A high-performance spaceborne hyperspectral remote sensing image compression method

Abstract

The application discloses a high-performance spaceborne hyperspectral remote sensing image compression method. The method is based on two classical compression algorithms of compression sensing and tensor decomposition, random sampling is performed at a star end, and the data quantity is directly reduced. Tensor decomposition eliminates data redundancy among spectrums, successfully realizes hyperspectral remote sensing image compression at low operation and low storage cost, and is very suitable for spaceborne hyperspectral remote sensing image compression.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of image processing, and particularly relates to a high-performance spaceborne hyperspectral remote sensing image compression method. BACKGROUND

[0002] Hyperspectral remote sensing technology is developed on the basis of imaging spectroscopy and is a frontier field of current remote sensing technology. Hyperspectral remote sensing technology can simultaneously obtain spatial information and spectral information of a ground object, and can reflect the characteristics of the ground object in continuous bands. The rich spatial, radiation and spectral triple information greatly improves the ground object identification capability of remote sensing technology. Hyperspectral remote sensing image belongs to three-dimensional cubic data, but the increase of the number of bands will cause a sharp increase in data volume, making data storage and transmission a big difficulty. Therefore, it is of important practical significance to study the efficient compression of hyperspectral remote sensing data.

[0003] Compared with traditional images, hyperspectral remote sensing images contain the energy emitted and reflected by ground objects in continuous bands, resulting in a large amount of redundancy in the data set, and being limited by channel bandwidth and star-end storage capacity. Therefore, it is necessary to compress hyperspectral remote sensing images. In addition to the spatial correlation of traditional images, hyperspectral remote sensing images also have spectral correlation, which provides a new idea for finding a compression method for hyperspectral remote sensing images.

[0004] Traditional Nyquist sampling produces too much redundancy, and the use of compressed sensing algorithm can solve this problem through unequal interval sampling. Compressed sensing algorithm is to obtain a small amount of observation data s by sampling the original signal in a certain transform domain with sparsity y through an observation matrix which is irrelevant to the transform basis, to complete data compression, which can be expressed by formula as follows: x wherein, s is the sparse signal in a certain sparse domain, is a sparse transform matrix, and satisfies , let the sensing matrix , and the underdetermined equation group can be obtained. By solving the optimization problem through the observation value, the original signal in the sparse domain can be reconstructed by using the small amount of observation data y to obtain the sparse signal x to complete decompression. However, the current compressed sensing algorithm is mainly used for the compression of two-dimensional images, and the research on hyperspectral remote sensing images is relatively less, and the reconstruction quality effect is relatively poor.

[0005] The commonly used methods in tensor decomposition include CP decomposition and Tucker decomposition, wherein the CP decomposition is to decompose a three-dimensional tensor into a sum of outer products of a series of vectors, and the Tucker decomposition is to decompose a three-dimensional tensor into a smaller three-dimensional core tensor and three mode matrices, and the tensor can be restored by multiplying the tensor core and the mode matrices.

[0006] The traditional compression algorithms have advantages and disadvantages in terms of operation speed and compression effect, and therefore, it is of great significance to find a balance between speed and effect and propose a compression algorithm combining the advantages of classical algorithms. SUMMARY

[0007] Considering the limited hardware conditions, insufficient operation and storage capacity of the satellite end and the characteristics of the spectral data redundancy of the hyperspectral remote sensing image, a high-performance satellite-borne hyperspectral remote sensing image compression method based on compression sensing and tensor decomposition is proposed. The compression sensing algorithm involved in the method has the characteristics of more simplified compression and the workload can be mainly concentrated on the decompression end, and the tensor decomposition algorithm involved in the method has the characteristics of eliminating the data redundancy of the image within and between the spectra at the same time, so that the workload of the satellite end can be reduced, and the quality of the reconstructed image after compression can be improved.

[0008] In order to achieve the above purpose, the technical scheme provided by the present application is a high-performance satellite-borne hyperspectral remote sensing image compression method, comprising the following steps:

[0009] Step 1, sparse processing is performed on the input hyperspectral remote sensing image to obtain a sparse image;

[0010] Step 2, a sensing matrix is constructed according to the size of the single-band image and the compression multiple;

[0011] Step 3, the sparse image of each band is respectively multiplied by the sensing matrix, and the data with a value of 0 is deleted, and the deletion operation parameters are saved, and the sensing matrix sampling is completed;

[0012] Step 4, the image after the sensing matrix sampling is further compressed by using the CP decomposition algorithm in tensor decomposition to obtain a mode matrix and a core tensor;

[0013] Step 5, tensor multiplication is performed on the mode matrix and the core tensor to complete decompression and obtain reconstructed data;

[0014] Step 6, according to the deletion operation parameters saved in step 3, the data with a value of 0 in the reconstructed data is restored, and the sparse domain image data is obtained by using the fast composite splitting algorithm in compression sensing for reconstruction;

[0015] Step 7, the inverse sparse processing is respectively performed on each band image obtained in step 6 to complete data decompression.

[0016] Furthermore, the step 1 input hyperspectral remote sensing image , x, y, z respectively, the length, width, and band number of the image, and the single image of each band is subjected to fast Fourier transform to generate a sparse domain image .

[0017] Furthermore, the step 2 of constructing the sensing matrix comprises the following steps:

[0018] Step 2.1, setting the image length as x , the width as y , the compression multiple as r , the maximum number of iterations as T , and the iteration step as Step, , constructing a zero matrix with the size of , SPF and Record , a random sensing matrix Mask , and setting the iteration number as , ;

[0019] Step 2.2, if and are true, setting ; otherwise, executing step 2.3;

[0020] Step 2.3, judging whether and are true, if true, executing step 2.5, otherwise, executing step 2.4;

[0021] Step 2.4, judging whether is true, if true, setting , ; otherwise, setting , and returning to step 2.2;

[0022] Step 2.5, judging whether is true, if not true, executing step 2.6, otherwise, setting , and returning to step 2.2;

[0023] Step 2.6, judging whether is true, if not true, executing step 2.7; otherwise, setting , and returning to step 2.2;

[0024] Step 2.7, setting , , and judging whether is true, if true, executing step 2.8, otherwise, returning to step 2.2;

[0025] Step 2.8, let , traverse the matrix , if , let ; otherwise let ;

[0026] Step 2.9, end.

[0027] Moreover, the step 3 lets each band image of the sparse domain image respectively multiply the sensing matrix Mask , to obtain the compressed result compressed by compressed sensing ; find , , , , delete data , wherein or , to obtain data , wherein , .

[0028] Moreover, the step 4 compresses the data using the CP decomposition algorithm in the tensor decomposition algorithm to obtain the mode matrix , , and the core tensor , and completes the compression.

[0029] Moreover, the step 5 performs tensor multiplication on the data mode matrix , , and the core tensor to complete decompression, and obtains the reconstructed data .

[0030] Moreover, the step 6 creates a full zero tensor , and traverses in a loop to let , wherein , , and the FCSA algorithm in compressed sensing is used to reconstruct to obtain the sparse domain image data .

[0031] Moreover, the step 7 uses inverse Fourier transform on the sparse domain image data to obtain the final reconstructed hyperspectral remote sensing image data .

[0032] Compared with the prior art, the present application has the following advantages:

[0033] 1) The first step of compression is performed using the compressed sensing algorithm, which greatly reduces the amount of original data and alleviates the storage burden of the tensor decomposition algorithm. At the same time, the presence of a large number of zeros in the data reduces the rank of the data, allowing the tensor decomposition results to converge better and reducing the computational burden caused by algorithm iteration.

[0034] 2) The tensor decomposition algorithm is used for the second step of compression, which significantly improves the reconstruction quality of the compressed sensing algorithm while keeping the total compression ratio unchanged. Attached Figure Description

[0035] Figure 1 This is a flowchart of an embodiment of the present invention.

[0036] Figure 2 This is a flowchart for creating a sensing matrix. Detailed Implementation

[0037] This invention provides a high-performance method for compressing spaceborne hyperspectral remote sensing images. The technical solution of this invention will be further explained below with reference to the accompanying drawings, taking the data compression of Pavia hyperspectral remote sensing images as an example.

[0038] like Figure 1 As shown, the process of this embodiment of the invention includes the following steps:

[0039] Step 1: Perform sparsification processing on the input hyperspectral remote sensing image to obtain a sparse image.

[0040] Input hyperspectral remote sensing images , x, y, z These represent the length, width, and number of bands of the image, and a single image for each band. Perform Fast Fourier Transform to generate sparse domain image .

[0041] Step 2, constructing the sensing matrix based on the single-band image size and compression factor, includes the following steps:

[0042] Step 2.1, set the image length to x Width is y Compression factor is r The maximum number of iterations is T The iteration step size is Step, Build size is All-zero matrix SPF and Record random sensing matrix Mask Let the number of iterations be... , ;

[0043] Step 2.2, if and If established, then... Otherwise, proceed to step 2.3;

[0044] Step 2.3, Determine and If the condition is met, proceed to step 2.5; otherwise, proceed to step 2.4.

[0045] Step 2.4, Determine Is it true? If it is true, then let , Otherwise, let Then return to step 2.2;

[0046] Step 2.5, Determine Check if it is true. If not, proceed to step 2.6; otherwise, let... Return to step 2.2;

[0047] Step 2.6, Determine Check if it is true. If not, proceed to step 2.7; otherwise, let... Return to step 2.2;

[0048] Step 2.7, let , ,judge Check if the condition is met. If it is met, proceed to step 2.8; otherwise, return to step 2.2.

[0049] Step 2.8, let traverse the matrix ,like Then let Otherwise, ;

[0050] Step 2.9, End.

[0051] Step 3: Multiply the sparsed image of each band by the sensing matrix, delete the data with a value of 0, save the deletion operation parameters, and complete the sensing matrix sampling.

[0052] Let each band of the sparse domain image be Dot product of the sensing matrix Mask The result of compressed sensing is obtained. Searching , , , Delete data ,in or , obtain data ,in , .

[0053] Step 4, further compress the image after sampling the sensing matrix using CP decomposition algorithm in tensor decomposition to obtain the mode matrix and core tensor.

[0054] The data is compressed using CP decomposition algorithm in tensor decomposition algorithm to obtain the mode matrix , , and core tensor , and the compression is completed.

[0055] Step 5, the mode matrix and core tensor are subjected to tensor multiplication to complete decompression and obtain the reconstructed data.

[0056] The data mode matrix , , and core tensor are subjected to tensor multiplication to complete decompression and obtain the reconstructed data .

[0057] Step 6, according to the deletion operation parameters saved in step 3, the data with value 0 in the reconstructed data is recovered, and the fast composite splitting algorithm in compressed sensing is used for reconstruction to obtain the sparse domain image data.

[0058] A full zero tensor is created, and the loop is traversed to let , wherein , The FCSA algorithm in compressed sensing is used to reconstruct to obtain the sparse domain image data .

[0059] Step 7, the image of each band reconstructed in step 6 is respectively subjected to inverse sparsification to complete data decompression.

[0060] The sparse domain image data is subjected to inverse Fourier transform to obtain the final reconstructed hyperspectral remote sensing image data .

[0061] The specific embodiments described herein are merely illustrative of the spirit of the present application. Those skilled in the art of the present application can make various modifications or supplements to the described specific embodiments or use similar ways to replace, but will not deviate from the spirit of the present application or exceed the scope defined by the appended claims.

Claims

1. A high-performance spaceborne hyperspectral remote sensing image compression method, characterized in that, Includes the following steps: Step 1: Perform sparsification processing on the input hyperspectral remote sensing image to obtain a sparse image; Input hyperspectral remote sensing images , x, y, z These represent the length, width, and number of bands of the image, and a single image for each band. Perform Fast Fourier Transform to generate sparse domain image ; Step 2: Construct a sensing matrix based on the single-band image size and compression factor; Step 2.1, set the image length to x Width is y Compression factor is r The maximum number of iterations is T The iteration step size is Step, Build size is All-zero matrix SPF and Record random sensing matrix Mask Let the number of iterations be... , ; Step 2.2, if and If established, then... ; Otherwise, proceed to step 2.3; Step 2.3, Determine and If the condition is met, proceed to step 2.5; otherwise, proceed to step 2.

4. Step 2.4, Determine Is it true? If it is true, then let , Otherwise, let Then return to step 2.2; Step 2.5, Determine Check if it is true. If not, proceed to step 2.6; otherwise, let... Return to step 2.2; Step 2.6, Determine Check if it is true. If not, proceed to step 2.7; otherwise, let... Return to step 2.2; Step 2.7, let , ,judge Check if the condition is met. If it is met, proceed to step 2.8; otherwise, return to step 2.

2. Step 2.8, let traverse the matrix ,like Then let Otherwise, ; Step 2.9, End; Step 3: Multiply the sparsed image of each band by the sensing matrix, delete the data with a value of 0, save the deletion operation parameters, and complete the sensing matrix sampling. Let each band of the sparse domain image be Dot product of the sensing matrix Mask The result of compressed sensing is obtained. Search , , , Delete data ,in or , obtain data ,in , ; Step 4: Use the CP decomposition algorithm in tensor decomposition to further compress the image after the sensor matrix is ​​sampled to obtain the mode matrix and core tensor. Step 5: Decompress the pattern matrix and the core tensor by performing tensor multiplication to obtain the reconstructed data; Step 6: Based on the deletion operation parameters saved in Step 3, restore the data with a value of 0 in the reconstructed data, and use the fast composite splitting algorithm in compressed sensing to reconstruct the sparse domain image data. Step 7: Perform inverse sparsification on the images of each band obtained from the reconstruction in Step 6 to complete the data decompression.

2. The high-performance spaceborne hyperspectral remote sensing image compression method as described in claim 1, characterized in that: In step 4, the data Compression was performed using the CP decomposition algorithm from the tensor decomposition algorithm to obtain the pattern matrix. , , and core tensor Compression complete.

3. The high-performance spaceborne hyperspectral remote sensing image compression method as described in claim 2, characterized in that: In step 5, the data pattern matrix will be... , , and core tensor Tensor multiplication is performed to decompress the data and obtain the reconstructed data. .

4. The high-performance spaceborne hyperspectral remote sensing image compression method as described in claim 3, characterized in that: In step 6, create an all-zero tensor. , loop traversal ,in , Using the FCSA algorithm in compressed sensing to... Sparse domain image data obtained through reconstruction .

5. The high-performance spaceborne hyperspectral remote sensing image compression method as described in claim 4, characterized in that: Step 7 involves processing sparse domain image data. The inverse Fourier transform is used to obtain the final reconstructed hyperspectral remote sensing image data. .

Images