A Remote Sensing Image Compression Method and Apparatus Based on Frequency Domain Enhancement and Adaptive Optimization

Abstract

This invention discloses a remote sensing image compression method and apparatus based on frequency domain enhancement and adaptive optimization. The method includes: compressing and decompressing a remote sensing image using a pre-trained first model deployed simultaneously at the compression and decompression ends; the first model includes an encoder, an entropy model, and a decoder; both the encoder and decoder perform frequency domain analysis on the input through a selective state space module with frequency domain enhancement, and model long-distance dependencies and frequency domain structure based on the state space mechanism; the first model is trained according to a preset training set and a loss function to obtain a pre-trained first model; compression includes: extracting the main features of the input image through the encoder; probabilistically modeling the distribution information of the main features through the entropy model to generate a bitstream, and sending the bitstream to the decompression end; decompression includes: receiving the bitstream at the decompression end; restoring the main features of the bitstream through the entropy model; and inputting the main features into the decoder to obtain a reconstructed image.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a remote sensing image compression method and apparatus based on frequency domain enhancement and adaptive optimization. Background Technology

[0002] Remote sensing imaging, by observing ground features across multiple narrow bands, enables precise identification of material composition, surface structure, and even underlying hidden features. Compared to traditional RGB images, remote sensing imaging contains richer and more continuous spectral information, and has wide applications in remote sensing detection, agricultural monitoring, geological exploration, refined urban management, and biomedical imaging. However, the massive data volume generated by remote sensing images also poses significant challenges to their storage, transmission, and processing efficiency.

[0003] How to effectively compress remote sensing images while ensuring image reconstruction quality has become an important issue of widespread concern in academia and industry. Current research on remote sensing image compression can be divided into two main categories: traditional compression methods and deep learning-based compression methods. Traditional compression methods mainly rely on hand-designed feature extractors and combine prediction or transformation strategies to compress redundant information in the image. For example, prediction-based coding methods infer spectral dimensions by constructing a correlation model between pixels, a representative example being the previously released hyperspectral compression standard CCSDS 123.0-B-2 (…). Lossless Multispectral&Hyperspectral Image CompressionIssue 2, November 2019, Consultative Committee for Space Data Systems (CCSDS). This approach offers advantages such as simple implementation, low algorithm complexity, and ease of hardware deployment. Traditional transform-based compression methods (such as discrete cosine transform and discrete wavelet transform) map the original image to the transform domain to remove spectral and spatial redundancy, as exemplified by JPEG2000. Although traditional compression methods are widely used in engineering practice, their manually designed feature extraction mechanisms lack data-driven adaptive capabilities, often making it difficult to guarantee spectral fidelity of the reconstructed image at high compression ratios, thus limiting compression performance. With the development of deep learning, image compression methods based on convolutional neural networks have matured and are gradually being extended to remote sensing image compression tasks. A typical deep learning compression framework employs an encoder-entropy model-decoder structure, where the encoder extracts high-dimensional image features, the entropy model probabilistically models the features and performs entropy encoding, and the decoder is used for image reconstruction. These methods can achieve end-to-end training and have strong representational capabilities in image feature extraction and distribution learning, becoming the mainstream research direction for remote sensing image compression in recent years. However, most existing models use two-dimensional convolutional neural networks to process spatial information, ignoring the frequency domain structure of the image. In particular, they lack the ability to model high-frequency details and texture edges, resulting in blurry or lost details in the reconstructed image. More importantly, most models treat the context modeling loss (such as causal context adjustment loss) as a static term, and cannot dynamically adjust its optimization degree according to the importance of the image content, which limits the performance of the compression model on different image structures.

[0004] Therefore, a new technical solution is urgently needed to address the technical problem of effectively compressing remote sensing images while ensuring image reconstruction quality. Summary of the Invention

[0005] This invention provides a remote sensing image compression method and apparatus based on frequency domain enhancement and adaptive optimization, which solves the technical problem of how to effectively compress remote sensing images while ensuring image reconstruction quality.

[0006] To achieve the above objectives, this invention provides a remote sensing image compression method based on frequency domain enhancement and adaptive optimization. This method compresses and decompresses remote sensing images using a pre-trained first model simultaneously deployed at both the compression and decompression ends, comprising:

[0007] The first model includes an encoder, an entropy model, and a decoder; both the encoder and decoder perform frequency domain analysis on the input through a frequency-domain enhanced selective state-space module, and model long-distance dependencies and frequency domain structures based on the state-space mechanism.

[0008] The first model is trained using a pre-set training set and a loss function that considers pixel distortion, structure preservation, frequency domain error, perceptual quality, bit rate constraint, context autoregressive regularization term, and auxiliary supervision loss to obtain a pre-trained first model.

[0009] Compression includes: extracting the main features of the input image through an encoder; probabilistically modeling the distribution information of the main features using an entropy model to generate a bitstream; and sending the bitstream to the decompression end.

[0010] Decompression includes: receiving the bitstream at the decompression end; restoring the bitstream to its main features using an entropy model; and inputting the main features into the decoder to obtain the reconstructed image.

[0011] Preferably, both the encoder and decoder perform frequency domain analysis on the input through a frequency-domain enhanced selective state-space module, and model long-distance dependencies and frequency domain structures based on state-space mechanisms, including:

[0012] The encoder comprises three sequentially connected first modules; each first module comprises a 5×5 convolution, three bottleneck residual modules, and four frequency-domain enhanced selective state-space modules connected in sequence.

[0013] The decoder consists of three sets of second modules connected in sequence; each second module consists of a 5×5 convolution, four frequency-domain enhanced selective state space modules, and three bottleneck residual modules connected in sequence.

[0014] In the encoder, a frequency-domain enhanced selective state-space module is used to enhance the high-frequency details and global contextual representation of input features through frequency domain analysis and remote dependency modeling.

[0015] In the decoder, a frequency-domain enhanced selective state-space module is used to enhance high-frequency details and spatial consistency in the decoded features through frequency-domain attention and long-range dependency modeling.

[0016] Preferably, the frequency-domain enhanced selective state-space module includes:

[0017] When input features After inputting a frequency-domain enhanced selective state-space module, the initial features are obtained through a 1×1 convolution. Initial features The frequency domain importance weights are obtained from the input frequency domain analysis submodule. Initial features Frequency domain importance weights Inputting a selective state space submodule and performing state updates yields context-enhanced features that fuse frequency domain information and state-aware information. ; Enhance features based on context Frequency domain importance weights and initial features Obtain features ;feature Feature extraction is performed by inputting a free block of nonlinear activation function to obtain features. ;feature The output features are obtained after a 1×1 convolution. .

[0018] Preferably, the frequency domain analysis submodule includes:

[0019] Initial features After inputting into the frequency domain analysis submodule, the high-frequency, mid-frequency, low-frequency, and DC components are extracted through four parallel frequency domain channels. The high-frequency component is extracted using a 3×3 grouped convolution, while the mid-frequency, low-frequency, and DC components are each extracted using a 1×1 convolution. Intermediate frequency components Low-frequency components and DC component ; to high frequency components Intermediate frequency components Low-frequency components and DC component The fused frequency domain features are obtained by channel splicing. ; frequency domain characteristics of fusion The frequency domain importance evaluation network then passes through average pooling, 1×1 convolution, ReLU activation function layer, 1×1 convolution, and Sigmoid function layer in sequence to obtain frequency domain importance weights consistent with the original channels. .

[0020] Preferably, the selective state space submodule includes:

[0021] Initial features Frequency domain importance weights After inputting the selective state-space submodule, the frequency domain importance weights are used. For initial features Channel-by-channel weighting is performed to obtain the frequency domain weighted feature map. ; Weighted feature map in the frequency domain The global feature vector is obtained through average pooling. Global feature vectors The three-layer perceptron is input to obtain the selectivity coefficients used to control the state update process. and The three-layer perceptron includes a first linear transformation layer connected in sequence. ReLU activation function layer and second linear transformation layer Global feature vectors The state-space representation is obtained by adjusting the channel dimension through linear mapping. According to the selectivity coefficient and State space representation using lightweight gating mechanisms Perform selective updates, selection coefficients Selectivity coefficient is used to control the proportion of the original state retained. The nonlinear perturbation amplitude used to regulate the state, and the selectivity coefficient. These three elements—namely, the global mean information and the frequency domain—work together to achieve dynamic state adjustment and fusion guided by the frequency domain, resulting in an adaptively updated state representation. Representing the state By restoring the original channel dimension through linear mapping, context-enhanced features that fuse frequency domain information and state-aware information are obtained. .

[0022] Preferably, the main features of the input image are extracted by an encoder; the distribution information of the main features is probabilistically modeled using an entropy model to generate a bitstream, including:

[0023] The entropy model includes a super-prior encoder, a super-prior decoder, a context modeling module, a first-stage modeling module, and a second-stage modeling module. The encoder extracts the main features of the input image; the super-prior encoder models the spatial and channel statistics of the main features to obtain the first super-prior feature; the first super-prior feature is quantized to obtain the second super-prior feature; and the super-prior decoder obtains the first probability modeling parameters based on the second super-prior feature.

[0024] After quantizing the main features, the non-uniform channels are grouped into two sub-features based on a preset ratio, and the first probability modeling parameters are also divided into two sub-parameters. Sub-features and sub-parameters with the same channel size are paired and randomly assigned as the first data and the second data.

[0025] The first probability distribution is obtained by modeling based on the sub-parameters in the first data through the first-stage modeling module; the second probability modeling parameters are obtained by extracting spatial and channel-dimensional context information based on the sub-features in the first data through the context modeling module; and the second probability distribution is obtained by modeling based on the second probability modeling parameters and the sub-parameters in the second data through the second-stage modeling module.

[0026] The first bit stream is generated by entropy encoding of the sub-features in the first data according to the first probability distribution.

[0027] The sub-features in the second data are entropy encoded according to the second probability distribution to generate the second bit stream.

[0028] The second hyperprior feature is modeled using a Gaussian distribution model to obtain the third probability distribution; the second hyperprior feature is then entropy-encoded based on the third probability distribution to generate the third bitstream.

[0029] Preferably, the bitstream is restored to its main features using an entropy model; the main features are then input into the decoder to obtain the reconstructed image, which includes:

[0030] The decompression end uses the third bitstream to perform probability modeling parameters to obtain the third probability distribution.

[0031] The second hyperprior feature is obtained by entropy decoding of the third bitstream based on the third probability distribution.

[0032] The first probability modeling parameter is obtained based on the second super-prior feature by the super-prior decoder; the first probability modeling parameter is divided into two sub-parameters based on the non-uniform channel grouping of the preset ratio, and matched with the channel size of the first data and the second data to obtain the first sub-parameter and the second sub-parameter respectively.

[0033] The first probability distribution is obtained by modeling based on the first sub-parameter through the first-stage modeling module.

[0034] Entropy decoding is performed on the first bit stream according to the first probability distribution to obtain the sub-features in the first data.

[0035] The context modeling module extracts spatial and channel-dimensional context information from the sub-features in the first data to obtain the second probability modeling parameters.

[0036] The second probability distribution is obtained by modeling based on the second probability modeling parameters and the second sub-parameters using the two-stage modeling module.

[0037] Entropy decoding is performed on the second bitstream based on the second probability distribution to obtain the sub-features in the second data.

[0038] The sub-features in the first data and the sub-features in the second data are concatenated to obtain the quantized main features.

[0039] The reconstructed image is obtained by restoring the quantized main features using a decoder.

[0040] Preferably, the loss function includes:

[0041] ;

[0042] in, This represents the distortion term, used to measure the reconstruction error between the compressed image and the original image; Indicates bit rate loss; This represents the causal context-adjusted loss with frequency domain awareness and adaptive loss weighting, which guides the model to adaptively adjust the focus of the loss function according to the frequency domain complexity of different images. This indicates the auxiliary monitoring loss item.

[0043] Distortion item include:

[0044] ;

[0045] in, This represents the squared error loss, used to optimize pixel-level reconstruction accuracy. Total number of pixels; This represents the reconstructed image output by the model; Indicates the input image; This represents perceptual loss, which enhances the perception of subjective quality by comparing the distances of images in the deep feature space. For the first in the pre-trained network Feature maps extracted from layers; This represents edge loss, used to enhance the preservation of image edges and structural contours. Represents the gradient of an image; This represents the frequency domain loss, and the image restoration performance in the high, mid, and low frequency regions is evaluated through spectral analysis. , , Indicates the spectral amplitude. Indicates Fourier transform; This represents structural similarity loss, used to model the structural information and local consistency of an image. Represents a multi-scale structural similarity index; , , and These are the learnable weight coefficients for perceptual loss, edge loss, frequency domain loss, and structural similarity loss, respectively.

[0046] Bit rate loss include:

[0047] ;

[0048] This indicates the second prior feature in the compression phase. The bit rate loss term; during model training, based on the first probability modeling parameters of the compression phase. Constructing the second super-prior features in the compression stage The probability distribution is calculated, and the negative log-likelihood value at each feature location is estimated and summed to obtain the result. .

[0049] , representing the sub-features in the first data of the compression stage. The bit rate loss term; during model training, based on the sub-parameters in the first data of the compression phase. Sub-features in the first data of the compression stage The probability distribution is calculated, and the negative log-likelihood value at each feature location is estimated and summed to obtain the result. .

[0050] , representing the sub-features in the second data of the compression stage. The bit rate loss term; during model training, based on the sub-parameters in the second data of the compression phase. The fusion results of the second probability modeling parameters are used to construct sub-features in the second data of the compression stage. The probability distribution is calculated, and the negative log-likelihood value at each feature location is estimated and summed to obtain the result. .

[0051] Auxiliary monitoring of loss items include:

[0052] During model training, based on the second super-prior features of the compression phase... Predicting sub-features in the second data of the compression stage probability distribution And calculate the negative log-likelihood of the features, and obtain .

[0053] Preferably, the causal context-adjusted loss with frequency-domain-aware adaptive loss weighting includes:

[0054] Obtain the first loss ,include:

[0055] ;

[0056] in, Represents information entropy; and All are true distributions.

[0057] , indicating that only given Under the condition of, calculate In probability distribution The expected value of the negative log probability under context-free prediction is the cross-entropy.

[0058] In the given and from context modeling Under the condition of, calculate Second probability distribution The expected value of the negative logarithmic probability under contextual prediction is the cross-entropy.

[0059] After grayscale processing of the input image, a two-dimensional fast Fourier transform is performed to obtain the image amplitude. ;Calculation based on a preset high-frequency mask The left and right quarter regions of the image are also marked as high-frequency regions, including:

[0060] ;

[0061] High-frequency energy Total energy The ratio of high-frequency energy to total energy is used as an indicator of frequency domain importance. ,include:

[0062] ;

[0063] in, This indicates the elimination of zero minute parameters.

[0064] Frequency domain importance index As a representation of the current image frequency domain complexity, a learnable parameter is introduced. and frequency domain importance index Added as a learnable weight parameter to the first loss. In this process, we obtain the causal context-adjusted loss with an adaptive loss weighting mechanism incorporating frequency domain awareness. ,include:

[0065] .

[0066] The present invention also provides a remote sensing image compression device based on frequency domain enhancement and adaptive optimization, which is used in the method of the present invention. The device includes a first unit and a second unit.

[0067] The first unit is used to train the first model based on a preset training set and a loss function that considers pixel distortion, structure preservation, frequency domain error, perceptual quality, bit rate constraint, context autoregressive regularization term and auxiliary supervision loss, to obtain a pre-trained first model.

[0068] The second unit is used to compress and decompress remote sensing images using a pre-trained first model deployed simultaneously at the compression and decompression ends. The first model includes an encoder, an entropy model, and a decoder. Both the encoder and decoder perform frequency domain analysis on the input through a frequency-domain enhanced selective state space module and model long-distance dependencies and frequency domain structures based on the state space mechanism.

[0069] Compression includes: extracting the main features of the input image through an encoder; probabilistically modeling the distribution information of the main features using an entropy model to generate a bitstream; and sending the bitstream to the decompression end.

[0070] Decompression includes: receiving the bitstream at the decompression end; restoring the bitstream to its main features using an entropy model; and inputting the main features into the decoder to obtain the reconstructed image.

[0071] The present invention has the following beneficial effects:

[0072] This invention presents a remote sensing image compression method based on frequency domain enhancement and adaptive optimization. By introducing a selective state-space module for frequency domain enhancement, it fully considers the expressive importance of different frequency components during feature modeling, achieving explicit perception of image frequency domain characteristics and efficient modeling of long-distance dependencies. This effectively improves the problem of insufficient expression in compressing high-frequency information in traditional convolutional networks. By introducing a frequency domain-aware adaptive loss weighting mechanism, the loss weights are dynamically adjusted based on the causal context of the training phase according to the spectral energy distribution of the input image. This allows the model to adaptively optimize the compression strategy when facing image content with varying frequency domain complexity, strengthening the protection of high-frequency details and suppressing overfitting in low-frequency regions. Furthermore, this invention enhances the model's explicit frequency domain and perceptual optimization capabilities by combining multiple frequency domain enhancement rate-distortion loss functions. Compared to traditional methods, this invention improves the modeling ability for high-frequency details and long-distance dependencies. This invention can significantly improve the compression fidelity of remote sensing images in complex texture regions and regions with detailed information. This invention can effectively compress remote sensing images while ensuring image reconstruction quality.

[0073] The remote sensing image compression device based on frequency domain enhancement and adaptive optimization of the present invention, when used in the method of the present invention, has the same beneficial effects as the method of the present invention.

[0074] In addition to the objectives, features, and advantages described above, the present invention has other objectives, features, and advantages. The invention will now be described in further detail with reference to the accompanying drawings. Attached Figure Description

[0075] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0076] Figure 1 This is a schematic diagram of the method flow of a preferred embodiment of the present invention.

[0077] Figure 2 This is a schematic diagram of an encoder according to a preferred embodiment of the present invention.

[0078] Figure 3This is a schematic diagram of a decoder according to a preferred embodiment of the present invention.

[0079] Figure 4 This is a schematic diagram of the internal flow of the selective state space module for frequency domain enhancement according to a preferred embodiment of the present invention.

[0080] Figure 5 This is a schematic diagram of the internal flow of the frequency domain analysis submodule in a preferred embodiment of the present invention.

[0081] Figure 6 This is a schematic diagram of the internal flow of the selective state space submodule in a preferred embodiment of the present invention.

[0082] Figure 7 This is a schematic diagram of loss function data acquisition according to a preferred embodiment of the present invention.

[0083] Figure 8 This is a schematic diagram of the compression stage in a preferred embodiment of the present invention.

[0084] Figure 9 This is a schematic diagram of a priori encoder according to a preferred embodiment of the present invention.

[0085] Figure 10 This is a schematic diagram of a preferred embodiment of the priori decoder of the present invention.

[0086] Figure 11 This is a schematic diagram of the decompression stage in a preferred embodiment of the present invention. Detailed Implementation

[0087] The embodiments of the present invention will be described in detail below with reference to the accompanying drawings, but the present invention can be implemented in many different ways as defined and covered by the claims.

[0088] See Figure 1 In a preferred embodiment of the present invention, a remote sensing image compression method based on frequency domain enhancement and adaptive optimization is provided, comprising:

[0089] S1. Construct the first model; the first model includes an encoder, an entropy model and a decoder; both the encoder and decoder perform frequency domain analysis on the input through a frequency-domain enhanced selective state space module, and model long-distance dependencies and frequency domain structures based on the state space mechanism.

[0090] In a preferred embodiment of the present invention, both the encoder and decoder perform frequency domain analysis on the input through a frequency-domain enhanced selective state-space module, and model long-distance dependencies and frequency domain structures based on a state-space mechanism, including:

[0091] See Figure 2 The encoder includes three sets of first modules connected in sequence; the first module includes a 5×5 convolution, three bottleneck residual modules and four frequency domain enhanced selective state space modules connected in sequence.

[0092] See Figure 3 The decoder includes three sets of second modules connected in sequence; each second module includes a 5×5 convolution, four frequency-domain enhanced selective state space modules, and three bottleneck residual modules connected in sequence.

[0093] In the encoder, the frequency-domain enhanced selective state-space module is used to enhance the high-frequency details and global contextual representation of input features through frequency domain analysis and remote dependency modeling, thereby improving the preservation of structural details and semantic representation of compressed features and reducing reconstruction errors.

[0094] In the decoder, the frequency-domain enhanced selective state-space module is used to enhance the high-frequency details and spatial consistency in the decoded features through frequency-domain attention and long-range dependency modeling, thereby achieving high-quality restoration of structural information and detail reconstruction, and significantly improving the visual quality and fidelity of the reconstructed image.

[0095] See Figure 4 In a preferred embodiment of the present invention, the frequency-domain enhanced selective state-space module includes:

[0096] When input features After inputting a frequency-domain enhanced selective state-space module, the initial features are obtained through a 1×1 convolution. Initial features The frequency domain importance weights are obtained from the input frequency domain analysis submodule. Initial features Frequency domain importance weights Inputting a selective state space submodule and performing state updates yields context-enhanced features that fuse frequency domain information and state-aware information. ; Enhance features based on context Frequency domain importance weights and initial features Obtain features ;feature Feature extraction is performed by inputting a free block of nonlinear activation function to obtain features. ;feature The output features are obtained after a 1×1 convolution. .

[0097] See Figure 5 In a preferred embodiment of the present invention, the frequency domain analysis submodule includes:

[0098] Initial features After inputting into the frequency domain analysis submodule, the high-frequency, mid-frequency, low-frequency, and DC components are extracted through four parallel frequency domain channels. The high-frequency component is extracted using a 3×3 grouped convolution, while the mid-frequency, low-frequency, and DC components are each extracted using a 1×1 convolution. Intermediate frequency components Low-frequency components and DC component ; to high frequency components Intermediate frequency components Low-frequency components and DC component The fused frequency domain features are obtained by channel splicing. ; frequency domain characteristics of fusion The frequency domain importance evaluation network then passes through average pooling, 1×1 convolution, ReLU activation function layer, 1×1 convolution, and Sigmoid function layer in sequence to obtain frequency domain importance weights consistent with the original channels. .

[0099] In a preferred embodiment of the present invention, the extraction of high-frequency, mid-frequency, low-frequency, and DC components through four parallel frequency domain channels includes:

[0100] ;

[0101] See Figure 6 In a preferred embodiment of the present invention, the selective state space submodule includes:

[0102] Initial features Frequency domain importance weights After inputting the selective state-space submodule, the frequency domain importance weights are used. For initial features Channel-by-channel weighting is performed to obtain the frequency domain weighted feature map. This yields a spectrum of higher importance, enabling frequency-domain driven feature recalibration. The frequency-domain weighted feature map is then processed. The global feature vector is obtained through average pooling. Global feature vectors The three-layer perceptron is input to obtain the selectivity coefficients used to control the state update process. and The three-layer perceptron includes a first linear transformation layer connected in sequence. ReLU activation function layer and second linear transformation layer Global feature vectors The state-space representation is obtained by adjusting the channel dimension through linear mapping. According to the selectivity coefficient and State space representation using lightweight gating mechanisms Perform selective updates, selection coefficients Selectivity coefficient is used to control the proportion of the original state retained. The nonlinear perturbation amplitude used to regulate the state, and the selectivity coefficient. These three elements—namely, the global mean information and the frequency domain—work together to achieve dynamic state adjustment and fusion guided by the frequency domain, resulting in an adaptively updated state representation. Representing the state By restoring the original channel dimension through linear mapping, context-enhanced features that fuse frequency domain information and state-aware information are obtained. .

[0103] In a preferred embodiment of the present invention, the state representation is adaptively updated. include:

[0104] ;

[0105] in, Represents the Sigmoid function; Represents the hyperbolic tangent function; ,right Calculate the average along the latitude of the passage. S for The number of channels.

[0106] S2. Train the first model based on the preset training set and a loss function that considers pixel distortion, structure preservation, frequency domain error, perceptual quality, bit rate constraint, context autoregressive regularization term and auxiliary supervision loss to obtain the pre-trained first model.

[0107] In a preferred embodiment of the present invention, pixel distortion includes mean square error (MSE) and multi-scale structural similarity (MS-SSIM); structural preservation includes edge information and structural similarity (SSIM).

[0108] See Figure 7 In a preferred embodiment of the present invention, the loss function is obtained from the compression stage of the first model of the input image. Second super-prior features Sub-features in the first data Sub-features in the second data Second probability distribution Obtain the reconstructed image from the model output during the decompression stage of the first model. The specific parameters are explained in step S3.

[0109] In a preferred embodiment of the present invention, the loss function includes:

[0110] ;

[0111] in, This represents the distortion term, used to measure the reconstruction error between the compressed image and the original image; Indicates bit rate loss; This represents the causal context-adjusted loss with frequency domain awareness and adaptive loss weighting, which guides the model to adaptively adjust the focus of the loss function according to the frequency domain complexity of different images. This indicates the auxiliary monitoring loss item.

[0112] Distortion item include:

[0113] ;

[0114] in, This represents the squared error loss, used to optimize pixel-level reconstruction accuracy. Total number of pixels; This represents the reconstructed image output by the model; Indicates the input image; This represents perceptual loss, which enhances the perception of subjective quality by comparing the distances of images in the deep feature space. For the first in the pre-trained network Feature maps extracted from layers; This represents edge loss, used to enhance the preservation of image edges and structural contours. Represents the gradient of an image; This represents the frequency domain loss, and the image restoration performance in the high, mid, and low frequency regions is evaluated through spectral analysis. , , Indicates the spectral amplitude. Indicates Fourier transform; This represents structural similarity loss, used to model the structural information and local consistency of an image. Represents a multi-scale structural similarity index; , , and These are the learnable weight coefficients for perceptual loss, edge loss, frequency domain loss, and structural similarity loss, respectively.

[0115] Bit rate loss include:

[0116] ;

[0117] This indicates the second prior feature in the compression phase. The bit rate loss term; during model training, based on the first probability modeling parameters of the compression phase. Constructing the second super-prior features in the compression stage The probability distribution is calculated, and the negative log-likelihood value at each feature location is estimated and summed to obtain the result. .

[0118] , representing the sub-features in the first data of the compression stage. The bit rate loss term; during model training, based on the sub-parameters in the first data of the compression phase. Sub-features in the first data of the compression stage The probability distribution is calculated, and the negative log-likelihood value at each feature location is estimated and summed to obtain the result. .

[0119] , representing the sub-features in the second data of the compression stage. The bit rate loss term; during model training, based on the sub-parameters in the second data of the compression phase. and compression phase second probability modeling parameters The fusion results construct sub-features in the second data of the compression stage. The probability distribution is calculated, and the negative log-likelihood value at each feature location is estimated and summed to obtain the result. .

[0120] Auxiliary monitoring of loss items include:

[0121] During model training, based on the second super-prior features of the compression phase... Predicting sub-features in the second data of the compression stage probability distribution And calculate the negative log-likelihood of the features, and obtain .

[0122] In a preferred embodiment of the present invention, based on the second prior feature of the compression stage Predicting sub-features in the second data of the compression stage probability distribution When using a noisy Gaussian distribution, the output parameters are... , The mean, Standard deviation, specifically including:

[0123] ;

[0124] in, Indicates a Gaussian distribution. This represents Gaussian noise.

[0125] In a preferred embodiment of the present invention, the causal context-adjusted loss with frequency-domain-aware adaptive loss weighting mechanism includes:

[0126] Obtain the first loss ,include:

[0127] ;

[0128] in, Represents information entropy; and All are true distributions.

[0129] , indicating that only given Under the condition of, calculate In probability distribution The expected value of the negative log probability under context-free prediction is the cross-entropy.

[0130] , and from context modeling Under the condition of, calculate Second probability distribution The expected value of the negative logarithmic probability under contextual prediction is the cross-entropy.

[0131] After grayscale processing of the input image, a two-dimensional fast Fourier transform is performed to obtain the image amplitude. ;Calculation based on a preset high-frequency mask The left and right quarter regions of the image are also marked as high-frequency regions, including:

[0132] ;

[0133] High-frequency energy Total energy The ratio of high-frequency energy to total energy is used as an indicator of frequency domain importance. ,include:

[0134] ;

[0135] in, This indicates the elimination of zero minute parameters.

[0136] Frequency domain importance index As a representation of the current image frequency domain complexity, a learnable parameter is introduced. and frequency domain importance index Added as a learnable weight parameter to the first loss. In this process, we obtain the causal context-adjusted loss with an adaptive loss weighting mechanism incorporating frequency domain awareness. ,include:

[0137] ;

[0138] In a preferred embodiment of the present invention, during model training, a frequency-domain-aware adaptive loss weighting mechanism assists the context modeling module in performing causal autoregressive modeling of the main features; and a frequency-domain enhancement rate-distortion loss function with multiple combinations is introduced into the loss function for joint optimization.

[0139] In a preferred embodiment of the present invention, several remote sensing images are selected as the training set, and the remaining images are used as the test set. The training set images are augmented using a sliding window cropping method to obtain multi-scale image patch samples. The first model is constructed under a deep learning framework. The training set is input into the first model for compression and decompression. The Adam optimizer is used to minimize the above loss function for parameter optimization training. The training results are verified using the test set, and finally the pre-trained first model is obtained.

[0140] In a preferred embodiment of the invention, a frequency-aware adaptive loss weighting mechanism is incorporated to dynamically adjust the context loss weights based on the frequency energy distribution, enabling content-adaptive rate-distortion optimization. By analyzing the frequency domain characteristics of the input image, the weight allocation of the training loss is dynamically adjusted. For images rich in high-frequency details, the compression loss's focus on details is automatically increased, thereby improving compression performance. Conversely, for images dominated by low frequencies, the influence of auxiliary losses is automatically reduced to prevent over-optimization. This allows the model to adaptively adjust the focus of the loss function according to the frequency domain complexity of different images.

[0141] In a preferred embodiment of the present invention, during model training, uniform noise is added to simulate the quantization effect, as shown below: Indicates in Uniform noise within the interval. During the compression and decompression of the input image using the model after training, the following hard quantization method is employed: , This represents the rounding function.

[0142] S3. Deploy the pre-trained first model to both the compression and decompression ends simultaneously, and compress and decompress the remote sensing image using the pre-trained first model at both ends.

[0143] Compression includes: extracting the main features of the input image through an encoder; probabilistically modeling the distribution information of the main features using an entropy model to generate a bitstream; and sending the bitstream to the decompression end.

[0144] Decompression includes: receiving the bitstream at the decompression end; restoring the bitstream to its main features using an entropy model; and inputting the main features into the decoder to obtain the reconstructed image.

[0145] See Figure 8In a preferred embodiment of the present invention, the main features of the input image are extracted by an encoder; the distribution information of the main features is probabilistically modeled using an entropy model to generate a bitstream, including:

[0146] In a preferred embodiment of the present invention, the entropy model includes a super-prior encoder, a super-prior decoder, a context modeling module, a one-stage modeling module, and a two-stage modeling module.

[0147] See Figure 9 In a preferred embodiment of the present invention, the super prior encoder is composed of three consecutive convolutional layers and two activation function layers cascaded together. The kernel size of the first convolutional layer is 3×3, followed by a GELU activation function layer; the kernel size of the second convolutional layer is 5×5, followed by a GELU activation function layer; and the kernel size of the third convolutional layer is 5×5.

[0148] See Figure 10 In a preferred embodiment of the present invention, the advanced prior decoder consists of three consecutive convolutional layers and two activation function layers cascaded together. The kernel size of the first convolutional layer is 5×5, followed by a GELU activation function layer; the kernel size of the second convolutional layer is 5×5, followed by a GELU activation function layer; and the kernel size of the third convolutional layer is 3×3.

[0149] Extract the input image using an encoder. Main features y ; Using a hyper-prior encoder to process the main features y The spatial and channel statistical information is used to model the first hyper-prior feature. z ; the first super-prior feature z Quantization is performed to obtain the second super-prior feature. ;Based on the second super-prior feature through the super-prior decoder Obtain the first probability modeling parameters , The mean, The standard deviation is denoted as .

[0150] Main features y Quantization is performed to obtain the quantized main features. ,Will The non-uniform channel groups are divided into two feature groups based on a preset ratio. and (e.g., before) Channel as ,back Channel as ), and the first probability modeling parameters Two sub-parameters are obtained by performing the same partitioning. and Sub-features and sub-parameters with the same channel size are paired and randomly assigned as first data and second data. In a preferred embodiment of the present invention, it is assumed that the first data includes sub-features. and sub-parameters The second data includes sub-features and sub-parameters .

[0151] The first-stage modeling module uses the sub-parameters in the first data... Modeling is performed to obtain the first probability distribution. Using a noisy Gaussian distribution:

[0152] ;

[0153] in, Indicates a Gaussian distribution. This represents Gaussian noise.

[0154] The context modeling module uses sub-features from the first data. Contextual information is extracted in the spatial and channel dimensions to obtain the second probability modeling parameters. .

[0155] The second probability distribution is obtained by modeling based on the second probability modeling parameters and the sub-parameters in the second data through the two-stage modeling module, including:

[0156] The second probability modeling parameters Sub-parameters in the second data By fusing, we obtain:

[0157] ;

[0158] The second probability distribution is obtained by modeling based on the fusion results using the two-stage modeling module. Using a noisy Gaussian distribution:

[0159] ;

[0160] According to the first probability distribution Sub-features in the first data Entropy encoding is performed to generate the first bit stream. .

[0161] According to the second probability distribution Sub-features in the second data Entropy encoding is performed to generate a second bitstream. .

[0162] The second hyperprior feature is analyzed using a Gaussian distribution model. Modeling is performed to obtain the third probability distribution. According to the third probability distribution For the second super-prior feature Perform entropy encoding to generate a third bit stream. .

[0163] See Figure 11 In a preferred embodiment of the present invention, the bitstream is restored to its main features using an entropy model; the main features are then input into a decoder to obtain the reconstructed image, including:

[0164] The decompression end is based on the third bit stream By performing probability modeling parameters, a third probability distribution is obtained. .

[0165] According to the third probability distribution For the third bit stream Entropy decoding is performed to obtain the second super-prior feature. .

[0166] Based on the second super-prior feature using the super-prior decoder Obtain the first probability modeling parameters ; model the first probability parameters Based on the same preset ratio of the non-uniform channel grouping mentioned above, it is divided into two sub-parameters. and And match it with the channel size of the first data and the second data to obtain the first sub-parameter. Second sub-parameter .

[0167] Based on the first sub-parameter, the first-stage modeling module is used. Modeling is performed to obtain the first probability distribution. .

[0168] According to the first probability distribution For the first bit stream Entropy decoding is performed to obtain the sub-features in the first data. .

[0169] The context modeling module uses sub-features from the first data. Contextual information is extracted in the spatial and channel dimensions to obtain the second probability modeling parameters. .

[0170] Based on the second probability modeling parameters, the two-stage modeling module is used. Second sub-parameter Modeling is performed to obtain the second probability distribution. ;

[0171] According to the second probability distribution For the second bit stream Entropy decoding is performed to obtain sub-features from the second data. .

[0172] Sub-features in the first data Sub-features in the second data The main features are then concatenated to obtain the quantized main features. ;

[0173] The reconstructed image is obtained by recovering the quantized main features using a decoder. .

[0174] This invention presents a remote sensing image compression method based on frequency domain enhancement and adaptive optimization. By introducing a selective state-space module for frequency domain enhancement, it fully considers the expressive importance of different frequency components during feature modeling, achieving explicit perception of image frequency domain characteristics and efficient modeling of long-distance dependencies. This effectively improves the problem of insufficient expression in compressing high-frequency information in traditional convolutional networks. By introducing a frequency-domain-aware adaptive loss weighting mechanism, the loss weights are dynamically adjusted based on the causal context of the training phase according to the spectral energy distribution of the input image. This allows the model to adaptively optimize the compression strategy when facing image content with varying frequency domain complexity, strengthening the protection of high-frequency details and suppressing overfitting in low-frequency regions. Furthermore, this invention enhances the model's explicit frequency domain and perceptual optimization capabilities by combining multiple frequency domain enhancement rate-distortion loss functions. Compared to traditional methods, this invention improves the modeling ability for high-frequency details and long-distance dependencies. This invention can significantly improve the compression fidelity of remote sensing images in complex texture regions and regions with detailed information. This method can effectively compress remote sensing images while ensuring image reconstruction quality.

[0175] In a preferred embodiment of the present invention, a remote sensing image compression device based on frequency domain enhancement and adaptive optimization is also provided for use in the method of the present invention. The device includes a first unit and a second unit.

[0176] The first unit is used to train the first model based on a preset training set and a loss function that considers pixel distortion, structure preservation, frequency domain error, perceptual quality, bit rate constraint, context autoregressive regularization term and auxiliary supervision loss, to obtain a pre-trained first model.

[0177] The second unit is used to compress and decompress remote sensing images using a pre-trained first model deployed simultaneously at the compression and decompression ends. The first model includes an encoder, an entropy model, and a decoder. Both the encoder and decoder perform frequency domain analysis on the input through a frequency-domain enhanced selective state space module and model long-distance dependencies and frequency domain structures based on the state space mechanism.

[0178] Compression includes: extracting the main features of the input image through an encoder; probabilistically modeling the distribution information of the main features using an entropy model to generate a bitstream; and sending the bitstream to the decompression end.

[0179] Decompression includes: receiving the bitstream at the decompression end; restoring the bitstream to its main features using an entropy model; and inputting the main features into the decoder to obtain the reconstructed image.

[0180] The remote sensing image compression device based on frequency domain enhancement and adaptive optimization of the present invention, when used in the method of the present invention, has the same beneficial effects as the method of the present invention.

[0181] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A remote sensing image compression method based on frequency domain enhancement and adaptive optimization, characterized in that, The remote sensing images are compressed and decompressed using a pre-trained first model deployed simultaneously at both the compression and decompression ends, including: The first model includes an encoder, an entropy model, and a decoder; both the encoder and the decoder perform frequency domain analysis on the input through a frequency-domain enhanced selective state space module, and model long-distance dependencies and frequency domain structures based on the state space mechanism. The first model is trained using a preset training set and a loss function that considers pixel distortion, structure preservation, frequency domain error, perceptual quality, bit rate constraint, context autoregressive regularization term and auxiliary supervision loss to obtain a pre-trained first model. The compression includes: extracting the main features of the input image through the encoder; probabilistically modeling the distribution information of the main features through the entropy model to generate a bitstream, and sending the bitstream to the decompression terminal; The decompression includes: the decompression terminal receiving the bitstream; restoring the bitstream to the main feature using the entropy model; and inputting the main feature into the decoder to obtain the reconstructed image. Both the encoder and decoder perform frequency domain analysis on the input using a frequency-domain enhanced selective state-space module, and model long-distance dependencies and frequency domain structures based on a state-space mechanism, including: The encoder includes three sets of first modules connected in sequence; the first module includes a 5×5 convolution, three bottleneck residual modules and four frequency domain enhanced selective state space modules connected in sequence. The decoder includes three sets of second modules connected in sequence; the second module includes a 5×5 convolution, four frequency-domain enhanced selective state space modules, and three bottleneck residual modules connected in sequence. In the encoder, the frequency-domain enhanced selective state-space module is used to enhance the high-frequency details and global contextual representation of the input features through frequency domain analysis and long-range dependency modeling. In the decoder, the frequency-domain enhanced selective state-space module is used to enhance high-frequency details and spatial consistency in the decoded features through frequency-domain attention and long-range dependency modeling. The frequency-domain enhanced selective state-space module includes: When input features After inputting the frequency-domain enhanced selective state-space module, the initial features are obtained through 1×1 convolution. The initial feature The frequency domain importance weights are obtained from the input frequency domain analysis submodule. The initial feature and the frequency domain importance weights Inputting a selective state space submodule and performing state updates yields context-enhanced features that fuse frequency domain information and state-aware information. Enhance features based on the context The frequency domain importance weights and the initial features Obtain features The features Feature extraction is performed by inputting a free block of nonlinear activation function to obtain features. The features The output features are obtained after a 1×1 convolution. .

2. The remote sensing image compression method based on frequency domain enhancement and adaptive optimization according to claim 1, characterized in that, The frequency domain analysis submodule includes: The initial features After input to the frequency domain analysis submodule, the high-frequency, mid-frequency, low-frequency, and DC components are extracted through four parallel frequency domain channels. The high-frequency component is extracted using a 3×3 grouped convolution, while the mid-frequency, low-frequency, and DC components are each extracted using a 1×1 convolution. Intermediate frequency components Low-frequency components and DC component ; the high-frequency components Intermediate frequency components Low-frequency components and DC component The fused frequency domain features are obtained by channel splicing. The fused frequency domain features The frequency domain importance evaluation network is then passed through average pooling, 1×1 convolution, ReLU activation function layer, 1×1 convolution, and Sigmoid function layer in sequence to obtain the frequency domain importance weights consistent with the original channels. .

3. The remote sensing image compression method based on frequency domain enhancement and adaptive optimization according to claim 2, characterized in that, The selective state space submodule includes: The initial features and the frequency domain importance weights After inputting the selective state space submodule, the frequency domain importance weights are used. For the initial features Channel-by-channel weighting is performed to obtain the frequency domain weighted feature map. The frequency domain weighted feature map The global feature vector is obtained through average pooling. The global feature vector The three-layer perceptron is input to obtain the selectivity coefficients used to control the state update process. and The three-layer perceptron includes a first linear transformation layer connected in sequence. ReLU activation function layer and second linear transformation layer The global feature vector The state-space representation is obtained by adjusting the channel dimension through linear mapping. According to the selectivity coefficient and The state space representation is combined with a lightweight gating mechanism. Perform selective updates, selection coefficients Selectivity coefficient is used to control the proportion of the original state retained. The nonlinear perturbation amplitude used to regulate the state, and the selectivity coefficient. These three elements—namely, the global mean information and the frequency domain—work together to achieve dynamic state adjustment and fusion guided by the frequency domain, resulting in an adaptively updated state representation. Representing the state By restoring the original channel dimension through linear mapping, context-enhanced features that fuse frequency domain information and state-aware information are obtained. .

4. The remote sensing image compression method based on frequency domain enhancement and adaptive optimization according to claim 3, characterized in that, The encoder extracts the main features of the input image; The generation of a bitstream involves probabilistically modeling the distribution information of the main features using the entropy model, and then performing the following: The entropy model includes a super-prior encoder, a super-prior decoder, a context modeling module, a one-stage modeling module, and a two-stage modeling module. The encoder extracts the main features of the input image; the super-prior encoder models the spatial and channel statistics of the main features to obtain a first super-prior feature; the first super-prior feature is quantized to obtain a second super-prior feature; and the super-prior decoder obtains a first probability modeling parameter based on the second super-prior feature. After quantizing the main feature, it is divided into two sub-features based on a preset ratio of non-uniform channel grouping, and the first probability modeling parameter is divided into two sub-parameters in the same way; the sub-features and sub-parameters with the same channel size are paired and randomly assigned as the first data and the second data. The first-stage modeling module performs modeling based on the sub-parameters in the first data to obtain a first probability distribution; the context modeling module extracts spatial and channel-dimensional context information based on the sub-features in the first data to obtain a second probability modeling parameter; and the second-stage modeling module performs modeling based on the second probability modeling parameter and the sub-parameters in the second data to obtain a second probability distribution. Based on the first probability distribution, the sub-features in the first data are entropy encoded to generate a first bit stream; The sub-features in the second data are entropy encoded according to the second probability distribution to generate a second bit stream; The second prior feature is modeled using a Gaussian distribution model to obtain a third probability distribution; the second prior feature is then entropy-encoded according to the third probability distribution to generate a third bit stream.

5. The remote sensing image compression method based on frequency domain enhancement and adaptive optimization according to claim 4, characterized in that, The bitstream is restored to the main features using the entropy model. Inputting the main features into the decoder to obtain the reconstructed image includes: The decompression end performs probability modeling parameters based on the third bit stream to obtain the third probability distribution; The third bitstream is entropy-decoded according to the third probability distribution to obtain the second prior feature; The first probability modeling parameter is obtained by the super-prior decoder based on the second super-prior feature; the first probability modeling parameter is divided into two sub-parameters based on the non-uniform channel grouping of the preset ratio, and matched with the channel size of the first data and the second data to obtain the first sub-parameter and the second sub-parameter respectively. The first probability distribution is obtained by modeling based on the first sub-parameters using the first-stage modeling module. Entropy decoding is performed on the first bitstream according to the first probability distribution to obtain sub-features in the first data; The context modeling module extracts spatial and channel-dimensional context information based on the sub-features in the first data to obtain the second probability modeling parameters. The second probability distribution is obtained by modeling based on the second probability modeling parameters and the second sub-parameters using the two-stage modeling module. Entropy decoding is performed on the second bitstream according to the second probability distribution to obtain the sub-features in the second data; The sub-features in the first data and the sub-features in the second data are concatenated to obtain the quantized main feature; The decoder recovers the quantized main features to obtain the reconstructed image.

6. The remote sensing image compression method based on frequency domain enhancement and adaptive optimization according to claim 5, characterized in that, The loss function includes: ； in, This represents the distortion term, used to measure the reconstruction error between the compressed image and the original image; Indicates bit rate loss; This represents the causal context-adjusted loss with frequency domain awareness and adaptive loss weighting, which guides the model to adaptively adjust the focus of the loss function according to the frequency domain complexity of different images. Indicates the auxiliary monitoring loss item; The distortion item include: ； in, This represents the squared error loss, used to optimize pixel-level reconstruction accuracy. Total number of pixels; This represents the reconstructed image output by the model; Indicates the input image; This represents perceptual loss, which enhances the perception of subjective quality by comparing the distances of images in the deep feature space. For the first in the pre-trained network Feature maps extracted from layers; This represents edge loss, used to enhance the preservation of image edges and structural contours. Represents the gradient of an image; This represents the frequency domain loss, and the image restoration performance in the high, mid, and low frequency regions is evaluated through spectral analysis. , , Indicates the spectral amplitude. Indicates Fourier transform; This represents structural similarity loss, used to model the structural information and local consistency of an image. Represents a multi-scale structural similarity index; , , and These are the learnable weight coefficients for perceptual loss, edge loss, frequency domain loss, and structural similarity loss, respectively. The bit rate loss include: ； This indicates the second prior feature during the compression phase. The bit rate loss term; during model training, based on the first probability modeling parameters from the compression phase. Constructing the second super-prior feature in the compression phase The probability distribution is calculated, and the negative log-likelihood value at each feature location is estimated and summed to obtain the result. ; , representing the sub-features in the first data during the compression stage. The bit rate loss term; during model training, based on the sub-parameters in the first data during the compression phase. Sub-features in the first data during the compression phase The probability distribution is calculated, and the negative log-likelihood value at each feature location is estimated and summed to obtain the result. ; , representing the sub-features in the second data during the compression stage. The bit rate loss term; during model training, based on the sub-parameters in the second data during the compression phase. The fusion result of the second probability modeling parameters and the second probability modeling parameters are used to construct the sub-features in the second data during the compression stage. The probability distribution is calculated, and the negative log-likelihood value at each feature location is estimated and summed to obtain the result. ; The auxiliary monitoring loss item include: During model training, based on the second hyperprior features from the compression phase... Predicting sub-features in the second data during the compression phase probability distribution And calculate the negative log-likelihood of the features, and obtain .

7. The remote sensing image compression method based on frequency domain enhancement and adaptive optimization according to claim 6, characterized in that, The causal context-adjusted loss of the adaptive loss weighting mechanism incorporating frequency domain awareness includes: Obtain the first loss ,include: ； in, Represents information entropy; and All are true distributions; , indicating that only given Under the conditions, calculate In probability distribution The expected value of the negative log probability under context-free prediction, i.e., the cross-entropy under context-free prediction; In the given and from context modeling Under the conditions, calculate Second probability distribution The expected value of the negative log probability under contextual prediction, i.e., the cross-entropy under contextual prediction; After grayscale processing of the input image, a two-dimensional fast Fourier transform is performed to obtain the image amplitude. ;Calculation based on a preset high-frequency mask The left and right quarter regions of the image are also marked as high-frequency regions, including: ； High-frequency energy Total energy The ratio of high-frequency energy to total energy is used as an indicator of frequency domain importance. ,include: ； in, This indicates the prevention of zero-valued parameters; Frequency domain importance index As a representation of the current image frequency domain complexity, a learnable parameter is introduced. and frequency domain importance index Added as a learnable weight parameter to the first loss. In this process, the causal context-adjusted loss of the adaptive loss weighting mechanism incorporating frequency domain awareness is obtained. ,include: 。 8. A remote sensing image compression device based on frequency domain enhancement and adaptive optimization, used in the method described in any one of claims 1 to 7, characterized in that, The device includes a first unit and a second unit; The first unit is used to train the first model based on a preset training set and a loss function that considers pixel distortion, structure preservation, frequency domain error, perceptual quality, bit rate constraint, context autoregressive regularization term and auxiliary supervision loss, so as to obtain a pre-trained first model. The second unit is used to compress and decompress remote sensing images using a pre-trained first model deployed simultaneously at the compression and decompression ends; the first model includes an encoder, an entropy model, and a decoder; both the encoder and decoder perform frequency domain analysis on the input through a frequency domain-enhanced selective state space module, and model long-distance dependencies and frequency domain structures based on the state space mechanism; The compression includes: extracting the main features of the input image through the encoder; probabilistically modeling the distribution information of the main features through the entropy model to generate a bitstream, and sending the bitstream to the decompression terminal; The decompression includes: the decompression terminal receiving the bitstream; restoring the bitstream to the main feature using the entropy model; and inputting the main feature into the decoder to obtain the reconstructed image.

Images