A Remote Sensing Image Compression Method Based on Dual-Tree Complex Wavelet Convolution and Frequency Dictionary Entropy Model

Abstract

This invention relates to the field of remote sensing image compression technology, specifically to a remote sensing image compression method based on dual-tree complex wavelet convolution and a frequency dictionary entropy model. The method includes: compressing remote sensing images using a remote sensing image compression model based on dual-tree complex wavelet convolution and a frequency dictionary entropy model; the remote sensing image compression model includes a dual-tree complex wavelet convolution module, a frequency dictionary entropy model module, and a residual feature extraction module; the dual-tree complex wavelet convolution module is used to perform downsampling and upsampling feature processing on the remote sensing image, effectively removing frequency domain redundancy in the latent representation; the frequency dictionary entropy model module is used to perform frequency division on the obtained latent representation of the remote sensing image to establish a probability model; the residual feature extraction module is used to extract features from the remote sensing image, effectively capturing long-range contextual information. This invention enhances the model's image compression effect by introducing dual-tree complex wavelet convolution and a frequency dictionary entropy model, significantly improving the fidelity of remote sensing images at high compression ratios.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image compression technology, specifically to a remote sensing image compression method based on dual-tree complex wavelet convolution and frequency dictionary entropy model. Background Technology

[0002] Remote sensing images typically contain richer and more complex geospatial information than natural images. Specifically, remote sensing images have richer spectral dimensions, more complex ground features, and ground features that are spatially intertwined, with potentially blurred boundaries. They also contain a large amount of high-frequency details and complex textures, making it particularly difficult to efficiently reconstruct high-frequency information at high compression ratios. Furthermore, remote sensing images exhibit numerous and complex correlations in the frequency domain, further increasing the challenge of achieving high fidelity while maintaining a high compression ratio in remote sensing image compression.

[0003] The core challenge of remote sensing image compression technology to reduce frequency domain redundancy lies in how to effectively preserve crucial frequency domain details in geospatial images at high compression ratios, thereby achieving high-fidelity reconstruction. This raises three technical questions: First, remote sensing images have high frequency domain correlation. Compression methods focus on spatial domain feature extraction and fail to make full use of frequency domain information, resulting in a lot of frequency domain redundancy in the potential representation. Second, the features of the latent representation of remote sensing images are highly complex, making it impossible to accurately predict the probability distribution of the latent representation; Third, remote sensing images often contain large-scale features and complex scenes, making it difficult to effectively capture long-distance contextual information, resulting in poor compression performance. Summary of the Invention

[0004] To address the technical problem of poor compression performance of remote sensing images, this invention proposes a remote sensing image compression method based on dual-tree complex wavelet convolution and frequency dictionary entropy model.

[0005] This invention provides a remote sensing image compression method based on dual-tree complex wavelet convolution and frequency dictionary entropy model, the method comprising: A remote sensing image compression network is constructed, comprising a main encoder network, a main decoder network, a prior encoder network, a prior decoder network, and a frequency dictionary entropy model. The main encoder network extracts features from the input remote sensing image to generate a latent representation of the image in a compressed dimension. The main decoder network is used to reconstruct the image and includes a dual-tree complex wavelet convolution module for downsampling and upsampling feature processing, a residual feature extraction module for feature extraction from the remote sensing image, and a standard residual block. The prior encoder network and the prior decoder network are used to capture spatial dependencies in the latent representation. The frequency dictionary entropy model is used to estimate the probability of the latent representation of the remote sensing image appearing in the true distribution. The remote sensing image compression network is trained to obtain a trained remote sensing image compression model; The remote sensing image to be compressed is input into the trained remote image compression model to obtain the compression result; The compression result is input into the trained remote sensing image compression model to obtain the decompression result.

[0006] Optionally, the data processing flow in the dual-tree complex wavelet convolution module includes: The dual-tree complex wavelet convolution module includes a DDTCWConv subnetwork for downsampling in the encoder network and a UDTCWConv subnetwork for upsampling in the decoder network. The first layer of the dual-tree complex wavelet convolution module is a convolutional layer for downsampling, while the first layer of DDTCWConv is a sub-pixel convolutional layer for upsampling, in order to adapt to the dimensionality transformation requirements of the encoder and decoder. After the input feature map is processed in the first layer, it is transformed into the wavelet domain through dual-tree complex wavelet transform. The low-frequency subband coefficients and high-frequency subband coefficients of the output are then subjected to convolution filtering to achieve separation of frequency domain features of remote sensing images. The high-frequency subband coefficients after convolution are decomposed inversely and used together with the low-frequency subband coefficients after convolution to perform inverse dual-tree complex wavelet transform, thereby achieving the conversion from the frequency domain to the spatial domain.

[0007] Optionally, the formula corresponding to the frequency domain coefficient convolution filtering process is: in, These are the image features after the first layer of processing; This is a dual-tree complex wavelet transform; For low-frequency subband coefficients; For high-frequency sub-band coefficients; is the real part of the high-frequency subband coefficient; This represents the imaginary part of the high-frequency subband coefficients; Indicates feature splicing; This represents downsampling convolution.

[0008] Optionally, the data processing flow in the frequency dictionary entropy model includes: Discrete wavelet transform is used to perform frequency division processing on the latent representation, and five non-uniform groups are used to non-uniformly group the low-frequency and high-frequency components from fine to coarse. Five low-frequency slices Consisting of five slice coding networks Encode the data and use the side information generated by the super-prior network. and the output of the previous slice To reduce redundancy between slices, high-frequency slice coding networks The low-frequency component information of the reconstructed image, which contains the overall image structure and main content, was referenced. The value range of i is {0, ..., 4}; For the reconstructed low-frequency component information and high-frequency component information Inverse discrete wavelet transform is used.

[0009] Optionally, slice coding network and high-frequency slice coding network The data processing flow includes: Side information generated by the prior network and the output of the previous slice Splicing along the channel dimension, where, It also includes the reconstructed low-frequency component information. ; The concatenated features are then input into the dictionary attention module; The input and output of the dictionary attention module are concatenated along the channel dimension, and the Gaussian distribution mean prediction network and Gaussian distribution variance prediction network are input to build a probability model for the slice; The slices are quantized, and the quantization results are optimized using the latent residual prediction module.

[0010] Optionally, the data processing flow in the dictionary attention module includes: The input features are processed by multiple NAF modules for basic feature extraction, and then enter the group attention module to refine the features and obtain multi-scale textures in order to query the prior information stored in the dictionary. The features output by the group attention module enter the dictionary attention module.

[0011] Optionally, the data processing flow in the NAF module includes: The process proceeds sequentially through LayerNorm, Conv1x1, DWConv3x3, SimpleGate, SCA, Conv1x1, the first residual join, LayerNorm, Conv1x1, SimpleGate, Conv1x1, and the second residual join.

[0012] Optionally, the data processing flow in the group attention module includes: The features are spliced ​​and fused after being processed by attention branches, and then passed through a multilayer perceptron. The attention branches for the features include point attention, local attention, and mid-range attention.

[0013] Optionally, the data processing flow in the dictionary attention module includes: Features are converted into query tensors via a Linear layer. Call a predefined dictionary Generated by two Linear layers respectively and ,calculate and The similarity is used to obtain the attention weights through softmax, and then compared with... Weighted fusion is used to output dictionary-enhanced features.

[0014] Optionally, the formula corresponding to the data processing flow in the residual feature extraction module is: in, For channel alignment, For depthwise separable convolution, For channel priority layer normalization, For tensor dimension permutation, For pointwise convolution, For GeLU activation function, This is for global response normalization.

[0015] The present invention has the following beneficial effects: First, to address the challenges of feature extraction, blurred high-frequency details, and poor coordination between details and the subject when compressing remote sensing images, this invention proposes a remote sensing image compression method called RSWeLIC, which reduces frequency domain redundancy. By considering and reducing the frequency domain redundancy of potential representations in remote sensing images, it can achieve relatively good visual quality at high compression ratios.

[0016] Second, the present invention designs a DTCWConv module, which improves the sparsity of the wavelet domain by embedding convolution between DTCWT and IDTCWT, and performs filtering and sparsification directly in the frequency domain, thereby obtaining a more compact and less redundant potential representation at the encoding end.

[0017] Third, this invention designs an FDEM module that uses discrete wavelet transform to decompose the latent representation into high- and low-frequency components, and then refines each component using a learnable dictionary. This conditional probability estimation based on frequency components is more accurate than a single model, thereby reducing the bit rate used for encoding.

[0018] Fourth, in order to alleviate the problem of insufficient feature extraction capability of remote sensing images, this invention designs ResConvNeXt to extract more representative and robust features from complex remote sensing images. Integrating ResConvNeXt into the encoding and decoding network improves the model's ability to extract discriminative features from remote sensing images with complex spatial structures, laying the foundation for generating high-quality compressed images.

[0019] Fifth, the method proposed in this invention has been validated on different remote sensing image datasets. Experimental results show that the overall performance of the RSWeLIC model is often superior to many other existing algorithms, achieving better visual quality even with high compression ratios. Attached Figure Description

[0020] To more clearly illustrate the technical solutions and advantages in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0021] Figure 1 This is a flowchart of a remote sensing image compression method based on dual-tree complex wavelet convolution and frequency dictionary entropy model according to the present invention; Figure 2 This is a schematic diagram of the overall structure of RSWeLIC of the present invention; Figure 3 This is a schematic diagram of the overall structure of the DTCWConv of the present invention; Figure 4 This is a schematic diagram of the overall structure of the FDEM of the present invention; Figure 5 For the present invention Overall structural diagram; Figure 6 This is a schematic diagram of the overall structure of ResConvNeXt of the present invention; Figure 7 The figures show the RD curves of different LIC methods of this invention on the ZY303 dataset; Figure 8 The figures show the RD curves of different LIC methods of this invention on the GF7 dataset; Figure 9 The figures show the RD curves of different LIC methods of this invention on the GF1B dataset. Figure 10 This is an example image of the visualization reconstructed based on the ZY303 dataset according to the present invention; Figure 11 This is an example image of the visualization reconstructed based on the GF7 dataset according to the present invention; Figure 12 This is an example image of the visualization reconstructed based on the GF1B dataset according to the present invention; Figure 13 The graph shows the RD performance of the method of the present invention on the ZY303 dataset under different configurations. Detailed Implementation

[0022] To further illustrate the technical means and effects adopted by the present invention to achieve its intended purpose, the specific implementation methods, structures, features, and effects of the technical solution proposed according to the present invention are described in detail below with reference to the accompanying drawings and preferred embodiments. In the following description, different "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics in one or more embodiments can be combined in any suitable form.

[0023] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0024] This invention provides a remote sensing image compression model based on dual-tree complex wavelet convolution and a frequency dictionary entropy model. The remote sensing image compression model includes a dual-tree complex wavelet convolution module (DTCWConv), a frequency dictionary entropy model module (FDEM), and a residual feature extraction module (ResConvNeXt). The DTCWConv module performs downsampling and upsampling feature processing on the remote sensing image, effectively removing frequency domain redundancy in the latent representation. The FDEM module performs frequency division on the obtained latent representation of the remote sensing image to establish a probability model. The ResConvNeXt module extracts features from the remote sensing image, effectively capturing long-range contextual information. This invention enhances the model's image compression capability by introducing dual-tree complex wavelet convolution and a frequency dictionary entropy model, significantly improving the fidelity of remote sensing images at high compression ratios.

[0025] refer to Figure 1 This document illustrates the flowchart of some embodiments of a remote sensing image compression method based on dual-tree complex wavelet convolution and frequency dictionary entropy model according to the present invention. The remote sensing image compression method based on dual-tree complex wavelet convolution and frequency dictionary entropy model includes the following steps: Step S1: Construct a remote sensing image compression network.

[0026] The aforementioned remote sensing image compression network includes a main encoder network, a main decoder network, a prior encoder network, a prior decoder network, and a frequency dictionary entropy model. The main encoder network extracts features from the input remote sensing image to generate a latent representation of the image in a compressed dimension. The main decoder network is used to reconstruct the image and includes a dual-tree complex wavelet convolution module for downsampling and upsampling feature processing, a residual feature extraction module for feature extraction from the remote sensing image, and a standard residual block. The prior encoder network and the prior decoder network are used to capture spatial dependencies in the latent representation. The frequency dictionary entropy model is used to estimate the probability of the latent representation of the remote sensing image appearing in the real distribution.

[0027] Step S2: Train the remote sensing image compression network to obtain a trained remote sensing image compression model.

[0028] Step S3: Input the remote sensing image to be compressed into the trained remote sensing image compression model to obtain the compression result.

[0029] Step S4: Input the compression result into the trained remote sensing image compression model to obtain the decompression result.

[0030] like Figure 2 As shown, to achieve high fidelity while maintaining a high compression ratio in remote sensing image compression, this invention proposes a remote sensing image compression method (RSWeLIC) based on dual-tree complex wavelet convolution and a frequency domain dictionary entropy model. First, a dual-tree complex wavelet convolution (DTCWConv) module is designed, which combines dual-tree complex wavelet transform with convolution, reducing the frequency domain correlation of remote sensing images while retaining more information. Second, a frequency domain dictionary entropy (FDEM) module is designed, which establishes a more accurate latent representation model by utilizing the image information retained by the dictionary. Finally, a feature extraction block (ResConvNeXt Block) is designed, implemented using residual connections and positionally upregulated large-kernel depthwise convolution, and effectively collaborates with RSWeLIC.

[0031] like Figure 3 As shown, in remote sensing image compression tasks, the large and complex correlations in the frequency domain severely interfere with the compression and reconstruction of remote sensing images. To address the problem that existing networks fail to effectively remove frequency domain redundancy from latent representations, a dual-tree complex wavelet convolution module (DTCWConv) is proposed. This module introduces dual-tree complex wavelet transform, decomposing the image more clearly and stably into sub-bands of different directions and scales, providing a frequency domain representation with less information loss and higher separation for subsequent processing. Secondly, by embedding convolution between the dual-tree complex wavelet transform and its inverse transform, the module adaptively and selectively filters out redundant information in the frequency domain while enhancing useful sparse components, effectively reducing the frequency domain redundancy of latent representations.

[0032] The DTCWConv module includes a DDTCWConv subnetwork in the encoder network for downsampling and a UDTCWConv subnetwork in the decoder network for upsampling. For example... Figure 3 As shown in (a) and (b), UDTCWConv and DDTCWConv have similar structures. Specifically, the first layer of UDTCWConv is a convolutional layer for downsampling, while the first layer of DDTCWConv is a sub-pixel convolutional layer for upsampling. After the input feature map is processed by the first layer, it is transformed into the wavelet domain through dual-tree complex wavelet transform. Subsequently, the low-frequency subband coefficients of the output are... Filtering is performed using convolution, targeting the six highly uncorrelated and already very sparse high-frequency subband coefficients. Then the real part of each high-frequency subband is used. and the virtual part A 1×1 convolutional filter is then applied to leverage its sparsity and reduce computational complexity. Afterward, the high-frequency subband coefficients from the convolutional processing are inversely decomposed and used with the low-frequency subband coefficients from the convolutional processing via inverse dual-tree complex wavelet transform to achieve a transformation from the frequency domain to the spatial domain, allowing the latent representation of the feature map to be processed by subsequent convolutional layers. The formula for this stage is shown below: in, These are the image features after the first layer of processing; This is a dual-tree complex wavelet transform; For low-frequency subband coefficients; For high-frequency sub-band coefficients; is the real part of the high-frequency subband coefficient; This represents the imaginary part of the high-frequency subband coefficients; Indicates input features, Indicates feature splicing, Indicates feature separation, Indicates downsampling convolution. This represents transposed convolution. This represents the inverse dual-tree complex wavelet transform. This indicates the output features.

[0033] In remote sensing image compression tasks, the core function of the entropy model is to accurately capture the data distribution characteristics of the latent representation of remote sensing images, and it is the core module for achieving efficient encoding and maximizing the compression ratio. Entropy models generally use a unified, complex model to predict the probability distribution of the entire latent representation. However, this may still be insufficiently precise for remote sensing images with highly complex features. Therefore, this invention focuses on how to fully predict the probabilistic model of the latent representation of remote sensing images. For example... Figure 4 As shown, a Frequency Dictionary Entropy Model (FDEM) was designed as the entropy model for RSWeLIC. FDEM mainly uses frequency bands and introduces a "dictionary" prior for refined modeling, which can more accurately estimate the probability of the latent representation of remote sensing images appearing in the real distribution.

[0034] The low-frequency and high-frequency components of the latent representation of remote sensing images exhibit drastically different statistical characteristics and significant frequency domain redundancy. Using a single model to fit all features is insufficiently accurate; therefore, frequency-band modeling can be employed. First, discrete wavelet transform is used to perform frequency-band processing on the latent representation. FDEM encodes the low-frequency and high-frequency components using five non-uniform groups. Taking the low-frequency component as an example, this embodiment employs a non-uniform grouping structure from fine to coarse, dividing the 320-channel feature map along the channel dimension into five progressively scaled blocks, corresponding to 16, 16, 32, 64, and 192 channels respectively. This finer channel grouping scheme further improves the performance of the entropy estimation module and reduces runtime by rebalancing the number of channels in different groups. The formula for this stage is shown below: in, Representing the input features (latent representation), Represents discrete wavelet transform. Indicates quantification, This indicates feature separation along the channel dimension.

[0035] Five low-frequency slices Consisting of five slice coding networks (i=0,...,4) are encoded, and side information generated by the super-prior network is used. (Assuming it follows a single Gaussian distribution) and the output of the previous slice To reduce redundancy between slices, using only the output of the previous slice as prior information can save runtime and does not affect performance, compared to using the output of all previous slices as prior information. Figure 5 Demonstrates slice coding network In detail, when building a latent representation model for high-frequency components, the network also references low-frequency component information that contains the overall image structure and main content after reconstruction. This can enhance the harmony between model details and the main subject, reducing visual artifacts. (Regarding the reconstructed low-frequency component information...) and high-frequency component information Inverse discrete wavelet transform is employed. This is guided by more stable low-frequency information. It can also better distinguish between real details and noise, enhancing reconstruction robustness. This invention uses lightweight group attention (LGA) that preserves multi-scale understanding and optimizes computational efficiency, including PA point attention, LA local attention, and MRA mid-range attention. This invention defines two learnable parameters as dictionaries for the low-frequency entropy model and the high-frequency entropy model, respectively. These dictionaries preserve typical textures during training, allowing for a more accurate construction of the latent representation model by querying the dictionaries that retain complete information. Taking the low-frequency component as an example, the formula for this stage is as follows: in, The mean and variance of the prior network predictions are given. A dictionary of learnable parameters for low-frequency components. This is a dictionary-based cross-attention module. It is a Gaussian mean prediction network. For Gaussian variance prediction networks, To enable direct quantitative operations, For potential residual prediction networks, Let be the low-frequency reconstruction features of the i-th slice.

[0036] In summary, the data processing flow in the frequency dictionary entropy model can include: performing frequency division processing on the latent representation using discrete wavelet transform, and dividing the low-frequency and high-frequency components into five non-uniform groups from fine to coarse; five low-frequency slices Consisting of five slice coding networks Encode the data and use the side information generated by the super-prior network. and the output of the previous slice To reduce redundancy between slices, high-frequency slice coding networks The low-frequency component information of the reconstructed image, which contains the overall image structure and main content, was referenced. The value of i ranges from {0, ..., 4}; for the reconstructed low-frequency component information and high-frequency component information Inverse discrete wavelet transform is used.

[0037] Slice coding network and high-frequency slice coding network The data processing flow may include: processing the side information generated by the prior network. and the output of the previous slice Splicing along the channel dimension, where, It also includes the reconstructed low-frequency component information. The concatenated features are input into the dictionary attention module; the input and output of the dictionary attention module are concatenated along the channel dimension and input into the Gaussian mean prediction network and the Gaussian variance prediction network to build a probability model for the slice; the slice is quantized and the quantization result is optimized using the latent residual prediction module.

[0038] The data processing flow in the dictionary attention module can include: input features undergoing basic feature extraction through multiple NAF modules, followed by feature refinement in the group attention module to obtain multi-scale textures for querying prior information stored in the dictionary; the features output by the group attention module then enter the dictionary attention module. The data processing flow in the NAF module can include: sequentially passing through LayerNorm (layer normalization), Conv1x1 (1×1 convolution), DWConv3x3 (depthseparable 3×3 convolution), SimpleGate (simple gate), SCA (spatial attention), Conv1x1 (1×1 convolution), the first residual connection, LayerNorm, Conv1x1, SimpleGate, Conv1x1, and the second residual connection. The data processing flow in the group attention module can include: features processed by attention branches, then concatenated and fused, and passed through a multilayer perceptron, where the attention branches include point attention, local attention, and mid-range attention. The data processing flow in the dictionary attention module can include: features being converted into query tensors via a linear layer. ( ), call the predefined dictionary ( Generate by two Linear layers respectively ( )and ( ),calculate and The similarity is used to obtain the attention weights through softmax, and then compared with... Weighted fusion is used to output dictionary-enhanced features.

[0039] Standard residual blocks typically use stacked 3x3 convolutions, which have limited receptive fields. When processing large-scale features (such as farmland and large building complexes) and complex scenes in remote sensing images, they may not effectively capture long-range contextual information. To improve the ability to extract features from remote sensing images, a more modern and powerful residual feature extraction block (ResConvNeXtBlock) was designed to address the unique challenges of remote sensing images. First, a 7x7 upshifted depthwise convolution is used as the core operation to expand the receptive field of each neuron, allowing the network to integrate a wider range of contextual information early on, which helps in understanding spatially widespread and structurally complex remote sensing scenes. This invention also introduces techniques such as Global Response Normalization (GRN) to adaptively suppress redundant channels and enhance the responses of important channels, thereby enhancing the discriminative power of features. Finally, carefully designed residual connections ensure the effective propagation of gradients in the deep network and promote information interaction between different channel groups, enabling the network to be trained deeper and more stable. Detailed structure is shown below. Figure 6 As shown. The calculation process is as follows: in, For channel alignment, For depthwise separable convolution, For channel priority layer normalization, For tensor dimension permutation, For pointwise convolution, For GeLU activation function, This is for global response normalization.

[0040] In summary, a remote sensing image compression model based on dual-tree complex wavelet convolution and a frequency dictionary entropy model is proposed. This model includes a dual-tree complex wavelet convolution module (DTCWConv), a frequency dictionary entropy model module (FDEM), and a residual feature extraction module (ResConvNeXt). The DTCWConv module performs downsampling and upsampling feature processing on the remote sensing image, effectively removing frequency domain redundancy in the latent representation. The FDEM module performs frequency division on the obtained latent representation of the remote sensing image to establish a probability model. The ResConvNeXt module extracts features from the remote sensing image, effectively capturing long-range contextual information. This invention enhances the model's image compression capability by introducing dual-tree complex wavelet convolution and a frequency dictionary entropy model, significantly improving the fidelity of remote sensing images at high compression ratios.

[0041] To verify the effectiveness of the present invention, the following experiment was conducted: The experimental environment used in this embodiment of the invention is as follows: the training phase was accelerated by an NVIDIA® RTX 4090 GPU, and the validation phase was accelerated by an NVIDIA® RTX 4080 Super GPU. Furthermore, the model training used the Python deep learning framework, with PyTorch versions 2.1.0 and Python version 3.8, and CUDA version 11.8. During training, randomly cropped image segments were 256×256 pixels in size, the Adam optimizer was used, the initial learning rate was 0.0001, and the model was trained in approximately 300 batches.

[0042] Extensive experiments were conducted on three datasets from the ZY303, GF7, and GF1B satellites. Each dataset contains 640x640 images and includes three visible light RGB bands and one near-infrared band. Specifically, the ZY303 dataset, from China's ZY303 remote sensing satellite, contains 6250 multispectral images with a spatial resolution of 5.8 meters, of which 5000 were used for training, 625 for testing, and 625 for validation. The GF7 dataset, from China's Gaofen-7 remote sensing satellite, contains 2400 multispectral images with a spatial resolution of 2.6 meters, of which 2000 were used for training, 200 for testing, and 200 for validation. The GF1B dataset comes from China's Gaofen-02 remote sensing satellite and contains 2,400 multispectral images with a spatial resolution of 8 meters. Of these, 2,000 images were used for training, 200 for testing, and 200 for validation.

[0043] Experiments were conducted on the ZY303 dataset and compared with other image compression methods. The experimental results are as follows: Figure 7 As shown, RSWeLIC achieved a significant advantage in compression metrics, exhibiting superior performance across all bpp levels. Particularly at high PSNR levels, compared to Fu2024, RSWeLIC showed a 0.2% improvement in PSNR and a significant 18.58% reduction in bpp at approximately 42 dB. When bpp exceeded 0.3, its Bjøntegaard Delta ratio (BD ratio) decreased by approximately 5.49%.

[0044] Experiments were conducted on the GF7 dataset and compared with other image compression methods. The experimental results are as follows: Figure 8As shown. In terms of PSNR, the RSWeLIC method still achieves the best compression performance in the range of BPP less than 0.4 and greater than 0.8, while in the other range it performs comparably to the Fu2024 method. Overall, compared with Fu2024, RSWeLIC reduces the BD-rate by approximately 4.12%.

[0045] Various compression models were also validated on GF1Bset, and the experimental results are as follows: Figure 9 As shown, among all compression models, our proposed RSWeLIC method further expands its performance advantage on high-resolution remote sensing images, outperforming other methods in almost all intervals and comparable to the best method in a few intervals. Compared to the Fu2024 model, which has the best overall performance among other LIC methods, RSWeLIC reduces the BD rate by approximately 7.35% when bpp is greater than 0.35.

[0046] To more intuitively demonstrate the advantages of RSWeLIC compared to other models, this study conducted a visual comparison experiment, comparing RSWeLIC with several state-of-the-art models. The experimental results are as follows: Figure 10 , Figure 11 , Figure 12 As shown. The reconstructed remote sensing image contains three visible light bands and one infrared band; the first three visible light bands are selected for display. The regions of interest (ROIs) are marked with red boxes.

[0047] Within the regions of interest (ROIs) marked with red boxes, Fu2024's method failed to accurately reconstruct the gaps and colors between objects, while Jiang2023's method did not effectively distinguish between realistic details and noise under similar bpp conditions as RSWeLIC. Compared to Xiang2024, RSWeLIC improved PSNR by approximately 2.66 dB with a similar bpp. Among all evaluated methods, the method of this invention achieved the highest PSNR and MSSSIM scores after reconstruction, and it correctly recovered the gaps between objects, achieving a high compression ratio while maintaining high fidelity.

[0048] Similarly, Figure 11 The visual contrast results of reconstructed images on the GF7 dataset are presented. Compared to the He2022 method, this method only increases BPP by 0.2%, but improves PSNR by approximately 0.48 dB, effectively restoring the fine contours and color details of the orange roofs within the ROI. Overall, the method of this invention more effectively preserves contour and color information.

[0049] Figure 12 This section presents a visual comparison of reconstructed images using different methods on the GF1B dataset. For example... Figure 12As shown in the ROI region, while Jiang2023 recovers the overall contour features of the image, it still loses a large amount of high-frequency information, resulting in severe detail smearing. The method of this invention achieves higher accuracy in recovering the detailed contours of the image, retaining a large amount of detail and reducing noise artifacts. Compared with Fu2024, RSWeLIC achieves higher PSNR and MS-SSIM while reducing bpp by approximately 10.8%.

[0050] To explore the effectiveness of different modules and the impact of their combinations on model performance, this invention uses the Balle2018 method based on the original architecture of the proposed model. It employs ResBlock to replace part of the encoder-decoder network, the activation function in the super-prior network, and a single DEM for entropy encoding to design a baseline model. In each case, the entire system is typically retrained to obtain optimal performance. Figure 13 The RD performance curves for Ballé2018, the base model, the base model + FDEM, the base model + FDEM + DTCWTConv, and the base model + DTCWTConv + FDEM + ResConvNeXt (RSWeLIC) are listed (where "+" indicates that the original components are replaced with the components proposed in this paper). Compared with Ballé2018, the PSNR is improved by approximately 0.66 dB across all bpp levels in the BASE model. Replacing a single DEM with the frequency dictionary entropy model FDEM reduces the BD rate and improves the PSNR by approximately 1.68 dB across the entire range. Replacing some convolutional layers with DTCWTConv for sampling further improves the PSNR to 1.95 dB. If ResConvNeXt is enabled to replace some ResBlocks, it can improve the PSNR by another 0.14 dB, achieving compression performance superior to other LIC methods across the entire range, especially at high bit rates, where the PSNR improvement reaches 0.31 dB when the bpp is greater than 0.3.

[0051] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention, and should all be included within the protection scope of the present invention.

Claims

1. A remote sensing image compression method based on dual-tree complex wavelet convolution and frequency dictionary entropy model, characterized in that, Includes the following steps: A remote sensing image compression network is constructed, comprising a main encoder network, a main decoder network, a prior encoder network, a prior decoder network, and a frequency dictionary entropy model. The main encoder network extracts features from the input remote sensing image to generate a latent representation of the image in a compressed dimension. The main decoder network is used to reconstruct the image and includes a dual-tree complex wavelet convolution module for downsampling and upsampling feature processing, a residual feature extraction module for feature extraction from the remote sensing image, and a standard residual block. The prior encoder network and the prior decoder network are used to capture spatial dependencies in the latent representation. The frequency dictionary entropy model is used to estimate the probability of the latent representation of the remote sensing image appearing in the true distribution. The remote sensing image compression network is trained to obtain a trained remote sensing image compression model; The remote sensing image to be compressed is input into the trained remote image compression model to obtain the compression result; The compression result is input into the trained remote sensing image compression model to obtain the decompression result.

2. The remote sensing image compression method based on dual-tree complex wavelet convolution and frequency dictionary entropy model according to claim 1, characterized in that, The data processing flow in the dual-tree complex wavelet convolution module includes: The dual-tree complex wavelet convolution module includes a DDTCWConv subnetwork for downsampling in the encoder network and a UDTCWConv subnetwork for upsampling in the decoder network. The first layer of the dual-tree complex wavelet convolution module is a convolutional layer for downsampling, while the first layer of DDTCWConv is a sub-pixel convolutional layer for upsampling, in order to adapt to the dimensionality transformation requirements of the encoder and decoder. After the input feature map is processed in the first layer, it is transformed into the wavelet domain through dual-tree complex wavelet transform. The low-frequency subband coefficients and high-frequency subband coefficients of the output are then subjected to convolution filtering to achieve separation of frequency domain features of remote sensing images. The high-frequency subband coefficients after convolution are decomposed inversely and used together with the low-frequency subband coefficients after convolution to perform inverse dual-tree complex wavelet transform, thereby achieving the conversion from the frequency domain to the spatial domain.

3. The remote sensing image compression method based on dual-tree complex wavelet convolution and frequency dictionary entropy model according to claim 2, characterized in that, The formula corresponding to the frequency domain coefficient convolution filtering process is: in, These are the image features after the first layer of processing; This is a dual-tree complex wavelet transform; For low-frequency subband coefficients; For high-frequency sub-band coefficients; is the real part of the high-frequency subband coefficient; This represents the imaginary part of the high-frequency subband coefficients; Indicates feature splicing; This represents downsampling convolution.

4. The remote sensing image compression method based on dual-tree complex wavelet convolution and frequency dictionary entropy model according to claim 1, characterized in that, The data processing flow in the frequency dictionary entropy model includes: Discrete wavelet transform is used to perform frequency division processing on the latent representation, and five non-uniform groups are used to non-uniformly group the low-frequency and high-frequency components from fine to coarse. Five low-frequency slices Consisting of five slice coding networks Encode the data and use the side information generated by the super-prior network. and the output of the previous slice To reduce redundancy between slices, high-frequency slice coding networks The low-frequency component information of the reconstructed image, which contains the overall image structure and main content, was referenced. The value range of i is {0, ..., 4}; For the reconstructed low-frequency component information and high-frequency component information Inverse discrete wavelet transform is used.

5. The remote sensing image compression method based on dual-tree complex wavelet convolution and frequency dictionary entropy model according to claim 4, characterized in that, Slice coding network and high-frequency slice coding network The data processing flow includes: Side information generated by the prior network and the output of the previous slice Splicing along the channel dimension, where, It also includes the reconstructed low-frequency component information. ; The concatenated features are then input into the dictionary attention module; The input and output of the dictionary attention module are concatenated along the channel dimension, and the Gaussian distribution mean prediction network and Gaussian distribution variance prediction network are input to build a probability model for the slice; The slices are quantized, and the quantization results are optimized using the latent residual prediction module.

6. The remote sensing image compression method based on dual-tree complex wavelet convolution and frequency dictionary entropy model according to claim 5, characterized in that, The data processing flow in the dictionary attention module includes: The input features are processed by multiple NAF modules for basic feature extraction, and then enter the group attention module to refine the features and obtain multi-scale textures in order to query the prior information stored in the dictionary. The features output by the group attention module enter the dictionary attention module.

7. The remote sensing image compression method based on dual-tree complex wavelet convolution and frequency dictionary entropy model according to claim 6, characterized in that, The data processing flow in the NAF module includes: The process proceeds sequentially through LayerNorm, Conv1x1, DWConv3x3, SimpleGate, SCA, Conv1x1, the first residual join, LayerNorm, Conv1x1, SimpleGate, Conv1x1, and the second residual join.

8. The remote sensing image compression method based on dual-tree complex wavelet convolution and frequency dictionary entropy model according to claim 6, characterized in that, The data processing flow in the group attention module includes: The features are spliced ​​and fused after being processed by attention branches, and then passed through a multilayer perceptron. The attention branches for the features include point attention, local attention, and mid-range attention.

9. A remote sensing image compression method based on dual-tree complex wavelet convolution and frequency dictionary entropy model according to claim 6, characterized in that, The data processing flow in the dictionary attention module includes: Features are converted into query tensors via a Linear layer. Call a predefined dictionary Generated by two Linear layers respectively and ,calculate and The similarity is used to obtain the attention weights through softmax, and then compared with... Weighted fusion is used to output dictionary-enhanced features.

10. The remote sensing image compression method based on dual-tree complex wavelet convolution and frequency dictionary entropy model according to claim 1, characterized in that, The formula corresponding to the data processing flow in the residual feature extraction module is: in, For channel alignment, For depthwise separable convolution, For channel priority layer normalization, For tensor dimension permutation, For pointwise convolution, For GeLU activation function, This is for global response normalization.

Images