Double-branch remote sensing image compression method based on global context feature guidance

Abstract

The invention discloses a double-branch remote sensing image compression method based on global context feature guidance, and relates to a remote sensing image compression method. The invention provides a double-branch remote sensing image compression method based on global context feature guidance in order to solve the problem that an obtained reconstructed image is poor in quality due to the fact that the global feature action range is small and the hierarchy is low in an existing conventional remote sensing image compression network. The method comprises the following steps of: 1, constructing a global context feature-guided double-branch remote sensing image compression network; the global context feature guided double-branch remote sensing image compression network comprises a PVTCB module, a fourth-order compression branch FSCB, a fourth-order reconstruction decoder FSRD, a global context feature guided module GVGM and a traditional codec. And step 2, firstly compressing the to-be-detected remote sensing image and then reconstructing the to-be-detected remote sensing image based on the global context feature guided double-branch remote sensing image compression network to obtain the reconstructed to-be-detected remote sensing image. The invention relates to the field of remote sensing image compression.

Description

Technical Field

[0001] The invention relates to a remote sensing image compression method. Background Art

[0002] Remote sensing images have many unique features of land objects, such as land cover, topography, landforms and temperature, which are usually not shown in natural images. Therefore, remote sensing images have been widely used in many fields such as environmental monitoring, meteorology, geological science, etc. [1][2][3][4]([1]Tang W,He F,Bashir AK,et al.Aremote sensing imagerotationobject detection approach for real-time environmental monitoring[J].Sustainable EnergyTechnologies and Assessments,2023,57:103270.[2]Gama PHT,Oliveira HN,Marcato J,etal.Weakly supervised few-shot segmentation via meta-learning[J].IEEE Transactions onMultimedia, 2022.[3]Han W, Zhang geological environment carrying capacity and sustainable development under remote sensing interpretation combined with integrated model[J].Ecotoxicology and Environmental Safety,2023,257:114956.). However, remote sensing images are usually captured by satellites in the high altitude through the atmosphere, which inevitably have more background noise. [5]([5]Kang J, Fernandez-Beltran R, Kang X, et al. Noise-tolerant deep neighborhood embedding for remotely sensed images with label noise [J]. IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing, 2021, 14: 2551-2562.) In addition, remote sensing images are often taken from high altitudes and contain complete object or scene information. Compared with natural images, their global context features are richer, that is, they contain significant global context information. [6] ([6] Zheng P, Jiang J, Zhang Y, et al. CGC-Net: AContext-GuidedConstrained Network for Remote-Sensing Image Super Resolution[J]. Remote Sensing, 2023, 15(12): 3171.). This information has an important impact on the compression of texture features and global context features. Secondly, with the rapid development of remote sensing technology, the spatial resolution and spectral resolution of remote sensing images are constantly improving, and the amount of data is also growing exponentially. [7] ([7] J, Fors O, Otazu X, et al. A wavelet-based method for the determination of the relative resolution between remotely sensed images [J]. IEEE Transactions on Geoscience and Remote Sensing, 2006, 44 (9): 2539-2548.). Based on the above reasons, a specialized compression method suitable for the characteristics of remote sensing images is urgently needed.

[0003] At present, common traditional image compression methods have achieved some results [8][9]([8] Sun C, Fan X, Zhao D. Lossless recompression of JPEG images using transform domain intraprediction [J]. IEEE Transactions on Image Processing, 2022, 32: 88-99. [9] De Cea-Dominguez C, Moure-Lopez JC, Bartrina-Rapesta J, et al. GPU-oriented architecture for an end-to-end image / video codec based on JPEG2000 [J]. IEEE Access, 2020, 8: 68474-68487.) For example, Báscones et al. proposed a method to compress hyperspectral image data by combining principal component analysis and JPEG2000.

[10] (

[10] Báscones D, González C, Mozos D. Hyperspectral image compression using vector quantization, PCAandJPEG2000[J]. Remote sensing, 2018, 10(6):907.). Classic JPEG

[11] (

[11] Wallace G K.TheJPEG still picture compressionstandard[J].Communications of the ACM,1991,34(4):30-44.) and JPEG2000

[12] (

[12] JPEG2000 official softwareOpenJPEG,https: / / jpeg.org / jpeg2000 / software.html) It mainly consists of three parts: image transformation, quantization and entropy coding. First, the image is transformed and dequantized; then, important information is retained through quantization; finally, entropy coding is used to compress the de-correlation coefficients. In addition, BPG, which has better performance

[13]

[14] (

[13] Kovalenko B,Lukin V,KryvenkoS,et al.BPG-Based Automatic Lossy Compression ofNoisy Images with the Prediction of anOptimal Operation Existence and ItsParameters[J].Applied Sciences,2022,12(15):7555.

[14] Li F,Lukin V,Ieremeiev O,et al.Quality control for the BPG lossy compression of three-channel remotesensing images[J].Remote Sensing,2022,14(8):1824.) and WebP

[15] (

[15] Maldonado M, WebP Ja new web oriented image format. Universitat Oberta de Catalunya, 2010[J]) was also born in the field of image compression. Li et al. used MDSI as a quality evaluation indicator to improve the BPG compression algorithm, and provided more accurate remote sensing image quality control through a two-step compression strategy, achieving consistency in compression efficiency and image quality.

[14] Traditional image compression methods can be classified into two types: vector quantization

[16] (

[16] Wang Z,NasrabadiN M,Huang T S.Spatial-spectral classification of hyperspectral imagesusingdiscriminative dictionary designed by learning vector quantization[J].IEEE Transactions onGeoscience and Remote Sensing,2013,52(8):4808-4822.), based on predictive coding

[17] (

[17] Hu Y, Yang W, Ma Z, et al. Learning end-to-end lossy image compression: A benchmark [J]. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2021, 44 (8): 4194-4211.) and transform coding-based algorithms

[18] (

[18] Aulí-Llinàs F,Marcellin MW,Serra-Sagrista J,et al.Lossy-to-lossless 3D image coding through prior coefficient lookup tables[J].Information Sciences,2013,239:266-282.) Qian proposed a fast vector quantization compression algorithm for multispectral images. The core of this algorithm is to replace the input vector with a codeword index that matches the codebook, thereby optimizing the data transmission and storage efficiency.

[19] (

[19] Qian SE. Hyperspectral data compression using a fast vector quantization algorithm [J]. IEEE Transactions on Geoscience and Remote Sensing, 2004, 42 (8): 1791-1798.) 3D-MBLP uses prediction technology to first eliminate the spatial redundant information in the image, then predict the current frequency band, and finally encode the prediction residual with the help of an entropy decoder.

[20] (

[20] Pizzolante R, Carpentieri B. Multiband and lossless compression of hyperspectral images [J]. Algorithms, 2016, 9 (1): 16.) In addition, 3D-SPIHT, as a transform compression method for three-dimensional images, achieves efficient image compression by applying 3D wavelet transform in spatial and spectral domains.

[21] (

[21] Thornton L, Soraghan J, Kutil R, et al. Unequally protected SPIHT video codec for low bit rate transmission over highly error-prone mobile channels [J]. Signal Processing: Image Communication, 2002, 17 (4): 327-335.). However, traditional image compression methods have certain limitations. For example, JPEG, JPEG2000 and BPG did not take the characteristics of remote sensing images into consideration when they were designed, so the compression performance of remote sensing images is relatively limited. Secondly, at high compression ratios, most traditional image compression methods show relatively poor rate-distortion performance. Finally, for the characteristics of remote sensing images such as high information entropy, complex background noise, and rich global context information, common traditional methods cannot make efficient adaptive adjustments. Therefore, for remote sensing images, it is difficult to provide good compression performance using traditional compression methods. In order to seek breakthroughs, researchers have focused on deep learning technology, which has been popular in recent years. The classic deep learning-based image compression framework mainly includes autoencoders (AE)

[22]

[23] (

[22] La Grassa R, Re C, Cremonese G, et al. Hyperspectral data compression using fully convolutional autoencoder [J]. Remote Sensing, 2022, 14 (10): 2472.

[23] Liu J, Yuan F, Xue C, et al. An Efficient and Robust Underwater Image Compression Scheme Based on Autoencoder [J]. IEEE Journal of Oceanic Engineering, 2023.) and Variational Autoencoder (VAE)

[24]

[25] (

[24] Alves de Oliveira V,Chabert M,Oberlin T,et al.Reduced-complexity end-to-end variational autoencoder for on board satellite image compression[J].Remote Sensing,2021,13(3):447.

[25] Xu Q,Xiang Y,Di Z,etal.Synthetic acquisition radar image compression based on a variational autoencoder[J].IEEE Geoscience and Remote Sensing Letters,2021,19:1-5.). Riccardo et al. proposed a compression network (SSCNet) based on a deep convolutional autoencoder to solve the problem of processing large amounts of data generated by complex satellite instruments in the field of space science and satellite imaging. The network shows superiority in both compression ratio and spectral signal reconstruction, and is also robust to data types larger than 8 bits.

[22] Alves et al. designed a reduced complexity variational autoencoder to meet the computational resource constraints in satellite image compression. By reducing the network dimension and simplifying the entropy model, the encoder reduces complexity while maintaining compression performance.

[24] However, compared with the AE-based framework, the VAE-based framework has a more powerful image reconstruction capability. The reason is that VAE has a continuous mapping space that AE does not have, which can reconstruct images with smooth transitions between pixels. In recent years, many VAE-based baseline networks have been published.

[26]

[27]

[28]

[29]

[30]

[31]

[32] (

[26] BalléJ,Laparra V,Simoncelli E P.End-to-end optimization of nonlinear transform codes for perceptual quality[C] / / 2016Picture Coding Symposium(PCS).IEEE,2016:1-5.

[27] Minnen D,BalléJ,TodericiG D.Joint autoregressive and hierarchical priors for learned imagecompression[J].Advances in neural information processing systems,2018,31.

[28] BalléJ,Minnen D,Singh S,et al.Variational image compression with a scalehyperprior[J].arXiv preprint arXiv:1802.01436,2018.

[29] Cheng Z,Sun H,TakeuchiM,et al.Learned image compression with discretized gaussian mixturelikelihoods and attention modules[C] / / Proceedings of the IEEE / CVF conferenceon computer vision and pattern recognition.2020:7939-7948.

[30] Guo Z,Zhang Z,Feng R,et al.Causal contextual prediction for learned image compression[J].IEEE Transactions on Circuits and Systems for Video Technology,2021,32(4):2329-2341.

[31] Chen T,Liu H,Ma Z,et al.End-to-end learnt image compression vianon-local attention optimization and improved context modeling[J].IEEE Transactions on Image Processing, 2021, 30: 3179-3191.

[32] Cao M, Dai W, Li S, et al. End-to-end optimized image compression with deep Gaussian process regression [J]. IEEE Transactions on Circuits and Systems for Video Technology, 2022.), showing superior rate-distortion performance over traditional image compression methods. These VAE-based image compression networks usually consist of three components: encoder, entropy encoder, and decoder. First, the image data block is preliminarily compressed using a neural network; then, the compressed pixel data is mapped to a quantized representation; finally, the data is further compressed into a bitstream using traditional coding techniques. In addition, in order to model more accurately, some compression models introduce entropy models such as Laplace models, single-core Gaussian models, mixed Gaussian models, and hierarchical entropy models into the framework to make full use of prior information.

[33]

[34]

[35]

[36]

[37]

[38] (

[33] Liu D,Sun X,Wu F,et al.Edge-oriented uniformintraprediction[J].IEEE Transactions on Image Processing,2008,17(10):1827-1836.

[34] Kong F,Cao T,Li Y,et al.Multi-scale spatial-spectral attention networkfor multispectral image compression based on variational autoencoder[J].Signal Processing,2022,198:108589.

[35] Fu C,Du B,Zhang L.Sar imagecompression based on multi-resblock and global context[J].IEEE Geoscience andRemote Sensing Letters,2023,20:1-5.

[36] Zhang L,Hu X,Pan T,et al.Global Priorswith Anchored-stripe Attention and MultiScale Convolution for Remote SensingImages Compression[J].IEEE Journal of Selected Topics in Applied EarthObservations and Remote Sensing,2023.

[37] Gao J,Teng Q,He X,et al.Mixedentropy model enhanced residual attention network for remote sensing imagecompression[J].Neural Processing Letters,2023,55(7):10117-10129.

[38] Xiang S,Liang Q,Fang L.Discrete wavelet transform-based Gaussian mixture model forremote sensing image compression[J].IEEE Transactions on Geoscience and Remote Sensing, 2023.). Based on the above theory, some scholars have studied many remote sensing image compression networks based on deep learning and achieved good rate-distortion performance.

[39]

[40]

[41]

[42]

[43]

[44]

[45]

[46] (

[39] Zhang L,Hu X,Pan T,et al.Global Priors with Anchored-stripe Attention and MultiScale Convolutionfor Remote Sensing Images Compression[J].IEEE Journal ofSelected Topics inApplied Earth Observations and Remote Sensing,2023.

[40] Guo Y,Chong Y,Ding Y,et al.Learned hyperspectral compression using a student’s T hyperprior[J].Remote Sensing,2021,13(21):4390.

[41] Zhao M,Yang R,Hu M,et al.Deep Learning-Based Technique for Remote Sensing Image Enhancement Using Multiscale FeatureFusion[J].Sensors,2024,24(2):673.

[42] Sumbul G,Xiang J,Demir B.Towardssimultaneous image compression and indexing for scalable content-basedretrieval in remote sensing[J].IEEE Transactions on Geoscience and RemoteSensing,2022,60:1-12.

[43] Wang H,Liao L,Xiao J,et al.Uplink-Assist DownlinkRemote Sensing Image Compression via Historical Referecing[J].IEEETransactions on Geoscience and Remote Sensing,2023.

[44] Liu M,Tang L,Fan L,etal.CARNet:Context-aware residual learning for JPEG-LS compressed remote sensing image restoration[J].Remote Sensing,2022,14(24):6318.

[45] Ye W,Lei W,Zhang W,et al.GFSCompNet:remote sensing image compression network based onglobal feature-assisted segmentation[J].Multimedia Tools and Applications,2024:1-25.

[46] Xiang S,Liang Q,Tang P.Task-oriented compression framework forremote sensing satellite data transmission[J].IEEE Transactions on IndustrialInformatics,2023.). Although the above methods have achieved good compression performance, the quality of the global context features extracted by these methods is relatively weak, that is, the global context features have a limited scope of application and lack multi-level nature. These shortcomings lead to the suboptimal rate-distortion performance of the model. .

[0004] Compared with natural images, remote sensing images contain rich global context features due to the influence of high-altitude photography. However, most of the current mainstream methods extract global context information by introducing a global feature module in the backbone network. There are two problems with this approach: first, the global context information obtained is relatively single and lacks diversity; second, the scope of global context information is limited and is only limited to the feature extraction process of the local network. Therefore, how to comprehensively and effectively extract global context information from remote sensing images, increase the diversity of global context features, and expand the scope of global context features has become a severe challenge that needs to be solved in the field of remote sensing image compression. Summary of the invention

[0005] The purpose of the present invention is to solve the problem that the global features in the existing conventional remote sensing image compression network have a small scope and low hierarchy, resulting in poor quality of the reconstructed image, and to propose a dual-branch remote sensing image compression method based on global context feature guidance.

[0006] The specific process of the dual-branch remote sensing image compression method based on global context feature guidance is as follows:

[0007] Step 1: Construct a dual-branch remote sensing image compression network GDRNet guided by global context features;

[0008] The dual-branch remote sensing image compression network guided by global context features includes a PVTCB module, a fourth-order compression branch FSCB, a fourth-order reconstruction decoder FSRD, a global context feature guided module GVGM, and a traditional codec;

[0009] The PVTCB module includes stage 1, stage 2, stage 3, and filter 5 in sequence;

[0010] Stage 1 consists of a patch embedding layer and a Transformer encoder in sequence;

[0011] Stage 2 consists of a patch embedding layer and a Transformer encoder in sequence;

[0012] Stage 3 consists of a patch embedding layer and a Transformer encoder in sequence;

[0013] The fourth-order compression branch FSCB includes compression module 1, QFMR-AM, compression module 2, compression module 3, and compression module 4;

[0014] The compression module 1 includes the first convolution layer and the first GDN in sequence. The convolution kernel size of the first convolution layer is 7×7, the number of input channels is 3, and the number of output channels is N / 4.

[0015] The compression module 2 includes a second convolutional layer and a second GDN in sequence. The convolution kernel size of the second convolutional layer is 3×3, the number of input channels is N / 4, and the number of output channels is N / 2.

[0016] The compression module 3 includes a third convolutional layer and a third GDN in sequence. The convolution kernel size of the third convolutional layer is 3×3, the number of input channels is N / 2, and the number of output channels is 3N / 4.

[0017] The compression module 4 includes a fourth convolutional layer and a fourth GDN in sequence. The convolution kernel size of the fourth convolutional layer is 3×3, the number of input channels is 3N / 4, and the number of output channels is N.

[0018] The GDN stands for generalized divisive normalization function; N is the number of channels;

[0019] The fourth-order reconstruction decoder FSRD includes a reconstruction module 1, a QFMR-AM, a reconstruction module 2, a reconstruction module 3, and a reconstruction module 4;

[0020] The reconstruction module 1 includes the eighth convolution layer and the first IGDN in sequence; the convolution kernel size of the eighth convolution is 3×3, the number of input channels is N / 4, and the number of output channels is 3;

[0021] The reconstruction module 2 includes the ninth convolution layer and the second IGDN in sequence; the convolution kernel size of the ninth convolution is 3×3, the number of input channels is N / 2, and the number of output channels is N / 4;

[0022] The reconstruction module 3 includes the tenth convolution layer and the third IGDN in sequence; the convolution kernel size of the tenth convolution is 3×3, the number of input channels is 3N / 4, and the number of output channels is N / 2;

[0023] The reconstruction module 4 includes the eleventh convolution layer and the fourth IGDN in sequence; the convolution kernel size of the eleventh convolution is 3×3, the number of input channels is N, and the number of output channels is 3N / 4;

[0024] The IGDN represents the inverse GDN operation;

[0025] The global context feature guidance module GVGM includes MCEB for extracting multi-channel information, MHSA for extracting global context features, and loss Loss for calculating the difference in global context features between the Compression part and the Reconstruction part. GVF ;

[0026] The specific working process of the global context feature guidance module GVGM is as follows:

[0027] InputA inputs MCEB, MCEB outputs feature input MHSA, MHSA outputs feature OutputA;

[0028] InputB inputs MCEB, MCEB outputs features input MHSA, MHSA outputs features OutputB;

[0029] Calculate the loss between feature OutputA and feature OutputB GVF ;

[0030] InputA represents the output feature map of compression module 3, and InputB represents the output feature map of reconstruction module 4;

[0031] Traditional codecs contain a probability model and a Q quantizer;

[0032] The specific working process of the dual-branch remote sensing image compression network GDRNet guided by the global context feature is as follows:

[0033] The remote sensing image is input into the PVTCB module, and the PVTCB module outputs the feature G;

[0034] The remote sensing image is sequentially input into compression module 1, QFMR-AM, compression module 2, and compression module 3 in the fourth-order compression branch FSCB, and compression module 3 outputs feature InputA;

[0035] The compression module 3 outputs the feature InputA which is sequentially input into MCEB and MHSA, and MHSA outputs the feature OutputA;

[0036] The MHSA output feature OutputA and the compression module 3 output feature InputA are added element by element to obtain the added feature, and the added feature is input into the compression module 4, and the compression module 4 outputs the feature;

[0037] Add the output feature of the compression module 4 and the output feature G of the PVTCB module element by element to obtain the added feature y;

[0038] Feature y is input into the probability model Probability Model, and the probability model Probability Model outputs arithmetic encoding AE and arithmetic decoding AD respectively;

[0039] Feature y is input into the Q quantizer, and the Q quantizer outputs feature feature Input AE arithmetic coding and AD arithmetic decoding in sequence, and AD arithmetic decoding outputs features

[0040] AD arithmetic decoding output characteristics Input reconstruction module 4, reconstruction module 4 outputs feature InputB;

[0041] The reconstruction module 4 outputs the feature InputB which is sequentially input into MCEB and MHSA, and MHSA outputs the feature OutputB;

[0042] Add the MHSA output feature OutputB and the reconstruction module 4 output feature InputB element by element to obtain the added feature K;

[0043] The added feature K is sequentially input into reconstruction module 3, reconstruction module 2, QFMR-AM, and reconstruction module 1, and reconstruction module 1 outputs the reconstructed image;

[0044] Calculate the loss between feature OutputA and feature OutputB GVF ;

[0045] Step 2: Based on the dual-branch remote sensing image compression network GDRNet guided by global context features, the remote sensing image to be tested is first compressed and then reconstructed to obtain the reconstructed remote sensing image to be tested.

[0046] The beneficial effects of the present invention are:

[0047] Remote sensing images are often taken from high altitudes and contain complete objects or scene information, which makes their global context features richer than natural images. Therefore, in order to enhance the scope of action and multi-level features of global context features of remote sensing images, the present invention proposes a dual-branch remote sensing image compression network guided by global visual features (GDRNet) to optimize global feature processing and further improve compression efficiency and reconstructed image quality. First, a pyramid vision transformer compression branch (Pyramid vision transformer compression branch, PVTCB) is designed, which captures multi-level global context information through a three-stage pyramid transformer module (Three-stage pyramid transformer module for image compression, TPTC), and uses filters to accurately control the output of PVTCB, thereby optimizing the compression effect of remote sensing images. Secondly, a quadruple filtered multi-core noise reduction attention module (QFMR-AM) for denoising and enhancing multi-level information features is constructed in the four-stage compression branch (FSCB). Finally, a global visual feature guidance module (GVGM) is built in FSCB and four-stage reconstruction decoder (FSRD). The proposed global context feature loss Loss is calculated by GVGM. GVF , and use this to reconstruct the overall loss Loss of this network Total , so that the network is more focused on extracting global information. The global features in the compression network of the present invention have a large range of action and high hierarchy, which improves the quality of the reconstructed image. Experimental results show that compared with some advanced compression models, the proposed GDRNet method has achieved better compression performance in multiple compression evaluation indicators.

[0048] The present invention proposes a dual-branch remote sensing image compression network guided by global visual features (GDRNet). By increasing the multi-level nature of global context information and expanding the scope of global information, the global context features in remote sensing images are enhanced, thereby improving the quality of reconstructed images. GDRNet mainly consists of the following parts: a dual-branch compression structure, including a pyramid vision transformer compression branch (Pyramid vision transformer compression branch, PVTCB) and a four-stage compression branch (Four-stage compression branch, FSCB); entropy coding; and a four-stage reconstruction decoder (Four-stage reconstruction decoder, FSRD). GDRNet is mainly analyzed and optimized from three aspects. First, remote sensing images are usually captured by satellites in space through complex clouds and atmosphere, so they inevitably carry more background noise. In response to this problem, the present invention proposes a quadruple filtered multi-core noise reduction attention module (QFMR-AM) for denoising and enhancing multi-level information features. Through multiple filters and convolutions of different scales, it can effectively weaken complex background noise and efficiently extract, enhance and fuse features of different scales. Second, the scope of common global context feature enhancement modules is limited. To address this problem, the present invention constructs a global visual feature guidance module (GVGM) between the compression part and the reconstruction part, through which the proposed global context feature loss Loss is calculated. GVF , and reconstructed the overall loss of this network Loss Total. This design allows the global context features to act efficiently on the entire network in the form of loss, thereby greatly enhancing the quality of the global context features. Third, using a single method to enhance the extraction of global context information will lead to the simplification of global features. Therefore, the present invention introduces a pyramid visual Transformer compression branch. This branch captures multi-level global context information through a three-stage pyramid transformer module (Three-stage pyramid transformer module for image compression, TPTC), and uses filters to accurately control the output of PVTCB, thereby optimizing the compression effect of remote sensing images. In summary, the present invention is based on the proposed PVTCB, FSCB, FSRD, QFMR-AM, GVGM, and the reconstructed overall loss Loss Total , built a high-performance GDRNet. This invention was invented in San Francisco

[47] (

[47] https: / / resources.maxar.com / product-samples / analysis-ready-data-san-francisco-california)NWPU-RESISC45

[48] (

[48] Cheng G, HanJ, Lu

[49] (

[49] Y.Yang and S.Newsam, "Bag-of-visual-words and spatial extensions for land-use classification," in ACMSIGSPATIAL International Conference on Advances in Geographic Information Systems (ACM GIS), 2010.) and other remote sensing image datasets. The experimental results show that compared with the traditional image compression method JPEG2000

[12] , BPG

[50] (

[50] F.Bellard.Bpgimage format.[Online].Available:http: / / bellard.org / bpg / ) and WebP

[51] (

[51] Maldonado M, WebPJ. a new web oriented image format. Universitat Obertade Catalunya, 2010[J].), and image compression method based on deep learning Minnen ta al.

[27] 、Minnenet al.(mean)

[27] 、Balle et al.(hyperprior)

[28] , Balle et al.(factorized-relu)

[28] 、Tong2023

[52] (

[52] Tong K, Wu Y, Li Y, et al. QVRF: A Quantization-error-aware Variable Rate Framework for Learned Image Compression [J]. arXiv preprint arXiv: 2303.05744, 2023.), the GDRNet proposed in this paper performs better in evaluation indicators such as Peak signal-to-noise ratio (PSNR) and multiscale structural similarity index metric (MS-SSIM).

[0049] In summary, the main contributions of the present invention are as follows:

[0050] 1) The present invention proposes a quadruple filtered multi-core noise reduction attention module (QFMR-AM), which aims to achieve noise reduction and enhancement of multi-level information features. The module effectively weakens complex background noise by using multiple filters and multi-scale convolution technology, and efficiently extracts, enhances and fuses feature information of different scales. 2) The present invention designs a global visual feature guidance module (GVGM), which calculates the global context feature loss Loss between the compressed part and the reconstructed part. GVF , to reconstruct the overall loss Loss of the proposed network Total, so that the global context features are efficiently applied to the entire network in the form of loss. 3) The present invention constructs a pyramid vision transformer compression branch (PVTCB), which captures multi-level global context information through a three-stage pyramid transformer module (Three-stage pyramid transformer module for image compression, TPTC), and uses filters to accurately control the output of PVTCB, thereby enhancing the multi-level nature of global features in remote sensing images. 4) The present invention combines PVTCB, FSCB, FSRD, QFMR-AM, GVGM, and the overall reconstruction loss Loss Total The Factorized Entropy Model is effectively embedded together to construct a high-performance dual-branch remote sensing image compression network guided by global visual features (GDRNet). Through a large number of experiments on the SanFrancisco, NWPU-RESISC45 and UC-Merced datasets, the superior performance of GDRNet in multiple evaluation indicators is demonstrated. BRIEF DESCRIPTION OF THE DRAWINGS

[0051] Figure 1 It is the overall structure diagram of GDRNet of the present invention;

[0052] Figure 2 The schematic diagram of the structure of the QFMR-AM of the present invention, wherein Input represents the input feature map, Output represents the output feature map, Filter1-4 represents filters, and the filters are mainly used to limit the output power, that is, to map the pixel values ​​of the feature map to a reasonable range, respectively. The mapping value of Filter1 is 0.3, the mapping value of Filter is 0.7, the mapping value of Filter3 is 0.7, and the mapping value of Filter4 is 0.3. Conv2D 1×1N / 4N / 4 represents the parameter setting of the convolution, wherein 1×1 represents the convolution kernel size, N / 4 represents the number of input channels, and N / 4 represents the number of output channels. The parameter setting descriptions of other convolutions are the same as above;

[0053] Figure 3Schematic diagram of the structure of the GVGM of the present invention, wherein InputA represents the output feature map of Compression block 3, InputB represents the output feature map of Reconstruction block 4, OutputA represents the global context feature map processed by the Compression part, OutputB represents the global context feature map processed by the Reconstruction part, GVGM for Compression represents the GVGM used to calculate the global context feature of the compression part, GVGMfor Reconstruction represents the GVGM used to calculate the global context feature of the reconstruction part, and Loss GVF Represents the loss between the global context features of the compressed part and the global context features of the reconstructed part, and this loss is measured using MSE;

[0054] Figure 4 The schematic diagram of the structure of the PVTCB of the present invention is generally divided into two parts; the first part is TPTC (the light yellow part in the figure), which has three stages, each of which consists of a patch embedding layer and a Transformer encoder; according to the pyramid structure, the size of the data block changes from large to small, and the dimension changes from low dimension to high dimension; the second part is Filter5, and its mapping coefficient is set to 0.3;

[0055] Figure 5 The rate-distortion performance curves of different methods on San Francisco, (a) PSNR curve, (b) MS-SSIM curve, PSNR (dB) is the peak signal-to-noise ratio (decibel), Rate (bpp) is the bit rate (bits per pixel), MS-SSIM (dB) is the multi-scale structural similarity index; Figure 6 Rate-distortion performance curves of different methods on NWPU-RESISC45: (a) PSNR curve, (b) MS-SSIM curve; Figure 7 Rate-distortion performance curves of different methods on UC-Merced: (a) PSNR curve, (b) MS-SSIM curve;

[0056] Figure 8Visual comparison of the reconstructed images obtained by different methods on the San Francisco dataset, (a) original image, (b) Minnen et al. (bpp: 0.252; PSNR: 29.92; MS-SSIM: 7.91), (c) Minnen et al. (mean) (bpp: 0.249; PSNR: 29.38; MS-SSIM: 7.36), (d) Balle et al. (hyperprior) (bpp: 0.251 PSNR: 29.53 MS-SSIM: 7.47), (e) Balle et al. al. (factorized-relu) (bpp: 0.252; PSNR: 29.25; MS-SSIM: 7.49), (f) Tong2023 (bpp: 0.252; PSNR: 30.28; MS_SSIM: 8.10), (g) JPEG2000 (bpp: 0.262; PSNR: 22.27; MS-S SIM:1.15), (h) Webp (bpp: 0.255; PSNR: 22.71; MS-SSIM: 1.19), (i) BPG (bpp: 0.270; PSNR: 23.96; MS-SSIM: 1.93), (j) GDRNet (bpp: 0.252; PSNR: 30.10; MS-SSIM: 8.26);

[0057] Fig. 9Visual comparison of the reconstructed images obtained by different methods on the dataset UC-Merced, (a) original image, (b) Minnen et al. (bpp: 0.249; PSNR: 30.37; MS-SSIM: 6.10), (c) Minnen et al. (mean) (bpp: 0.248; PSNR: 30.05; MS-SSIM: 5.69), (d) Balle et al. (hyperprior) (bpp: 0.249; PSNR: 30.39; MS-SSIM: 6.18), (e) Balle et al. al. (factorized-relu) (bpp: 0.250; PSNR: 28.93; MS-SSIM: 5.30), (f) Tong2023 (bpp: 0.251; PSNR: 29.59; MS-SSIM: 5.11), (g) JPEG2000 (bpp: 0.261; PSNR: 16.82; MS-S SIM: 0.75), (h) Webp (bpp: 0.486; PSNR: 19.93; MS-SSIM: 2.46), (i) BPG (bpp: 0.276; PSNR: 18.49; MS-SSIM: 0.97), (j) GDRNet (bpp: 0.249; PSNR: 30.76; MS-SSIM: 6.93);

[0058] Fig.10 Comparison of ablation test results of different methods on the San Francisco dataset: (a) PSNR, (b) MS-SSIM; Fig.11 Comparison of OA values ​​of reconstructed images obtained with different compression methods in remote sensing scene classification. DETAILED DESCRIPTION

[0059] Specific implementation method 1: The specific process of the dual-branch remote sensing image compression method based on global context feature guidance in this implementation method is as follows:

[0060] Step 1: Construct a dual-branch remote sensing image compression network GDRNet (Dual-branch remote sensing image compression network guided by global visual features);

[0061] The dual-branch remote sensing image compression network guided by global context features includes PVTCB module, four-stage compression branch FSCB (Four-stage compression branch), four-stage reconstruction decoder FSRD (Four-stage reconstruction decoder), global context feature guidance module GVGM (Global visual feature guidance module), and traditional codecs.

[0062] The PVTCB module includes stage 1, stage 2, stage 3, and filter 5 in sequence;

[0063] Stage 1 consists of a patch embedding layer and a Transformer encoder in sequence;

[0064] Stage 2 consists of a patch embedding layer and a Transformer encoder in sequence;

[0065] Stage 3 consists of a patch embedding layer and a Transformer encoder in sequence;

[0066] The fourth-order compression branch FSCB includes compression module 1, QFMR-AM, compression module 2, compression module 3, and compression module 4;

[0067] The compression module 1 includes the first convolutional layer (two-dimensional) and the first GDN in sequence. The convolution kernel size of the first convolutional layer is 7×7, the number of input channels is 3, and the number of output channels is N / 4.

[0068] The compression module 2 includes a second convolutional layer (two-dimensional) and a second GDN in sequence. The convolution kernel size of the second convolutional layer is 3×3, the number of input channels is N / 4, and the number of output channels is N / 2.

[0069] The compression module 3 includes a third convolutional layer (two-dimensional) and a third GDN in sequence. The convolution kernel size of the third convolutional layer is 3×3, the number of input channels is N / 2, and the number of output channels is 3N / 4.

[0070] The compression module 4 includes a fourth convolutional layer (two-dimensional) and a fourth GDN in sequence. The convolution kernel size of the fourth convolutional layer is 3×3, the number of input channels is 3N / 4, and the number of output channels is N.

[0071] The GDN stands for generalized divisive normalization function; N is the number of channels;

[0072] The fourth-order reconstruction decoder FSRD includes a reconstruction module 1, a QFMR-AM, a reconstruction module 2, a reconstruction module 3, and a reconstruction module 4;

[0073] The reconstruction module 1 includes the eighth convolution layer (two-dimensional) and the first IGDN in sequence; the convolution kernel size of the eighth convolution is 3×3, the number of input channels is N / 4, and the number of output channels is 3;

[0074] The reconstruction module 2 includes the ninth convolution layer (two-dimensional) and the second IGDN in sequence; the convolution kernel size of the ninth convolution is 3×3, the number of input channels is N / 2, and the number of output channels is N / 4;

[0075] The reconstruction module 3 includes the tenth convolution layer (two-dimensional) and the third IGDN in sequence; the convolution kernel size of the tenth convolution is 3×3, the number of input channels is 3N / 4, and the number of output channels is N / 2;

[0076] The reconstruction module 4 includes the eleventh convolution layer (two-dimensional) and the fourth IGDN in sequence; the convolution kernel size of the eleventh convolution is 3×3, the number of input channels is N, and the number of output channels is 3N / 4;

[0077] The IGDN represents the inverse GDN operation;

[0078] The global context feature guidance module GVGM includes MCEB for extracting multi-channel information, MHSA for extracting global context features, and loss Loss for calculating the difference in global context features between the Compression part and the Reconstruction part. GVF ;

[0079] The specific working process of the global context feature guidance module GVGM is as follows:

[0080] InputA inputs MCEB, MCEB outputs feature input MHSA, MHSA outputs feature OutputA;

[0081] InputB inputs MCEB, MCEB outputs features input MHSA, MHSA outputs features OutputB;

[0082] Calculate the loss between feature OutputA and feature OutputB GVF ;

[0083] InputA represents the output feature map of compression module 3, and InputB represents the output feature map of reconstruction module 4;

[0084] Traditional codecs include probability models and Q quantizers;

[0085] The specific working process of the dual-branch remote sensing image compression network GDRNet guided by the global context feature is as follows:

[0086] The remote sensing image is input into the PVTCB module, and the PVTCB module outputs the feature G;

[0087] The remote sensing image is sequentially input into compression module 1, QFMR-AM, compression module 2, and compression module 3 in the fourth-order compression branch FSCB, and compression module 3 outputs feature InputA;

[0088] The compression module 3 outputs the feature InputA which is sequentially input into MCEB and MHSA, and MHSA outputs the feature OutputA;

[0089] The MHSA output feature OutputA and the compression module 3 output feature InputA are added element by element to obtain the added feature, and the added feature is input into the compression module 4, and the compression module 4 outputs the feature;

[0090] Add the output feature of the compression module 4 and the output feature G of the PVTCB module element by element to obtain the added feature y;

[0091] Feature y is input into the probability model Probability Model, and the probability model Probability Model outputs arithmetic encoding AE and arithmetic decoding AD respectively;

[0092] Feature y is input into the Q quantizer, and the Q quantizer outputs feature feature Input AE arithmetic coding and AD arithmetic decoding in sequence, and AD arithmetic decoding outputs features

[0093] AD arithmetic decoding output characteristics Input reconstruction module 4, reconstruction module 4 outputs feature InputB;

[0094] The reconstruction module 4 outputs the feature InputB which is sequentially input into MCEB and MHSA, and MHSA outputs the feature OutputB;

[0095] Add the MHSA output feature OutputB and the reconstruction module 4 output feature InputB element by element to obtain the added feature K;

[0096] The added feature K is sequentially input into reconstruction module 3, reconstruction module 2, QFMR-AM, and reconstruction module 1, and reconstruction module 1 outputs the reconstructed image;

[0097] Calculate the loss between feature OutputA and feature OutputB GVF ;

[0098] Step 2: Based on the dual-branch remote sensing image compression network GDRNet guided by global context features, the remote sensing image to be tested is first compressed and then reconstructed to obtain the reconstructed remote sensing image to be tested.

[0099] The overall framework of GDRNet: The GDRNet proposed in this invention enhances the global context features in remote sensing images from the perspective of increasing the multi-level nature of global information and expanding the scope of global information, thereby comprehensively improving the quality of reconstructed images. It mainly achieves high-quality remote sensing image compression through QFMR-AM for denoising and enhancing multi-scale information, GVGM for efficiently applying global context features to the entire network in the form of loss, and PVTCB for enhancing the multi-level nature of global features in remote sensing images. Among them, QFMR-AM involves convolutions of different scales and four filters (Filter1-4) for controlling output power. Among them, GVGM involves three sub-modules, including a multi-channel enhancement block (MCEB), a multi-head self-attention module (MHSA), and a global context feature loss Loss GVF Among them, PVTCB involves two submodules, including TPTC and Filter5. In addition, the present invention uses Loss GVF Reconstructed the overall loss function Loss Total , thereby coordinating the compression and reconstruction work between the base network and the proposed module, thereby achieving higher compression efficiency.

[0100] The present invention designs four Compression blocks in FSCB and four Reconstruction blocks in FSRD by cleverly selecting the convolution kernel size and reallocating the number of channels, thereby achieving excellent compression performance under low complexity. The specific parameters of Compression block1-4 and Reconstruction block1-4 are shown in Table 1. Probabilitymodel is the probability model part, which mainly includes a hyper-prior network (hyper encoder, hyper decoder), Q quantizer, AE arithmetic coding and AD arithmetic decoding. The specific parameters of Hyper encoder and Hyperdecoder are shown in Table 2. Between AE and AD is the minimum form (bit stream) of data in this model. In the present invention, the hyper-prior network is used to learn the probability model (i.e., entropy model) on which entropy coding depends, and is also used to generate parameters of the entropy model (i.e., mean parameter μ i and scale parameter s i2 ), the entropy model is modeled as a conditional Gaussian. GVGM is used to calculate the global context feature loss Loss between the compressed part and the reconstructed part GVF , Loss GVF Represents the difference between the global features of the compressed part and the global features of the reconstructed part. The smaller the loss value, the smaller the difference between the global features on both sides, that is, the higher the quality of the extracted and reconstructed global features. The loss here uses the Mean squared error (MSE), which can be expressed as Formula 1. In the Rate-Distortion Optimization part, R represents the entropy rate, and λ represents the penalty coefficient used to control different bit rates. Indicates potential information. Represents side information.

[0101] GDRNet uses VAE as the basic framework to achieve a smoother transition between pixels. GDRNet mainly consists of the following parts: FSCB, PVTCB, GVGM, FSRD, Probability Model and traditional encoding and decoding. The image compression part adopts a dual-branch structure, namely FSCB and PVTCB. FSCB combines QFMR-AM and GVGM to first remove noise from remote sensing image blocks, and then extract the noise used to calculate Loss. GVF The global context features are finally mapped to the preliminarily compressed potential representation features. PVTCB first obtains the potential representation features rich in multi-level global context information through TPTC, and then constrains the output through filter Filter5 to obtain the corresponding potential representation features. Then, the present invention fuses the potential representation features obtained by the two branches. After that, statistical redundancy is further removed through quantization, arithmetic coding, and Probability model to obtain the minimum bit stream after data processing by this model. Image decoding part: The model combines the obtained bit stream with the mean μ in the Probability model i and scale s i 2 Parameters, through QFMR-AM in FSRD, multiple reconstruction blocks, and GVGM, the image is reconstructed with high quality.

[0102] Where m represents the number of pixels. represents the predicted image and X represents the original image.

[0103] Table 1 Specific parameters of Compression block and Reconstruction block

[0104]

[0105] Table 2 Specific parameters of Hyper encoder and Hyper decoder

[0106]

[0107] In Tables 1 and 2, N represents the number of channels, ↓ represents downsampling, ↑ represents upsampling, and RELU represents the linear rectification function. GDN represents the generalized split normalization function, and IGDN represents its inverse operation. They are nonlinear activation functions and are more suitable for normalization of image data than other normalization functions.

[0108] Specifically, in view of the rich background noise in remote sensing images, this study designed a multi-scale feature enhancement module QFMR-AM based on noise removal. In view of the small scope of common global feature extraction modules, the present invention proposes a global context feature guidance module GVGM that applies global context features to the entire network through loss. In view of the problem that the global context features extracted by conventional networks are relatively single, the present invention introduces the pyramid Transformer branch PVTCB that can extract multi-level global context information. Overall, GDRNet efficiently explores the potential relationship between global context features in the process of remote sensing image compression and reconstruction, as well as the multi-level nature of global context features through QFMR-AM, GVGM and PVTCB, thereby achieving high efficiency in compression and reconstruction.

[0109] Specific implementation method 2: This implementation method is different from the specific implementation method 1 in that: the QFMR-AM includes a fifth convolution layer (two-dimensional), a sixth convolution layer (two-dimensional), a seventh convolution layer (two-dimensional), an average pooling layer, a maximum pooling layer, a filter 1, a filter 2, a filter 3, and a filter 4;

[0110] The working process of the QFMR-AM is:

[0111] The input feature F is input into the fifth convolutional layer, the sixth convolutional layer, and the seventh convolutional layer respectively. The output feature of the fifth convolutional layer is input into the average pooling layer, and the average pooling layer outputs feature A.

[0112] The sixth convolutional layer outputs feature B;

[0113] The output feature of the seventh convolutional layer is input into the maximum pooling layer, and the maximum pooling layer outputs feature C;

[0114] Perform Hadamard product on feature A and feature B to obtain feature D;

[0115] Perform Hadamard product on feature B and feature C to obtain feature E;

[0116] Feature D is input into filter 1, feature B is input into filter 2, feature C is input into filter 3, and feature F is input into filter 4;

[0117] The output features of filter 1, filter 2, filter 3, and filter 4 are added element by element to obtain the QFMR-AM output features.

[0118] The other steps and parameters are the same as those in the first embodiment.

[0119] Specific implementation method three: This implementation method is different from specific implementation methods one or two in that the working process of the QFMR-AM is expressed as follows:

[0120] X=Conv 3′3 (Input)

[0121] In the formula, Input represents the input image data block, Conv 3′3 represents a convolution layer with a convolution kernel size of 3, and X represents the output data of filter 2;

[0122] The working process of the branch of filter 1 is expressed as:

[0123]

[0124] In the formula, Conv 1′1 Represents a convolution layer with a convolution kernel size of 1, Avgpool2D 3′3 represents the average pooling layer, ⊙ represents the Hadamard product, Filter1 represents filter 1, Represents the output of Filter1 branch;

[0125] The working process of the branch of filter 2 is expressed as:

[0126] I Filter2 =Filter2(Conv 3′3 (Input)

[0127] In the formula, Conv 3′3 represents the convolution layer with a convolution kernel size of 3, Filter2 represents filter 2, and I Filter2 Represents the output of the Filter2 branch;

[0128] The working process of the branch of filter 3 is expressed as:

[0129]

[0130] In the formula, Conv 5′5 Represents a convolution layer with a convolution kernel size of 5, Maxpool2D 3′3represents the maximum pooling layer, Filter3 represents filter 3, Represents the output of Filter3 branch;

[0131] The working process of the branch of filter 4 is expressed as:

[0132] I Filter4 =Filter4(Input)

[0133] Where Filter4 represents filter 4, I Filter4 Represents the output of Filter4 branch;

[0134] The working process of QFMR-AM can be expressed as:

[0135]

[0136] In the formula, represents element-by-element addition, I QFMR-AM Represents the output of QFMR-AM.

[0137] QFMR-AM: Remote sensing images are usually captured by satellites in space through complex clouds and atmosphere, and inevitably carry more background noise. In addition, remote sensing images do not have as many bands as hyperspectral images or multispectral images, and image features can be extracted through bands with less noise to compensate for the impact of bands with more noise. Therefore, the noise in remote sensing images will have a more serious negative impact on application tasks, such as target detection, scene classification, image compression, etc. This study designed a QFMR-AM module to target the complex background noise in remote sensing images. The module uses multiple filters and multi-scale convolution technology to weaken the background noise of remote sensing images, while efficiently extracting, enhancing and fusing feature information of different scales.

[0138] Common noises in remote sensing images include speckle noise and salt and pepper noise. The essence of this type of noise is the sudden change of values, that is, the sharp change of pixel values. The present invention constructs a four-branch QFMR-AM to effectively eliminate the influence of noise. The structure of QFMR-AM is as follows: Figure 2 shown.

[0139] Specifically, the Filter1 branch uses point convolution and small kernel convolution to extract a high-frequency feature. This type of high-frequency feature will contain more noise information, so Filter1 limits it to a smaller pixel value by mapping, and the mapping coefficient here is 0.3. The convolution of the Filter2 and Filter3 branches selects larger convolution kernels, 3×3 and 5×5 respectively, to obtain multi-scale features in remote sensing images, and is mapped to larger pixel values ​​through Filter2 and Filter3 as the main branch of this module. The mapping coefficients of Filter2 and Filter3 are 0.7. The Filter4 branch is similar to a residual structure, which is used to accelerate network training and prevent overfitting. Unlike the residual, the present invention adds a filter to this branch and maps it to a smaller pixel value through the filter, which prevents the noise in the original data block from being introduced into the final feature map again with the same high impact. In addition, this module introduces Avgpool2D between Filter1 and Filter2, and Maxpool2D between Filter3 and Filter4 to strengthen the interaction between multi-scale features. Finally, the feature maps of Filter1, Filter2, Filter3, and Filter4 are fused together by element-by-element addition. Due to the reasonable restrictions and proportional distribution of the output of each branch by the filter, the final output feature value is still within a reasonable value range without the problem of value explosion.

[0140] The other steps and parameters are the same as those in the first or second embodiment.

[0141] Specific embodiment 4: This embodiment is different from any one of the specific embodiments 1 to 3 in that the working process of the PVTCB module is expressed as follows:

[0142]

[0143] In the formula, B represents the batch number, C1, C2, C3, C4 represent the number of channels, H represents the image height, and W represents the image width; Stage1, Stage2, Stage3 represent stage 1, stage 2, and stage 3 respectively; Filter5 represents filter 5; Input represents input; I PVTCB Represents the output of the PVTCB branch.

[0144] The embedding dimension of the patch embedding layer of Stage1 is set to

[64] , the embedding dimension of the patch embedding layer of Stage2 is set to

[160] , and the embedding dimension of the patch embedding layer of Stage3 is set to

[256] ; the number of heads of the multi-head self-attention of the Transformer encoder of Stage1 is set to [2], the number of heads of the multi-head self-attention of the Transformer encoder of Stage2 is set to [4], and the number of heads of the multi-head self-attention of the Transformer encoder of Stage3 is set to [8]; the stacking depth of the Transformer encoder of Stage1 is set to [2], the stacking depth of the Transformer encoder of Stage2 is set to [2], and the stacking depth of the Transformer encoder of Stage3 is set to [2];

[0145] The common Transformer architecture has achieved great success in the field of natural language processing (NLP), but there are still some shortcomings in the application of computer vision: (1) The self-attention mechanism of the standard Transformer will cause significant computational and memory overhead when processing large-sized images, because it needs to perform pairwise attention calculations on all elements in the input sequence. (2) Transformers usually output feature maps of a single scale, which limits their application in tasks that require high-resolution output (such as pixel-level tasks). However, the emergence of the Pyramid Vision Transformer (PVT) solves these problems well.

[53] (

[53] Wang W, Xie E, Li X, et al. Pyramid vision transformer: A versatile backbone for dense prediction without convolutions [C] / / Proceedings of the IEEE / CVF international conference on computer vision. 2021: 568-578.). First, PVT uses a pyramid structure to generate multi-scale feature maps, which enables the model to adapt to downstream tasks of different resolutions, especially suitable for dense prediction tasks such as object detection and semantic segmentation. Secondly, it uses fine-grained image blocks (i.e., each 4 × 4 pixels) as input to learn high-resolution representations. Finally, it introduces Spatial-reduction attention (SRA), which effectively reduces the computation and memory consumption by reducing the input spatial dimension before attention calculation, so that PVT can handle high-resolution tasks more efficiently.

[0146] Based on the above advantages, the present invention introduces PVT into the field of remote sensing image compression for the first time and improves it to be more suitable for remote sensing image compression tasks. The present invention constructs a PVTCB compression branch to increase the multi-level nature of global features. PVTCB consists of two core components, including TPTC and Filter5. The original PVT contains four stages. Analogously to the field of image compression, these four stages can be regarded as four more complex downsamplings. As the stage deepens, the size of the data block gradually decreases and the number of channels continues to increase. However, PVT reduces the size of the data block too much and has a large number of parameters. Therefore, the present invention reconstructs the number of stages and the key parts of the model to construct a TPTC that is more suitable for remote sensing image compression. The number of stages is reconstructed to three, which greatly reduces the complexity of the model on the one hand. On the other hand, the output of the PVTCB branch and the output of the FSCB branch are matched. In addition, the number of embedding dimensions of each stage is set to [64, 160, 256], thereby improving the model's ability to acquire channel features. The number of multi-head self-attention of each stage is set to [2,4,8] to strengthen the fusion of information on different subspaces. The coefficient of the hidden layer in the multi-layer perceptron of Stage1 is set to [8], the coefficient of the hidden layer in the multi-layer perceptron of Stage2 is set to [4], and the coefficient of the hidden layer in the multi-layer perceptron of Stage3 is set to [4], so as to reduce a certain amount of parameters. The encoder stacking depth of each stage is set to [2,2,2]. The spatial reduction ratio of the sub-layer of Stage1 is set to [4], the spatial reduction ratio of the sub-layer of Stage2 is set to [2], and the spatial reduction ratio of the sub-layer of Stage3 is set to [1], so as to retain more channel features. Finally, at the end of TPTC, the present invention adds Filter5 to limit the pixel values ​​of the multi-level global feature map to a reasonable range, so as to facilitate effective fusion with the features of the FSCB branch.

[0147] The other steps and parameters are the same as those in Specific Embodiments 1 to 3.

[0148] Specific implementation mode 5: This implementation mode is different from any one of the specific implementation modes 1 to 4 in that: the MCEB includes a twelfth convolutional layer (two-dimensional), a thirteenth convolutional layer (two-dimensional), and a fourteenth convolutional layer (two-dimensional);

[0149] The convolution kernel size of the twelfth convolutional layer is 1×1, the number of input channels is 3N / 4, and the number of output channels is 3N / 4;

[0150] The convolution kernel size of the thirteenth convolutional layer is 1×1, the number of input channels is 3N / 4, and the number of output channels is 3N / 4;

[0151] The convolution kernel size of the fourteenth convolutional layer is 1×1, the number of input channels is 3N / 4, and the number of output channels is 3N / 4;

[0152] The specific working process of the MCEB is as follows:

[0153] InputA inputs the twelfth convolutional layer, the thirteenth convolutional layer, and the fourteenth convolutional layer respectively, and adds the output features of the twelfth convolutional layer, the thirteenth convolutional layer, and the fourteenth convolutional layer element by element. The features after element-by-element addition are the output features of MCEB;

[0154] InputB inputs the twelfth convolutional layer, the thirteenth convolutional layer, and the fourteenth convolutional layer respectively, and adds the output features of the twelfth convolutional layer, the thirteenth convolutional layer, and the fourteenth convolutional layer element by element. The features after element-by-element addition are the output features of MCEB.

[0155] The other steps and parameters are the same as those in Specific Embodiments 1 to 4.

[0156] Specific implementation method 6: This implementation method is different from the specific implementation methods 1 to 5 in that the working process of the MCEB is expressed as:

[0157] I MCEB =Conv 1′1 (Input)+Conv 1′1 (Input)+Conv 1′1 (Input)

[0158] In the formula, Input represents the input image data block, Conv 1′1 represents a convolution with a kernel size of 1×1, I MCEB Represents the output of MCEB.

[0159] The other steps and parameters are the same as those in Specific Implementations 1 to 5.

[0160] Specific implementation method 7: This implementation method is different from any one of the specific implementation methods 1 to 6 in that: the loss Loss between the feature OutputA and the feature OutputB GVF The working process is expressed as:

[0161] Loss GVF =L MSE (I MHSA(Compression) ,I MHSA(Reconstruction) )

[0162] Where, L MSE represents the loss measured using MSE, I MHSA(Compression) Represents the MHSA output of the compression part, I MHSA(Reconstruction) Represents the MHSA output of the reconstruction part, LossGVF Represents the global context feature loss.

[0163] GVGM: Remote sensing images contain rich global information, but the common global feature acquisition methods mainly exist in a certain part of the network in the form of a module. The global context features extracted by these methods either act on a certain part or a certain branch, and their influence range is very limited. Therefore, the present invention constructs a global context feature guidance module GVGM, which can apply the global context features of the image to the entire network in the form of loss. The overall structure of GVGM is shown in the figure below. Figure 3 It mainly includes three core components, namely MCEB for extracting multi-channel information, MHSA for extracting global context features, and Loss for calculating the difference in global context features between the Compression part and the Reconstruction part. GVF .

[0164] GVGM for Compression represents the GVGM used to calculate the global context features of the compression part, GVGMforReconstruction represents the GVGM used to calculate the global context features of the reconstruction part, and Loss GVF It represents the loss between the global context features of the compressed part and the global context features of the reconstructed part. This loss is measured by MSE.

[0165] Among them, the main body of the MCEB part uses three point convolutions to extract and enhance multi-channel information. The reason for using multiple channels to extract spatial information is that the size of the MCEB input feature map is relatively large, which leads to more spatial information to be extracted; therefore, the low information capacity of a single channel will lead to the loss of some spatial information; and the reason for using element-by-element addition to process multi-channel information is that the feature map processed by the MCEB module is input into the MHSA to extract global features. Here, the pixel values ​​of the three channels are superimposed to increase the overall pixel value, thereby increasing the difference between different features and transforming the output features into a data form more suitable for MHSA.

[0166] Another core component is MHSA: In recent years, Vision transformer (ViT) has been widely used in the field of computer vision. Its powerful ability to capture long-distance contextual information enables the ViT model to perform well in various tasks. This ability is mainly attributed to its core module, the self-attention mechanism. MHSA introduces a multi-head mechanism based on self-attention, which not only improves the training speed, but also realizes the fusion of information from different subspaces. MHSA maps remote sensing image data to different projection spaces through Feature1, Feature2 and Feature3, and fuses information through tensor multiplication, thereby strengthening the global context features.

[0167] The working process of MHSA can be expressed as:

[0168]

[0169] In the formula, Conv 3′3 represents convolution, Linear represents linear layer, Dropout represents random dropout layer, Softmax represents Softmax layer, Represents matrix multiplication.

[0170] The last core component is Loss GVF , which calculates the difference between the global features of the compressed part and the global features of the reconstructed part, and inputs this difference in global features into Loss Total The reason is that: (1) this network can reduce the loss GVF The difference between the global features of the compressed part and the global features of the reconstructed part is reduced by using a numerical value, so as to improve the similarity between the global features of the compressed part and the global features of the reconstructed part, thereby improving the quality of the global context features. (2) This global feature loss is introduced into the overall loss, so that the scope of the global context features is radiated to the entire network, thereby solving the limitation of the small scope of the ordinary global feature extraction module.

[0171] The other steps and parameters are the same as those in Specific Implementations 1 to 6-1.

[0172] Specific implementation eight: This implementation differs from any one of specific implementations one to seven in that: the mapping value of filter 1 is 0.3, the mapping value of filter 2 is 0.7, the mapping value of filter 3 is 0.7, and the mapping value of filter 4 is 0.3; the mapping coefficient of filter 5 is set to 0.3.

[0173] The other steps and parameters are the same as those in Specific Implementation 1 to 7-1.

[0174] Specific implementation method 9: This implementation method is different from any one of the specific implementation methods 1 to 8 in that the total loss function of the dual-branch remote sensing image compression network GDRNet guided by global context features is Loss Total :

[0175] The training process is: argminProposedLoss Total =R+l(D+yLoss GVF )

[0176] In the formula, R represents the entropy rate, λ represents the penalty coefficient, D represents the degree of distortion between the original image and the reconstructed image, y represents the weight coefficient of the global context feature loss, and ProposedLoss Total Represents the total loss of the dual-branch remote sensing image compression network GDRNet guided by global context features;

[0177] Obtain the trained dual-branch remote sensing image compression network GDRNet guided by global context features.

[0178] E: Rate-Distortion Optimization: The core training goal of the compression framework is to achieve a balance between compression rate and distortion. To achieve this goal, a rate-distortion optimization strategy is usually added to the compression framework to guide the model to train efficiently. In short, this strategy aims to ensure that while compressing data, information loss is minimized as much as possible to achieve high-quality compression effects. The rate-distortion optimization strategy can be expressed as:

[0179] argminLoss Totel =R+lD

[0180] Where R represents the entropy rate, i.e., the cross entropy between the potential marginal distribution and the learned entropy model. D represents the degree of distortion between the original image and the reconstructed image, and different bit rates can be controlled by adjusting the penalty coefficient λ.

[0181]

[0182] In the formula, the bit rate is represented by the potential information and side information Together form.

[0183]

[0184] In the formula, is a learnable entropy model, It is a super encoder.

[0185] In order to further improve the efficiency and quality of image compression, this study proposes a new rate-distortion optimization strategy.Total The global context feature loss Loss is introduced in GVF , that is, to spread the scope of global features to the entire network, thereby improving the quality of global features of the entire network. This new rate-distortion optimization strategy can be expressed as:

[0186] argminProposedLoss Total =R+l(D+yLoss GVF )

[0187] Where y represents the weight coefficient of the global context feature loss.

[0188] The other steps and parameters are the same as those in Specific Implementation 1 to 8-1.

[0189] Specific implementation method ten: This implementation method is different from any one of specific implementation methods one to nine in that: in step two, the dual-branch remote sensing image compression network GDRNet guided by global context features first compresses the remote sensing image to be measured and then reconstructs it to obtain the reconstructed remote sensing image to be measured; the specific process is:

[0190] Based on the trained dual-branch remote sensing image compression network GDRNet guided by global context features, the remote sensing image to be tested is first compressed and then reconstructed to obtain the reconstructed remote sensing image to be tested.

[0191] The other steps and parameters are the same as those in Specific Implementation Methods 1 to 9.

[0192] The following examples are used to verify the beneficial effects of the present invention:

[0193] Embodiment 1:

[0194] In San Francisco

[47] 、NWPU-RESISC45

[48] and UC-Merced

[49] Adequate experiments were conducted on remote sensing image datasets. The selected datasets contain rich ground object information, which can effectively evaluate the effectiveness of the proposed GDRNet method. This paper compares the proposed GDRNet method with some excellent compression methods, including traditional codecs and deep learning-based compression models, to verify the superiority of the proposed method. Traditional image compression methods include JPEG2000

[12] , BPG

[50] and WebP

[51] . Compression models based on deep learning, including Minnen ta al.

[27] 、Minnen et al.(mean)

[27] 、Balle et al.(hyperprior)

[28] , Balle et al.(factorized-relu)

[28] Tong2023

[52] The experimental results show that the proposed GDRNet method has the best performance in terms of evaluation indicators such as PSNR and MS-SSIM. In addition, this experiment also evaluates the quality of reconstructed images obtained by different compression methods from the perspective of remote sensing image classification to further verify the superiority of the GDRNet method.

[0195] A. Dataset

[0196] 1) Dataset San Francisco: from

[47] A remote sensing image dataset. It is a remote sensing image with a resolution of 17408×17408, covering various ground features such as building areas, coasts, roads, ports, lakes, etc. This paper cuts it into 256×256 pixel images and selects 3000 valid images to form a dataset. These images are divided into training set, validation set and test set in a ratio of 8:1:1.

[0197] 2) Dataset NWPU-RESISC45: NWPU-RESISC45 is provided by Northwestern Polytechnical University (NWPU) in China. It contains 45 categories of remote sensing image scenes, 700 images per category, and 256×256 pixels per image. The dataset has diverse scenes, including airports, deserts, churches, forests, etc. 140 images per category are selected to form a dataset of 6,300 images, which are divided into training set, validation set, and test set in a ratio of 8:1:1.

[0198] 3) Dataset UC-Merced: UC-Merced is provided by the University of California, Merced. It contains 21 categories of remote sensing images, 100 images per category, a total of 2100 images, with a resolution of 256×256 pixels. The images cover a variety of landforms such as farmland, airports, and forests. The dataset is divided into training set, validation set, and test set in a ratio of 8:1:1.

[0199] B. Evaluation Indicators

[0200] In terms of image quality assessment, the present invention uses two common indicators, peak signal-to-noise ratio and multi-scale structural similarity index. In the remote sensing image scene classification part, the present invention uses overall accuracy and confusion matrix to measure the classification performance.

[0201] 1) Peak signal-to-noise ratio: PSNR compares the reconstructed image with the original image from the perspective of mean square error. The larger the PSNR value, the higher the fidelity of the reconstructed image. The peak signal-to-noise ratio can be expressed as

[0202]

[0203] Among them, MSE represents the mean square error between the original image and the reconstructed image, and max 2 (X i ) represents i The square of the maximum pixel in a band, C represents the number of bands.

[0204] 2) Multi-scale structural similarity index measurement: MS-SSIM is a multi-scale structural similarity index that measures the difference between the original image and the reconstructed image by merging image details at different resolutions. Its value range is 0 to 1, and the higher the value, the higher the similarity, that is, the higher the quality of the reconstructed image. The formula of MS-SSIM can be expressed as:

[0205]

[0206] Among them, M represents different resolutions, μ X , Represent the mean of the original image and the reconstructed image, s X , represents the standard deviation between the original image and the reconstructed image, represents the covariance between the original image and the reconstructed image, α m 、z m It indicates the relative importance between the two terms. C1 and C2 are constant terms used to prevent the divisor from being zero.

[0207] In order to more clearly compare the difference in MS-SSIM values, it is converted into decibel values, which can be expressed as:

[0208] MS-SSIM=-10log 10 (1-D MS-SSIM ) (18)

[0209] 3) Remote sensing scene classification index: This paper selects two widely used classification evaluation indexes to measure the quality of reconstructed images, including overall accuracy (OA) and confusion matrix (CM). The OA value is obtained by dividing the number of correctly classified images by the total number of test images, which reflects the overall performance of a classification model. CM reflects the detailed classification error and confusion degree between different scene categories. Each row in CM represents the true category, and each column represents the predicted category.

[0210] C. Experimental setup: In this study, the proposed GDRNet method is implemented by the PyTorch framework. The optimizer used is the Adam optimizer. There are two optimizers in this network, one is the main optimizer between the main encoder (PVTCB, FSCB) and the main decoder (FSRD), and the other is the auxiliary optimizer between the super encoder and the super decoder. For the main optimizer, the initial learning rate is set at 10 -4 During network training, when the learning rate decays to 10 -6 The best training model of GDRNet will be stored. For the auxiliary optimizer, its initial learning rate is set at 10 -3 . The batch size is set to 8 during training. For fairness in the experiment, the neural network models are all trained on NVIDIA GeForceRTX 3090, and the traditional encoding and decoding parts are all performed on the same CPU (i9-9900K CPU@3.60GHz). The present invention selects eight comparison methods, three traditional image compression methods based on encoding and decoding, and five image compression methods based on deep learning. The penalty coefficients commonly used in the present invention λ The values ​​are [0.660, 0.508, 0.211, 0.072, 0.033, 0.013, 0.007]. The mapping coefficients of filters 1-5 are [0.3, 0.7, 0.7, 0.3, 0.3]. The number of multi-heads in MHSA in GVGM is set to 4. In the proposed rate-distortion optimization strategy, the global context feature loss Loss GVF The weight coefficient y Set to 0.65. N is set to 256 in Compression block, Reconstruction block, Hyper encoder and Hyper decoder. For remote sensing scene classification, the benchmark model used for testing is EMTCAL (Efficient multiscale transformer and cross-level attention learning)

[54] (

[54] Tang X, Li M, Ma J, et al. EMTCAL: Efficient multiscale transformer and cross-level attention learning for remote sensing scene classification [J]. IEEE Transactions on Geoscience and Remote Sensing, 2022, 60: 1-15.), the training dataset used is NWPU-RESISC45, and the training-test ratio is 10%-90%. The images used for compression and the images used for remote sensing scene classification training are not crossed, and the reconstructed images are only used for testing the classification performance, not for training the classification network.

[0211] D. Rate-distortion performance curve: This experiment uses PSNR and MS-SSIM to evaluate the rate-distortion performance of the model. The present invention selects eight comparison methods, including three traditional codec image compression methods and five deep learning-based image compression methods. Figure 5 , Figure 6 and Figure 7The PSNR and MS-SSIM rate-distortion performance curves of all model experiments on the San Francisco, NWPU-RESISC45, and UC-Merced datasets are shown respectively. Among the traditional codec-based image compression methods, BPG shows better rate-distortion performance than WebP and JPEG2000 in most cases. This is mainly due to BPG's multi-channel encoding technology, which allows independent encoding of different color channels, thereby achieving fine control of detail features, which helps to reconstruct high-quality images. Among the image compression methods based on deep learning, Balle et al. (factorized-relu) shows relatively poor rate-distortion performance. This is mainly because it only uses simple convolutional layers, which has limited feature extraction capabilities. Although the feature extraction capability can be improved to a certain extent by increasing the number of stacked convolutional layers, this will significantly increase the number of model parameters and prolong the inference time. On the NWPU-RESISC45 and UC-Merced datasets, the Tong2023 method achieved the highest PSNR and MS-SSIM rate-distortion performance except for GDRNet. This is mainly due to its excellent attention mechanism design and more reasonable residual convolution module. However, on the San Francisco dataset, the rate-distortion performance of the Tong2023 method is poor, which indicates that the compression model is less robust. The other neural network-based comparison methods perform generally in rate-distortion performance, mainly because they lack a strong attention mechanism and excellent rate-distortion optimization strategy. The GDRNet method proposed in the present invention achieves the highest PSNR and MS-SSIM rate-distortion performance on three datasets at the same time. This superior rate-distortion performance not only strongly proves the robustness of the GDRNet method, but also strongly proves the effectiveness of PVTCB, QFMR-AM, GVGM and the newly proposed rate-distortion optimization strategy in the GDRNet method.

[0212] E. Visualization comparison experiment of reconstructed images: To further verify the visual effect of GDRNet reconstructed images, the present invention conducted a visualization comparison experiment. Figure 8 and Fig. 9 The reconstructed images on the San Francisco dataset and the UC-Merced dataset, as well as their local magnified images, are shown in the visualization experiment. The images in the visualization experiment are all reconstructed images when the bit rate is 0.25bpp. Figure 8As an example, they are the original image, the reconstructed images of eight comparison methods and the GDRNet method, and their local method diagrams. Among the traditional image compression methods based on codecs, the visualization effect of BPG is significantly better than that of JPEG2000 and WebP. In the enlarged image of the reconstructed image on the lower right, the roof of the BPG method retains more texture information compared to JPEG2000 and WebP. The reconstructed areas of JPEG2000 and WebP have lost most of the detailed features and appear blurred. The main reason for this phenomenon is that the BPG method has multi-channel encoding technology, which has a stronger reconstruction capability for detailed features. The image compression comparison method based on deep learning generally achieves better visualization effects than traditional image compression methods, but is still worse than the GDRNet method. Figure 8 In the reconstructed images of methods such as Minnen et al., Minnenet al. (mean)., Balle et al., and Balle et al. (factorized-relu), there are generally some artifacts and noise. This causes the local image to be blurred. A careful observation shows that the pixel transitions of the reconstructed images of these four comparison methods are too rough, which leads to color flattening and distortion. Finally, the Tong2023 and GDRNet methods with better visualization effects among the comparison methods are compared. The tree of the local magnification image in the GDRNet method retains more texture features and clearer object edges. Therefore, GDRNet achieved the best visualization effect on the San Francisco dataset. In addition, Fig. 9 In the experiment, GDRNet also achieved the best visualization effect on the UC-Merced dataset. This also verifies the strong robustness of GDRNet. From the perspective of visualization, the above experiments fully prove the rationality of the three working principles of GDRNet (removing complex background noise, enhancing the multi-level nature of global context features, and expanding the scope of global context features), as well as the efficiency of the new rate-distortion optimization strategy.

[0213] F. Ablation Experiment: The present invention conducts sufficient ablation experiments to verify the effectiveness of the proposed PVTCB, QFMR-AM and GVGM components. Fig.10is the ablation experiment result obtained on the San Francisco dataset. Among them, (1) Baseline represents the baseline network. (2) GDRNet (PVTCB) represents the integration of PVTCB on the basis of Baseline. (3) GDRNet (QFMR-AM) represents the integration of QFMR-AM on the basis of Baseline. (4) GDRNet (GVGM) represents the integration of GVGM on the basis of Baseline. (5) GDRNet represents the integration of PVTCB, QFMR-AM and GVGM on the basis of Baseline. It should be noted that GVGM here includes the proposed new rate-distortion optimization strategy. By Fig.10 It can be seen that the rate-distortion performance of the Baseline is at the lowest in most cases. At the same bit rate, the performance of GDRNet (PVTCB) is better than that of the Baseline, which verifies the effectiveness of the method of improving the quality of reconstructed images by improving the multi-level nature of global context features. At the same bit rate, the rate-distortion performance of GDRNet (QFMR-AM) is also better than that of the Baseline, which shows that removing complex background noise in remote sensing images is of great significance to improving the quality of image reconstruction. At the same bit rate, the rate-distortion performance of GDRNet (GVGM) is better than that of the Baseline, which fully demonstrates the positive impact of expanding the scope of global context features on improving the quality of reconstructed images. In addition, GDRNet basically achieves the best rate-distortion performance at the same bit rate. This phenomenon shows that PVTCB, QFMR-AM and GVGM, under the guidance of the reconstructed rate-distortion optimization strategy, achieve efficient feature extraction by removing background noise, enhancing the multi-level nature of global features and expanding the scope of global features.

[0214] H. Remote sensing image scene classification: The present invention uses reconstructed images obtained by different compression methods for remote sensing scene classification to verify the effectiveness of the proposed GDRNet method from an application perspective. The selected dataset is NWPU-RESISC45. The compared image compression methods include Minnen ta al., Minnen et al. (mean), Balle et al. (hyperprior), Balle et al. (factorized-relu), Tong2023, and the proposed GDRNet method. The remote sensing scene classification benchmark model is EMTCAL. To ensure the fairness of the experiment, the reconstructed images of different methods are obtained at a bit rate of 0.6bpp. Fig.11The overall accuracy (OA) of the reconstructed images obtained by different methods when classifying remote sensing image scenes. Among them, the proposed GDRNet achieved the highest OA value and the best classification performance, exceeding 0.37% compared with Minnen ta al.; exceeding 0.37% compared with Minnen et al. (mean); exceeding 0.37% compared with Balle et al. (hyperprior); exceeding 0.73% compared with Balle et al. (factorized-relu); and exceeding 0.19% compared with Tong2023.

[0215] I Complexity Analysis: The present invention performs complexity analysis, and the test parameters mainly include Parameter, FLOPs, GPU Memory, Compression time and Inference time. Here, the floating point library used for calculation is thop, the library used for calculating Parameter is torchsummary, and the input image format is 3×256×256. By comparison, it is found that the Parameter of GDRNet is the third least among all methods; it is worth mentioning that although the Parameter of Balle et al. (factorized-relu) is less than that of the proposed GDRNet, its PSNR and MS-SSIM rate-distortion performance are much lower than our method under the premise of bit rate. By comparing FLOPs, it can be found that GDRNet has achieved the smallest FLOPs parameter amount, which is only 44.27%, 44.70%, 45.19%, 46.13%, and 17.81% of the FLOPs parameter amount of Minnen ta al., Minnen et al. (mean), Balle et al. (hyperprior), Balle et al. (factorized-relu), Tong2023 and other methods, respectively. This fully demonstrates the superiority of the GDRNet method. Compared with GPU Memory, GDRNet consumes the most among all methods. The reason is that the designed dual-branch structure requires a large amount of parallel computing, which leads to a large GPUMemory overhead. Comparing Compression time and Inference time, it can be found that GDRNet has a moderate time consumption, but at the same bit rate, GDRNet's PSNR and MS-SSIM are much better than other comparison methods. These experiments strongly prove that GDRNet can achieve excellent rate-distortion performance at a relatively low complexity.

[0216] Table 3 Complexity experiments of different compression methods

[0217]

[0218] The present invention may also have many other embodiments. Without departing from the spirit and essence of the present invention, those skilled in the art may make various corresponding changes and modifications based on the present invention, but these corresponding changes and modifications should all fall within the scope of protection of the claims attached to the present invention.

Claims

1. A dual-branch remote sensing image compression method based on global context feature guidance, characterized by: The specific process of the method is: Step 1: Construct a dual-branch remote sensing image compression network GDRNet guided by global context features; The dual-branch remote sensing image compression network guided by global context features includes a PVTCB module, a fourth-order compression branch FSCB, a fourth-order reconstruction decoder FSRD, a global context feature guided module GVGM, and a traditional codec; The PVTCB module includes stage 1, stage 2, stage 3, and filter 5 in sequence; Stage 1 consists of a patch embedding layer and a Transformer encoder in sequence; Stage 2 consists of a patch embedding layer and a Transformer encoder in sequence; Stage 3 consists of a patch embedding layer and a Transformer encoder in sequence; The fourth-order compression branch FSCB includes compression module 1, QFMR-AM, compression module 2, compression module 3, and compression module 4; The compression module 1 includes the first convolution layer and the first GDN in sequence. The convolution kernel size of the first convolution layer is 7×7, the number of input channels is 3, and the number of output channels is N / 4. The compression module 2 includes a second convolutional layer and a second GDN in sequence. The convolution kernel size of the second convolutional layer is 3×3, the number of input channels is N / 4, and the number of output channels is N / 2. The compression module 3 includes a third convolutional layer and a third GDN in sequence. The convolution kernel size of the third convolutional layer is 3×3, the number of input channels is N / 2, and the number of output channels is 3N / 4. The compression module 4 includes a fourth convolutional layer and a fourth GDN in sequence. The convolution kernel size of the fourth convolutional layer is 3×3, the number of input channels is 3N / 4, and the number of output channels is N. The GDN stands for generalized divisive normalization function; N is the number of channels; The fourth-order reconstruction decoder FSRD includes a reconstruction module 1, a QFMR-AM, a reconstruction module 2, a reconstruction module 3, and a reconstruction module 4; The reconstruction module 1 includes the eighth convolution layer and the first IGDN in sequence; the convolution kernel size of the eighth convolution is 3×3, the number of input channels is N / 4, and the number of output channels is 3; The reconstruction module 2 includes the ninth convolution layer and the second IGDN in sequence; the convolution kernel size of the ninth convolution is 3×3, the number of input channels is N / 2, and the number of output channels is N / 4; The reconstruction module 3 includes the tenth convolution layer and the third IGDN in sequence; the convolution kernel size of the tenth convolution is 3×3, the number of input channels is 3N / 4, and the number of output channels is N / 2; The reconstruction module 4 includes the eleventh convolution layer and the fourth IGDN in sequence; the convolution kernel size of the eleventh convolution is 3×3, the number of input channels is N, and the number of output channels is 3N / 4; The IGDN represents the inverse GDN operation; The global context feature guidance module GVGM includes MCEB for extracting multi-channel information, MHSA for extracting global context features, and loss Loss for calculating the difference in global context features between the Compression part and the Reconstruction part. GVF ; The specific working process of the global context feature guidance module GVGM is as follows: InputA inputs MCEB, MCEB outputs feature input MHSA, MHSA outputs feature OutputA; InputB inputs MCEB, MCEB outputs features input MHSA, MHSA outputs features OutputB; Calculate the loss between feature OutputA and feature OutputB GVF ; InputA represents the output feature map of compression module 3, and InputB represents the output feature map of reconstruction module 4; Traditional codecs contain probability models and Q quantizers; The specific working process of the dual-branch remote sensing image compression network GDRNet guided by the global context feature is as follows: The remote sensing image is input into the PVTCB module, and the PVTCB module outputs the feature G; The remote sensing image is sequentially input into compression module 1, QFMR-AM, compression module 2, and compression module 3 in the fourth-order compression branch FSCB, and compression module 3 outputs feature InputA; The compression module 3 outputs the feature InputA which is sequentially input into MCEB and MHSA, and MHSA outputs the feature OutputA; The MHSA output feature OutputA and the compression module 3 output feature InputA are added element by element to obtain the added feature, and the added feature is input into the compression module 4, and the compression module 4 outputs the feature; Add the output feature of the compression module 4 and the output feature G of the PVTCB module element by element to obtain the added feature y; Feature y is input into the probability model Probability Model, and the probability model Probability Model outputs arithmetic encoding AE and arithmetic decoding AD respectively; Feature y is input into the Q quantizer, and the Q quantizer outputs feature feature Input AE arithmetic coding and AD arithmetic decoding in sequence, and AD arithmetic decoding outputs features AD arithmetic decoding output characteristics Input reconstruction module 4, reconstruction module 4 outputs feature InputB; The reconstruction module 4 outputs the feature InputB which is sequentially input into MCEB and MHSA, and MHSA outputs the feature OutputB; Add the MHSA output feature OutputB and the reconstruction module 4 output feature InputB element by element to obtain the added feature K; The added feature K is sequentially input into reconstruction module 3, reconstruction module 2, QFMR-AM, and reconstruction module 1, and reconstruction module 1 outputs the reconstructed image; Calculate the loss between feature OutputA and feature OutputB GVF ; Step 2: Based on the dual-branch remote sensing image compression network GDRNet guided by global context features, the remote sensing image to be tested is first compressed and then reconstructed to obtain the reconstructed remote sensing image to be tested.

2. The dual-branch remote sensing image compression method based on global context feature guidance according to claim 1, characterized in that: The QFMR-AM includes a fifth convolutional layer, a sixth convolutional layer, a seventh convolutional layer, an average pooling layer, a maximum pooling layer, a filter 1, a filter 2, a filter 3, and a filter 4; The working process of the QFMR-AM is: The input feature F is input into the fifth convolutional layer, the sixth convolutional layer, and the seventh convolutional layer respectively. The output feature of the fifth convolutional layer is input into the average pooling layer, and the average pooling layer outputs feature A. The sixth convolutional layer outputs feature B; The output feature of the seventh convolutional layer is input into the maximum pooling layer, and the maximum pooling layer outputs feature C; Perform Hadamard product on feature A and feature B to obtain feature D; Perform Hadamard product on feature B and feature C to obtain feature E; Feature D is input into filter 1, feature B is input into filter 2, feature C is input into filter 3, and feature F is input into filter 4; The output features of filter 1, filter 2, filter 3, and filter 4 are added element by element to obtain the QFMR-AM output features.

3. The dual-branch remote sensing image compression method based on global context feature guidance according to claim 2, characterized in that: The working process of the QFMR-AM is expressed as: X=Conv 3′3 (Input) In the formula, Input represents the input image data block, Conv 3′3 represents a convolution layer with a convolution kernel size of 3, and X represents the output data of filter 2; The working process of the branch of filter 1 is expressed as: In the formula, Conv 1′1 Represents a convolution layer with a convolution kernel size of 1, Avgpool2D 3′3 represents the average pooling layer, ⊙ represents the Hadamard product, Filter1 represents filter 1, Represents the output of the Filter1 branch; The working process of the branch of filter 2 is expressed as: I Filter2 =Filter2(Conv 3′3 (Input)) In the formula, Conv 3′3 represents the convolution layer with a convolution kernel size of 3, Filter2 represents filter 2, and I Filter2 Represents the output of the Filter2 branch; The working process of the branch of filter 3 is expressed as: In the formula, Conv 5′5 Represents a convolution layer with a convolution kernel size of 5, Maxpool2D 3′3 represents the maximum pooling layer, Filter3 represents filter 3, Represents the output of Filter3 branch; The working process of the branch of filter 4 is expressed as: I Filter4 =Filter4(Input) Where Filter4 represents filter 4, I Filter4 Represents the output of the Filter4 branch; The working process of QFMR-AM can be expressed as: In the formula, represents element-by-element addition, I QFMR-AM Represents the output of QFMR-AM.

4. The dual-branch remote sensing image compression method based on global context feature guidance according to claim 3 is characterized in that: The working process of the PVTCB module is expressed as follows: Where B represents the batch number, C1, C2, C3, C4 represent the number of channels, H represents the image height, and W represents the image width; Stage1, Stage2, Stage3 represent stage 1, stage 2, and stage 3 respectively; Filter5 represents filter 5; Input stands for input; PVTCB Represents the output of the PVTCB branch.

5. The dual-branch remote sensing image compression method based on global context feature guidance according to claim 4 is characterized in that: The MCEB includes a twelfth convolutional layer, a thirteenth convolutional layer, and a fourteenth convolutional layer; The convolution kernel size of the twelfth convolutional layer is 1×1, the number of input channels is 3N / 4, and the number of output channels is 3N / 4; The convolution kernel size of the thirteenth convolutional layer is 1×1, the number of input channels is 3N / 4, and the number of output channels is 3N / 4; The convolution kernel size of the fourteenth convolutional layer is 1×1, the number of input channels is 3N / 4, and the number of output channels is 3N / 4; The specific working process of the MCEB is as follows: InputA inputs the twelfth convolutional layer, the thirteenth convolutional layer, and the fourteenth convolutional layer respectively, and adds the output features of the twelfth convolutional layer, the thirteenth convolutional layer, and the fourteenth convolutional layer element by element. The features after element-by-element addition are the output features of MCEB; InputB inputs the twelfth convolutional layer, the thirteenth convolutional layer, and the fourteenth convolutional layer respectively, and adds the output features of the twelfth convolutional layer, the thirteenth convolutional layer, and the fourteenth convolutional layer element by element. The features after element-by-element addition are the output features of MCEB.

6. The dual-branch remote sensing image compression method based on global context feature guidance according to claim 5, characterized in that: The expression of the working process of the MCEB is: I MCEB =Conv 1′1 (Input)+Conv 1′1 (Input)+Conv 1′1 (Input) In the formula, Input represents the input image data block, Conv 1′1 represents a convolution with a kernel size of 1×1, I MCEB Represents the output of MCEB.

7. The dual-branch remote sensing image compression method based on global context feature guidance according to claim 6, characterized in that: The loss between the feature OutputA and feature OutputB is Loss GVF The working process is expressed as: Loss GVF =L MSE (I MHSA(Compression) ,I MHSA(Reconstruction) ) Where, L MSE represents the loss measured using MSE, I MHSA(Compression) Represents the MHSA output of the compression part, I MHSA(Reconstruction) Represents the MHSA output of the reconstruction part, Loss GVF Represents the global context feature loss.

8. The dual-branch remote sensing image compression method based on global context feature guidance according to claim 7, characterized in that: The mapping value of filter 1 is 0.3, the mapping value of filter 2 is 0.7, the mapping value of filter 3 is 0.7, and the mapping value of filter 4 is 0.3; the mapping coefficient of filter 5 is set to 0.

3.

9. The dual-branch remote sensing image compression method based on global context feature guidance according to claim 8, characterized in that: The total loss function of the dual-branch remote sensing image compression network GDRNet guided by global context features is Loss Total : The training process is: argminProposedLoss Total =R+l(D+yLoss GVF ) In the formula, R represents the entropy rate; λ represents the penalty coefficient; D represents the degree of distortion between the original image and the reconstructed image; y represents the weight coefficient of the global context feature loss; ProposedLoss Total Represents the total loss of the dual-branch remote sensing image compression network GDRNet guided by global context features; Obtain the trained dual-branch remote sensing image compression network GDRNet guided by global context features.

10. The dual-branch remote sensing image compression method based on global context feature guidance according to claim 9, characterized in that: In the step 2, the dual-branch remote sensing image compression network GDRNet guided by global context features first compresses the remote sensing image to be tested and then reconstructs it to obtain the reconstructed remote sensing image to be tested; the specific process is: Based on the trained dual-branch remote sensing image compression network GDRNet guided by global context features, the remote sensing image to be tested is first compressed and then reconstructed to obtain the reconstructed remote sensing image to be tested.

Images