CCSDS algorithm for improving compression ratio by using correlation between spectrums of hyperspectral images

Abstract

The invention discloses a CCSDS algorithm for improving a compression ratio by using hyperspectral image inter-spectral correlation, and belongs to the technical field of hyperspectral image processing. The algorithm takes a maximum inter-spectral correlation coefficient as a criterion, performs inter-spectral correlation coefficient sorting processing and adaptively selects a spectral band having the highest correlation with a to-be-predicted spectral band as a reference spectral band; and the overall correlation between the reference spectrum segment and the to-be-predicted spectrum segment is effectively improved. The algorithm can well solve the problem of trade-off selection between the original CCSDS compression ratio and the prediction spectrum segment number P influencing the calculation complexity, the prediction precision is improved and the prediction residual error is reduced under the condition that the calculation complexity and the storage requirement are hardly increased, the complex ground feature is obviously improved, the uniform ground feature can be slightly improved, and the method is suitable for large-scale popularization and application. The compression performance is more stable. The algorithm is simple, the application range is wider, the reference value is higher, hardware implementation is easy, an effective research approach is provided for hyperspectral data compression, and the method is suitable for satellite-borne hyperspectral data compression application.

Description

Technical Field

[0001] The present invention relates to the technical field of hyperspectral image processing, and in particular to a CCSDS algorithm for improving compression ratio by utilizing inter-spectral correlation of hyperspectral images. Background Art

[0002] Hyperspectral imaging technology is increasingly used in military, civil and aerospace fields. The application fields of hyperspectral remote sensing images are mainly geological exploration, agricultural research and satellite remote sensing image research. With the substantial improvement of spectral resolution, spatial resolution and width of hyperspectral images, the image data rate has increased exponentially. The improvement of in-orbit data storage, downlink channel and other technologies requires more hardware resources. In addition, the data transmission bandwidth of the equipment is limited. Image data compression in orbit can greatly reduce the burden on satellite hardware and save satellite development, manufacturing and launch costs.

[0003] Hyperspectral image compression can be divided into lossy compression and lossless compression. Lossy compression can achieve a high compression ratio, but there is information loss. Lossless compression has a lower compression ratio and can retain all the information of the data. Therefore, lossless compression is the preferred solution for spectral imager data compression. There are roughly three types of lossless compression of hyperspectral images: transformation-based methods, vector quantization-based methods, and prediction-based methods.

[0004] Transformation-based image compression algorithms mainly include KLT transform, discrete cosine transform, discrete wavelet transform, etc. Real number transformation algorithms can only obtain floating-point results. Floating-point numbers are not convenient for quantization and storage. Therefore, only integer-to-integer transformations can be used in lossless compression.

[0005] The vector quantization compression algorithm directly quantizes the data blocks, which can balance the relationship between compression rate and distortion rate. However, the computational complexity is extremely high and grows exponentially with the dimension, making it difficult to apply in practice.

[0006] Prediction-based image compression algorithms are one of the hot research directions in the field of image compression. They have the advantages of simple structure and low hardware development difficulty, and have been widely used in practice. Typical prediction algorithms include context-based adaptive image coding (CALIC), JPEG-LS with LOCO-I as the core, lookup table (LUT), differential pulse code modulation DPCM, and reverse search method.

[0007] Among them, in May 2012, the Consultative Committee for Space Data Systems (CCSDS) proposed an adaptive three-dimensional prediction lossless compression standard CCSDS123.0-B-1 algorithm for hyperspectral images. The core of the algorithm predictor is the FL algorithm, which is simple to calculate and has low complexity. It can achieve good prediction results for most hyperspectral images and is currently the international standard for lossless compression of multi- / hyperspectral images on board space. However, the algorithm uses the same prediction mode for all pixels, does not distinguish pixel positions, does not fully utilize the correlation between image spectra, and the compression ratio still needs to be improved. For this reason, He Shuxian proposed a prediction method based on edge detection in the implementation of the CCSDS hyperspectral image lossless compression algorithm based on the Tilera platform in the literature. The edge detection is performed on the image, and different weights are given to the pixels around the predicted pixel according to the edges in different directions, thereby improving the prediction accuracy. Wang Liping analyzed the influence of 11 parameters such as the number of predicted spectral segments and weight resolution in the CCSDS123.0-B-1 standard on the compression ratio in the research on lossless compression technology of hyperspectral images based on prediction, and proposed a set of optimized parameters, which significantly improved the compression performance. Zhang Lei and Yang Yang proposed the optimized combined prediction coding OCPC algorithm in the optimization of CCSDS hyperspectral image compression algorithm in the literature. The predictor adaptively selects the prediction mode according to the pixel position, and obtains a better compression effect.

[0008] Chinese patent CN101742299A discloses a system and method for image lossless compression processing that complies with the CCSDS standard. The processing system includes: a control module, a JPEG-LS preprocessing module, a RICE encoding module and a code stream splicing module. The processing process of the system and method is: the input image data is grouped according to 16 sampling points by the control module; the grouped data enters the preprocessing module for decorrelation processing to obtain a group of mutually independent data, and then passes through the RICE encoding module and the code stream splicing module to obtain a compressed code stream. The lossless compression coding efficiency of the lossless compression processing system and method of the present invention is high, and because the core design of the processing system is to perform operations on the original data after grouping, the decoding resynchronization mechanism can prevent the propagation of errors in space data transmission, and the power consumption is low; in addition, the improved coding option calculation method greatly improves the processing speed of the hardware and can meet the requirements of lossless compression of satellite-borne images.

[0009] It can be seen that in the CCSDS123.0-B-1 algorithm, the correlation between image spectra is not fully utilized, and the trade-off between the number of predicted reference spectrum segments P and the compression ratio that affects the computational complexity still needs to be solved. The compression ratio still has room for further improvement, and the smaller the mapping prediction residual, the shorter the compressed code length will be, and the higher the compression ratio will be. Summary of the invention

[0010] Based on the above problems, the present invention aims at the characteristics that the spectral correlation of hyperspectral images is much higher than the spatial correlation, and the fact that the spectral correlation is not fully utilized in the CCSDS123.0-B-1 algorithm, which affects the trade-off between the number of predicted reference spectrum segments P and the compression ratio, and the compression ratio needs to be improved. A CCSDS algorithm that utilizes the spectral correlation of hyperspectral images to improve the compression ratio is proposed. The contents of the present invention are as follows:

[0011] The object of the present invention is to provide a CCSDS algorithm for improving the compression ratio by using the inter-spectral correlation of hyperspectral images, comprising a predictor module and an encoder module, wherein the hyperspectral image data is subjected to the predictor to reduce the redundancy of the hyperspectral image data to obtain a mapping prediction residual, and then the encoder module is used to perform Golomb-Rice encoding to obtain compressed hyperspectral image data, wherein the predictor module comprises the following steps:

[0012] Step 1, inter-spectral correlation analysis: Calculate the inter-spectral correlation coefficient according to the pixel value of the hyperspectral image data. The specific algorithm of the present invention is as follows:

[0013] Cor(r, c) represents the inter-spectral correlation coefficient between the rth spectral segment and the cth spectral segment, x i,j,r and x i,j,c Respectively represent the pixel values ​​of the rth spectral image and the cth spectral image at the spatial position (i, j), and Respectively represent the average pixel value of the corresponding spectral band;

[0014] Step 2, sorting of inter-spectral correlation coefficients: According to the inter-spectral correlation coefficients in step 1, the inter-spectral correlation coefficients are sorted with the maximum inter-spectral correlation coefficient as the criterion, and the inter-spectral correlation coefficients are simply sorted in descending order using an algorithm to locate the position of the highest correlation spectrum segment, that is, firstly the spectrum segment with the largest inter-spectral correlation coefficient is used as the first spectrum segment before the current spectrum segment to be predicted, and then the spectrum segment with the largest correlation coefficient with the current spectrum segment to be predicted is selected from the unsorted spectrum segments as the first P spectrum segments. The specific algorithm of the present invention is as follows:

[0015]

[0016] Step 3, selection of predictor: for the hyperspectral image data samples without non-uniform correction, select the simple mode predictor; for the hyperspectral image data samples with non-uniform correction, select the full mode predictor;

[0017] Step 4, local sum calculation: calculate the weighted sum of the samples adjacent to the sample in the spectrum segment. If the sample is a hyperspectral image data without non-uniform correction, select the local sum algorithm calculation in the local direction; if the sample is a hyperspectral image data with non-uniform correction, select the column-wise local sum algorithm calculation;

[0018] Step 5, local difference calculation: According to the result of the sorting process of the inter-spectral correlation coefficient in step 2 and the result of the local sum calculation in step 4, the local difference calculation introduces an adaptive spectrum segment selection method, and uses the first P spectrum segments with the highest inter-spectral correlation with the current spectrum segment to be predicted as the reference spectrum segments for calculating the central local difference, and obtains the highest overall correlation. The specific algorithm of the present invention is as follows:

[0019]

[0020] Step 6: Calculate the predicted value: Adopt adaptive linear prediction;

[0021] 1. Step 7, prediction residual: It is the difference between the actual sample value and the predicted sample. The prediction residual also needs to be updated by the weight vector to obtain the initialization weight vector and input into the prediction value calculation of step 6.

[0022] Compared with the prior art, the CCSDS algorithm of the present invention, which utilizes the correlation between the spectra of hyperspectral images to improve the compression ratio, can achieve the following beneficial effects:

[0023] The present invention aims at the characteristics that the inter-spectral correlation of hyperspectral images is much higher than the spatial correlation, and the fact that the inter-spectral correlation is not fully utilized in the CCSDS123.0-B-1 algorithm, the trade-off between the number of predicted reference spectrum segments P that affects the computational complexity and the compression ratio, and the problem that the compression ratio needs to be improved. A CCSDS algorithm that uses the inter-spectral correlation of hyperspectral images to improve the compression ratio is proposed. The algorithm uses the maximum inter-spectral correlation coefficient as a criterion, performs inter-spectral correlation coefficient sorting processing, and adaptively selects the spectrum segment with the highest correlation with the spectrum segment to be predicted as the reference spectrum segment, which effectively improves the overall correlation between the reference spectrum segment and the spectrum segment to be predicted. The experimental results show that the algorithm can well solve the problem of the trade-off between the original CCSDS compression ratio and the number of predicted spectrum segments P that affects the computational complexity, improves the prediction accuracy, and reduces the prediction residual without almost increasing the computational complexity and storage requirements. Compared with CCSDS123.0-B-1, it has a more obvious improvement effect on complex objects, and can also slightly improve uniform objects, and the compression performance is more stable. This algorithm is simple, has a wider range of applications, has high reference value, and is easy to implement in hardware. It provides an effective research approach for hyperspectral data compression and is suitable for spaceborne hyperspectral data compression applications. BRIEF DESCRIPTION OF THE DRAWINGS

[0024] The drawings described herein are used to provide a further understanding of the present invention and constitute a part of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings:

[0025] Figure 1 The CCSDS compression algorithm model of the present invention;

[0026] Figure 2 The prediction points for calculating the local sum of the present invention are selected as (a) neighborhood-oriented local sum (local direction neighbor-oriented) and (b) column-oriented;

[0027] Figure 3 Flow chart of the CCSDS algorithm of the present invention for improving compression ratio by using the correlation between the spectra of hyperspectral images;

[0028] Figure 4 Correlation analysis of the spectra corresponding to the 9 images [(a) ZY1E_AHSI-389497 (b) ZY1E_AHSI-393510 (c) ZY1E_AHSI-411921 (d) ZY1E_AHSI-411939 (e) ZY1E_AHSI-411940 (f) ZY1E_AHSI-412476 (g) ZY1E_AHSI-415642 (h) ZY1E_AHSI-426365 (i) ZY1E_AHSI-426619];

[0029] Figure 5 Variation of compression ratio with the number of predicted spectrum segments P (a) CCSDS123.0-B-1 (b) algorithm of the present invention;

[0030] Figure 6 Variation of compression ratio with weight resolution;

[0031] Figure 7 Comparison of residuals between the proposed algorithm and CCSDS123.0-B-1 (a) ZY1E_AHSI-389497 (b) ZY1E_AHSI-393510 (c) ZY1E_AHSI-411921;

[0032] Figure 8 Comparison of residuals between the proposed algorithm and CCSDS123.0-B-1 (a) ZY1E_AHSI-411939 (b) ZY1E_AHSI-411940 (c) ZY1E_AHSI-412476

[0033] Fig. 9Comparison of residuals between the proposed algorithm and CCSDS123.0-B-1 (a) ZY1E_AHSI-415642 (b) ZY1E_AHSI-426365 (c) ZY1E_AHSI-426619;

[0034] Fig.10 Comparison of compression ratios between the algorithm of the present invention and CCSDS123.0-B-1 (a) ZY1E_AHSI-389497 (b) ZY1E_AHSI-393510 (c) ZY1E_AHSI-411921;

[0035] Fig.11 Comparison of compression ratios between the proposed algorithm and CCSDS123.0-B-1 (a) ZY1E_AHSI-411939 (b) ZY1E_AHSI-411940 (c) ZY1E_AHSI-412476

[0036] Fig.12 Comparison of compression ratios between the proposed algorithm and CCSDS123.0-B-1 (a) ZY1E_AHSI-415642 (b) ZY1E_AHSI-426365 (c) ZY1E_AHSI-426619 DETAILED DESCRIPTION

[0037] In order to make the purpose, technical solution and advantages of the present invention clearer, the technical solution of the present invention will be clearly and completely described below in conjunction with the specific embodiments of the present invention and the corresponding drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative work are within the scope of protection of the present invention.

[0038] The present invention provides a CCSDS algorithm for improving the compression ratio by utilizing the correlation between the spectra of a hyperspectral image, which is applied to the compression of hyperspectral image data. The specific implementation includes the following steps:

[0039] like Figure 1 As shown, the CCSDS123.0-B-1 compression system consists of a predictor and an encoder. The predictor module is used to reduce the redundancy of the hyperspectral image data; the encoder module performs Golomb-Rice encoding on the data after the redundancy is reduced. The flowchart of the CCSDS algorithm of the present invention for improving the compression ratio by using the correlation between the spectra of the hyperspectral image is shown in FIG. Figure 3 As shown, the specific operations are as follows:

[0040] 1. Predictor

[0041] (1) Calculation of inter-spectral correlation and arrangement of inter-spectral correlation coefficients

[0042] The inter-spectral correlation coefficient between the rth spectral segment and the cth spectral segment is defined as shown in formula (1), x i,j,r and x i,j,c Respectively represent the pixel values ​​of the rth spectral image and the cth spectral image at the spatial position (i, j), and Respectively represent the average pixel value of the corresponding spectral band.

[0043]

[0044] Based on this, the algorithm idea of ​​the present invention is to solve the problem of underutilization of inter-spectral correlation in CCSDS123.0-B-1 by arranging the inter-spectral correlation coefficients based on the maximum inter-spectral correlation coefficient. The present invention uses the inter-spectral correlation coefficients to perform a simple descending arrangement process in order to locate the position of the highest correlation spectrum segment, that is, firstly taking the spectrum segment with the largest inter-spectral correlation coefficient as the first spectrum segment of the current spectrum segment to be predicted, and then selecting the spectrum segment with the largest correlation coefficient with the current spectrum segment to be predicted from the unsorted spectrum segments as the first P spectrum segments.

[0045]

[0046] (2) Local and Calculation

[0047] Local and σ z,y,x is the spectrum segment z with sample S z,y,x The weighted sum of adjacent samples. Figure 2 The method used to calculate the local sum and the neighboring samples are given. The user can choose to use local (neighbor-oriented) or column-oriented (column-oriented) prediction for the image.

[0048] The local sum of the local directions is calculated as shown in (3) of the present invention.

[0049]

[0050] The column-wise local sum calculation is as shown in (4) of the present invention.

[0051]

[0052] (3) Local difference calculation

[0053] The original CCSDS has the problem that the higher the number of predicted spectral segments P, the better. In particular, when the reference spectral segments contain more low-correlation spectral segments, the prediction accuracy will be reduced. In addition, the increase of P will also affect the prediction speed. In view of this problem, after sorting the inter-spectral correlation coefficients in the above manner, when calculating the central local difference, the algorithm of the present invention introduces an adaptive spectral segment selection method, which is improved to use the calculation of the present invention (5), that is, using the first P spectral segments with the highest inter-spectral correlation with the current spectral segment to be predicted as the reference spectral segments for calculating the central local difference, so as to obtain the highest overall correlation.

[0054]

[0055] For images without non-uniformity correction, i.e., with more obvious sensor stripe noise, using the simple mode and column-wise local sum calculation method often results in higher compression ratios. For corrected images, using the full mode and local sum calculation method often results in higher compression ratios.

[0056] (4) Calculation of predicted values

[0057] The predictor uses adaptive linear prediction to proportionally predict sample values is an integer, defined as:

[0058]

[0059] Among them, the register size R range is max{32, D+Ω+2}≤R≤64. Properly increasing the size of R can reduce the possibility of overflow when calculating the proportional prediction sample value.

[0060] When the sample is an unsigned integer, the upper and lower limits and the middle value S of the sample min , S max , S mid They are defined as:

[0061] S min =0,S max =2 D -1,S mid =2 D-1 (7)

[0062] When the samples are signed integers, S min , S max , S mid Defined as:

[0063] S min =-2 D-1 ,S max =2 D-1 -1,S mid =0 (8)

[0064] function Defined as:

[0065]

[0066] The predicted value of the current sample value is calculated as shown in (10) of the present invention.

[0067]

[0068] (5) Weight vector update calculation

[0069] Weight vector W z (t) is used to predict the sample value C z A vector of weight values. Each weight value is a signed integer coefficient that can be represented by Ω+3 bits. The resolution of the weight value is controlled by the user-specified parameter Ω, which is an integer in the range of 4≤Ω≤19. The higher the Ω, the higher the weight composition resolution, and the more bits are required to store it. The standard has two ways of initializing weights: default and custom. This algorithm uses the default method to select the initial weight vector W z (1). For each spectral segment z, the weight vector is initialized using the same weight initialization method. The weights in the three directions are initialized The initial weight values ​​of the local differences at the center of each spectral segment are as shown in (11) of the present invention.

[0070]

[0071] The weight update scaling exponent ρ(t) is used to control the convergence speed and is defined as:

[0072]

[0073] Among them, the parameter v min 、v max and weight update change interval t inc , specified by the user, determines the rate at which the predictor adapts to the image content statistics. min and v max The value range is -6≤v min ≤v max ≤9. Weight update factor t inc The value range is 2 4 ≤t inc ≤2 11 ,ρ(t) according to t inc The initial and final values ​​are defined as ρ(1)=v min +D-Ω and v max +D-Ω.

[0074] Proportional prediction error e z (t) is an integer, according to The next weight vector W can be calculated z (t+1) is:

[0075]

[0076] Among them, ω min and ω max They are: min =-2 Ω+2 ,ω max =2 Ω+2 -1

[0077] (5) Mapping prediction residuals

[0078] Prediction residual Δ z (t) is the actual sample value S z (t) and predicted sample The difference, The residual mapping converts the signed prediction residual into a D-bit unsigned integer to obtain the mapped prediction residual δ z (t) is:

[0079]

[0080] Among them, θ z (t) is defined as:

[0081]

[0082] 2 Encoder

[0083] The present invention encodes the mapped prediction residuals in the BIP order, and uses a sample adaptive entropy encoder to encode each mapped prediction residual δ z (t) Encoding with length-limited Golomb-Rice. The encoding method is similar to JPEG-LS, except that a shorter code length is used to encode smaller residuals, and a longer code length is used to encode larger residuals. This method is suitable for the case where all residuals are small. However, the data with residuals approaching 0 in the present invention is more suitable for this type of encoding and is more conducive to improving the compression ratio. For each mapping residual δ z (t), the maximum codeword length generated is U max +D, where the unary length limit U max 8≤Umax≤32.

[0084] The adaptive coding selection statistic is calculated by an accumulator Σ z (t) and a counter Γ(t), which is adaptively updated during the encoding process. z(t) / Γ(t) provides an estimate of the average mapping prediction residual value within the spectral segment. This ratio is used to select k z (t) value, the initial value of the counter and accumulator is defined as:

[0085]

[0086] Among them, the initial counting index γ0 ranges from 1≤γ0≤8, and the accumulator initialization constant k z ′ is 0≤k z ' is an integer in the range of D-2. The updating calculation of the accumulator of the spectrum segment z is calculated as shown in (17) of the present invention, and the counter Γ(t) is incremented, and the calculation is shown in (18) of the present invention.

[0087]

[0088] The rescaling counter parameter γ* controls the maximum value of Γ(t), and the value range of γ* is max{4,γ0+1}≤γ * ≤9.

[0089] The predicted residual δ after mapping z The codeword of (t) depends on k z (t) and u z (t), parameter k z (t) is defined as the largest non-negative integer k that satisfies the following inequality z (t) = max{1≤k z (t)≤D-2}.

[0090]

[0091] When k z (t) satisfies the following equation, k z (t)=0.

[0092]

[0093] For an input residual δ z (t) and the defined parameter U max , the calculation of the quotient and the residual is as shown in (21) of the present invention:

[0094]

[0095] Parameter k z (t), r z (t) and u z (t) is used for codeword generation. When t>0, δ z The codeword of (t) is determined by the following rules

[14] :

[0096] (1) When u z (t) max When δ z The codeword of (t) is u z (t) “0” + 1 “1”, followed by δ z The least significant remainder r of (t) z (t) k z (t)bit binary code.

[0097] (2) When u z (t)≥U max When z The codeword of (t) is given by U max "0", followed by δ z (t) is composed of D bit binary numbers.

[0098] (3) When t = 0, the code word is δ z (t) is represented by a D bit unsigned binary integer, that is, the first mapping prediction residual of each spectral segment z is not encoded, δ z The codeword for (0) is δ z (0) is a D-bit unsigned binary integer representation of

[0099] Example

[0100] The hyperspectral image data of the present invention are derived from hyperspectral images of different complexity of ground objects and very representative scenes currently acquired on orbit. Table 1 shows the specific information of 9 test images.

[0101] Table 1 Test image information

[0102]

[0103]

[0104] In order to verify the above-mentioned problem, the present invention uses MATLAB to perform spectral correlation analysis on 9 corrected hyperspectral images of AHSI_ZY1E. The results are as follows: Figure 4 By calculating the inter-spectral correlation coefficient of each image, this paper finds that the hyperspectral images with the highest correlation with the current spectral segment to be predicted are not all previous spectral segments or adjacent spectral segments. By comparison, it can be seen that the correlation between adjacent spectral segments of the four images with lower complexity, ZY1E_AHSI-389497, ZY1E_AHSI-411921, ZY1E_AHSI-411939 and ZY1E_AHSI-415642, is relatively high, proving that the CCSDS123.0-B-1 algorithm does not fully utilize the correlation between image spectra. ​

[0105] In order to verify that this algorithm can effectively solve the trade-off between the number of predicted reference spectral segments P and the compression ratio in the CCSDS123.0-B-1 algorithm, which does not fully utilize the spectral correlation of hyperspectral images and affects the computational complexity, the algorithm model of this algorithm CCSDS123.0-B-1 and JPEG-LS was established in MATLAB, and the AHSI_ZY1E hyperspectral image data set with different ground object complexity was selected for algorithm testing. From the experimental results, the compression ratio of this algorithm is significantly better than that of the CCSDS123.0-B-1 algorithm.

[0106] First of all, since many parameters in CCSDS123.0-B-1 are user-defined, the parameters that affect the compression ratio, such as the number of predicted spectral bands and weight resolution, are optimized and analyzed, and a set of optimization parameters suitable for the hyperspectral test image in this article is proposed. In the present invention, the number of predicted spectral bands and weight resolution that have the greatest impact on the results are shown. Finally, the predictor parameters of the algorithm of the present invention are shown in Table 2, and the encoder parameters are shown in Table 3.

[0107] Table 2 Predictor parameters

[0108]

[0109] Table 3 Encoder parameters

[0110]

[0111] (1) Predicting the impact of the number of spectral segments P on the compression ratio

[0112] The number of predicted spectral segments P determines the number of previous reference spectral segments used when predicting the current spectral segment, which in turn affects the compression performance. When the P value is small, as the number of predicted spectral segments increases, the overall correlation will become higher, and the compression ratio will be significantly improved. However, the number of predicted spectral segments P is not necessarily larger, as P increases, the overall correlation between the current spectral segment and the reference spectral segment will decrease, resulting in a decrease in the compression ratio, and will increase the algorithm calculation complexity.

[0113] In order to verify that this algorithm can effectively solve the trade-off between the number of predicted reference spectrum segments P and the compression ratio that affects the computational complexity, the algorithm and CCSDS123.0-B-1 compression ratio are simulated as the number of predicted spectrum segments changes, and P is set to 1, 3, 5, ..., 13, 15 to compress 9 images. Figure 5(a) It can be seen that with the continuous increase of P, the compression ratio does not increase continuously, but has obvious fluctuations. The compression ratio of ZY1E_AHSI-389497, ZY1E_AHSI-411939 and ZY1E_AHSI-415642 reaches the maximum when P=3. The ground objects of ZY1E_AHSI-411921 are the flattest among the 9 images, and the correlation between adjacent spectral bands is relatively high, so it reaches the maximum when P=1. The five images of ZY1E_AHSI-393510, ZY1E_AHSI-411940, ZY1E_AHSI-412476, ZY1E_AHSI-426365 and ZY1E_AHSI-426619 reach the highest when P=15. It can be found that the value of P is related to the complexity of the ground objects in the image. In addition, the value of P also affects the complexity of the algorithm. The larger the value of P, the greater the amount of calculation and the higher the complexity. Here, considering the compression ratio issue comprehensively, the optimized parameter value of the predicted spectrum number P is 15. Figure 5 (b) It can be seen that the CCSDS hyperspectral image lossless compression algorithm based on adaptive spectral band selection based on inter-spectral correlation coefficient can achieve a slight improvement in the compression ratio of uniform objects and a significant improvement in the compression ratio of complex objects. At the same time, the value of P for the 9 images is only 2, which can well solve the problem that the original CCSDS123.0-B-1 cannot take into account both compression ratio and computational complexity.

[0114] (2) Impact of weight resolution Ω on compression ratio

[0115] Each element of the weight vector is represented by Ω+3 bits. The weight resolution Ω is not only related to the bit width of each element of the weight vector, but also to the value range of the weight update factor ρ(t). The higher the Ω, the higher the weight composition resolution, and the more bits are required to store it.

[0116] from Figure 6 It can be seen that with the increase of weighted resolution, the compression ratio shows a trend of first rising and then falling until it stabilizes at around Ω = 10. For images with relatively flat terrain and high correlation between adjacent spectral bands, the larger the Ω, the higher the compression ratio. For images with complex terrain and low correlation between adjacent spectral bands, the compression ratio has the highest value when Ω is 7 and 8. Taking into account the compression ratio of complex terrain, the weighted resolution is set to 7.

[0117] (3) Comparison of algorithm compression performance

[0118] To further verify the compression performance of the CCSDS algorithm that uses the spectral correlation of hyperspectral images to improve the compression ratio, it can be seen from Tables 4 and 5 that the images compressed by the algorithm in the BIP compression order are decompressed. There is no difference in visual effect between the images before and after compression. By comparing the original image and the decompressed image in MATLAB software, it can be seen that the pixel values ​​of the two are exactly the same. This algorithm can achieve lossless compression of hyperspectral images.

[0119] Table 4 ZY1E_AHSI-411939 medium and short wave infrared compressed and decompressed images

[0120]

[0121] Table 5. Visible near infrared compressed and decompressed images in ZY1E_AHSI-389497

[0122]

[0123] In addition, in order to prove the role of this algorithm in improving prediction accuracy and reducing prediction residuals, this paper also conducts experimental processing on CCSDS123.0-B-1. The experimental results are as follows: Figure 7 , Figure 8 , Fig. 9 As shown in the figure. The comparative experiment shows that the proposed algorithm uses the spectrum with the highest correlation for prediction. Compared with CCSDS123.0-B-1, the prediction residuals of most pixels have decreased, and the proportion of prediction residuals approaching 0 has increased by 4% to 16.0%. The data with residuals approaching 0 has become more, and the rationality of the encoding method used has been verified. Therefore, the proposed algorithm has significantly improved the prediction accuracy and reduced the prediction residual, thereby shortening the compressed code length and improving the compression ratio, which is significantly better than CCSDS123.0-B-1.

[0124] Finally, in order to verify the impact of the CCSDS algorithm proposed in this paper on improving the compression ratio by using the correlation between the spectra of hyperspectral images, this paper also simulated the compression ratio of 45 consecutive frames of CCSDS123.0-B-1 and compared it with the algorithm in this paper. The compression ratio simulation test results are shown in the figure. Fig.10 , Fig.11 , Fig.12 As shown in the figure. By comparing the experimental results, it can be seen that the algorithm proposed in this paper improves the compression ratio of all the tested hyperspectral images when the parameters remain unchanged, and the complex objects have a more obvious improvement effect. The flattest objects ZY1E_AHSI-389497 and ZY1E_AHSI-411921 can also be slightly improved while achieving a high compression ratio, and the compression performance is more stable. Therefore, the algorithm proposed in this paper can make better use of the inter-spectral correlation information of the hyperspectral image, make more accurate predictions of the image, and improve the compression ratio of the algorithm.

[0125] In order to more objectively evaluate the lossless compression performance of the algorithm in this paper, the compression performance of the algorithm in this paper is compared with the CCSDS123.0-B-1 algorithm and the classic hyperspectral image lossless compression JPEG-LS algorithm. All experimental results are obtained from the simulation test of CCSDS123.0-B-1 and JPEG-LS algorithms in this paper. The compression effect comparison of the three algorithms is shown in Tables 6 and 7.

[0126] Table 6 Comparison of compression ratios of three algorithms for corrected images

[0127]

[0128] Note: The image in the above figure selects the 1000th to 1050th frame data for compression

[0129] Table 7. Comparison of compression ratios of three algorithms for uncorrected images

[0130]

[0131] As shown in Tables 6 and 7, compared with JPEG-LS and CCSDS123.0-B-1, the compression ratio of this algorithm has achieved the highest compression performance. The corrected image has a more obvious improvement, with an increase of 0.32% to 8.4%, and the uncorrected image has a smaller improvement, with an increase of 1.48%. This is due to the large strip noise and low inter-spectral correlation. The highest correlation coefficient of all spectral segments is between 0.2 and 0.85, and the inter-spectral correlation coefficient of all selected reference spectral segments of the corrected image is above 0.95. Therefore, the image is given a preliminary non-uniformity correction before compression, which will effectively improve the compression ratio. Among them, the JPEG-LS algorithm has a better effect in processing spatial correlation, targeting the compression of a single spectral segment of the image, but the inter-spectral correlation of the hyperspectral image is much higher than the spatial correlation. Therefore, the compression ratio of the JPEG-LS algorithm here is quite different from that of CCSDS 123.0-B-1 and the algorithm of this paper.

[0132] Among the improved algorithms for CCSDS123.0-B-1, He Shuxian proposed a prediction method based on edge detection in the paper Implementation of CCSDS hyperspectral image lossless compression algorithm based on Tilera platform, with an improvement of 1.5% to 11.2%. Wang Liping proposed the MHDC algorithm in the paper Research on lossless compression technology of hyperspectral images based on prediction, with an improvement of 0.7% to 3.6%. Zhang Lei and Yang Yang proposed the OCPC algorithm in the paper Optimization of CCSDS hyperspectral image compression algorithm, with an improvement of 2.4% to 5.2%. Deng Cun, Fan Ci'en, and Deng Dexiang proposed the CCSD algorithm based on V-type scanning in the paper Lossless compression algorithm of satellite-borne multispectral images based on V-type scanning, with an improvement of 1.1% to 4.2%. Li Jiaying, Zhu Wenquan, and Meng Fanyun proposed the RMPR algorithm in the paper Hyperspectral image compression algorithm based on reducing mapping prediction residuals, with an improvement of 0.5% to 3.8%. The above algorithms all process spatial correlation and perform compression of a single spectrum segment, namely, the BSQ compression order. This mode consumes a lot of subsequent FPGA cache resources and compression time. Among them, the prediction method based on edge detection proposed by He Shuxian is slightly higher than that of this article, but it uses edge detection-based prediction and dynamically adjusts the threshold value calculation, and its complexity is much higher than that of this article.

[0133] In summary, the simulation results show that the inter-spectral correlation calculation only uses a simple whole image for inter-spectral analysis, and the arrangement of the inter-spectral correlation coefficients is only a simple descending arrangement, which will not significantly increase the complexity, and basically does not affect the processing time, and it can well solve the problem of the trade-off between the original CCSDS compression ratio and the number of predicted spectrum segments P that affects the computational complexity. Therefore, the prediction accuracy is improved and the prediction residual is reduced without almost increasing the computational complexity and storage requirements. While maintaining a high compression ratio for uniform ground feature images and slightly improving it, the compression ratio of complex ground feature images is significantly improved, which has a wider range of applications and a high reference value.

[0134] In summary: The present invention aims at the characteristics that the inter-spectral correlation of hyperspectral images is much higher than the spatial correlation, and the fact that the inter-spectral correlation is not fully utilized in the CCSDS123.0-B-1 algorithm, and the trade-off between the number of predicted reference spectrum segments P that affects the computational complexity and the compression ratio, and the compression ratio needs to be improved. A CCSDS algorithm that uses the inter-spectral correlation of hyperspectral images to improve the compression ratio is proposed. The algorithm uses the maximum inter-spectral correlation coefficient as a criterion, performs inter-spectral correlation coefficient sorting processing, and adaptively selects the spectrum segment with the highest correlation with the spectrum segment to be predicted as the reference spectrum segment, which effectively improves the overall correlation between the reference spectrum segment and the spectrum segment to be predicted. The experimental results show that the algorithm can well solve the problem of the trade-off between the original CCSDS compression ratio and the number of predicted spectrum segments P that affects the computational complexity, improves the prediction accuracy, and reduces the prediction residual without almost increasing the computational complexity and storage requirements. Compared with CCSDS123.0-B-1, it has a more obvious improvement effect on complex objects, and can also slightly improve uniform objects, and the compression performance is more stable. This algorithm is simple, has a wider range of applications, has high reference value, and is easy to implement in hardware. It provides an effective research approach for hyperspectral data compression and is suitable for spaceborne hyperspectral data compression applications.

[0135] The above description is only an embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention may have various modifications and variations. Any modification, equivalent substitution, improvement, etc. made within the spirit and principle of the present invention should be included in the scope of the claims of the present invention.

Claims

1. A CCSDS algorithm for improving compression ratio by using inter-spectral correlation of hyperspectral images, comprising a predictor module and an encoder module, wherein hyperspectral image data is subjected to the predictor to reduce the redundancy of the hyperspectral image data to obtain a prediction residual map, and the encoder module is used to perform Golomb-Rice encoding to obtain compressed hyperspectral image data, characterized in that: The predictor module comprises the following steps: Step 1, inter-spectral correlation analysis: Calculate the inter-spectral correlation coefficient according to the pixel value of the hyperspectral image data. The specific algorithm of the present invention is as follows: Cor(r, c) represents the inter-spectral correlation coefficient between the rth spectral segment and the cth spectral segment, x i,j,r and x i,j,c Respectively represent the pixel values ​​of the rth spectral image and the cth spectral image at the spatial position (i, j), and Respectively represent the average pixel value of the corresponding spectral band; Step 2, sorting of inter-spectral correlation coefficients: According to the inter-spectral correlation coefficients in step 1, the inter-spectral correlation coefficients are sorted with the maximum inter-spectral correlation coefficient as the criterion, and the inter-spectral correlation coefficients are simply sorted in descending order using an algorithm to locate the position of the highest correlation spectrum segment, that is, firstly the spectrum segment with the largest inter-spectral correlation coefficient is used as the first spectrum segment before the current spectrum segment to be predicted, and then the spectrum segment with the largest correlation coefficient with the current spectrum segment to be predicted is selected from the unsorted spectrum segments as the first P spectrum segments. The specific algorithm of the present invention is as follows: Step 3, selection of predictor: for the hyperspectral image data samples without non-uniform correction, select the simple mode predictor; for the hyperspectral image data samples with non-uniform correction, select the full mode predictor; Step 4, local sum calculation: calculate the weighted sum of the samples adjacent to the sample in the spectrum segment. If the sample is a hyperspectral image data without non-uniform correction, select the local sum algorithm calculation in the local direction; if the sample is a hyperspectral image data with non-uniform correction, select the column-wise local sum algorithm calculation; Step 5, local difference calculation: According to the result of the sorting process of the inter-spectral correlation coefficient in step 2 and the result of the local sum calculation in step 4, the local difference calculation introduces an adaptive spectrum segment selection method, and uses the first P spectrum segments with the highest inter-spectral correlation with the current spectrum segment to be predicted as the reference spectrum segments for calculating the central local difference, and obtains the highest overall correlation. The specific algorithm of the present invention is as follows: Step 6: Calculate the predicted value: Adaptive linear prediction is used; Step 7, prediction residual: It is the difference between the actual sample value and the predicted sample. The prediction residual also needs to be updated by the weight vector to obtain the initialization weight vector and input into the prediction value calculation of step 6.

2. The CCSDS algorithm for improving compression ratio by using inter-spectral correlation of hyperspectral images according to claim 1 is characterized in that: The local sum σ of the local direction in step 4 z,y,x is the spectrum segment z with sample S z,y,x The weighted sum of adjacent samples, the specific algorithm of the present invention is:

3. The CCSDS algorithm for improving compression ratio by using inter-spectral correlation of hyperspectral images according to claim 1 is characterized in that: In step 4, the local sum σ z,y,x is the spectrum segment z with sample S z,y,x The weighted sum of adjacent samples, the specific algorithm of the present invention is:

4. The CCSDS algorithm for improving compression ratio by using inter-spectral correlation of hyperspectral images according to claim 1 is characterized in that: The encoder module encodes the mapped prediction residuals in BIP order, and uses a sample adaptive entropy encoder to encode each mapped prediction residual δ z(t) Encoding is done using length-restricted Golomb-Rice.

Images