A satellite image compression method based on chirp modulation

Abstract

The present invention discloses a satellite image compression method based on chirp, which includes: S1 obtaining a satellite image and performing block division on the satellite image by using an adaptive block division method to separate different projective transformations in the satellite image and obtain a plurality of image blocks; S2 inputting all the image blocks into a trained neural network model to predict the chirp transform parameters of each image block; S3 performing discrete chirp transform on each image block according to the chirp transform parameters corresponding to each image block to realize the conversion from the image domain to the chirp frequency spectrum domain; S4 in the chirp frequency spectrum domain, quantizing the amplitude of the frequency components to map the continuous spectrum coefficients into a discrete set; S5 compressing the quantization results of all the image blocks in step S4 by using a lossless algorithm, and then encoding the compressed spectrum data by using an encoder to complete the compression of the satellite image.

Description

Technical Field

[0001] The present invention relates to image compression technology, and specifically to a satellite image compression method based on chirp. Background Art

[0002] As a key link in computer image processing, image compression faces the challenge of the increasing demand for image information storage and transmission. With the development of technology, neural networks have provided new ideas for the research direction of image compression methods. The prior art CN 118741145A discloses a performance-balanced compression method for satellite images. This method arranges images in combination, balances the excellent compression performance and poor compression performance of images from the perspective of a multi-image system, overcomes the impact of the barrel effect on the practicality of image compression, overcomes the defect that the JPEG2000 standard compression algorithm has a large difference in compression quality for multiple images, improves the performance of the standard compression algorithm without increasing complexity, and has the characteristics of being easy to upgrade and implement in software and hardware, and has practical value in satellite image compression and transmission systems.

[0003] Although the above invention has many advantages, it still has the following disadvantages in use: This invention is essentially a traditional compression method of JPEG2000. Traditional image compression technologies similar to JPEG are limited by their algorithm architectures and will inevitably produce significant blocky distortion phenomena in the reconstructed images. Summary of the Invention

[0004] Aiming at the above deficiencies in the prior art, the satellite image compression method based on chirp provided by the present invention solves the problem of distortion existing in the prior art during image compression.

[0005] In order to achieve the above invention purpose, the technical solution adopted by the present invention is as follows:

[0006] Provide a satellite image compression method based on chirp, which includes the steps of:

[0007] S1. Obtain a satellite image, and use an adaptive block method to block the satellite image to separate different projective transformations in the satellite image, and obtain a number of image blocks;

[0008] S2. Input all the image blocks into a trained neural network model, and predict the chirp transform parameters of each image block;

[0009] S3. According to the chirp transform parameters corresponding to each image block, perform a discrete chirp transform on each image block to realize the conversion from the image domain to the chirp frequency spectrum domain;

[0010] S4. In the chirp frequency spectrum domain, quantize the amplitudes of the frequency components to map the continuous spectral coefficients into a discrete set;

[0011] S5. Use a lossless algorithm to compress the quantization results of all image blocks in step S4, and then use an encoder to encode the compressed spectral data to complete the compression of the satellite image.

[0012] Furthermore, the method of using an adaptive block division method to divide the satellite image includes:

[0013] S11. Divide the satellite image into a number of non-overlapping image blocks and set the initial value of the matching degree parameter;

[0014] S12. Calculate the two-dimensional projective entropy of each image block using a two-dimensional projective entropy model, and calculate the average value of the two-dimensional projective entropies of all image blocks corresponding to the divided blocks. The two-dimensional projective entropy model is:

[0015]

[0016] ,

[0017] where G is the two-dimensional projective entropy; L is the gray level of the satellite image; is the average gray value in the image block for the probability of occurrence; is the matching degree parameter between the frequency change trend of the pixels in the image block and the projective transformation parameter; is the frequency of occurrence of pixel points with gray value i and neighborhood average gray value j in the image block; is the total number of gray value pairs in the image block; is the projective transformation parameter difference of the q-th gray value pair; is the size of the satellite image, is the length of the satellite image, is the width of the satellite image;

[0018] S13. Determine whether the two-dimensional projective entropy of each image block is greater than its corresponding two-dimensional projective entropy average value or whether the size of the image block is less than the set size. If either is satisfied, go to step S14; otherwise, stop dividing the current image block.

[0019] S14. Re-divide the current image block to obtain a number of image blocks with a size equal to the preset size, and then return to step S12.

[0020] Furthermore, when performing the first division of the satellite image, the expression for the number b of the obtained image blocks is ; When performing each image block division, the size of the obtained image block is , where k is the number of times the current image block is divided.

[0021] Furthermore, the expression for performing the discrete chirp transform on each image block is:

[0022] ,

[0023] where, is the discrete chirp transform; is the mapping in the frequency domain; is the two-dimensional distribution of the image gray values after projective transformation; is the two-dimensional distribution of the image gray values; is the chirp transform parameter predicted by the neural network model; e is the natural logarithm; is the j-th pixel point in the image block; is the area of the image block; N is the total number of pixel points in the image block.

[0024] Furthermore, the satellite image compression method based on chirp also includes decompressing the compressed satellite image:

[0025] S6. Perform entropy decoding on the compressed satellite image to recover the compressed data;

[0026] S7. Perform the inverse discrete chirp transform on the data after entropy decoding to convert the processed spectral data back to the image space, and then perform the reconstruction of the satellite image; the expression for performing the inverse discrete chirp transform is:

[0027] .

[0028] Furthermore, the neural network model is an autoregressive network, a recurrent neural network, a long short-term memory network, a gated recurrent unit, or a gated attention mechanism Transformer;

[0029] When the neural network model is trained, the data set includes several image blocks and the corresponding chirp transform parameters for each image block. The several image blocks are the input of the neural network model, and the chirp transform parameters are the output of the neural network model.

[0030] Furthermore, when dividing the satellite image into blocks, it also includes:

[0031] S15. Group all the image blocks, divide the image blocks of the same size into the same group to obtain several groups, and calculate the proportion of each group in all the image blocks;

[0032] S16. Determine the sampling frequency and the number of sampling times of the image blocks in each group according to the sampling rate of the satellite image and the ratio of each group;

[0033] S17. According to the number of sampling times of each image block, construct a corresponding measurement matrix based on the block compressive sensing theory for compressive measurement, and all image blocks within the same group use the same measurement matrix;

[0034] S18. Generate a measurement value vector of each image block as the final image block according to the measurement matrix corresponding to the image block and the gray value vector corresponding to the image block.

[0035] Further, step S16 further includes:

[0036] S161. Calculate the sampling frequency and the number of sampling times of each image block within the group according to the sampling frequency of the satellite image and the ratio of each group:

[0037] ,

[0038] wherein, is the sampling frequency of the image blocks in each group; is the sampling rate of the satellite image; is the ratio of each group; is the number of sampling times of the image blocks in each group; B is the size of the satellite image;

[0039] S162. Determine whether the number of sampling times of each image block is greater than the sampling upper bound. If so, go to step S163; otherwise, obtain the number of sampling times of each image block;

[0040] S163. Evenly distribute the number of times exceeding the sampling upper bound to other groups until the number of sampling times of each image block in each group is less than or equal to the sampling upper bound;

[0041] Sampling upper bound The expression of .

[0042] The beneficial effects of the present invention are as follows: The purpose of image block division in this solution is different from that in traditional compression methods such as JPEG. The purpose of block division in traditional compression methods is to perform independent parallel processing on each block, thereby improving the algorithm efficiency of the entire transform coding; the adaptive block division in this solution is to separate different local distortions in the image, that is, different projective transformations, so as to fit the corresponding chirp transformation on each different local distortion through a neural network model, so as to achieve a better compression ratio and prevent the problem of distortion in the compressed and reconstructed image, so as to ensure the quality of the reconstructed image.

[0043] The two-dimensional projective entropy calculated by this method can not only reflect the complexity of the image block, but also reflect whether the image block is suitable for chirp transform, and can help determine the optimal chirp transform parameters. It should be able to evaluate the projective change difference between regions, maximize the distance and difference of projective transform parameters between different image blocks, so as to ensure a better compression ratio. Brief Description of the Drawings

[0044] Figure 1 It is a flowchart of a satellite image compression method based on chirp.

[0045] Figure 2 It is a flowchart of another embodiment of a satellite image compression method based on chirp. Detailed Description of the Invention

[0046] The following describes the specific embodiments of the present invention to facilitate those skilled in the art to understand the present invention. However, it should be clear that the present invention is not limited to the scope of the specific embodiments. For those of ordinary skill in the art, as long as various changes are within the spirit and scope of the present invention defined and determined by the appended claims, these changes are obvious, and all inventions made using the concept of the present invention are within the scope of protection.

[0047] Refer to Figure 1 , Figure 1 which shows a flowchart of a satellite image compression method based on chirp; as Figure 1 shown, the method S includes steps S1 to S5.

[0048] In step S1, a satellite image is obtained, and the satellite image is segmented by an adaptive segmentation method to separate different projective transforms in the satellite image, obtaining a number of image segments.

[0049] In an embodiment of the present invention, the method for segmenting a satellite image by an adaptive segmentation method includes:

[0050] S11. Divide the satellite image into a number of non-overlapping image segments, and set an initial value of the matching degree parameter;

[0051] S12. Calculate the two-dimensional projective entropy of each image segment using a two-dimensional projective entropy model, and calculate the average value of the two-dimensional projective entropy of all image segments corresponding to the segmented segments. The two-dimensional projective entropy model is:

[0052]

[0053] ,

[0054] where G is the two-dimensional projective entropy; L is the gray level of the satellite image; is the mean gray value pair in the image block the probability of occurrence; is the matching degree parameter between the frequency change trend of pixels in the image block and the projective transformation parameters; is the frequency of pixel points with gray value i and neighborhood mean gray value j in the image block; is the total number of mean gray value pairs in the image block; is the projective transformation parameter difference of the q-th mean gray value pair; is the size of the satellite image, is the length of the satellite image, is the width of the satellite image;

[0055] The block description of the divided blocks here. Assume that when the image is divided for the first time, it is divided into 4 blocks. Among them, the two-dimensional projective entropy of two blocks is less than or equal to the mean two-dimensional projective entropy, and the two blocks are greater than the mean two-dimensional projective entropy. Then these two blocks need to be divided again. Assume that each block is divided into 4 image blocks again. Then the divided block is the image block with a two-dimensional projective entropy greater than the mean two-dimensional projective entropy when divided for the first time.

[0056] S13. Determine whether the two-dimensional projective entropy of each image block is greater than its corresponding mean two-dimensional projective entropy or whether the size of the image block is less than the set size. If either is satisfied, go to step S14; otherwise, stop dividing the current image block.

[0057] The set size is , because experiments have shown that there is no obvious difference in the texture detail distribution at this time, and the smaller the image block, the higher the total number of samplings. At this time, the meaning of compression is lost; furthermore, defining the minimum value can prevent serious block effects.

[0058] S14. Divide the current image block again to obtain several image blocks with a size equal to the preset size, and then return to step S12.

[0059] When implementing, in this solution, preferably when the satellite image is divided for the first time, the expression for the number b of image blocks obtained is ; when dividing the image each time, the size of the obtained image block is , where k is the number of times the current image block is divided; this solution preferably sets the preset size to .

[0060] In step S2, input all the image blocks into the trained neural network model to predict the chirp transformation parameters of each image block.

[0061] The neural network model of this solution is preferably an autoregressive network, a recurrent neural network, a long short-term memory network, a gated recurrent unit, or a gated attention mechanism Transformer; the dataset for training the neural network model includes several image patches and the corresponding chirp transform parameters for each image patch. The several image patches are the input of the neural network model, and the chirp transform parameters are the output of the neural network model.

[0062] In step S3, according to the chirp transform parameters corresponding to each image block, perform a discrete chirp transform on each image block to achieve the conversion from the image domain to the chirp frequency domain.

[0063] During implementation, the expression for preferably performing a discrete chirp transform on each image block in this solution is:

[0064] ,

[0065] where, is the discrete chirp transform; is the mapping in the frequency domain; is the two-dimensional distribution of the image gray values after projective transformation; is the two-dimensional distribution of the gray values of the image; is the chirp transform parameter predicted by the neural network model; e is the natural logarithm; is the j-th pixel point in the image block; is the area of the image block; N is the total number of pixel points in the image block.

[0066] In step S4, in the chirp frequency domain, quantize the amplitude of the frequency components to map the continuous spectral coefficients into a discrete set;

[0067] In step S5, use a lossless algorithm to compress the quantization results of all image blocks in step S4, and then use an encoder to encode the compressed spectral data to complete the compression of the satellite image.

[0068] As Figure 2 shown, in an embodiment of the present invention, the satellite image compression method based on chirp also includes performing a decompression operation on the compressed satellite image:

[0069] S6. Perform entropy decoding on the compressed satellite image to recover the compressed data;

[0070] S7. Perform an inverse discrete chirp transform on the data after entropy decoding to convert the processed spectral data back to the image space, and then perform the reconstruction of the satellite image; the expression for performing the inverse discrete chirp transform is:

[0071] 。

[0072] In one embodiment of the present invention, when the satellite image compression method based on chirp is used to divide the satellite image into blocks, it further includes:

[0073] S15. Group all the image blocks, divide the image blocks of the same size into the same group to obtain several groups, and calculate the proportion of each group in all the image blocks;

[0074] S16. Determine the sampling frequency and the number of sampling times of the image blocks in each group according to the sampling rate of the satellite image and the proportion of each group;

[0075] S17. Based on the number of sampling times of each image block, construct a corresponding measurement matrix according to the block compressive sensing theory for compressive measurement, and all the image blocks within the same group use the same measurement matrix;

[0076] S18. Generate a measurement value vector of each image block as the final image block according to the measurement matrix corresponding to the image block and the gray value vector corresponding to the image block (that is, multiplying the measurement matrix by the gray value vector).

[0077] During implementation, step S16 of this solution preferably further includes:

[0078] S161. Calculate the sampling frequency and the number of sampling times of each image block within the group according to the sampling frequency of the satellite image and the proportion of each group:

[0079] ,

[0080] wherein, is the sampling frequency of the image blocks in each group; is the sampling rate of the satellite image; is the proportion of each group; is the number of sampling times of the image blocks in each group; B is the size of the satellite image;

[0081] S12. Judge whether the number of sampling times of each image block is greater than the sampling upper bound. If so, go to step S13; otherwise, obtain the number of sampling times of each image block;

[0082] S13. Evenly distribute the number of times exceeding the sampling upper bound to other groups until the number of sampling times of each image block in each group is less than or equal to the sampling upper bound;

[0083] Sampling upper bound The expression of 。

[0084] In traditional image compression methods, the block compression process often first divides the image into blocks of the same size, while this scheme uses an adaptive block method, which can group several image blocks of different sizes. In order to verify the effectiveness of the adaptive block technology on the model, the ablation experiment is set as follows:

[0085] The experimental group is set to use the model of ordinary equal-size image block division technology, and the control group is set to use the model of adaptive image block division technology. Experiments are conducted under high bit rate and low bit rate conditions of relevant satellite images, and the average test results are shown in Table 1.

[0086] Table 1 Ablation experiment results of adaptive block module

[0087] Model PSNR / dB at high bitrate (1.00 bpp) PSNR / dB at low bitrate (0.50 bpp) Experimental model 36.230 32.308 Control model 37.532 33.674

[0088] In Table 1, PSNR is implemented by calculating the mean square error (MSE) between the original image and the reconstructed image to quantify this difference. The smaller the MSE, the closer the two images are, and thus the higher the PSNR value, the better the image quality.

[0089] The results show that in both high and low bit rate cases, the overall model performance using adaptive blocking technology is improved in terms of PSNR indicators compared to the model performance using ordinary blocking technology. Specifically, the average improvement is 1.302dB in high bit rate conditions and 1.366dB in low bit rate conditions. It can be seen that the compression effect of the compression method of this scheme is significantly better than the compression effect of the existing equal-size blocking technology.

Claims

1. A satellite image compression method based on linear frequency modulation, characterized in that: Includes steps: S1, obtaining a satellite image, and using an adaptive blocking method to block the satellite image, so as to separate different projective transformations in the satellite image, and obtain a plurality of image blocks; S2, input all image blocks into the trained neural network model to predict the chirp transformation parameters of each image block; S3, performing discrete chirp transformation on each image block according to the chirp transformation parameters corresponding to each image block, so as to realize the conversion from image domain to chirp spectrum domain; S4, in the chirp spectrum domain, quantize the amplitude of the frequency component to map the continuous spectrum coefficients into a discrete set; S5, compressing the quantization results of all the image blocks in step S4 using a lossless algorithm, and then encoding the compressed spectrum data using an encoder to complete the compression of the satellite image; Methods for segmenting satellite images using an adaptive segmentation method include: S11, dividing the satellite image into a number of non-overlapping image blocks, and setting an initial value of a matching degree parameter; S12, using a two-dimensional projective entropy model to calculate the two-dimensional projective entropy of each image block, and calculating the mean of the two-dimensional projective entropy of all image blocks corresponding to the divided blocks, the two-dimensional projective entropy model is: ， Among them, G is the two-dimensional projective entropy; L is the gray level of the satellite image; is the grayscale mean value pair in the image block Probability of occurrence; is the matching parameter between the frequency variation trend of pixels in the image block and the projection transformation parameter; is the frequency of occurrence of pixels with gray value i and neighborhood gray mean j in the image block; is the total number of gray mean pairs in the image block; is the difference in projection transformation parameters of the qth grayscale mean pair; is the size of the satellite image, is the length of the satellite image, is the width of the satellite image; S13, judging whether the two-dimensional projective entropy of each image block is greater than the corresponding two-dimensional projective entropy mean value or whether the size of the image block is less than the set size, if any of the conditions are met, proceeding to step S14, otherwise the current image block stops being segmented; S14, re-dividing the current image block to obtain a number of image blocks with a size equal to a preset size, and then returning to step S12; The expression for discrete chirp transform of each image block is: ， in, is the discrete chirp transform; is the mapping in frequency domain; is the two-dimensional distribution of image grayscale value after projection transformation; is the two-dimensional distribution of the grayscale value of the image; is the chirp transformation parameter predicted by the neural network model; e is the natural logarithm; is the jth pixel in the image block; is the area of ​​the image block; N is the total number of pixels in the image block.

2. The satellite image compression method based on linear frequency modulation according to claim 1, characterized in that: When the satellite image is divided into blocks for the first time, the expression of the number of image blocks b is: ; When each image is divided into blocks, the size of the image block is , k is the number of times the current image is partitioned.

3. The satellite image compression method based on linear frequency modulation according to claim 1, characterized in that: It also includes decompression of compressed satellite images: S6. performing entropy decoding on the compressed satellite image to restore the compressed data; S7. Perform a discrete chirp inverse transform on the entropy decoded data, convert the processed spectrum data back to the image space, and then reconstruct the satellite image; the expression for the discrete chirp inverse transform is: 。 4. The satellite image compression method based on linear frequency modulation according to any one of claims 1 to 3, characterized in that: The neural network model is an autoregressive network, a recurrent neural network, a long short-term memory network, a gated recurrent unit or a gated attention mechanism Transformer; The data set for training the neural network model includes a number of image blocks and chirp transformation parameters corresponding to each image block. The number of image blocks are inputs of the neural network model, and the chirp transformation parameters are outputs of the neural network model.

5. The satellite image compression method based on linear frequency modulation according to claim 1, characterized in that: When satellite images are segmented, the following are also included: S15, grouping all image blocks, dividing image blocks of the same size into the same group, obtaining a plurality of groups, and calculating the proportion of each group to all image blocks; S16, determining the sampling frequency and sampling times of the image blocks in each group according to the sampling rate of the satellite image and the proportion of each group; S17, constructing a corresponding measurement matrix for compression measurement based on the block compression sensing theory according to the sampling times of each image block, and all image blocks in the same group use the same measurement matrix; S18. Generate a measurement value vector for each image block as the final image block according to the measurement matrix corresponding to the image block and the gray value vector corresponding to the image block.

6. The satellite image compression method based on linear frequency modulation according to claim 5, characterized in that: Step S16 further comprises: S161. Calculate the sampling frequency and sampling times of each image block in the group according to the sampling frequency of the satellite image and the proportion of each group: ， in, is the sampling frequency of the image blocks in each group; is the sampling rate of satellite images; is the proportion of each group; is the number of sampling times of the image blocks in each group; B is the size of the satellite image; k is the number of times the current image block is segmented; S162, determine whether the sampling number of each image block is greater than the sampling upper limit, if so, proceed to step S163, otherwise obtain the sampling number of each image block; S163, evenly distributing the number of times exceeding the sampling upper bound to other groups until the sampling number of each image block in each group is less than or equal to the sampling upper bound; Sampling upper bound The expression is: .

Images