CN111107360B - A method and system for lossless compression of hyperspectral images combined with spectral-spatial dimension - Google Patents

Abstract

The invention discloses a spectral-spatial dimension combined hyperspectral image lossless compression method and system. First, the original hyperspectral image is preprocessed to correct nonlinear factors caused by imaging instruments; Reversible integer transform to remove spectral correlation; calculate image uniformity index channel by channel for hyperspectral images after reversible integer transformation; use FILF compressor to encode relatively uniform image channels, and switch to arithmetic encoder for image channels with poor uniformity Coding; decoding through corresponding decoding methods, removing the correlation between spectra, performing reversible integer inverse transformation on the image that has removed the correlation between spectra, and restoring the original hyperspectral image before compression. The invention removes the correlation between the spectra by performing a reversible integer transformation on the hyperspectral image, so as to reduce the information entropy of the original data, remove the redundant information between the spectra, and adopt FILF compression coding to extract a high compression ratio. Arithmetic encoder speeds up compression.

Description

A method and system for lossless compression of hyperspectral images combined with spectral-spatial dimension

technical field

The invention belongs to the field of image processing, and more particularly, relates to a method and system for lossless compression of hyperspectral images combined with spectral-spatial dimensions.

Background technique

The spectral resolution rate of hyperspectral remote sensing reaches the nanometer level, and the number of bands in the visible light to short-wave infrared spectral range is as many as tens to hundreds. Higher spectral resolution enables hyperspectral images to provide more detailed information on ground features, and has been widely used in geological surveys, mineral deposit detection, precision agriculture and marine remote sensing, environmental and disaster monitoring, and military reconnaissance. In the development of hyperspectral remote sensing technology, with the continuous improvement of spectral resolution and spatial resolution, the amount of data acquired by imaging spectrometers has expanded rapidly, which has brought enormous pressure to data storage and transmission.

Since remote sensing image information is very valuable, lossless compression should be used as much as possible. Lossless compression does not allow any loss of original image information, and there is no error between the image restored after decoding and the original image. Lossless compression is the compression of the file itself. Like the compression of other data files, it is to optimize the data storage method of the file. A certain algorithm is used to represent the repeated data information. The file can be completely restored without affecting the content of the file. In terms of data, there will be no loss of image details. Therefore, lossless compression can be regarded as a reversible process.

In the field of still visible light image compression, a unified international compression standard has been formulated, while in the field of hyperspectral image compression, no compression standard has yet been formed. At present, for the compression of spaceborne multispectral or hyperspectral images, most popular visible light compression systems such as PNG, JPEG-LS, and JPEG2000 are directly applied. These visible light image compression systems can only compress single-channel grayscale images or three-channel RGB images. Only the spatial correlation in the image spectrum can be eliminated, and the correlation between the bands is not considered, the compression performance is low, and the data transmission bandwidth cannot be effectively reduced. Regarding how to eliminate the correlation between spectra, there are many prediction-based methods in the field of hyperspectral image compression. Using simple predictors (such as linear prediction) often cannot reduce the prediction error, and the compression ability is weak, while using complex predictors (such as BP neural network) ) can guarantee a small prediction error, but the size of the model itself is relatively large, and the volume of the compressed image file may not decrease but increase, including the size of the model file. In addition, simply applying a popular visible light compressor to compress hyperspectral images will waste a lot of compression time. Therefore, an effective and practical method for lossless compression of hyperspectral image data is urgently needed in the art.

SUMMARY OF THE INVENTION

In view of the defects of the prior art, the purpose of the present invention is to provide a method and system for lossless compression of hyperspectral images combined with spectrum-spatial dimension, aiming to solve the problem of FILF in the field of existing visible light images still used in the lossless compression technology of hyperspectral image data. Therefore, it is impossible to remove the strong inter-spectral correlation of hyperspectral images, resulting in the problem of limited compression performance.

In order to achieve the above object, according to an aspect of the present invention, there is provided a method for lossless compression of hyperspectral images combined with spectrum-spatial dimension, comprising the following steps:

(1) Preprocess the original hyperspectral image to correct the nonlinear factors caused by imaging;

(2) Perform reversible integer transformation on the preprocessed hyperspectral image, remove the correlation between spectra, and save the transformation matrix;

(3) Calculate the image uniformity channel-by-channel for the hyperspectral image after reversible integer transformation, determine the uniformity evaluation index and divide it into a relatively uniform image channel and an image channel with poor uniformity;

(4) adopting FILF compressor encoding for relatively uniform image channels, and adopting arithmetic encoder encoding for image channels with poor uniformity to realize lossless compression of hyperspectral images;

(5) decoding the compressed hyperspectral image channel-by-channel, adopting arithmetic encoder encoding and adopting the image channel to be decoded of FILF compressor encoding to decode the image that removes the correlation between spectra by corresponding decoding mode;

(6) Read the transformation matrix obtained in step (2), perform reversible integer inverse transformation on the image whose spectral correlation has been removed, and restore the original hyperspectral image before compression.

Further, step (2) specifically includes:

(2.1) Calculate the Pearson correlation matrix Coeff between the spectra of the preprocessed hyperspectral image data:

Among them, ρ(p _i , p _j ) represents the Pearson correlation coefficient between the i-th spectral segment and the j-th spectral segment, and the specific calculation formula is as follows:

Among them, Cov(pi , p _j ) represents the covariance between spectral segments, D( _pi ), D(p _j ) are the variances of the i- _th spectral segment and the j-th spectral segment;

(2.2) Perform principal component analysis on the Pearson correlation matrix Coeff to obtain the transformation matrix A. The transformation matrix A above can only realize the transformation from real numbers to real numbers, and the transformation matrix A=PLUS is obtained through decomposition to realize the mapping from integers to integers, where L and U are respectively Unit lower triangular and upper triangular matrix, P is a permutation matrix, S is a unit-element invertible matrix. The specific decomposition process of A is as follows:

Suppose A is a non-singular matrix that satisfies the decomposition conditions:

(2-3) First, there is a row transformation matrix P ₁ such that

不为零，可以找到s₁使

即

于是构造一个单位元素可逆矩阵S₁：The first element of the Nth column of

is not zero, s ₁ can be found such that

which is

Then construct a unit-element invertible matrix S ₁ :

then there are

Then use the traditional Gaussian elimination method to define the elementary Gaussian matrix:

Then there are:

(2-4) Repeat the above process N-1 times to obtain:

L _N-1 P _N-1 …L ₂ P ₂ L ₁ P ₁ AS ₁ S ₂ …S _N-1 ＝D _R U

where k=1, 2,...N-1; _DR is the twiddle factor, _DR =±1

(2-5) The transformation matrix A can be finally decomposed into the following formula:

A=PLUS

in:

P=(P _N-1 …P ₂ P ₁ ) ^T

U＝±L _N-1 P _N-1 …L ₂ P ₂ L ₁ P ₁ AS ₁ S ₂ …S _N-1

(2-6) Since L is a lower triangular matrix, U is an upper triangular matrix, and only the last row of the S matrix has valid data, the LUS can be combined into a real matrix and saved.

(2-7) Multiply the hyperspectral image data data with the S, U, L, P transformation matrix to the left in turn, and round the matrix after each multiplication to realize the integer KLT transformation. The specific formula is as follows:

Y=P*round(L*round(U*round(S*data)))

Among them, round represents the rounding and rounding operation.

Further, step (3) specifically includes:

(3.1) Calculate the minimum value of the hyperspectral image after reversible integer transformation channel by channel and save it, and then subtract the minimum value of the corresponding channel channel by channel, so that the channel image data is a positive integer greater than or equal to 0;

(3.2) Check the channel image data obtained after the processing in step (3.1) whether there is a maximum overflow problem one by one. If there is an overflow in the channel image data, split the channel into 2 channels, one saves the high-order data, and the other save low-order data;

(3.3) Calculate the image uniformity of the channel image data obtained after the processing in step (3.2), and then divide the uniformity index into a relatively uniform image channel and an image channel with poor uniformity.

Further, an image channel has M rows and N columns, and the average value of row pixels is u _i , and the determination process of the uniformity evaluation index is as follows:

The autocorrelation coefficient ρ _x of adjacent row pixels in the image channel, the specific formula is as follows:

Image channel row autocorrelation coefficient mean

The autocorrelation coefficient ρ _y of adjacent column pixels in the image channel, the specific formula is as follows:

Image channel column autocorrelation coefficient mean

和图像通道列自相关系数均值

定义图像均匀性评价指标psr，具体公式如下：Combined image channel row autocorrelation coefficient mean

and image channel column autocorrelation coefficient mean

Define the image uniformity evaluation index psr, and the specific formula is as follows:

Further, step (4) specifically includes:

Use the FILF compressor to encode the more uniform image channels, calculate the gray value distribution for the image channels with poor uniformity, use the Gaussian function to fit the gray value distribution, and save the fitted Gaussian function parameters. The combined Gaussian function parameters are used to entropy encode the channel image.

Further, step (5) specifically includes:

Decode the compressed hyperspectral image channel by channel. For the image channel to be decoded encoded by the arithmetic encoder, first read the fitted Gaussian function parameters, and combine the arithmetic encoder and the fitted Gaussian function parameters for the channel image data. Decoding: The image channel to be decoded encoded by the FILF compressor is decoded by the FILF compressor.

Further, step (6) specifically includes:

Read the transformation matrix obtained in step (2), and perform reversible integer inverse transformation on the image that has removed the inter-spectral correlation;

Check the records of the dismantled channels, and combine the dismantled channels according to the high order and low order;

Read the minimum value of the image data of each channel after merging, and add the minimum value to the corresponding channel to restore the original hyperspectral image before compression.

The traditional method of directly using visible light compressors to compress hyperspectral images usually compresses adjacent three-channel groups of hyperspectral images. Although visible light compressors also have the ability to eliminate intra-group spectral redundancy, they cannot eliminate grouping. Inter-spectral redundancy information. There are many ways to remove the correlation between spectra, but most of the transformations are real transformations, which are not suitable for image lossless compression, and can only be used for image lossy compression. To achieve lossless compression of data, in addition to requiring the algorithm to be an integer transformation, the algorithm must also be reversible. At present, there are only reversible integer KLT transform and integer wavelet transform in the reversible integer transform method. However, the integer wavelet transform belongs to the frequency domain transform and is more suitable for eliminating the spatial information correlation in the image spectrum. It cannot guarantee the complete removal of the inter-spectral correlation. It belongs to orthogonal transformation. After transformation, the data between spectra are orthogonal, and the ability to eliminate the correlation between spectra is stronger. The reason why FILF compressor is selected is because it is the best compressor in the field of visible light image compression, and its performance exceeds popular compressors such as PNG, JPEG-LS and JPEG2000. Regardless of the spectral band of the hyperspectral image, the imaging bands of adjacent channel images are very close, and generally have strong inter-spectral correlation.

According to another aspect of the present invention, there is provided a lossless compression system suitable for full-spectrum hyperspectral images, characterized in that it includes a preprocessing module, a de-correlation module, an image uniformity evaluation module, a FILF encoder, and a Arithmetic encoder, FILF decoder and arithmetic decoder, image restoration module;

The preprocessing module is used to preprocess the original hyperspectral image to correct the nonlinear factors caused by imaging;

The inter-spectral correlation removal module is used to perform reversible integer transformation on the preprocessed hyperspectral image, remove the inter-spectral correlation, and save the transformation matrix;

The image uniformity evaluation module is used to calculate the image uniformity channel-by-channel for the hyperspectral image after reversible integer transformation, and after determining the uniformity evaluation index, it is divided into relatively uniform image channels and poor uniformity image channels;

The FILF encoder and the arithmetic encoder are respectively used for sampling the relatively uniform image channel and encoding the image channel with poor uniformity, so as to realize lossless compression of hyperspectral images;

The FILF decoder and the arithmetic decoder are used for channel-by-channel decoding of the compressed hyperspectral image, and the to-be-decoded image channels encoded by the arithmetic encoder and encoded by the FILF compressor are decoded by corresponding decoding methods to remove spectral noise. Relevant images;

The image restoration module is used to read the transformation matrix obtained by the de-inter-spectral correlation module, perform reversible integer inverse transformation on the image from which the inter-spectral correlation has been removed, and restore the original hyperspectral image before compression.

Further, the image uniformity evaluation module includes:

The high and low bit splitting unit is used to calculate the minimum value of the hyperspectral image after reversible integer transformation channel by channel and save it, and then subtract the minimum value of the corresponding channel channel by channel, so that the channel image data is a positive integer greater than or equal to 0. Check whether there is a maximum overflow problem in the channel. If there is an overflow in the image data of the channel, split the channel into 2 channels, one saves the high-order data, and the other saves the low-order data;

The uniformity determining unit is used for calculating the image uniformity for the obtained channel image data, and after determining the uniformity index, it is divided into a relatively uniform image channel and an image channel with poor uniformity.

Further, the FILF compressor is used to encode the relatively uniform image channel, the gray value distribution is calculated for the image channel with poor uniformity, the Gaussian function is used to fit the gray value distribution, and the fitted Gaussian function parameters are saved, combined with The arithmetic encoder and the fitted Gaussian function parameters perform entropy encoding on the channel image;

Decode the compressed hyperspectral image channel by channel. For the image channel to be decoded encoded by the arithmetic encoder, first read the fitted Gaussian function parameters, and combine the arithmetic encoder and the fitted Gaussian function parameters for the channel image data. Decoding: The image channel to be decoded encoded by the FILF compressor is decoded by the FILF compressor.

Through the above technical solutions conceived by the present invention, compared with the prior art, the following beneficial effects can be achieved:

1. The spectral-spatial dimension combined hyperspectral image lossless compression method proposed by the present invention is suitable for full-band hyperspectral images, and the strong correlation between all spectral segments of the hyperspectral image is removed through the reversible integer KLT transformation, reducing the hyperspectral image. The information entropy can eliminate the information redundancy between spectra and improve the performance of the FILF compressor; it also solves the possible integer overflow problem and improves the robustness of the FILF compressor.

2. The FILF compressor used in the present invention belongs to a complex compression system, with strong compression performance, but the internal algorithm is complex and time-consuming; while the arithmetic encoder belongs to a simple compression system, with high speed but weak compression capability. In order to shorten the compression time without losing the compression performance as much as possible, the present invention adopts a strategy of flexibly switching between the FILF compressor and the arithmetic encoder according to the uniformity of the image, and the FILF compressor should be selected for a uniform image to improve the compression ability. ; But for the non-uniform image with large noise, the performance of FILF compressor is comparable to that of arithmetic coder, and at this time, a faster arithmetic coder is selected to shorten the compression time.

Description of drawings

Fig. 1 is the flow chart of the practical hyperspectral image lossless compression of improving FILF provided by the present invention;

Fig. 2 is the public hyperspectral data Salinas image provided by the present invention;

Fig. 3 (a) is the first channel diagram of image Salinas after KLT transformation provided by the present invention;

Fig. 3 (b) is the 2nd channel diagram of image Salinas after KLT transformation provided by the present invention;

Fig. 3 (c) is the 203rd channel diagram of image Salinas after KLT transformation provided by the present invention;

Fig. 3 (d) is the 204th channel diagram of the image Salinas after KLT transformation provided by the present invention;

Fig. 4 (a) is the correlation curve diagram of the adjacent spectral section before transformation provided by the present invention;

Fig. 4 (b) is the correlation curve diagram of adjacent spectral sections after transformation provided by the present invention;

FIG. 5 is a trend diagram of the maximum dynamic range of pixel values of each channel after the Salinas processing of the image provided by 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. In addition, the technical features involved in the various embodiments of the present invention described below can be combined with each other as long as they do not conflict with each other.

The present invention provides a method for removing the spectral correlation of hyperspectral images. Different from the common real number domain KLT analysis, the method uses a reversible KLT algorithm, which can ensure that the transformed data is still an integer. Therefore, the hyperspectral image lossless compression method designed in the present invention is especially suitable for the application scenario of the hyperspectral image data lossless compression.

As shown in FIG. 1, the present invention proposes a method for lossless compression of hyperspectral images combined with spectral-spatial dimension, which includes the following steps:

(1) Preprocess the original hyperspectral image to correct nonlinear factors such as image deformation and response drift caused by imaging;

(2) Perform reversible integer transformation on the preprocessed hyperspectral image, remove the correlation between spectra, and save the transformation matrix;

(3) Calculate the image uniformity channel-by-channel for the hyperspectral image after reversible integer transformation, determine the uniformity evaluation index and divide it into a relatively uniform image channel and an image channel with poor uniformity;

(4) adopting FILF compressor encoding for relatively uniform image channels, and adopting arithmetic encoder encoding for image channels with poor uniformity to realize lossless compression of hyperspectral images;

(5) decoding the compressed hyperspectral image channel-by-channel, adopting arithmetic encoder encoding and adopting the image channel to be decoded of FILF compressor encoding to decode the image that removes the correlation between spectra by corresponding decoding mode;

(6) Read the transformation matrix obtained in step (2), perform reversible integer inverse transformation on the image whose spectral correlation has been removed, and restore the original hyperspectral image before compression.

The present invention also provides a lossless compression system suitable for full-spectrum hyperspectral images, which is characterized in that it includes a preprocessing module, a de-inter-spectral correlation module, an image uniformity evaluation module, a FILF encoder and an arithmetic encoder, and a FILF encoder. Decoder and arithmetic decoder, image restoration module;

The preprocessing module is used to preprocess the original hyperspectral image to correct nonlinear factors such as image deformation and response drift caused by imaging;

The inter-spectral correlation removal module is used to perform reversible integer transformation on the preprocessed hyperspectral image, remove the inter-spectral correlation, and save the transformation matrix;

The image uniformity evaluation module is used to calculate the image uniformity channel-by-channel for the hyperspectral image after reversible integer transformation, and after determining the uniformity evaluation index, it is divided into relatively uniform image channels and poor uniformity image channels;

The FILF encoder and the arithmetic encoder are respectively used for sampling the relatively uniform image channel and encoding the image channel with poor uniformity, so as to realize lossless compression of hyperspectral images;

The FILF decoder and the arithmetic decoder are used for channel-by-channel decoding of the compressed hyperspectral image, and the to-be-decoded image channels encoded by the arithmetic encoder and encoded by the FILF compressor are decoded by corresponding decoding methods to remove spectral noise. Relevant images;

The image restoration module is used to read the transformation matrix obtained by the de-inter-spectral correlation module, perform reversible integer inverse transformation on the image from which the inter-spectral correlation has been removed, and restore the original hyperspectral image before compression.

Specifically, the image uniformity evaluation module includes:

The high and low bit splitting unit is used to calculate the minimum value of the hyperspectral image after reversible integer transformation channel by channel and save it, and then subtract the minimum value of the corresponding channel channel by channel, so that the channel image data is a positive integer greater than or equal to 0. Check whether there is a maximum overflow problem in the channel. If there is an overflow in the image data of the channel, split the channel into 2 channels, one saves the high-order data, and the other saves the low-order data;

The uniformity determining unit is used for calculating the image uniformity for the obtained channel image data, and after determining the uniformity index, it is divided into a relatively uniform image channel and an image channel with poor uniformity.

Specifically, the FILF compressor is used to encode the relatively uniform image channels, the gray value distribution is calculated for the image channels with poor uniformity, and the Gaussian function is used to fit the gray value distribution, and the fitted Gaussian function parameters are saved. The arithmetic encoder and the fitted Gaussian function parameters perform entropy encoding on the channel image;

Decode the compressed hyperspectral image channel by channel. For the image channel to be decoded encoded by the arithmetic encoder, first read the fitted Gaussian function parameters, and combine the arithmetic encoder and the fitted Gaussian function parameters for the channel image data. Decoding: The image channel to be decoded encoded by the FILF compressor is decoded by the FILF compressor.

The present invention improves the system compression performance and shortens the compression time by using the integer KLT to remove the inter-spectral correlation and flexibly switch the compressor strategy. In a specific embodiment, there are the following main steps:

(1) According to the corresponding pre-processing procedure of the imaging equipment, the original hyperspectral image data is pre-processed to correct nonlinear factors such as image deformation and response drift caused by the imaging instrument. Specifically, the input image is the public hyperspectral data image Salinas. As shown in Figure 2, the image Salinas is a 16-bit grayscale image with a size of 512*217 with 204 channels. The following experimental results are all completed under this condition. The original image Salinas was preprocessed with ENVI software tool and saved as Salinas_corrected. The preprocessing step is very critical, and the correlation between the data spectra without preprocessing is poor, which will reduce the compression performance of the algorithm.

(2) Perform reversible integer KLT forward transformation on the preprocessed hyperspectral image Salinas_corrected, remove the correlation between spectra, and save the transformation matrix.

Specifically, first calculate the Pearson correlation matrix Coeff between the image Salinas_corrected spectra:

Among them, ρ(p _i , p _j ) represents the Pearson correlation coefficient between the i-th spectral segment and the j-th spectral segment, and the specific calculation formula is as follows:

Among them, Cov(pi , p _j ) represents the covariance between spectral segments, and D( _pi ) and D(p _j ) represent the variances of the i- _th spectral segment and the j-th spectral segment.

Next, perform principal component analysis on the inter-spectral Pearson correlation matrix Coeff to obtain the transformation matrix A.

The above-mentioned transformation matrix A can only realize the transformation from real numbers to real numbers, and the transformation matrix A ^* = PLUS is obtained through decomposition to realize the mapping of integers to integers, wherein L and U are the unit lower triangular and upper triangular matrices respectively, P is the permutation matrix, and S is the A unit-element invertible matrix. The specific decomposition process of A is as follows:

Suppose A is a non-singular matrix that satisfies the decomposition conditions:

First, there is a row transformation matrix P ₁ such that

不为零，可以找到s₁使

即

于是构造一个单位元素可逆矩阵S₁：The first element of the Nth column of

is not zero, s ₁ can be found such that

which is

Then construct a unit-element invertible matrix S ₁ :

then there are

Then use the traditional Gaussian elimination method to define the elementary Gaussian matrix:

Then there are:

Repeat the above process N-1 times to get:

L _N-1 P _N-1 …L ₂ P ₂ L ₁ P ₁ AS ₁ S ₂ …S _N-1 ＝D _R U

where k=1, 2,...N-1; _DR is the twiddle factor, _DR =±1

The transformation matrix A can finally be decomposed into the following formula:

A=PLUS

in:

P=(P _N-1 …P ₂ P ₁ ) ^T

Since L is a lower triangular matrix, U is an upper triangular matrix, and only the last row of the S matrix has valid data, the LUS can be combined into a real matrix and saved.

Multiply the hyperspectral image data data by the S, U, L, P transformation matrix to the left in turn, and round the matrix after each multiplication to realize the integer KLT transformation. The specific formula is as follows:

Y=P*round(L*round(U*round(S*data)))

Among them, round represents the rounding and rounding operation.

(3) Calculate the minimum value of the hyperspectral data after KLT transformation channel by channel and save it, and then subtract the minimum value of the corresponding channel channel by channel, so that the image data is a positive integer greater than or equal to 0.

For the specific image Salinas_corrected, the 1st, 2nd, 203rd, and 204th channels of the transformed image are extracted as shown in Figure 3(a), Figure 3(b), Figure 3(c), and Figure 3(d) respectively; in order to illustrate the invertible integer KLT transformation removes the effect of correlation between spectra, and calculates the correlation between adjacent spectral segments before and after transformation, as shown in Figure 4(a) and Figure 4(b). It can be seen from Figure 4 that after KLT transformation The correlation between spectra basically disappeared.

(4) Check whether there is a maximum overflow problem on the data obtained after processing in step (3) channel by channel. If there is an overflow in the image data of this channel, split the channel data into 2 channels, one to save the high-order data, and the other to save the low-order data. data, and record the number of the data channel to be demolished, and repeatedly check the data until there is no maximum overflow problem.

For the specific image Salinas_corrected, the maximum value after KLT transformation can be calculated by the following formula:

Among them, BIT represents the number of image bits, and n represents the number of image spectral bands. Knowing that the image Salinas_corrected is 16 bits and has 204 spectral bands, the maximum value can be calculated to be 1872054. Therefore, splitting into 2 16-bit channels at most can solve the maximum overflow problem.

The maximum dynamic range of the processed pixel values of each channel is shown in Figure 5 below.

(5) Calculate the image uniformity index psr on a channel-by-channel basis on the data obtained after processing in step (4).

Calculate the autocorrelation coefficient ρ _x of the pixels of adjacent rows in the image, and the specific formula is as follows:

Calculate the mean of the autocorrelation coefficient of the image row

Calculate the autocorrelation coefficient ρ _y of adjacent column pixels in the image, and the specific formula is as follows:

Calculate the mean of the autocorrelation coefficient of the image column

和图像列自相关系数均值

定义图像均匀性评价指标psr，具体公式如下：Combined image row autocorrelation coefficient mean

and image column autocorrelation coefficient mean

Define the image uniformity evaluation index psr, and the specific formula is as follows:

(6) Use the FILF compressor to encode the image channel with strong uniformity; calculate the gray value distribution for the image channel data with poor uniformity, use the Gaussian function to fit the gray value distribution, and save the fitted Gaussian function parameters, The channel image data is entropy encoded by combining the arithmetic encoder and the fitted Gaussian distribution function.

For the specific image Salinas_corrected, it is generally considered that the uniformity index psr less than 0.3 is a non-uniform image.

Hyperspectral image decoding includes the following steps:

(7) Decode hyperspectral image data channel by channel. If the channel to be decoded is compressed by an arithmetic encoder, first read the parameters of the fitted Gaussian function, and then combine the arithmetic encoder and the fitted Gaussian distribution function to decode the image data of this channel.

(8) Check the records of the dismantled channels, and combine the dismantled channels according to the high order and low order.

(9) Read the minimum value of the image data of each channel, and add the minimum value to the corresponding channel.

(10) Read the reversible integer KLT transformation matrix, and perform inverse KLT transformation on the data obtained after processing in step (9). The specific transformation process is as follows:

First read the transformation matrix PLUS.

Assuming that the hyperspectral data after KLT transformation is Y, and the data before transformation is X, in order to recover X, solve the following equations in turn:

PLUSX=Y

Since P is an elementary matrix, it is itself an integer transformation, and LUSX can be solved:

LUSX=P ^T Y

L is a lower triangular matrix, and the integer implementation of the inverse transformation is calculated from top to bottom:

At this point, USX can be solved.

In the same way, U is an upper triangular matrix, and the integer of the inverse transformation is implemented as a bottom-up calculation, and SX can be solved.

Similarly, S is a lower triangular matrix, and the solution method is the same as above, and the hyperspectral image data X before transformation can be solved.

At this point, the entire compression and decompression process is as described above.

In order to compare the effect of the present invention, one uses the FILF compressor to directly compress the image Salinas_corrected, and the other uses the method of the present invention, and the compression ratio CR is defined as the ratio of the size after compression to the ratio before compression, and the final test results are shown in Table 1 below.

Table 1

Compression Ratio (CR) Compression time (seconds) FILF only 3.07 84 method of the invention 3.39 52

It can be seen from the above table that the present invention can effectively improve the image compression ratio, which is 0.3 higher than that of the FILF method; at the same time, the compression time is shortened by nearly half.

In the present invention, the input image is preprocessed first, and then the reversible integer KLT forward transformation is used to remove the inter-spectral correlation. Then, the uniformity index psr is calculated for each image channel, and the compressor is flexibly switched according to the uniformity of the image. By implementing the method of the present invention, the inter-spectral correlation of the hyperspectral image can be effectively removed, so that the compression performance in FILF is greatly improved. In addition, since the present invention flexibly switches to the arithmetic encoder, the complexity of the algorithm is further reduced, and the compression speed of the algorithm is improved.

Those skilled in the art can easily understand that the above 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, etc., All should be included within the protection scope of the present invention.

Claims (6)

1. A spectrum-space dimension combined hyperspectral image lossless compression method is characterized by comprising the following steps:

(1) preprocessing an original hyperspectral image, and correcting nonlinear factors caused by imaging;

(2) reversible integer transformation is carried out on the preprocessed hyperspectral images, the correlation among spectrums is removed, and a transformation matrix is stored; the method specifically comprises the following steps:

(2.1) calculating a Pearson correlation matrix Coeff among the preprocessed hyperspectral image data spectrums:

where ρ (p)_i，p_j) The Pearson correlation coefficients of the ith and jth spectral bands are expressed, and the specific calculation formula is as follows:

wherein, Cov (p)_i，p_j) Denotes the covariance between spectral bands, D (p)_i)、D(p_j) Is the variance of the ith and jth spectral bands;

(2.2) carrying out principal component analysis on the Pearson correlation matrix Coeff to obtain a transformation matrix A, and decomposing the transformation matrix A into the following formula:

A＝PLUS

wherein:

P＝(P_N-1…P₂P₁)^T

U＝±L_N-1P_N-1…L₂P₂L₁P₁AS₁S₂…S_N-1

the transformation matrix A can only realize the transformation from real numbers to real numbers, and the decomposed transformation matrix realizes the mapping from integers to integers;

(3) calculating image uniformity of the hyperspectral images subjected to reversible integer transformation channel by channel, and dividing the hyperspectral images after determining uniformity evaluation indexes into a first image channel and a second image channel which are different in uniformity; the method specifically comprises the following steps:

(3.1) calculating and saving the minimum value of the hyperspectral images subjected to reversible integer transformation channel by channel, and subtracting the minimum value of the corresponding channel by channel so that the channel image data are positive integers which are more than or equal to 0;

(3.2) checking channel-by-channel whether the channel image data obtained after the processing of the step (3.1) has the maximum overflow problem, and if the channel image data has overflow, disassembling the channel into 2 channels, wherein one channel stores high-bit data, and the other channel stores low-bit data;

(3.3) calculating the image uniformity of the channel image data obtained after the processing in the step (3.2), and dividing the channel image data into a first image channel and a second image channel with different uniformity after determining the uniformity index;

wherein one image channel has M rows and N columns, and the average value of the pixels in the rows is u_iThe process for determining the uniformity evaluation index comprises the following steps:

autocorrelation coefficients rho of adjacent rows of pixels within an image channel_xThe concrete formula is as follows:

mean value of image channel row autocorrelation coefficients

Autocorrelation coefficients rho of adjacent columns of pixels within an image channel_yThe concrete formula is as follows:

mean value of autocorrelation coefficients of image channel row

Combining image channel row autocorrelation coefficient mean values

And image channel column autocorrelation coefficient mean

Defining an image uniformity evaluation index psr, wherein a specific formula is as follows:

(4) the first image channel is coded by adopting a FILF compressor, and the second image channel is coded by adopting an arithmetic coder, so that lossless compression of the hyperspectral image is realized;

(5) decoding the compressed hyperspectral images channel by channel, and decoding the images without the inter-spectral correlation by adopting an arithmetic encoder to encode and adopting an image channel to be decoded encoded by a FILF compressor in a corresponding decoding mode;

(6) and (3) reading the transformation matrix obtained in the step (2), and performing reversible integer inverse transformation on the image without the inter-spectral correlation to restore the original hyperspectral image before compression.

2. The compression method according to claim 1, characterized in that said step (4) comprises in particular:

and encoding the first image channel by adopting a FILF compressor, calculating gray value distribution of the second image channel, fitting the gray value distribution by using a Gaussian function, storing the fitted Gaussian function parameters, and entropy encoding the channel image by combining the arithmetic encoder and the fitted Gaussian function parameters.

3. The compression method according to claim 1, characterized in that said step (5) comprises in particular:

decoding the compressed hyperspectral image channel by channel, reading a fitted Gaussian function parameter for an image channel to be decoded, which is coded by an arithmetic coder, and decoding channel image data by combining the arithmetic coder and the fitted Gaussian function parameter; the image channel to be decoded, which is encoded using the FILF compressor, is decoded by the FILF compressor.

4. The compression method according to claim 1, characterized in that said step (6) comprises in particular:

reading the transformation matrix obtained in the step (2), and performing reversible integer inverse transformation on the image without the inter-spectral correlation;

checking the record of the disassembled channel, and combining the disassembled channel according to the high order and the low order;

and reading the minimum value of the image data of each channel after combination, adding the minimum value to the corresponding channel, and recovering the original hyperspectral image before compression.

5. The spectrum-space dimension combined hyperspectral image lossless compression system is characterized by comprising a preprocessing module, a inter-spectrum correlation removing module, an image uniformity evaluation module, a FILF (first-class image frequency) encoder and arithmetic encoder, a FILF decoder and arithmetic decoder and an image restoration module;

the preprocessing module is used for preprocessing an original hyperspectral image and correcting nonlinear factors caused by imaging;

the inter-spectrum correlation removal module is used for performing reversible integer transformation on the preprocessed hyperspectral images, removing inter-spectrum correlation and storing a transformation matrix; the method specifically comprises the following steps: calculating a Pearson correlation matrix Coeff among the preprocessed hyperspectral image data spectrums:

where ρ (p)_i，p_j) The Pearson correlation coefficients of the ith and jth spectral bands are expressed, and the specific calculation formula is as follows:

wherein, Cov (p)_i，p_j) Denotes the covariance between spectral bands, D (p)_i)、D(p_j) Is the variance of the ith and jth spectral bands;

performing principal component analysis on the Pearson correlation matrix Coeff to obtain a transformation matrix A, and decomposing the transformation matrix A into the following formula:

A＝PLUS

wherein:

P＝(P_N-1…P₂P₁)^T

U＝±L_N-1P_N-1…L₂P₂L₁P₁AS₁S₂…S_N-1

the transformation matrix A can only realize the transformation from real numbers to real numbers, and the decomposed transformation matrix realizes the mapping from integers to integers;

the image uniformity evaluation module is used for calculating the image uniformity of the hyperspectral image subjected to reversible integer transform channel by channel, and dividing the hyperspectral image subjected to reversible integer transform into a first image channel and a second image channel with different uniformity after determining uniformity evaluation indexes; the image uniformity evaluation module comprises:

the high and low bit splitting unit is used for calculating and solving the minimum value of the hyperspectral image subjected to reversible integer transformation channel by channel and storing the hyperspectral image, then subtracting the minimum value of the corresponding channel by channel, so that the channel image data is a positive integer which is more than or equal to 0, checking whether the maximum value overflows channel by channel, and if the channel image data overflows, splitting the channel into 2 channels, wherein one channel stores high bit data and the other channel stores low bit data;

the uniformity determining unit is used for calculating the image uniformity of the obtained channel image data, and dividing the channel image data into a first image channel and a second image channel with different uniformity after determining the uniformity index; wherein one image channel has M rows and N columns, and the average value of the pixels in the rows is u_iThe process for determining the uniformity evaluation index comprises the following steps:

autocorrelation coefficients rho of adjacent rows of pixels within an image channel_xThe concrete formula is as follows:

mean value of image channel row autocorrelation coefficients

Autocorrelation coefficients rho of adjacent columns of pixels within an image channel_yThe concrete formula is as follows:

mean value of autocorrelation coefficients of image channel row

Combining image channel row autocorrelation coefficient mean values

And image channel column autocorrelation coefficient mean

Defining an image uniformity evaluation index psr, wherein a specific formula is as follows:

the FILF encoder and the arithmetic encoder are respectively used for encoding the first image channel and the second image channel to realize lossless compression of the hyperspectral image;

the FILF decoder and the arithmetic decoder are used for decoding the compressed hyperspectral images channel by channel, and the images without the inter-spectral correlation are decoded by adopting arithmetic encoder coding and adopting the image channel to be decoded coded by the FILF compressor in a corresponding decoding mode;

the image recovery module is used for reading the transformation matrix obtained by the inter-spectrum correlation removal module, and performing reversible integer inverse transformation on the image without the inter-spectrum correlation to recover the original hyperspectral image before compression.

6. The system of claim 5, wherein the first image channel is encoded using a FILF compressor, the second image channel is encoded using a gray value distribution, the gray value distribution is fitted using a Gaussian function, the fitted Gaussian function parameters are stored, and the channel image is entropy encoded using the arithmetic encoder and the fitted Gaussian function parameters;

decoding the compressed hyperspectral image channel by channel, reading a fitted Gaussian function parameter for an image channel to be decoded, which is coded by an arithmetic coder, and decoding channel image data by combining the arithmetic coder and the fitted Gaussian function parameter; the image channel to be decoded, which is encoded using the FILF compressor, is decoded by the FILF compressor.

Images