A low-light image compression method and system based on an intelligent diffusion model
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NINGBO KANGDA KAINENG MEDICAL TECH CO LTD
- Filing Date
- 2025-12-31
- Publication Date
- 2026-05-26
Smart Images
Figure CN122093585A_ABST
Abstract
Description
Technical Field
[0001] This application relates to the fields of image compression and computational photography, and in particular to a low-light image compression method and system based on an intelligent diffusion model. Background Technology
[0002] Traditional image compression standards (such as JPEG and HEVC) are designed for natural images under normal lighting conditions, and their statistical models are severely inconsistent with the source characteristics of low-light images. Low-light images have a low signal-to-noise ratio, and their high-frequency energy mainly comes from noise rather than real texture. Traditional encoders cannot effectively distinguish between the two, wasting a large amount of bitrate on compressing noise. This results in the loss of real image details due to insufficient bitrate allocation, leading to a poor visual effect with both blockiness and blurriness.
[0003] Currently, low-light image compression methods can be mainly divided into three categories:
[0004] 1. The preprocessing-based enhancement compression method adopts a concatenated architecture of enhancement before encoding, but it suffers from the problem of enhanced artifact interference and compression efficiency.
[0005] 2. The joint optimization method based on the transform domain can perform denoising and quantization simultaneously in the wavelet or curve domain, but it has poor adaptability to complex noise patterns.
[0006] 3. Feature compression methods based on traditional machine learning have limited model expressive power and weak generalization ability. These methods share the scientific challenge of accurately establishing statistical models for noise, and cannot effectively distinguish the energy distribution of signal and noise during the compression process. Summary of the Invention
[0007] To address the aforementioned issues, this application provides a low-light image compression method and system based on an intelligent diffusion model. By optimizing the adaptive quantization and conditional diffusion reconstruction model through an intelligent gain criterion, it collaboratively achieves efficient compression and high-quality generation, realizing high-fidelity compression and reconstruction of low-light images at extremely low bit rates, effectively suppressing noise and restoring details.
[0008] The first technical solution adopted in this application is: providing a low-light image compression method based on an intelligent diffusion model, including the following steps:
[0009] The encoding steps include:
[0010] The input low-light image is mapped to a low-dimensional latent space by an encoder to obtain a latent feature representation;
[0011] The latent feature representation is subjected to noise-aware adaptive quantization to obtain discretized quantization coefficients;
[0012] The quantization coefficients are entropy encoded to generate a compressed bitstream for storage or transmission;
[0013] The decoding process includes:
[0014] The received compressed bitstream is entropy decoded to recover the quantization coefficients;
[0015] Using the recovered quantization coefficients as conditional inputs, an iterative denoising process is performed through a pre-trained conditional diffusion model to reconstruct clear latent feature representations in the latent space.
[0016] The reconstructed sharp latent feature representation is transformed into pixel space by a decoder to generate the final reconstructed image.
[0017] In an optional embodiment, the information gain is obtained by evaluating the perceptual similarity between the reconstructed image and the corresponding sharp reference image; the system complexity is the weighted sum of the sum of the parameters of the encoder and the conditional diffusion model and the uncertainty of the model output.
[0018] In an optional embodiment, the adaptive quantization process is a noise-aware adaptive quantization process, specifically including:
[0019] Based on the brightness information or noise model of the low-light image, estimate the noise level distribution of the image;
[0020] Based on the noise level distribution, quantization step size is dynamically assigned to feature components at different spatial locations in the latent feature representation, wherein a larger quantization step size is assigned to regions with high estimated noise levels, and a smaller quantization step size is assigned to regions with low estimated noise levels.
[0021] The latent feature representation is scalar quantized using the assigned quantization step size to obtain the discretized compressed feature representation.
[0022] In an optional embodiment, the conditional diffusion model operates in the latent space of a pre-trained variational autoencoder, and the compressed feature representations, after being processed by a conditional embedding network, serve as conditional inputs to the conditional diffusion model during the denoising process.
[0023] In an optional embodiment, the conditional diffusion model performs an iterative denoising process specifically including:
[0024] When training the denoising network, starting from the clear latent feature representation corresponding to the clear image, Gaussian noise is gradually added through a forward diffusion process to generate a noisy latent feature representation;
[0025] Train a denoising network, taking the noise latent feature representation, the diffusion time step, and the conditional input as input, to learn to predict the added noise;
[0026] When performing image compression using the method, starting from random Gaussian noise and guided by the conditional input obtained after decoding the compressed feature representation, the denoising network performs multiple rounds of iterative denoising sampling to gradually reconstruct the clear latent feature representation.
[0027] In an optional embodiment, the method further includes end-to-end joint training of the encoder, the quantizer for implementing the adaptive quantization processing, and the denoising network in the conditional diffusion model;
[0028] The overall loss function for joint training is a linear weighted sum of bitrate loss, pixel- or feature-based distortion loss, and loss terms based on the intelligent gain criterion.
[0029] In an optional embodiment, when performing the method, the conditional diffusion model employs a deterministic sampling algorithm to perform the iterative denoising process, and the number of sampling steps in the iterative denoising process is configurable.
[0030] The second technical solution adopted in this application is: providing a low-light image compression system based on an intelligent diffusion model, used to implement any of the aforementioned methods, the system comprising:
[0031] The encoding module is used to map the input low-light image to a low-dimensional latent space and output a latent feature representation.
[0032] The quantization module, connected to the encoding module, is used to perform noise-aware adaptive quantization on the latent feature representation and output a discretized compressed feature representation.
[0033] The conditional diffusion reconstruction module, connected to the quantization module, is used to perform an iterative denoising process based on the compressed feature representation, and output a reconstructed clear latent feature representation.
[0034] The decoding module, connected to the conditional diffusion reconstruction module, is used to convert the clear latent feature representation to pixel space and output the reconstructed image.
[0035] In an optional embodiment, the quantization module includes:
[0036] A noise estimation unit is used to estimate the noise level distribution map based on the input image;
[0037] A step size mapping unit, connected to the noise estimation unit, is used to generate a quantization step size mapping map corresponding to the spatial location of the latent feature representation based on the noise level distribution map;
[0038] The quantization execution unit, connected to the step size mapping unit and the encoding module, is used to quantize the latent feature representation based on the quantization step size mapping map.
[0039] In an optional embodiment, the conditional diffusion reconstruction module includes:
[0040] A conditional embedding unit is used to process the compressed feature representation into a conditional vector;
[0041] A denoising network unit is configured to perform iterative denoising prediction from noise to clear latent features, guided by the conditional vector.
[0042] The sampling scheduling unit is used to control the total number of steps and sampling strategy of the iterative denoising process.
[0043] 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:
[0044] 1. By introducing a conditional diffusion model into the latent space for reconstruction, generative priors are effectively used to recover high-frequency details lost by quantization, while suppressing the dominant noise components in low-light images, thus obtaining reconstructed images with high visual fidelity even at extremely low bit rates.
[0045] 2. An adaptive quantization strategy based on noise level distribution is adopted to dynamically adjust the quantization step size in different regions—relaxing quantization in high-noise regions and fine quantizing in low-noise or structurally critical regions to improve compression efficiency.
[0046] 3. The number of sampling steps in the conditional diffusion reconstruction module can be dynamically configured. Combined with the intelligent gain feedback mechanism, the system can automatically select the optimal operating point based on bandwidth, computing power, or latency requirements during runtime, thereby improving adaptability.
[0047] 4. Tests were conducted on the UIEB low-light dataset. Experimental results show that the proposed method compresses the image from the original 24bpp (24 bits per pixel) to 0.67bpp, achieving a compression ratio of 36:1. The SSIM value is 0.5605, indicating that the reconstructed image maintains the structural information of the image in terms of structural similarity. Attached Figure Description
[0048] 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.
[0049] in:
[0050] Figure 1A flowchart illustrating a low-light image compression method based on an intelligent diffusion model provided in an embodiment of this application;
[0051] Figure 2 This is a schematic diagram of the framework of a low-light image compression system provided in an embodiment of this application. Detailed Implementation
[0052] 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.
[0053] 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.
[0054] 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.
[0055] Existing low-light image compression techniques generally suffer from severe rate-distortion performance degradation: traditional methods (such as enhancement before compression) cannot distinguish between real signals and noise in an image, resulting in a large amount of bitrate being wasted on encoding meaningless noise, while truly useful structural information is lost due to insufficient bitrate, and reconstructed images often exhibit blurring, blockiness, and enhancement artifacts; even if some methods attempt joint optimization in the transform domain or feature domain, they still lack the ability to adaptively model noise distribution and high-quality generation priors, making it difficult to recover details at extremely low bitrates.
[0056] In view of this, this application achieves intelligent bitrate allocation through noise-aware adaptive quantization, and combines it with a conditional diffusion model optimized based on intelligent gain criteria to accurately reconstruct sharp features in the latent space. This effectively preserves and restores image details while suppressing noise, significantly improving the compression quality and efficiency of low-light images. Figure 1 As shown, Figure 1 A flowchart illustrating a low-light image compression method based on a smart diffusion model provided in an embodiment of this application includes the following steps:
[0057] The input low-light image is mapped to a low-dimensional latent space by an encoder to obtain a latent feature representation; that is, the input low-light image is first processed by a CNN-based lightweight encoder, which maps it from a high-resolution pixel space to a highly compressed low-dimensional latent space to obtain a latent feature representation z.
[0058] Adaptive quantization is performed on the latent feature representation to obtain a discretized compressed feature representation; the adaptive quantization process is a noise-aware adaptive quantization process, specifically including:
[0059] Based on the brightness information or noise model of the low-light image, the noise level distribution of the image is estimated; the local noise level of the image is analyzed by the noise estimation sub-network, such as by estimating the brightness variance, and a noise level map σ(x) is generated accordingly; the adaptive quantizer dynamically adjusts the quantization step size Δ(x)∝σ(x) based on the map.
[0060] Based on the noise level distribution, the quantization step size is dynamically assigned to the feature components at different spatial locations in the latent feature representation. Specifically, a larger quantization step size is assigned to regions with high noise levels, and a smaller quantization step size is assigned to regions with low noise levels.
[0061] In this embodiment, the quantization step size is dynamically adjusted as follows:
[0062] △(x)=△ bass ·(1+γ·σ(x))
[0063] Among them, △ bass γ is the global base step size, and γ is an adjustable hyperparameter; in other embodiments, the formula for determining the quantization step size can be selected differently, and there are no limitations on this.
[0064] In regions with high Δ(x) values, a larger quantization step size will be allocated (e.g., Δ(x) = 2Δ). bass This achieves coarse quantization; in regions with low Δ(x) values, a smaller quantization step size will be allocated (e.g., Δ(x) = 1.1Δ). bass This enables precise quantization to generate a quantization step size mapping map.
[0065] The latent feature representation is scalar-quantized using an assigned quantization step size to obtain a discretized compressed feature representation; the scalar quantization formula is as follows:
[0066]
[0067] Where round is the rounding function to transform the latent feature representation z into a discretized compressed feature representation.
[0068] Entropy coding and bitstream generation:
[0069] The discretized compressed feature representation z is obtained ^ Then, it needs to be converted into a compressed bitstream that can be used for storage or transmission. In specific implementations, context-based adaptive binary arithmetic coding (CABAC) is used for entropy coding.
[0070] Bitstream parsing and reconstruction at the decoding end:
[0071] After receiving the compressed bitstream, the decoder first parses the header information to obtain the necessary parameters. Then, using a CABAC decoder symmetrical to the encoder, and under the same context model rules, it decodes the quantization coefficients. After that, enter into The diffusion reconstruction process is conditional.
[0072] During the end-to-end training phase, to ensure that the gradient of the quantization operation is backpropagable, a differentiable approximation with added uniform noise is used to simulate the above hard quantization process:
[0073]
[0074] in To represent a uniform distribution, it is used in the forward propagation during training. Alternative For subsequent diffusion reconstruction, the gradient can be obtained during backpropagation. Smooth propagation optimizes encoder and quantization strategies.
[0075] Using compressed feature representation as a condition, an iterative denoising process is performed through a conditional diffusion model to reconstruct a clear latent feature representation. The conditional diffusion model runs in the latent space of a pre-trained variational autoencoder, and the compressed feature representation, after being processed by a conditional embedding network, serves as the conditional input of the conditional diffusion model in the denoising process.
[0076] The conditional diffusion model is preferably run in the latent space of a pre-trained VQ-GAN (a variational autoencoder); the compressed feature representation of the encoder output... It can be directly converted into a conditional vector c, or after processing through a conditional embedding network. This conditional vector c will be used in the reverse denoising process of the diffusion model, along with the noise latent variable x. t The diffusion time step t is input together into the denoising network ∈ θ This serves as a guide for noise reduction.
[0077] The iterative denoising process of the conditional diffusion model specifically includes:
[0078] When training the denoising network, we start from the clear latent feature representation corresponding to the clear image and gradually add Gaussian noise through the forward diffusion process to generate the noisy latent feature representation.
[0079] Train a denoising network with latent noise features, diffusion time steps, and conditional inputs to learn how to predict added noise;
[0080] When performing image compression, the method starts from random Gaussian noise and uses the conditional input obtained after decoding the compressed feature representation as a guide. It then performs multiple rounds of iterative denoising sampling through a denoising network to gradually reconstruct a clear latent feature representation.
[0081] The iterative denoising process is specifically divided into two stages in implementation: training and inference.
[0082] 1. During the training phase, start with a clear reference image I high Starting with the corresponding latent variable x0, through the forward process formula Add noise to construct training sample pairs {x} t ,t,c,∈}, where ∈~N(0,I),α t =1-β t ,and Denoising network ∈ θ (x t (t,c) are trained to predict added noise ∈, and the objective is to minimize L. diff =E[∣∣∈-∈ θ || 2 ].
[0083] 2. From random Gaussian noise Departure. Based on compressed features. The corresponding condition c serves as guidance, utilizing the trained denoising network ∈ θ Perform T-step iterative sampling. For example, using the DDIM sampling rule: x t-1 =DDIM-Step(x t ,t,∈ θ (x t After T steps, the reconstructed clear latent variable x0 is obtained.
[0084] The method also includes end-to-end joint training of the encoder, the quantizer for adaptive quantization, and the denoising network in the conditional diffusion model;
[0085] The overall loss function for joint training is a linear weighted sum of bitrate loss, pixel- or feature-based distortion loss, and loss terms based on the smart gain criterion.
[0086] To optimize overall performance, the encoder, quantizer, and denoising network are jointly trained end-to-end. Specifically, the overall loss function is constructed as follows:
[0087]
[0088] in, This is the estimated bitrate loss, D(I) high ,I rec ) is a pixel- or feature-based distortion loss, L intropy-diff =-log(dL)=-log(δS)+log(R) is the loss term based on the smart gain criterion, λ rate ,λ dist ,λ intropy The weights for each loss term are used, and joint training achieves synergistic optimization of compression ratio, reconstruction quality, and model efficiency.
[0089] The reconstructed sharp latent feature representation is transformed into pixel space by a decoder to generate the final reconstructed image;
[0090] Among them, the parameters of the conditional diffusion model and encoder are optimized based on the intelligent gain criterion, which is the ratio of information gain to system complexity.
[0091] Information gain is obtained by evaluating the perceptual similarity between the reconstructed image and the corresponding sharp reference image; system complexity is the weighted sum of the number of parameters of the encoder and the conditional diffusion model and the uncertainty of the model output.
[0092] Information gain δS can be specifically reflected as the denoising mean square error loss L in the training of the diffusion model. diff =E[∣∣∈-∈ θ (x t ,t,c)∣∣ 2 This loss is directly related to the accuracy of reconstruction. Perceptual similarity can also be further evaluated on the final image using metrics such as LPIPS, and integrated into the distortion loss term D(I) during end-to-end training. high ,I rec In the context of system complexity R, it is implemented, for example, by defining R = R0. capacity +λ u ·R uncertainty , where R capacityFor the number of parameters in the denoising network, R uncertainty λ is obtained by calculating the variance of the output after sampling the same input condition multiple times. u These are the weighting coefficients.
[0093] The conditional diffusion model employs a deterministic sampling algorithm to perform an iterative denoising process, and the number of sampling steps in the iterative denoising process is configurable.
[0094] The conditional diffusion model employs a deterministic sampling algorithm, such as DDIM, during inference. Crucially, the total number of iterative denoising steps, T, is designed as a dynamically adjustable parameter. The system can be preset with multiple operating modes, such as a high-quality mode (using the maximum number of steps T). high Balanced mode (using standard step count T) default ), Fast mode (using minimum number of steps T) fast The system can make decisions based on the real-time calculation of the intelligent gain dL = δS / R, and automatically select the optimal number of steps before the marginal benefit of complexity R (which is positively correlated with T) decreases, thereby achieving the optimal quality-speed tradeoff under given computing power or latency constraints.
[0095] In summary, the low-light image compression method of this embodiment includes: mapping the low-light image to a low-dimensional latent space using an encoder to obtain a latent feature representation; performing adaptive quantization on the latent feature representation to obtain a discretized compressed feature representation; using the compressed feature representation as a condition, performing an iterative denoising process through a conditional diffusion model to reconstruct a clear latent feature representation; and converting the reconstructed clear latent feature representation to pixel space using a decoder to generate a reconstructed image. The parameters of the conditional diffusion model and the encoder are optimized based on a smart gain criterion, which is the ratio of information gain to system complexity. This embodiment achieves efficient compression and high-quality reconstruction of low-light images at extremely low bit rates, effectively suppressing noise and restoring details.
[0096] This application also provides a low-light image compression system based on an intelligent diffusion model, used to implement the method described in the above embodiments, such as... Figure 2 As shown, Figure 2 A schematic diagram of the framework of a low-light image compression system provided in an embodiment of this application includes:
[0097] The encoding module is used to map the input low-light image to a low-dimensional latent space and output a latent feature representation.
[0098] The quantization module, connected to the encoding module, is used to perform noise-aware adaptive quantization on the latent feature representation and output a discretized compressed feature representation.
[0099] The conditional diffusion reconstruction module, connected to the quantization module, is used to perform an iterative denoising process conditioned on the compressed feature representation, and output a clear reconstructed latent feature representation.
[0100] The decoding module, connected to the conditional diffusion reconstruction module, is used to convert the sharp latent feature representation to pixel space and output the reconstructed image.
[0101] The quantization module includes:
[0102] A noise estimation unit is used to estimate the noise level distribution map based on the input image;
[0103] The step size mapping unit, connected to the noise estimation unit, is used to generate a quantization step size mapping map corresponding to the spatial location of the latent feature representation based on the noise level distribution map.
[0104] The quantization execution unit, connected to the step size mapping unit and the encoding module, is used to quantize the latent feature representation based on the quantization step size mapping map.
[0105] The conditional diffusion reconstruction module includes:
[0106] Conditional embedding units are used to process compressed feature representations into conditional vectors;
[0107] The denoising network unit is used to perform iterative denoising prediction from noise to clear latent features under the guidance of conditional vectors;
[0108] The sampling scheduling unit is used to control the total number of steps and sampling strategy in the iterative denoising process.
[0109] 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.
[0110] 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.
[0111] 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.
[0112] 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 compression method based on an intelligent diffusion model, characterized in that, Includes the following steps: The encoding steps include: The input low-light image is mapped to a low-dimensional latent space by an encoder to obtain a latent feature representation; The latent feature representation is subjected to noise-aware adaptive quantization to obtain discretized quantization coefficients; The quantization coefficients are entropy encoded to generate a compressed bitstream for storage or transmission; The decoding steps include: The received compressed bitstream is entropy decoded to recover the quantization coefficients; Using the recovered quantization coefficients as conditional inputs, an iterative denoising process is performed through a pre-trained conditional diffusion model to reconstruct clear latent feature representations in the latent space. The reconstructed sharp latent feature representation is transformed into pixel space by a decoder to generate the final reconstructed image.
2. The method according to claim 1, characterized in that, The information gain is obtained by evaluating the perceptual similarity between the reconstructed image and the corresponding sharp reference image; the system complexity is the weighted sum of the parameters of the encoder and the conditional diffusion model and the uncertainty of the model output.
3. The method according to claim 1, characterized in that, The adaptive quantization process is a noise-aware adaptive quantization process, specifically including: Based on the brightness information or noise model of the low-light image, estimate the noise level distribution of the image; Based on the noise level distribution, quantization step size is dynamically assigned to feature components at different spatial locations in the latent feature representation, wherein a larger quantization step size is assigned to regions with high estimated noise levels, and a smaller quantization step size is assigned to regions with low estimated noise levels. The latent feature representation is scalar quantized using the assigned quantization step size to obtain the discretized compressed feature representation.
4. The method according to claim 1, characterized in that, The conditional diffusion model operates in the latent space of a pre-trained variational autoencoder, and the compressed feature representations, after being processed by a conditional embedding network, serve as the conditional inputs of the conditional diffusion model during the denoising process.
5. The method according to claim 4, characterized in that, The iterative denoising process performed by the conditional diffusion model specifically includes: When training the denoising network, starting from the clear latent feature representation corresponding to the clear image, Gaussian noise is gradually added through a forward diffusion process to generate a noisy latent feature representation; Train a denoising network, taking the noise latent feature representation, the diffusion time step, and the conditional input as input, to learn to predict the added noise; When performing image compression using the method, starting from random Gaussian noise and guided by the conditional input obtained after decoding the compressed feature representation, the denoising network performs multiple rounds of iterative denoising sampling to gradually reconstruct the clear latent feature representation.
6. The method according to claim 5, characterized in that, The method further includes end-to-end joint training of the encoder, the quantizer for implementing the adaptive quantization processing, and the denoising network in the conditional diffusion model. The overall loss function for joint training is a linear weighted sum of bitrate loss, pixel- or feature-based distortion loss, and loss terms based on the intelligent gain criterion.
7. The method according to claim 1, characterized in that, When the method is executed, the conditional diffusion model uses a deterministic sampling algorithm to perform the iterative denoising process, and the number of sampling steps in the iterative denoising process is configurable.
8. A low-light image compression system based on an intelligent diffusion model, characterized in that, The system for implementing the method as described in any one of claims 1 to 7 comprises: The encoding module is used to map the input low-light image to a low-dimensional latent space and output a latent feature representation. The quantization module, connected to the encoding module, is used to perform noise-aware adaptive quantization on the latent feature representation and output a discretized compressed feature representation. The conditional diffusion reconstruction module, connected to the quantization module, is used to perform an iterative denoising process based on the compressed feature representation, and output a reconstructed clear latent feature representation. The decoding module, connected to the conditional diffusion reconstruction module, is used to convert the clear latent feature representation to pixel space and output the reconstructed image.
9. The system according to claim 8, characterized in that, The quantization module includes: A noise estimation unit is used to estimate the noise level distribution map based on the input image; A step size mapping unit, connected to the noise estimation unit, is used to generate a quantization step size mapping map corresponding to the spatial location of the latent feature representation based on the noise level distribution map; The quantization execution unit, connected to the step size mapping unit and the encoding module, is used to quantize the latent feature representation based on the quantization step size mapping map.
10. The system according to claim 8, characterized in that, The conditional diffusion reconstruction module includes: A conditional embedding unit is used to process the compressed feature representation into a conditional vector; A denoising network unit is configured to perform iterative denoising prediction from noise to clear latent features, guided by the conditional vector. The sampling scheduling unit is used to control the total number of steps and sampling strategy of the iterative denoising process.