rank tensor decomposition according to claim 4, wherein: When obtaining the denoising result of the hyperspectral image: Based on the low-rank tensor decomposition strategy, representing the global spectral low-rank property of the hyperspectral image as a tensor train decomposition, and obtaining the final denoising result of the hyperspectral image according to the obtained spectral orthogonal bases and the spatial reduction factor: Among them, ×3 represents the modulo-3 tensor matrix product.

Images