Satellite-borne image compression method and system based on generative adversarial network

Abstract

The application discloses a kind of satellite-borne image compression method and system based on generative adversarial network, wherein the method comprises: obtaining compressed code stream and binary mask chart;Get logic gate control signal;Using neural network to train training data set obtains network weight parameter;The JPEG-LS lossy compressed code stream of real-time transmission on satellite is decompressed to obtain preliminary recovery image and satellite binary mask chart;Primary recovery image and satellite binary mask chart are input into neural network, and network weight parameter is used to repair the pixel value of the pixel value of satellite run-length encoding area to obtain intermediate recovery image data;Residual chart data is obtained;According to residual chart data and maximum allowable error value, satellite run-length encoding area optimization image data is obtained;According to satellite run-length encoding area optimization image data and preliminary recovery image, the final recovered image data is obtained.The application improves the quality of reconstructed image.

Description

Technical Field

[0001] This invention belongs to the field of image compression technology, and in particular relates to a spaceborne image compression method and system based on generative adversarial networks. Background Technology

[0002] The purpose of image compression is to represent or encode the original camera payload data using as few bytes as possible, and to recover image data with some loss by setting different compression ratios (2-8). Common static image compression algorithms include JPEG, JPEG-LS, JPEG-2000, PNG, and JPEG-XR. Considering the real-time requirements, algorithm complexity, and low power consumption of onboard remote sensing image hardware implementation, analysis shows that the JPEG-LS compression algorithm has significant advantages. LS is an algorithm for lossless or near-lossless compression of static continuous images. This algorithm has advantages such as low implementation complexity, high fidelity, and ease of hardware portability, and is widely used in many different types of image compression fields.

[0003] The algorithm achieves good visual effects and high-quality restored images in lossless and near-lossless modes. However, to adapt to the limitations of on-board data transmission bandwidth, the maximum tolerance error (Near) parameter is adjusted significantly and fluctuates greatly for larger compression ratios (i.e., when the compression ratio is ≥4:1). This results in poor visual effects of the decompressed on-board remote sensing images and affects image quality. Summary of the Invention

[0004] The technical problem solved by this invention is to overcome the shortcomings of the prior art and provide a satellite image compression method and system based on generative adversarial networks. This method performs local optimization and correction on the decompressed images on the satellite, which largely eliminates the "linear texture" phenomenon in the restored images, improves the quality of the reconstructed images, and the objective evaluation indicators are comparable to or slightly improved by the original algorithm.

[0005] The objective of this invention is achieved through the following technical solution: a satellite image compression method based on generative adversarial networks, comprising: compressing preset remote sensing image data to obtain a compressed bitstream and a binary mask; using the preset remote sensing image data and the binary mask as input to a training dataset for a neural network; comparing the absolute value of the residual between the restored pixel values ​​of the run-length encoded region of the binary mask and the initial pixel values ​​of the run-length encoded region with the maximum permissible error value to obtain a logic gating signal; training the training dataset using a neural network to obtain network weight parameters; and decompressing the lossy JPEG-LS compressed bitstream transmitted in real-time on satellite to obtain a primary restored image I. R0 Binary mask image on the star; primary recovery image I R0 The on-board binary mask image is input into a neural network, and the pixel values ​​of the on-board upstream coding region are repaired using the network weight parameters to obtain the intermediate restored image data I.R1 The residual map data D is obtained based on the intermediate recovered image data, the on-board binary mask map, and the initial image of the on-board upstream coding region. run Based on residual plot data D run The maximum permissible error value is used to obtain the satellite-based uplink coding region optimized image data; based on the satellite-based uplink coding region optimized image data and the primary restored image I... R0 The final recovered image data I reb .

[0006] In the above-mentioned spaceborne image compression method based on generative adversarial networks, the logic gate signal is obtained through the following formula:

[0007]

[0008] Where D_flag is the logic gate signal, I d Near is the residual between the repaired pixel value of the run-length encoded region of the binary mask and the initial pixel value of the run-length encoded region, where Near is the maximum allowable error value.

[0009] In the above-mentioned spaceborne image compression method based on generative adversarial networks, the logic gate signal is 1 during the training process.

[0010] In the above-mentioned spaceborne image compression method based on generative adversarial networks, the residual map data D run It can be obtained through the following formula:

[0011]

[0012] Among them, D run For residual plot data, I R1 To recover intermediate image data, I mask For the on-board binary mask, I run This is the initial image of the satellite's upstream coding region.

[0013] In the above-mentioned satellite image compression method based on generative adversarial networks, the optimized image data of the satellite upstream coding region is obtained by the following formula:

[0014] I R2 ={I R2 (i)}(i=1,2,3,…,M);

[0015] in,

[0016] I R2 To optimize image data for the satellite uplink coding region, I R2(i) represents the reconstructed pixel value of the i-th satellite uplink coding region after scaling, M is the total number of satellite uplink coding regions, i is the sequence number of the satellite uplink coding region, Near is the maximum allowable error value, and D run (i) represents the residual map data for the i-th satellite upstream coding region segment, I run (i) is the initial image of the i-th satellite upstream coding region segment, I R1 (i) represents the intermediate recovered image data of the i-th satellite upstream coding region segment.

[0017] In the above-mentioned spaceborne image compression method based on generative adversarial networks, the finally recovered image data is obtained by the following formula:

[0018]

[0019] Among them, I reb For the final recovered image data, I R0 For the initial image restoration, I mask For the on-board binary mask, I R2 Optimize image data for the satellite uplink coding region.

[0020] A spaceborne image compression system based on generative adversarial networks includes: a first module for compressing preset remote sensing image data to obtain a compressed bitstream and a binary mask; a second module for using the preset remote sensing image data and the binary mask as a training dataset for a neural network, and comparing the absolute value of the residual between the restored pixel values ​​of the run-length encoded region of the binary mask and the initial pixel values ​​of the run-length encoded region with the maximum allowable error value to obtain a logic gating signal; a third module for training the neural network on the training dataset to obtain network weight parameters; and a fourth module for decompressing the lossy JPEG-LS bitstream transmitted in real time on the satellite to obtain a primary restored image I. R0 And the binary mask image on the satellite; the fifth module, used to restore the primary image I R0 The on-board binary mask image is input into a neural network, and the pixel values ​​of the on-board upstream coding region are repaired using the network weight parameters to obtain the intermediate restored image data I. R1 The sixth module is used to obtain residual map data D based on the intermediate recovered image data, the on-board binary mask map, and the initial image of the on-board upstream coding region. run The seventh module is used to process residual map data D. run The maximum permissible error value is used to obtain optimized image data for the satellite's upstream coding region; the eighth module is used to optimize image data and primary reconstructed image I based on the satellite's upstream coding region. R0 The final recovered image data I reb .

[0021] In the aforementioned spaceborne image compression system based on generative adversarial networks, the logic gate signal is obtained through the following formula:

[0022]

[0023] Where D_flag is the logic gate signal, I d Near is the residual between the repaired pixel value of the run-length encoded region of the binary mask and the initial pixel value of the run-length encoded region, where Near is the maximum allowable error value.

[0024] In the aforementioned spaceborne image compression system based on generative adversarial networks, the logic gate signal is 1 during the training process.

[0025] In the aforementioned spaceborne image compression system based on generative adversarial networks, the residual map data D run It can be obtained through the following formula:

[0026]

[0027] Among them, D run For residual plot data, I R1 To recover intermediate image data, I mask For the on-board binary mask, I run This is the initial image of the satellite's upstream coding region.

[0028] Compared with the prior art, the present invention has the following advantages:

[0029] (1) This invention generates a 0-1 binary mask image (MASK) by extracting the image regions corresponding to run-length encoding and regular encoding in the JPEG-LS compression algorithm, corresponding to the MASK mask in step 1, so as to accurately repair the pixels in the run-length encoding region. Moreover, the MASK mask obtained by the encoding part is used to generate the training dataset of the adversarial network, and the MASK mask in the decoding part is used for real-time image repair processing after compression on the satellite.

[0030] (2) This invention modifies the convergence condition of the second-stage decision maker in step two, based on the absolute value of the residual |I d | The logical judgment signal D_flag is generated by comparing with Near, which works together with the decision loss value in the DeepFill-GAN network to ensure the convergence of the training process and reduce the error between the pixel reconstruction value and the original value in the area to be repaired.

[0031] (3) The present invention utilizes the residual map data D in step five. run After reconstructing the pixel values ​​from the intermediate data of the run-length region, scaling is performed to preserve the image data variation trend characteristics within the run-length region. Simultaneously, the reconstructed values ​​satisfy the Near range, which improves the image quality of the reconstructed image. rebPeak signal-to-noise ratio (PSNR);

[0032] (4) Step 1 of this invention performs a large amount of remote sensing data compression processing by setting different Near values, and obtains different weight parameters by classifying and training image data with different bit width precision (10-bit, 12-bit), which can match the run-length region pixel value repair of satellite image data with different bit widths; Steps 4 and 5 do not change the original JPEG-LS standard algorithm, but only need to add run-length region pixel value repair optimization processing to the sub-block image obtained by ground decompression processing, which is easy to implement and simple to operate;

[0033] (5) The proposed method generates a mask image based on different JPEG-LS encoding modes, and then uses a DeepFill-GAN network to repair and reconstruct pixels in local areas, effectively improving the linear false texture phenomenon in the high compression ratio mode of JPEG-LS. Furthermore, by combining Near to further optimize the reconstruction value of each run region, the reconstructed value is closer to the original data. This improves the visual effect and increases the image PSNR value. Attached Figure Description

[0034] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the invention. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings:

[0035] Figure 1 This is a flowchart of a spaceborne image compression method based on generative adversarial networks provided in an embodiment of the present invention;

[0036] Figure 2 This is a structural block diagram of the spaceborne image compression and transmission subsystem provided in an embodiment of the present invention;

[0037] Figure 3 This is a flowchart of the JPEG-LS algorithm provided in an embodiment of the present invention;

[0038] Figure 4 This is a flowchart of the LS run-length pixel repair process of the improved DeepFill-GAN network provided in this embodiment of the invention;

[0039] Figure 5 This is a schematic diagram showing the comparison of the results of decompressing a 10-bit image Near=8 original and improving SRCN provided in an embodiment of the present invention. Detailed Implementation

[0040] Exemplary embodiments of the present disclosure will now be described in more detail with reference to the accompanying drawings. While exemplary embodiments of the present disclosure are shown in the drawings, it should be understood that the present disclosure may be implemented in various forms and should not be limited to the embodiments set forth herein. Rather, these embodiments are provided to enable a more thorough understanding of the present disclosure and to fully convey the scope of the disclosure to those skilled in the art. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other. The present invention will now be described in detail with reference to the accompanying drawings and embodiments.

[0041] Improving the "linear texture" phenomenon caused by fixed values ​​in the run-length region of the restored image is crucial. Restoring pixel values ​​within the run-length region can be viewed as a problem of reconstructing pixel values ​​in small fluctuation areas, or as an image restoration process where the run-length region represents the missing portion. Here, a two-stage network is built and improved based on the DeepFill generative adversarial network structure for image inpainting. To further optimize the pixel values ​​in the restored region, a residual judgment mechanism between the restored region value and the initial run-length value is added and fed back to the second-stage local decision unit, resulting in the restored image after secondary reconstruction of the run-length region. This embodiment performs local optimization and correction on the decompressed image from the satellite; it largely eliminates the "linear texture" phenomenon in the restored image, significantly improves the subjective visual effect of the reconstructed image, and achieves objective evaluation metrics comparable to or slightly improved compared to the original algorithm.

[0042] Figure 1 This is a flowchart of a spaceborne image compression method based on generative adversarial networks provided in an embodiment of the present invention. Figure 1 As shown, the spaceborne image compression method based on generative adversarial networks includes: compressing preset remote sensing image data to obtain a compressed bitstream and a binary mask; using the preset remote sensing image data and the binary mask as a training dataset for a neural network, comparing the absolute value of the residual between the restored pixel values ​​of the run-length encoded region of the binary mask and the initial pixel values ​​of the run-length encoded region with the maximum allowable error value to obtain a logic gating signal; training the neural network on the training dataset to obtain network weight parameters; and decompressing the JPEG-LS lossy compressed bitstream transmitted in real time on the satellite to obtain a primary restored image I. R0 Binary mask image on the star; primary recovery image I R0 The on-board binary mask image is input into a neural network, and the pixel values ​​of the on-board upstream coding region are repaired using the network weight parameters to obtain the intermediate restored image data I. R1 The residual map data D is obtained based on the intermediate recovered image data, the on-board binary mask map, and the initial image of the on-board upstream coding region. run Based on residual plot data D run The maximum permissible error value is used to obtain the satellite-based uplink coding region optimized image data; based on the satellite-based uplink coding region optimized image data and the primary restored image I...R0 The final recovered image data I reb .

[0043] The logic gate signal is obtained by the following formula:

[0044]

[0045] Where D_flag is the logic gate signal, I d Near is the residual between the repaired pixel value of the run-length encoded region of the binary mask and the initial pixel value of the run-length encoded region, where Near is the maximum allowable error value.

[0046] Residual plot data D run It can be obtained through the following formula:

[0047]

[0048] Among them, D run For residual plot data, I R1 To recover intermediate image data, I mask For the on-board binary mask, I run This is the initial image of the satellite's upstream coding region.

[0049] The optimized image data for the satellite upstream coding region is obtained using the following formula:

[0050] I R2 ={I R2 (i)}(i=1,2,3,…,M);

[0051] in,

[0052] I R2 To optimize image data for the satellite uplink coding region, I R2 (i) represents the reconstructed pixel value of the i-th satellite uplink coding region after scaling, M is the total number of satellite uplink coding regions, i is the sequence number of the satellite uplink coding region, Near is the maximum allowable error value, and D run (i) represents the residual map data for the i-th satellite upstream coding region segment, I run (i) is the initial image of the i-th satellite upstream coding region segment, I R1 (i) represents the intermediate recovered image data of the i-th satellite upstream coding region segment.

[0053] The final recovered image data is obtained using the following formula:

[0054]

[0055] Among them, I reb For the final recovered image data, IR0 For the initial image restoration, I mask For the on-board binary mask, I R2 Optimize image data for the satellite uplink coding region.

[0056] Specifically, the method includes the following steps:

[0057] Step 1: Compress a large amount of remote sensing image data at different maximum tolerance (Near) values ​​using the JPEG-LS compression algorithm to obtain compressed bitstream and decompressed small image data blocks corresponding to the Near values ​​of the compressed bitstream; generate a MASK mask map using different encoding methods in the algorithm, i.e., the value corresponding to the pixel position in run-length encoding is 1, and the value corresponding to the pixel position in regular encoding is 0.

[0058] Step 2: Based on the original small image patches (a large amount of remote sensing image data) from Step 1 and the corresponding mask images obtained, perform sequential naming processing on pairs and store them to obtain the training dataset for the DeepFill-GAN network. Adjust the number of convolutional channels in the network, increase the convergence condition of the second-stage decision maker, and use the absolute value of the residual between the run-length encoded repaired pixel values ​​and the initial pixel values ​​of the current run-length encoded region |I d The D_flag signal is obtained by comparing it with the Near value, and the calculation formula is as follows:

[0059]

[0060] Step 3: Train the DeepFill-GAN network on the training dataset from Step 2. During training, if D_flag is 1 and the loss value of the second-stage decision maker meets the set conditions, stop the training iteration; otherwise, continue the decision iteration. If the set number of training iterations is exceeded during this process, stop the training iteration as well; obtain the final network weight parameters.

[0061] Step 4: Decompress the JPEG-LS lossy compressed bitstream transmitted in real time on the satellite to obtain the primary restored image I. R0 The corresponding mask image I is obtained based on the run-length encoding information of the JPEG-LS lossy compressed bitstream. mask ;

[0062] Step 5: Take the I from step 4 R0 The mask image is input into the DeepFill-GAN network, and the pixel values ​​of the run-length encoded region are repaired using the network weight parameters obtained in step three to obtain the intermediate restored image data I. R1 Furthermore, I R1 with I mask Corresponding pixel dot product After processing, the initial image I corresponding to the run-length encoded region is compared with the original image. run The residual plot data D is obtained by subtraction.run :

[0063]

[0064] Judge D run Whether the absolute value is less than Near is further calculated using the following formula for the repaired pixel values ​​of the run-length encoded region:

[0065]

[0066] I R2 ={I R2 (i)}(i=1,2,3,…,M)

[0067] In the formula, M represents the number of times the current recovered image enters the run-length encoding, and the index i takes values ​​sequentially up to M, D run (i) represents the residual data of the i-th run segment, max{|D run (i)|} represents the maximum absolute value of the residual of the i-th run segment, I run (i) represents the initial pixel value of the i-th run segment, I R2 (i) represents the pixel reconstruction value of the i-th run segment after scaling, I R2 This represents the reconstructed data after optimizing all travel regions;

[0068] Step Six: Optimize the I obtained in Step Five R2 with I R0 By performing a merge calculation and repairing only the pixel values ​​in the run-length encoded region, the final recovered image data I is obtained. reb The calculation formula is as follows:

[0069]

[0070] In the above formula The reconstructed image data corresponding to conventional encoding, and I R2 The complete image data is obtained by adding them together.

[0071] like Figure 2As shown, the satellite image compression data processing flow includes a real-time compression unit for onboard payload data and a ground decompression unit. Ground-based compression data is packetized and framed to obtain a two-level bitstream protocol format. Ground-based detection software then performs format parsing and compression algorithm decoding on this data to obtain fixed-size image sub-blocks. Finally, these sub-blocks are reassembled and stitched together to obtain the original-size image. The compression algorithm is JPEG-LS, with its core component being LOCO-I (low complexity lossless compression for images). Its main features include a predictive mechanism and a low-complexity coding mode, making it easy to implement in hardware and achieving compression efficiency similar to or even better than arithmetic-based compression algorithms.

[0072] like Figure 3 As shown, the original image is input into the LS encoder sequentially according to the raster scan order. For the current pixel x, the previously encoded neighboring pixels (a, b, c, d) are used to predict x, and the local gradient characteristics of the neighboring pixels are used as contextual model information to select between conventional coding and run-length coding. High compression ratio (greater than 4:1) compression of remote sensing images inevitably requires a small amount of bitstream for image encoding. Conventional coding performs pixel-by-pixel prediction encoding, which cannot achieve a high compression ratio. Therefore, the maximum allowable error value (Near) needs to be increased to make it easier to enter run-length coding mode. In this mode, the encoder does not perform prediction or error encoding; it only encodes the initial value (a) and the run length. Therefore, the larger the Near value, the more detailed information of the original image is lost in the corresponding run-length region, and the stronger the "regularity" of the run-length mechanism corresponding to the JPEG-LS algorithm becomes. This usually manifests as a linear pseudo-texture, significantly degrading the visual effect of the reconstructed image. To optimize pixel value reconstruction in this region, the initial decompressed image data is input into a generative adversarial network (GAN) for local region inpainting. Combined with the characteristics of the JPEG-LS algorithm, Near values ​​are extracted from the bitstream and fed back to the network, making the reconstructed values ​​of the run-length region closer to the original data. Finally, the locally improved data replaces some pixels in the original run-length region to obtain the restored image.

[0073] like Figure 4 As shown, this embodiment generates a 0-1 mask image (MASK image) based on different encoding modes in the compression algorithm. "1" corresponds to the run-length encoded region, representing the missing image region; "0" represents the conventionally encoded image region. The primary restored image I... R0 The corresponding mask image is used as input to DeepFill-GAN, and after processing by the trained network, intermediate restored image data I is obtained. R1 Then, based on the reconstructed values ​​and the initial run values, the residual data D is obtained. run I is calculated by comparing with the Near value and scaling the run-length encoded region. R2 This enables pixel value restoration in the run-length region.

[0074] This embodiment of the satellite image compression method based on generative adversarial networks includes MASK image generation based on the JPEG-LS compression feature coding mode and an optimization module with added Near feedback DeepFill-GAN network. First, the real-time compressed data on satellite is parsed to obtain sequential sub-block bitstream packets. These packets are then decompressed according to the JPEG-LS decoding algorithm to obtain a lossy restored image, a MASK mask image, and the current bitstream Near value parameters. These three data parts are then used as input to the improved DeepFill-GAN network system to repair the pixel grayscale values ​​of the lossy parts in the run-length region. The back-end network uses a Near threshold to judge the residual between the reconstructed value and the initial value of the segmented run-length region in real time. The result is fed back to the local discriminator in the fine network structure and affects its discrimination value. Next, after a finite number of iterations and convergence, residual scaling is used to optimize the pixel value reconstruction of the run-length region. Finally, the run-length reconstruction value optimized by the DeepFill-GAN network is merged with the restored data of the regular region into image sub-blocks. The sub-block images are then stitched together according to the sequence to obtain the restored image of the original size.

[0075] The following example verifies the effectiveness of this patent in reconstructing pixel values ​​in the run-length region of JPEG-LS compression. A 10-bit wide, 1024×1024 remote sensing image (including a flat sea surface area in the left half) is used for verification. For a tolerance error Near = 8, the partial compression effect of the original JPEG-LS algorithm is as follows: Figure 5 As shown in the left half, the linear texture of the decompressed sea surface image within the stretched box is very obvious; Figure 5 The pixel quality of the run-length portion within the same region in the right half, after local optimization using a generative adversarial network, has been greatly improved, and the linear texture phenomenon has disappeared.

[0076] Twelve small images with similar textures and containing flat regions were selected for testing to verify the recovery of the run length in flat regions. The maximum pixel difference (MAE), peak signal-to-noise ratio (PSNR), and image entropy (En) were used as objective evaluation metrics, as shown in Table 1.

[0077] Table 1. Evaluation Table of Objective Parameters for Original Decompression and Improved Decompression

[0078]

[0079] according to Figure 5Analysis of the test results shows that the JPEG-LS algorithm based on the improved DeepFill-GAN significantly improves the false linear texture caused by run-length errors, resulting in a better improvement in subjective visual perception. Meanwhile, as shown in Table 1, the PSNR values ​​of the original decompressed image and the improved decompressed image are basically equivalent, and some test results are better than the original decompressed image. The MAE value is better than the original decompressed image because a small deviation is introduced by the deep network optimization and repair of the run-length correction.

[0080] This embodiment also provides a spaceborne image compression system based on a generative adversarial network. The system includes: a first module for compressing preset remote sensing image data to obtain a compressed bitstream and a binary mask; a second module for using the preset remote sensing image data and the binary mask as a training dataset for a neural network, and comparing the absolute value of the residual between the restored pixel values ​​of the run-length encoded region of the binary mask and the initial pixel values ​​of the run-length encoded region with the maximum permissible error value to obtain a logic gating signal; a third module for training the training dataset using the neural network to obtain network weight parameters; and a fourth module for decompressing the lossy JPEG-LS bitstream transmitted in real-time onboard to obtain a primary restored image I. R0 And the binary mask image on the satellite; the fifth module, used to restore the primary image I R0 The on-board binary mask image is input into a neural network, and the pixel values ​​of the on-board upstream coding region are repaired using the network weight parameters to obtain the intermediate restored image data I. R1 The sixth module is used to obtain residual map data D based on the intermediate recovered image data, the on-board binary mask map, and the initial image of the on-board upstream coding region. run The seventh module is used to process residual map data D. run The maximum permissible error value is used to obtain optimized image data for the satellite's upstream coding region; the eighth module is used to optimize image data and primary reconstructed image I based on the satellite's upstream coding region. R0 The final recovered image data I reb .

[0081] This invention generates a 0-1 binary mask image (MASK) by extracting the corresponding image regions of run-length encoded and conventional encoded data in the JPEG-LS compression algorithm. This corresponds to the MASK mask image in step 1, achieving the effect of accurately repairing pixels in the run-length encoded region. Furthermore, the MASK mask image obtained in the encoding part is used to generate the training dataset for the adversarial network, and the MASK mask image obtained in the decoding part is used for real-time image inpainting processing after compression on satellite.

[0082] This invention modifies the convergence condition of the second-stage decision maker in step two, based on the absolute value of the residual |I dThe logical judgment signal D_flag is generated by comparing with Near, and works together with the decision loss value in the DeepFill-GAN network to ensure the convergence of the training process and reduce the error between the pixel reconstruction value and the original value in the area to be repaired.

[0083] This invention utilizes the residual map data D in step five. run After reconstructing the pixel values ​​from the intermediate data of the run-length region, scaling is performed to preserve the image data variation trend characteristics within the run-length region. Simultaneously, the reconstructed values ​​satisfy the Near range, which improves the image quality of the reconstructed image. reb Peak signal-to-noise ratio (PSNR).

[0084] Step one of this invention compresses a large amount of remote sensing data by setting different Near values, and obtains different weight parameters by classifying and training image data with different bit width precision (10-bit, 12-bit), which can match the run-length region pixel value repair of satellite image data with different bit widths. Steps four and five do not change the original JPEG-LS standard algorithm. They only need to add run-length region pixel value repair optimization processing to the sub-block image obtained by ground decompression processing, which is easy to implement and simple to operate.

[0085] This invention proposes generating mask images based on different JPEG-LS encoding modes. The DeepFill-GAN network is then used for local pixel inpainting and reconstruction, effectively improving the linear false texture phenomenon in the high compression ratio mode of JPEG-LS. Furthermore, combining Near filters further optimizes the reconstructed values ​​for each run-length region, making the reconstructed values ​​closer to the original data. This improves visual effects and increases the image's PSNR value.

[0086] Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make possible changes and modifications to the technical solutions of the present invention by utilizing the methods and techniques disclosed above without departing from the spirit and scope of the present invention. Therefore, any simple modifications, equivalent changes and alterations made to the above embodiments based on the technical essence of the present invention without departing from the content of the technical solutions of the present invention shall fall within the protection scope of the technical solutions of the present invention.

Claims

1. A spaceborne image compression method based on generative adversarial networks, characterized in that... include: The preset remote sensing image data is compressed to obtain a compressed bitstream and a binary mask image; The preset remote sensing image data and binary mask are used as the input to the training dataset of the neural network. The absolute value of the residual between the repaired pixel value of the run-length encoded region of the binary mask and the initial pixel value of the run-length encoded region is compared with the maximum allowable error value to obtain the logic gate signal. The network weight parameters are obtained by training the training dataset using a neural network. The JPEG-LS lossy compressed bitstream transmitted in real time on the satellite is decompressed to obtain the primary restored image and the on-board binary mask image; The primary restored image and the on-board binary mask are input into the neural network, and the pixel values ​​of the on-board roaming coding region are repaired using the network weight parameters to obtain intermediate restored image data. Residual map data is obtained based on intermediate recovered image data, on-board binary mask map, and initial image of the on-board upstream coding region; Based on the residual map data and the maximum allowable error value, the image data for the satellite's upstream coding region is optimized. The final recovered image data is obtained by optimizing the image data and the primary recovered image based on the satellite upstream coding region.

2. The spaceborne image compression method based on generative adversarial networks according to claim 1, characterized in that: The logic gate signal is obtained by the following formula: Where D_flag is the logic gate signal, I d Near is the residual between the repaired pixel value of the run-length encoded region of the binary mask and the initial pixel value of the run-length encoded region, where Near is the maximum allowable error value.

3. The spaceborne image compression method based on generative adversarial networks according to claim 1, characterized in that: During training, the logic gating signal is 1.

4. The spaceborne image compression method based on generative adversarial networks according to claim 1, characterized in that: Residual plot data D run It can be obtained through the following formula: Among them, D run For residual plot data, I R1 To recover intermediate image data, I mask For the on-board binary mask, I run This is the initial image of the satellite's upstream coding region.

5. The spaceborne image compression method based on generative adversarial networks according to claim 1, characterized in that: The optimized image data for the satellite upstream coding region is obtained using the following formula: I R2 ={I R2 (i)}(i=1,2,3,…,M); in, I R2 To optimize image data for the satellite uplink coding region, I R2 (i) represents the reconstructed pixel value of the i-th satellite uplink coding region after scaling, M is the total number of satellite uplink coding regions, i is the sequence number of the satellite uplink coding region, Near is the maximum allowable error value, and D run (i) represents the residual map data for the i-th satellite upstream coding region segment, I run (i) is the initial image of the i-th satellite upstream coding region segment, I R1 (i) represents the intermediate recovered image data of the i-th satellite upstream coding region segment.

6. The spaceborne image compression method based on generative adversarial networks according to claim 1, characterized in that: The final recovered image data is obtained using the following formula: Among them, I reb For the final recovered image data, I R0 For the initial image restoration, I mask For the on-board binary mask, I R2 Optimize image data for the satellite uplink coding region.

7. A spaceborne image compression system based on generative adversarial networks, characterized in that... include: The first module is used to compress the preset remote sensing image data to obtain a compressed bitstream and a binary mask image; The second module is used to take the preset remote sensing image data and binary mask as the input of the training dataset of the neural network, and compare the absolute value of the residual between the repaired pixel value of the run-length encoded region of the binary mask and the initial pixel value of the run-length encoded region with the maximum allowable error value to obtain the logic gate signal. The third module is used to train the training dataset using a neural network to obtain the network weight parameters; The fourth module is used to decompress the lossy JPEG-LS bitstream transmitted in real time on the satellite to obtain the primary restored image and the on-board binary mask image; The fifth module is used to input the primary restored image and the on-board binary mask into the neural network, and use the network weight parameters to repair the pixel values ​​of the on-board roaming coding region to obtain intermediate restored image data; The sixth module is used to obtain residual map data based on the intermediate recovered image data, the on-board binary mask map, and the initial image of the on-board upstream coding region; The seventh module is used to optimize image data for the satellite roaming coding region based on residual map data and the maximum permissible error value; The eighth module is used to optimize image data and primary restored images based on the satellite upstream coding region to obtain the final restored image data.

8. The spaceborne image compression system based on generative adversarial networks according to claim 7, characterized in that: The logic gate signal is obtained by the following formula: Where D_flag is the logic gate signal, I d Near is the residual between the repaired pixel value of the run-length encoded region of the binary mask and the initial pixel value of the run-length encoded region, where Near is the maximum allowable error value.

9. The spaceborne image compression system based on generative adversarial networks according to claim 7, characterized in that: During training, the logic gating signal is 1.

10. The spaceborne image compression system based on generative adversarial networks according to claim 7, characterized in that: Residual plot data D run It can be obtained through the following formula: Among them, D run For residual plot data, I R1 To recover intermediate image data, I mask For the on-board binary mask, I run This is the initial image of the satellite's upstream coding region.

Images