A low-light image generation compression method, system and storage medium

Abstract

The application discloses a low-light image generation type compression method, system and storage medium, relates to the technical field of image processing and machine learning compression. The method first establishes a physical imaging model containing scene radiation, transmittance, global background light and additive noise; the input low-light image is decomposed into a scene radiation component representing essential content and a degradation parameter component representing degradation factors through a physical perception encoder; joint optimization is performed to maximize the intelligent gain, wherein the intelligent gain is defined as the ratio of the information gain to the system complexity, the information gain is measured by the reconstruction error in the high-level semantic feature space, and the system complexity is jointly constituted by the model parameter quantity and the output uncertainty; the two types of components are quantized and entropy encoded at the encoding end to generate a bit stream, and only the scene radiation component is used at the decoding end to reconstruct an enhanced clear image. The application improves the visual quality of the reconstructed image at a low code rate.

Description

Technical Field

[0001] This application relates to the field of image processing and machine learning compression technology, and in particular to a generative compression method, system and storage medium for low-light images. Background Technology

[0002] Images created in low-light environments commonly suffer from insufficient brightness, low signal-to-noise ratio, color distortion, and blurred details, posing significant challenges to subsequent storage, transmission, and analysis. Traditional image compression methods (such as JPEG and HEVC) often fail to consider degradation mechanisms when processing such images, leading to amplified noise, further loss of detail in dark areas, and severely compromised rate-distortion performance.

[0003] In recent years, deep learning-based image compression methods have achieved joint optimization of enhancement and compression through end-to-end training, resulting in significant progress in subjective quality. Some studies have introduced physical imaging models (such as atmospheric scattering models) to model the degradation process, decomposing the image into a clean scene and degradation parameters, thereby achieving a more compact representation. However, existing methods generally suffer from the following shortcomings:

[0004] The lack of an explicit trade-off mechanism between compression efficiency and model complexity can easily lead to over-parameterization of the model or a sharp drop in performance at low bitrates.

[0005] Information measurement still relies on pixel-level loss, making it difficult to guarantee the integrity of high-level semantic structures;

[0006] The decoding end usually needs to combine the usage scenario and degradation parameters for reconstruction, which increases the computational burden on the receiving end.

[0007] Insufficient consideration of the uncertainty of the model output affects the robustness and generalization ability of the system. Summary of the Invention

[0008] To address the aforementioned issues, this application provides a low-light image generative compression method, system, and storage medium to improve the visual quality of reconstructed images at low bit rates; it also provides a low-light image generative compression system for implementing low-light image generative compression methods under different conditions.

[0009] The first technical solution adopted in this application is: providing a low-light image generative compression method, including the following steps:

[0010] A physical imaging model for low-light images is established, which expresses the observed degraded image as a function of four elements: scene radiance, transmittance, global background light, and additive noise.

[0011] The input low-light image is processed using a physical sensing encoder to separate the scene radiative component, which represents the essential content of the scene, and the degradation parameter component, which represents the degradation factors of the imaging environment.

[0012] The network is optimized with the goal of maximizing the intelligent gain, wherein the intelligent gain is the ratio of information gain to system complexity, the information gain is measured by the difference between the reconstructed intrinsic scene image and the real intrinsic scene image in the high-level semantic feature space, and the system complexity is a weighted combination of the inherent complexity of the model parameter quantity measure and the uncertainty of the model output variance measure.

[0013] At the encoding end, the scene radiative component and the degradation parameter component are quantized and entropy encoded to generate a compressed bitstream; at the decoding end, the compressed bitstream is entropy decoded, and the obtained scene radiative component is input to the intrinsic scene generator to directly output the enhanced clear image.

[0014] In an alternative embodiment, in the physical imaging model, the degraded image is formed by multiplying the scene radiation and the transmittance element-wise, adding the product of the global background light and a factor subtracting the transmittance, and then adding the additive noise.

[0015] In an optional embodiment, the physical perception encoder uses a shared backbone network for basic feature extraction, then outputs the scene radiation component through a structurally independent first branch network, and outputs the degradation parameter component through a structurally independent second branch network.

[0016] In an optional embodiment, the degradation parameter component output by the second branch network is further decoupled into a transmittance parameter vector for generating a transmittance map and a background light parameter vector for directly estimating the background light color.

[0017] In an optional embodiment, in the joint training step, the optimized overall objective function is composed of a weighted sum of a physical consistency loss term and a smart gain entropy loss term, wherein the smart gain entropy loss term is the negative logarithm of the smart gain.

[0018] In an optional embodiment, the physical consistency loss term includes:

[0019] The first loss function is used to ensure that the reconstructed image synthesized based on the physical model remains consistent with the original input image;

[0020] And a second loss function used to constrain the reconstruction of the intrinsic scene image to conform to the statistical properties of natural images.

[0021] In an optional embodiment, the inherent complexity of the model parameter quantity measure is quantified by calculating the total number of all trainable parameters in the encoder and generator; the uncertainty of the model output variance measure is estimated by performing multiple random forward propagations on the latent vector and statistically analyzing the pixel variance of the generated multiple images.

[0022] The second technical solution adopted in this application is: providing a low-light image generative compression system for implementing any of the aforementioned methods, comprising:

[0023] The physical sensing encoder module is configured to extract and decompose features from the input low-light image and output scene radiance components and degradation parameter components.

[0024] The intrinsic scene generator module is configured to reconstruct a visually enhanced intrinsic scene image based on the quantized scene radiative components.

[0025] The physical model calculation module is configured as a differentiable computation unit without trainable parameters, used to calculate the synthesized and reconstructed degraded image based on the intrinsic scene image and the degradation parameter components according to the physical imaging model;

[0026] The intelligent gain optimization module is configured to calculate the intelligent gain based on the information gain and the system complexity during the training phase, and adjust the network parameters accordingly to achieve the training objective.

[0027] In an optional embodiment, the operation of the physical model calculation module includes:

[0028] Convert the transmittance parameter in the degradation parameter component into a full-resolution transmittance map;

[0029] The background light estimate is read directly from the degradation parameter components;

[0030] The intrinsic scene image, the full-resolution transmittance map, and the background light estimate are substituted into the physical imaging model to calculate the reconstructed degraded image.

[0031] The third technical solution adopted in this application is: providing a computer-readable storage medium storing a computer program thereon, wherein when the computer program is executed by a processor, it implements the steps of the low-light image generative compression method described in any of the preceding claims.

[0032] Due to the adoption of the above technical solution, this application has at least one of the following beneficial effects compared with the prior art:

[0033] 1. By defining information gain as the reconstruction error in the high-level semantic feature space, rather than the traditional pixel-level error, the compression system can effectively preserve the key semantic structure and visual content of the image at extremely low bit rates, avoiding semantic distortion and improving imaging quality.

[0034] 2. Introducing intelligent gain as an optimization objective enables the network to learn the most efficient feature representation with limited resources, preventing overfitting and improving generalization ability and deployment robustness.

[0035] 3. The degradation process is explicitly modeled based on the optical physical imaging model, and the reconstructed intrinsic scene radiation image is trained by physical consistency loss to ensure that the reconstructed intrinsic scene radiation image conforms to the real illumination and scattering laws.

[0036] 4. During the decoding stage, only the scene radiative components are needed to directly output an enhanced and clear image through the generator, without the need to perform physical model inversion in real time or use degradation parameters in combination, thus simplifying the receiver process.

[0037] 5. Tested on the UIEB low-light dataset, this method achieves an average PSNR approximately 2.93 dB higher than JPEG, indicating more stable reconstructed image quality and lower noise, especially under low-light conditions. Furthermore, the bitrate (bpp) of this patented method is only about 1 / 483 that of JPEG, resulting in a compression efficiency improvement of approximately 483 times. Attached Figure Description

[0038] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0039] in:

[0040] Figure 1 A schematic flowchart of a low-light image generative compression method provided in an embodiment of this application;

[0041] Figure 2 A schematic diagram of the framework of a low-light image generative compression system provided in an embodiment of this application;

[0042] Figure 3 This is a schematic diagram of the structure of an embodiment of the computer-readable storage medium of this application. Detailed Implementation

[0043] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. It is understood that the specific embodiments described herein are only for explaining this application and not for limiting it. Furthermore, it should be noted that, for ease of description, only the parts related to this application are shown in the accompanying drawings, not all structures. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0044] The terms "first," "second," etc., used in this application are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to these processes, methods, products, or apparatuses.

[0045] In this document, the term "embodiment" means that a particular feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a separate or alternative embodiment mutually exclusive with other embodiments. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described herein can be combined with other embodiments.

[0046] Existing low-light image compression techniques generally suffer from the following drawbacks: traditional methods typically enhance before compressing, which easily introduces artifacts and results in low compression efficiency; while existing deep learning methods can jointly optimize enhancement and compression, they mostly rely on pixel-level reconstruction loss, neglecting high-level semantic fidelity, leading to severe loss of semantic information at low bitrates; furthermore, these methods lack an explicit trade-off mechanism between model complexity and information acquisition efficiency, making it difficult to achieve optimal representation with limited bits, and the decoding end often requires complex computation to recover the image, which is not conducive to practical deployment. In view of this, this application proposes a generative compression method based on physical modeling, feature decoupling, intelligent gain optimization, and lightweight decoding, such as... Figure 1 As shown, Figure 1 A flowchart illustrating a low-light image generative compression method according to an embodiment of this application includes the following steps:

[0047] A physical imaging model for low-light images is established, which expresses the observed degraded image as a function of four factors: scene radiance, transmittance, global background light, and additive noise. In a specific embodiment of the present invention, the physical imaging model for low-light images is constructed as follows:

[0048] The observed degradation pattern is denoted as I(x), which is determined by four physical factors:

[0049] Scene radiance J(x): represents the true, sharp image that the camera should capture under ideal, non-degradable conditions; transmittance t(x)∈(0,1]: reflects the degree of light attenuation due to absorption and scattering by the medium during propagation, and is related to the depth of field d(x) and the medium attenuation coefficient β, such as t(x)=e -βd(x) Global background light The uniform color shift component formed by the forward scattering of ambient light in the medium is usually modeled as an RGB three-channel constant; additive noise N(x) includes random interference such as sensor noise and quantization error.

[0050] In the physical imaging model, a degraded image is formed by element-wise multiplying the scene radiance and transmittance, adding the product of the global background light and the transmittance minus one, and then adding additive noise; that is:

[0051]

[0052] in, Represents the element-wise (Hadamard) product, used to synthesize large-scale paired training data (I is the degenerate input, J is the clear label), and can also be embedded as a differentiable module in end-to-end networks to provide physical constraints during training.

[0053] Clearly distinguishing between the essential content in the image and environmental degradation factors provides a theoretical basis for designing a dual-branch structure for the subsequent encoder (extracting scene radiation components and degradation parameter components separately). The generated scene radiation and estimated degradation parameters are recombined into a reconstructed image through the physical model calculation module and compared with the degradation input I to form a strong supervision signal, ensuring that the reconstructed image conforms to the real optical laws and avoiding color distortion or unnatural enhancement.

[0054] A physical sensing encoder is used to process the input low-light image to separate the scene radiance component, which represents the essential content of the scene, and the degradation parameter component, which represents the degradation factors of the imaging environment. In one embodiment, the physical sensing encoder uses a shared backbone network to extract basic features, and then outputs the scene radiance component through a structurally independent first branch network and the degradation parameter component through a structurally independent second branch network.

[0055] The degradation parameter components output by the second branch network are further decoupled into a transmittance parameter vector for generating a transmittance map and a background light parameter vector for directly estimating the background light color.

[0056] The physical sensing encoder is constructed as a shared trunk + two-branch decoupled architecture for processing the input low-light image.

[0057] Shared backbone network: A multi-layer convolutional neural network (such as ResNet-34 or EfficientNet) is used to extract general features from the input image and generate high-dimensional, multi-scale intermediate feature representations to ensure that the two branches are based on a consistent semantic foundation for subsequent specific modeling.

[0058] The first branch network, also known as the scene radiation branch, consists of several fully connected layers or global pooling + MLP, and outputs a compact latent vector. This vector encodes only the essential scene content that is unrelated to imaging degradation, i.e., the semantic and structural information of an ideal, sharp image.

[0059] The second branch network, also known as the degenerate parameter branch, is structurally completely independent of the first branch. Its output is further subdivided into two sub-components:

[0060] Transmittance parameter vector The data is fed into a lightweight deconvolutional network (e.g., 4 layers of transposed convolution) to generate a low-resolution transmittance map, which is then bilinearly upsampled to obtain the full-resolution t(x)∈[0,1]. H×W ;

[0061] Background light parameter vector B = (B) is directly used as the global background light estimate for the RGB three channels. R B G B B No spatial variation modeling is required.

[0062] During the training phase, Z J Input intrinsic scene generator G J Obtain the reconstructed image J rec And Z t and Z b With J rec Substitute them together into the physical imaging model End-to-end joint optimization is achieved by calculating the loss with the original input I.

[0063] By separating the structure, the network is forced to model scene content and environmental degradation separately, avoiding feature mixing and making the compressed representation more physically meaningful and interpretable. The degradation parameters are further decoupled into two independent sub-tasks: transmittance and background light, which is consistent with the optical imaging mechanism and helps the network to learn its respective parameters more focused and accurately, thus improving its robustness in complex underwater or hazy scenes.

[0064] The network is optimized with the goal of maximizing the intelligent gain, where the intelligent gain is the ratio of information gain to system complexity. Information gain is measured by the difference between the reconstructed intrinsic scene image and the real intrinsic scene image in the high-level semantic feature space. System complexity is a weighted combination of the inherent complexity of the model parameter quantity measure and the uncertainty of the model output variance measure.

[0065] In the joint training step, the overall objective function of optimization is composed of a weighted sum of a physical consistency loss term and a smart gain entropy loss term, where the smart gain entropy loss term is the negative logarithm of the smart gain.

[0066] The physical consistency loss term includes:

[0067] The first loss function is used to ensure that the reconstructed image synthesized based on the physical model remains consistent with the original input image;

[0068] And a second loss function used to constrain the reconstruction of the intrinsic scene image to conform to the statistical properties of natural images.

[0069] In the measurement of system complexity, the inherent complexity of the model parameter quantity is quantified by calculating the total number of all trainable parameters in the encoder and generator; the uncertainty of the model output variance is estimated by performing multiple random forward propagations on the latent vector and statistically analyzing the pixel variance of the generated images.

[0070] The intelligent gain dL is defined as the ratio of information gain δS to system complexity R. The core objective of training is to maximize the intelligence gain, which means pursuing the highest possible information recovery quality at the lowest possible system complexity cost.

[0071] Information gain does not use pixel-level MSE, but rather calculates the reconstructed intrinsic scene image J. rec Compared with real intrinsic scene images J gt Differences in high-level semantic feature spaces; for example, using a deep convolutional neural network (such as VGG19, ResNet50) pre-trained on a large image dataset (such as ImageNet) as a feature extractor. Then we have: This ensures that the optimization direction is semantic fidelity, rather than simply pixel alignment.

[0072] The system complexity R is the inherent complexity Rn. capacity With output uncertainty R uncertainty Weighted sum:

[0073] R = w1R capacity +w2R uncertainty

[0074] Among them, w1 and w2 are weighting coefficients. It should be noted that this application does not limit the specific values ​​of the weighting coefficients.

[0075] Inherent complexity R capacity During training, the total number of trainable parameters (weights, biases) in the physical perception encoder and intrinsic scene generator is dynamically counted; the output uncertainty R is... uncertainty To radiate the latent vector Z of the scene during training. J A series of reconstructed images are generated by performing K (e.g., K=5) random forward propagations (via Dropout or random noise injection), and the average of their pixel variances is calculated to estimate the gain. This definition collectively considers the expressive power of the model and the confidence of its predictions, providing a robust complexity benchmark for the computation of smart gains.

[0076] Total loss function L total It is a weighted sum of multiple constraints, and its expression is L. total =λ phys ·L phys +λ intropy ·L intropy Among them, L phys =L recon +L prior This represents a loss of physical consistency, while The loss term is derived from the intelligent gain. The training objective of the model is to minimize the total loss function. By optimizing this composite loss function, the system is guided to find an optimal compact latent representation Z that fully utilizes the physical laws and recovers the essential scene J from the degraded observed image I with the highest learning efficiency (i.e., maximizing the intelligent gain dL).

[0077] An optimization mechanism aimed at maximizing intelligent gain creatively solves the core challenge of co-optimizing quality, efficiency, and robustness in low-light image generative compression by defining the ratio of information gain to system complexity as the core optimization metric and constructing a joint objective function that integrates physical consistency constraints and intelligent gain entropy loss. This mechanism incorporates high-level semantic fidelity, inherent model complexity, and output uncertainty into a unified consideration, guiding the model to automatically learn the most information-efficient physical representation under limited resources.

[0078] At the encoding end, the scene radiative components and degradation parameter components are quantized and entropy encoded to generate a compressed bitstream; at the decoding end, the compressed bitstream is entropy decoded, and the obtained scene radiative components are input into the intrinsic scene generator to directly output an enhanced clear image.

[0079] Bitstream generation and syntax structure:

[0080] Quantized scene radiation components With degradation parameter components Entropy encoding is performed separately to generate the final compressed bitstream. The bitstream uses a layered syntax structure, which includes:

[0081] (1) Sequence header: contains image height H, width W, and quantization table identifier Q_index;

[0082] (2) Scene radiation data segment: for Use context-based adaptive binary arithmetic coding (CABAC), where the probabilistic model is updated based on the same type of components from the previous frame;

[0083] (3) Degradation parameter data segment: for Differential pulse code modulation (DPCM) followed by Huffman coding is used to improve compression efficiency by taking advantage of its time smoothness.

[0084] (4) End flag: indicates the end of the bit stream.

[0085] Quantization and entropy coding mechanisms:

[0086] The quantization employs a non-uniform quantization strategy: for scene radiative components The quantization step size is dynamically allocated based on the spatial importance of its feature map; for the degenerate parameter components... Uniform quantization with a fixed step size is adopted.

[0087] During the compression (transmitting end) process, the input low-light image I is first processed by the physical sensing encoder E. phys After processing, the scene radiative component Z is obtained. J and degradation parameter component Z tb Subsequently, quantization operations are performed on the two components respectively to obtain the discretized representation. and Next, entropy encoding is performed on the quantized vector to generate the final compressed bitstream for transmission or storage. It should be understood that the degradation parameter component Z... tb The dimension is significantly lower than that of the scene radiative component Z. J Furthermore, its values ​​have higher smoothness in the sequence, thus naturally possessing higher compressibility. This characteristic enables the system to achieve higher overall compression efficiency.

[0088] During the reconstruction (at the receiving end), entropy decoding is first performed on the received compressed bitstream to recover the quantized scene radiation components. and degradation parameter components Subsequently, the scene radiation components Input to the intrinsic scene generator G JIn the process, the reconstructed, visually enhanced intrinsically sharp image J is directly output. rec .

[0089] In summary, the low-light image generative compression method of this embodiment first establishes a physical imaging model including scene radiance, transmittance, global background light, and additive noise; then, through a physical perception encoder, decomposes the input low-light image into a scene radiance component representing the essential content and a degradation parameter component representing degradation factors; jointly optimizes the model with the goal of maximizing intelligent gain, where intelligent gain is defined as the ratio of information gain to system complexity. Information gain is measured by the reconstruction error in the high-level semantic feature space, and system complexity is composed of the number of model parameters and output uncertainty; at the encoding end, the two types of components are quantized and entropy encoded to generate bitstreams, and at the decoding end, the enhanced clear image can be reconstructed using only the scene radiance component; this application improves the visual quality of the reconstructed image at a low bit rate.

[0090] This application also provides a low-light image generative compression system that implements the low-light image production compression method of the above embodiments, such as... Figure 2 As shown, Figure 2 A schematic diagram of the framework of a low-light image generative compression system provided in an embodiment of this application includes:

[0091] The physical sensing encoder module is configured to extract and decompose features from the input low-light image and output scene radiance components and degradation parameter components.

[0092] The specific parameters and connection methods for the network structure of the physical sensing encoder are as follows:

[0093] Shared backbone network: using the first four layers of ResNet-34.

[0094] Scene radiation branch: connects to a global average pooling layer + a fully connected layer (512→256).

[0095] Degenerate parameter branch: connects to two parallel fully connected layers.

[0096] The intrinsic scene generator module is configured to reconstruct a visually enhanced intrinsic scene image based on the quantized scene radiative components.

[0097] Intrinsic Scene Generator G J Based on the U-Net architecture, it contains 4 upsampling layers, each using transposed convolution and LeakyReLU activation.

[0098] The physical model calculation module is configured as a differentiable computational unit without trainable parameters, used to calculate the synthesized and reconstructed degraded image based on the intrinsic scene image and degradation parameter components according to the physical imaging model;

[0099] The intelligent gain optimization module is configured to calculate intelligent gain based on information gain and system complexity during the training phase, and adjust network parameters accordingly to achieve the training objective.

[0100] The operation of the physics model calculation module includes:

[0101] Convert the transmittance parameter in the degradation parameter component into a full-resolution transmittance map;

[0102] The background light estimate is read directly from the degradation parameter components;

[0103] The original scene image, full-resolution transmittance map, and background light estimate are substituted into the physical imaging model to calculate the reconstructed degraded image.

[0104] The methods described in the above embodiments can be implemented as computer programs; therefore, this application proposes a computer-readable storage medium. Please refer to [link to relevant documentation]. Figure 3 , Figure 3 This is a schematic diagram of a computer-readable storage medium according to an embodiment of the present application. The computer-readable storage medium stores program data that can be executed by a processor to implement the steps of any embodiment of the low-light image production compression method described above.

[0105] In this embodiment, the computer-readable storage medium can be a USB flash drive, a portable hard drive, a read-only memory (ROM), a random access memory (RAM), a magnetic disk, or an optical disk, or a medium that can store program data. Alternatively, it can be a server that stores the program data. The server can send the stored program data to other devices for execution, or it can run the stored program data itself.

[0106] In the several embodiments provided in this application, it should be understood that the disclosed methods and devices can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of modules or units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed.

[0107] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0108] Furthermore, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit. The integrated unit can be implemented in hardware or as a software functional unit.

[0109] The above description is merely an embodiment of this application and does not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A low-light image generative compression method, characterized in that, Includes the following steps: A physical imaging model for low-light images is established, which expresses the observed degraded image as a function of four elements: scene radiance, transmittance, global background light, and additive noise. The input low-light image is processed using a physical sensing encoder to separate the scene radiative component, which represents the essential content of the scene, and the degradation parameter component, which represents the degradation factors of the imaging environment. The network is optimized with the goal of maximizing the intelligent gain, wherein the intelligent gain is the ratio of information gain to system complexity, the information gain is measured by the difference between the reconstructed intrinsic scene image and the real intrinsic scene image in the high-level semantic feature space, and the system complexity is a weighted combination of the inherent complexity of the model parameter quantity measure and the uncertainty of the model output variance measure. At the encoding end, the scene radiative component and the degradation parameter component are quantized and entropy encoded to generate a compressed bitstream; at the decoding end, the compressed bitstream is entropy decoded, and the obtained scene radiative component is input to the intrinsic scene generator to directly output the enhanced clear image.

2. The low-light image generative compression method according to claim 1, characterized in that, In the physical imaging model, the degraded image is formed by multiplying the scene radiation and the transmittance element-wise, adding the product of the global background light and a factor minus the transmittance, and then adding the additive noise.

3. The low-light image generative compression method according to claim 1, characterized in that, The physical perception encoder uses a shared backbone network to extract basic features, then outputs the scene radiation components through a structurally independent first branch network, and outputs the degradation parameter components through a structurally independent second branch network.

4. The low-light image generative compression method according to claim 3, characterized in that, The degradation parameter components output by the second branch network are further decoupled into a transmittance parameter vector for generating a transmittance map and a background light parameter vector for directly estimating the background light color.

5. The low-light image generative compression method according to claim 1, characterized in that, In the joint training step, the overall objective function of the optimization is composed of a weighted sum of a physical consistency loss term and a smart gain entropy loss term, wherein the smart gain entropy loss term is the negative logarithm of the smart gain.

6. The low-light image generative compression method according to claim 5, characterized in that, The physical consistency loss term includes: The first loss function is used to ensure that the reconstructed image synthesized based on the physical model remains consistent with the original input image; And a second loss function used to constrain the reconstruction of the intrinsic scene image to conform to the statistical properties of natural images.

7. The low-light image generative compression method according to claim 1, characterized in that, In the measurement of system complexity, the inherent complexity of the model parameter quantity measurement is quantified by calculating the total number of all trainable parameters in the encoder and generator; the uncertainty of the model output variance measurement is estimated by performing multiple random forward propagations on the latent vector and statistically analyzing the pixel variance of the generated multiple images.

8. A low-light image generative compression system implementing the method of any one of claims 1 to 7, characterized in that, include: The physical sensing encoder module is configured to extract and decompose features from the input low-light image and output scene radiance components and degradation parameter components. The intrinsic scene generator module is configured to reconstruct a visually enhanced intrinsic scene image based on the quantized scene radiative components. The physical model calculation module is configured as a differentiable computation unit without trainable parameters, used to calculate the synthesized and reconstructed degraded image based on the intrinsic scene image and the degradation parameter components according to the physical imaging model; The intelligent gain optimization module is configured to calculate the intelligent gain based on the information gain and the system complexity during the training phase, and adjust the network parameters accordingly to achieve the training objective.

9. The low-light image generation and compression system according to claim 8, characterized in that, The operation of the physical model calculation module includes: Convert the transmittance parameter in the degradation parameter component into a full-resolution transmittance map; The background light estimate is read directly from the degradation parameter components; The intrinsic scene image, the full-resolution transmittance map, and the background light estimate are substituted into the physical imaging model to calculate the reconstructed degraded image.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by a processor, it implements the steps of the low-light image generative compression method as described in any one of claims 1 to 7.

Images