Method and system for real degraded image super-resolution reconstruction based on noise schedule

Abstract

The application provides a real degraded image super-resolution reconstruction method and system based on a noise schedule, which comprises the following steps: obtaining a real degraded low-resolution image to be reconstructed, determining a DiT model that has completed pre-training; obtaining corresponding noise schedule parameters and bit budget constraints in the inference process of the DiT model; constructing a time step complexity parameter based on the noise schedule parameters; generating a time step precision table based on the time step complexity parameter under the bit budget constraint; configuring the quantization bits of the quantizer in the DiT model based on the time step precision table and the activation saturation rate; performing diffusion inference quantization processing on the real degraded low-resolution image by using the DiT model; completing the diffusion inference of the DiT model in all time steps, and outputting a high-resolution reconstructed image. The application can improve the low-bit inference stability and robustness under the condition of real degradation input or distribution offset, and improve the image reconstruction quality in the reconstruction process of the real degraded low-resolution image.

Description

Technical Field

[0001] This application relates to the fields of artificial intelligence and computer vision technology, and more specifically, to a method and system for super-resolution reconstruction of real-world degraded images based on noise patterns. Background Technology

[0002] Realistic degraded image super-resolution tasks aim to recover high-resolution images from low-resolution inputs containing complex degradation. Such inputs often contain multiple factors simultaneously, including noise, blur, compression artifacts, and non-ideal downsampling, leading to higher sensitivity of the model to detail texture, edge structure, and overall consistency during inference. In recent years, diffusion-based super-resolution methods, especially those using Diffusion Transformer (DiT) as the backbone, have achieved good results in realistic degraded super-resolution tasks. However, they suffer from high computational cost and latency, often requiring low-bit quantization at the deployment end to meet computational and storage constraints. The DiT model can employ the following techniques: 1) Fixed-precision quantization inference: Using the same number of activation quantization bits (e.g., 4 bits or 6 bits throughout) for all diffusion time steps; 2) Time-step mixed precision based on statistics / sensitivity evaluation: Assigning different precisions to different time steps by collecting activation statistics (e.g., extreme values, kurtosis, distribution shape, etc.) or calculating sensitivity scores; 3) Time-step scheduling based on optimization solution: Treating time-step bit allocation as a constrained optimization problem, using search, dynamic programming, or other solution strategies to obtain a piecewise or stepwise precision plan. However, the above techniques still have the following shortcomings in the quantization deployment of DiT real degenerate superresolution: Fixed-precision solutions struggle to balance quality and cost: the signal state and noise intensity differ significantly at different time steps in diffusion inference. Using uniformly low bits can easily introduce excessive quantization errors at critical time steps, causing texture loss, structural shifts, or artifacts. While using uniformly high bits can improve quality, it increases inference costs, making it difficult to meet the needs of edge or high-throughput deployments.

[0003] Statistics-driven approaches rely on data acquisition and have poor transferability: existing time-step mixed precision often requires the acquisition of activation statistics on specific datasets or specific degradation distributions, resulting in additional calibration costs; when the input degradation type, distribution, or model structure changes, the statistics and configuration may become invalid, requiring re-acquisition and parameter tuning, resulting in weak engineering reusability.

[0004] Optimizing the solution-based scheduling is complex and has a heavy deployment chain: time-step accuracy plans obtained based on methods such as search or dynamic programming usually require additional offline solution processes and multi-parameter constraint processing; in actual deployment, stability and budget control must also be taken into account, making the overall chain complex and difficult to quickly adapt to different models, samplers or inference platforms.

[0005] Lack of closed-loop adaptive mechanism in the inference process: Most existing solutions adopt static precision planning after deployment and cannot self-correct based on abnormal phenomena such as activation overflow / truncation during the inference process, which may lead to unstable outputs even with extremely degraded inputs or distribution offsets.

[0006] A literature search of existing technologies revealed a Chinese patent with publication number CN119722506A, which proposes a time-step adaptive diffusion model training post-quantization method. This method uses a weighted average to determine the reparameterization coefficients, reparameterizes the activation values, and then performs quantization to further adapt to the differences in activation distribution at different time steps, thus reducing quantization errors. However, this approach relies on additional calibration data, has weak engineering reusability, and lacks a closed-loop adaptive mechanism for the inference process.

[0007] Therefore, there is an urgent need for a real-degraded image super-resolution reconstruction method and system that can construct an accuracy plan by utilizing the inherent time step information in the diffusion inference process without relying on or minimizing additional calibration data, and support lightweight feedback adjustments during inference, thereby achieving high-quality, stable, and easily deployable real-degraded image super-resolution inference under low-bit conditions. Summary of the Invention

[0008] To address the shortcomings of existing technologies, the purpose of this application is to provide a method and system for super-resolution reconstruction of real degraded images based on noise schedules.

[0009] According to a first aspect of this application, a method for super-resolution reconstruction of real-world degraded images based on a noise schedule is provided, comprising: Obtain the real degraded low-resolution image to be reconstructed and determine the pre-trained DiT model; Obtain the noise schedule parameters and bit budget constraints corresponding to the inference process of the DiT model; Construct time step complexity parameters based on the noise schedule parameters; Under the bit budget constraint, a time step precision table is generated based on the time step complexity parameter; Based on the time step precision table and activation saturation rate, the quantization bits of the quantizer in the DiT model are configured, and the DiT model with the quantizer configured is used to perform diffusion inference quantization processing on the real degraded low-resolution image. Complete the DiT model diffusion inference at all time steps and output a high-resolution reconstructed image.

[0010] Optionally, obtaining the noise schedule parameters and bit budget constraints corresponding to the DiT model inference process includes: Define time step index set Where t is the time step number, T This represents the total number of time steps. Load a sampler configuration for the DiT model, the sampler being used to schedule the DiT model to perform noise prediction at each time step, the sampler configuration including noise schedule parameters for each time step; Define the target average activation bits and the preset bit budget set.

[0011] Optionally, constructing the time step complexity parameter based on the noise schedule parameter includes: At least two noise schedule features are extracted from the noise schedule parameters, including noise intensity, step intensity, and signal retention intensity. The extracted noise schedule features are used to construct a time step complexity parameter, the expression of which is as follows: or, or, in, Let be the noise intensity at time step t; The signal strength is preserved at time step t; Let be the step strength at time step t; Let t be the time step complexity parameter for time step t; This indicates a preset positive number.

[0012] Optionally, generating a time-step precision table based on the time-step complexity parameter under the bit budget constraint includes: A complexity parameter sequence is constructed using all time step complexity parameters, where each element corresponds to a time step. The elements in the complexity parameter sequence are sorted in descending order of value to obtain a descending complexity parameter sequence. The descending complexity parameter sequence is then divided into high-sensitivity intervals, medium-sensitivity intervals, and low-sensitivity intervals according to the sorting order, where each sensitivity interval is a continuous and non-overlapping numerical interval. Each element within the same sensitive interval is assigned the same activation bit value for its corresponding time step. The activation bit value is the largest for the high-sensitivity interval and the smallest for the low-sensitivity interval. The activation bit values ​​for each sensitive interval are selected from a preset bit budget set. Calculate the arithmetic mean of the activation bit values ​​corresponding to all time steps; when the arithmetic mean is greater than the target average activation bit, select the time step with the smallest time step complexity parameter that has not yet undergone activation bit adjustment, and adjust its corresponding activation bit value to the next lower bit value in the preset bit budget set; when the arithmetic mean is less than the target average activation bit, select the time step with the largest time step complexity parameter that has not yet undergone activation bit adjustment, and adjust its corresponding activation bit value to the previous higher bit value in the preset bit budget set; repeat the above activation bit adjustment process until the arithmetic mean of the activation bit values ​​corresponding to all time steps equals the target average activation bit; A time step precision table is constructed based on the correspondence between each time step and its corresponding activation bit value.

[0013] Optionally, configuring the quantization bits of the quantizer within the DiT model based on the time-step precision table and activation saturation rate, and performing diffusion inference quantization processing on the real degraded low-resolution image using the DiT model with the quantizer configured, includes: The real degraded low-resolution image is input into the DiT model. When the diffusion inference enters the current time step, the activation bit corresponding to the current time step is read from the time step precision table. The activation bit is used as the quantization bit of the quantizer in the DiT model to complete the real-time configuration of the quantizer. Initiate forward propagation computation of the DiT model at the current time step; during forward propagation, at the preset quantization point of the DiT model, perform activation quantization computation using the configured quantizer; During the activation quantization calculation process, the activation saturation rate corresponding to the current time step is obtained, and the quantization bits corresponding to the current time step are dynamically corrected according to the activation saturation rate and a preset adjustment rule. Based on the corrected quantization bits, complete the forward propagation and sampling update of the DiT model at the current time step.

[0014] Optionally, the quantization point includes a linear projection layer of a multi-head attention module, an output layer of a feedforward neural network, an input end of a residual connection branch, and an output end of a residual connection branch.

[0015] Optionally, the step of dynamically correcting the quantization bits corresponding to the current time step according to the activation saturation rate and a preset adjustment rule includes: When the activation saturation rate of the current time step is greater than the upper threshold of the preset saturation rate, and the quantization bit corresponding to the current time step is less than the maximum bit value in the preset bit budget set, the quantization bit corresponding to the current time step is adjusted to the higher bit value of the next level in the preset bit budget set; when the activation saturation rate of the current time step is less than the lower threshold of the preset saturation rate, and the quantization bit corresponding to the current time step is greater than the minimum bit value in the preset bit budget set, the quantization bit corresponding to the current time step is adjusted to the lower bit value of the next level in the preset bit budget set; when the activation saturation rate of the current time step is between the lower threshold of the preset saturation rate and the upper threshold of the preset saturation rate, the quantization bit corresponding to the current time step remains unchanged. When the quantization bit corresponding to the current time step is increased, the time step with the smallest time step complexity parameter and which has not yet performed quantization bit correction is selected in the subsequent time steps, and its corresponding activation bit is adjusted to the next lower bit value in the preset bit budget set; when the quantization bit corresponding to the current time step is decreased, the time step with the largest time step complexity parameter and which has not yet performed quantization bit correction is selected in the subsequent time steps, and its corresponding activation bit is adjusted to the previous higher bit value in the preset bit budget set, so that the arithmetic mean of the quantization bits corresponding to all time steps is equal to the target average activation bit.

[0016] According to a second aspect of this application, a real-degraded image super-resolution reconstruction system based on a noise schedule is provided, comprising: The acquisition module is used to acquire the real degraded low-resolution image to be reconstructed and to determine the completed pre-trained DiT model; The super-resolution parameter extraction module is used to obtain the noise schedule parameters and bit budget constraints corresponding to the inference process of the DiT model; A complexity metric construction module is used to construct time step complexity parameters based on the noise schedule parameters. The precision table generation module is used to generate a time step precision table based on the time step complexity parameter under the bit budget constraint. The quantization module is used to configure the quantization bits of the quantizer in the DiT model based on the time step precision table and activation saturation rate, and to perform diffusion inference quantization processing on the real degraded low-resolution image using the DiT model with the quantizer configured. The reconstruction module is used to complete the DiT model diffusion inference at all time steps and output a high-resolution reconstructed image.

[0017] According to a third aspect of this application, a non-transitory computer-readable storage medium is provided, on which a computer program is stored, which, when executed by a processor, implements the steps of a noise-based real-degraded image super-resolution reconstruction method provided in the first aspect of this application.

[0018] According to a fourth aspect of this application, an electronic device is provided, comprising: At least one memory for storing program instructions; At least one processor is configured to invoke program instructions stored in the memory and execute the steps of the real degraded image super-resolution reconstruction method based on noise schedule provided in the first aspect of this application, according to the obtained program instructions.

[0019] This application provides a super-resolution reconstruction method for real-degraded images based on a noise schedule. It utilizes the inherent noise schedule parameters of the diffusion sampling process to construct time-step complexity parameters, eliminating the need for activation data statistics, distribution assumptions, or dynamic programming solutions. This allows for rapid acquisition of time-step level accuracy data with limited calibration data. A time-step accuracy table is generated under a fixed bit budget constraint, and lightweight closed-loop correction is performed during inference through activation saturation rate feedback. This improves the stability and robustness of low-bit inference under real-degraded inputs or distribution offsets, solving the technical problems of existing static accuracy plans being unable to self-correct and prone to artifacts and unstable outputs under extreme degradation or abnormal inputs. Therefore, it improves image reconstruction quality during the reconstruction of real-degraded low-resolution images.

[0020] Other technical effects resulting from the additional features will be further illustrated in the corresponding embodiments. Attached Figure Description

[0021] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart of a real degraded image super-resolution reconstruction method in one embodiment of this application; Figure 2 Comparison of super-resolution reconstruction results of an embodiment of this application and a comparative scheme. Figure 1 ; Figure 3 Comparison of super-resolution reconstruction results of an embodiment of this application and a comparative scheme. Figure 2 ; Figure 4 This is a schematic diagram of a real degraded image super-resolution reconstruction system in one embodiment of this application. Detailed Implementation

[0022] The present application will now be described in detail with reference to specific embodiments. These embodiments will help those skilled in the art to further understand the present application, but do not limit the present application in any way. It should be noted that those skilled in the art can make several modifications and improvements without departing from the concept of the present application, and these all fall within the protection scope of the present application. Parts not described in detail in the following embodiments can be implemented using existing technology.

[0023] It should be noted that all information (including but not limited to user device information, user personal information, etc.) and data (including but not limited to data used for analysis, data stored, data displayed, etc.) involved in this application are information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of related data must comply with relevant regulations.

[0024] The task of super-resolution reconstruction of real-degraded images aims to recover high-resolution images from low-resolution inputs containing complex degradation. Such inputs often contain multiple factors simultaneously, including noise, blurring, compression artifacts, and non-ideal downsampling. DiT models still suffer from the following drawbacks in the quantization deployment of real-degraded super-resolution: fixed-precision schemes struggle to balance quality and cost; statistics-driven schemes rely on data acquisition and have poor transferability; optimization-based scheduling is complex and deployment is cumbersome; and there is a lack of closed-loop adaptive mechanisms for the inference process. Based on these problems, this application provides a real-degraded image super-resolution reconstruction method based on a noise schedule to address the aforementioned issues.

[0025] Reference Figure 1 As shown, this application provides a method for super-resolution reconstruction of real-world degraded images based on noise schedules, including: S1. Obtain the real degraded low-resolution image to be reconstructed and determine the pre-trained DiT model; S2. Obtain the noise schedule parameters and bit budget constraints corresponding to the DiT model inference process; S3. Construct time step complexity parameters based on noise schedule parameters; S4. Under bit budget constraints, generate a time step precision table based on the time step complexity parameter; S5. Based on the time step precision table and activation saturation rate, configure the quantization bits of the quantizer in the DiT model, and use the DiT model with the quantizer configured to perform diffusion inference quantization processing on the real degraded low-resolution image. S6. Complete the DiT model diffusion inference for all time steps and output a high-resolution reconstructed image.

[0026] Optionally, the realistically degraded low-resolution image has at least one realistic scene degradation type, which may include one or more combinations of camera shake blur, out-of-focus blur, compression distortion, noise interference, and chromatic aberration.

[0027] In some embodiments, the resolution of the real degraded low-resolution image may be no higher than 128×128 pixels, and the resolution of the high-resolution reconstructed image may be no lower than 512×512 pixels.

[0028] The embodiments described above utilize the inherent noise schedule parameters of the diffusion sampling process to construct time step complexity parameters, eliminating the need for activation data statistics, distribution assumptions, or dynamic programming solutions. This enables the rapid acquisition of time step-level accuracy data with limited calibration data. A time step accuracy table is generated under fixed bit budget constraints, and lightweight closed-loop correction is performed during inference through activation saturation rate feedback. This improves the stability and robustness of low-bit inference under real degradation inputs or distribution offsets, solving the technical problems of static accuracy plans being unable to self-correct and easily generating artifacts and unstable outputs under extreme degradation or abnormal inputs in the prior art. Consequently, it improves image reconstruction quality during the reconstruction of real degradation low-resolution images.

[0029] In some specific embodiments of this application, obtaining the noise schedule parameters and bit budget constraints corresponding to the DiT model inference process may further include: S21. Define the time step index set Where t is the time step number, T This represents the total number of time steps. S22. Load the sampler configuration for the DiT model. The sampler is used to schedule the DiT model to perform noise prediction at each time step. The sampler configuration includes the noise schedule parameters for each time step. S23. Define the target average activation bits and the preset bit budget set.

[0030] For example, the above steps are used to obtain the diffusion inference configuration and budget constraints of the DiT model as parameters for the diffusion-based super-resolution inference process to be deployed. The diffusion-based super-resolution inference process parameters include: 1) Total number of time steps and time step index set ; 2) Sampler or solver type and its corresponding noise schedule parameters The sampler / solver is independent of the DiT model and is used to schedule the DiT model to perform noise prediction at each time step during the diffusion inference process. The sampler or solver type is used to determine the inverse denoising iterative solution algorithm used for diffusion inference. The noise schedule parameters include the noise intensity, attenuation coefficient, step strength, signal retention strength, etc. at each time step. 3) Target average activation bits ,in, The possible value is 4 bits; 4) Preset bit budget set ,in, For the i-th element in the preset bit budget set, The minimum value in the preset bit budget set. The maximum value in the preset bit budget set, for example, the preset bit budget set. ,in, , ; Among them, noise schedule parameters It can be derived from the preset schedule of the DiT model or the calculation results inside the sampler, without the need for additional calibration data.

[0031] In some specific embodiments of this application, constructing time step complexity parameters based on noise schedule parameters may further include: S31. Extract at least two noise schedule features from the noise schedule parameters. The noise schedule features include noise intensity, step intensity, and signal retention intensity. S32. Construct the time step complexity parameter using the extracted noise schedule features. The expression for the time step complexity parameter is as follows: or, or, in, Let be the noise intensity at time step t; The signal strength is preserved at time step t; Let be the step strength at time step t; Let t be the time step complexity parameter for time step t; Indicates a preset positive number. The value of is usually between 1e-8 and 1e-10 (which can be fine-tuned depending on the scenario) to avoid the denominator being zero and to improve numerical stability.

[0032] For example, for each time step Based on the noise schedule parameters of that time step Calculate the time step complexity parameter To characterize the sensitivity of this time step to quantization error, specifically including: Step 3.1: Determine noise schedule characteristic quantities: from noise schedule parameters Extract at least one noise percentage feature and a step intensity feature, for example: Noise proportion feature: This indicates the noise intensity at that time step. Indicates the signal retention strength; Step intensity type features: This indicates the step strength at that time step, used to characterize the update step size or equivalent update strength.

[0033] Step 3.2: Constructing sortable time step complexity parameters: Constructing time step complexity parameters based on feature quantities To make it comparable and sortable, it can take any of the following forms: ; ; ; Multiple time step complexity parameters These can be combined to form a sequence of complexity parameters.

[0034] The embodiments described above address the problems of existing time-step mixed accuracy relying on statistical acquisition / calibration data and having poor transferability. For real-world degradation super-resolution scenarios where degradation types and data distributions are highly variable, and calibration data is difficult to obtain or reuse, this application directly uses the inherent parameters in diffusion inference (i.e., the noise schedule parameter) to construct the time-step complexity parameter. This eliminates the need for acquiring activation statistics and avoids the statistical activation data or distribution fitting commonly used in existing methods to characterize sensitivity. This application provides a time-step accuracy generation mechanism that requires or minimizes additional calibration through the noise schedule parameter, enabling more stable use of accuracy plans across different data distributions.

[0035] This application employs a time-step complexity construction technique based on diffused noise schedule (generating sensitivity indices for each time step directly from sampler schedule parameters, rather than relying on activation statistics or distribution assumptions). This technique enables the rapid acquisition of time-step level accuracy data even with little or no calibration data, solving the technical problems of existing technologies that require the collection of activation data / distribution statistics, have high calibration costs, and exhibit poor cross-data distribution mobility.

[0036] In some specific embodiments of this application, generating a time step precision table based on the time step complexity parameter under bit budget constraints may further include: S41. Construct a complexity parameter sequence using all time step complexity parameters. The elements in the complexity parameter sequence correspond to their respective time steps. Sort the elements in the complexity parameter sequence in descending order of value to obtain a descending complexity parameter sequence. Divide the descending complexity parameter sequence into high-sensitivity intervals, medium-sensitivity intervals, and low-sensitivity intervals in the sorting order. Each sensitivity interval is a continuous and non-overlapping numerical interval. S42. Assign the same activation bit value to the time step corresponding to each element in the same sensitive interval, with the activation bit value corresponding to the high sensitive interval being the largest and the activation bit value corresponding to the low sensitive interval being the smallest. The activation bit value corresponding to each sensitive interval is selected from the preset bit budget set. S43. Calculate the arithmetic mean of the activation bit values ​​corresponding to all time steps; when the arithmetic mean is greater than the target average activation bit, select the time step with the smallest time step complexity parameter that has not yet performed activation bit adjustment, and adjust its corresponding activation bit value to the next lower bit value in the preset bit budget set; when the arithmetic mean is less than the target average activation bit, select the time step with the largest time step complexity parameter that has not yet performed activation bit adjustment, and adjust its corresponding activation bit value to the previous higher bit value in the preset bit budget set; repeat the above activation bit adjustment process until the arithmetic mean of the activation bit values ​​corresponding to all time steps satisfies the target average activation bit constraint; A time step precision table is constructed based on the correspondence between each time step and its corresponding activation bit value.

[0037] Specifically, the complexity parameter sequence is formed by obtaining the time step complexity parameters corresponding to each time step and maintaining the correspondence between each time step and its corresponding time step complexity parameters to form a complexity parameter sequence. The high-sensitivity interval, medium-sensitivity interval, and low-sensitivity interval correspond to the starting segment (e.g., the first 20%), the middle segment (e.g., the middle 60%), and the ending segment (e.g., the last 20%) in the descending complexity parameter sequence, respectively. The activation bit value corresponding to the high-sensitivity interval is greater than the activation bit value corresponding to the medium-sensitivity interval, and the activation bit value corresponding to the medium-sensitivity interval is greater than the activation bit value corresponding to the low-sensitivity interval. For example, a preset bit budget set The next lower bit value of bit 5 is 4, and the previous higher bit value is 6. Considering the need for subsequent adjustment of the activation bit value, the boundary values ​​of the preset bit budget set are generally not allocated during the initial allocation of activation bit values ​​in S42; that is, no allocation is made in S42. , As the activation bit value.

[0038] For example, the above steps are based on a sequence of complexity parameters. and target average activation bits Generate time step precision table ,in, It must contain at least the activation bit for each time step. Optionally, it includes a quantization threshold / scaling parameter for that time step, specifically including the following steps: Step 4.1: Time step sorting or segmentation: sorting the complexity parameter sequence Sort by size from largest to smallest or segment by quantile to obtain several sensitivity ranges, such as high sensitivity range, medium sensitivity range, and low sensitivity range.

[0039] Step 4.2: Mapping different tiers to bit sets: Map different tiers to different active bits, for example: High-sensitivity intervals are mapped to higher bits (e.g., 6 bits); Medium-sensitive intervals are mapped to medium-sized bits (such as 4 or 5 bits). Low-sensitivity intervals are mapped to lower bits (such as 3 or 4 bits).

[0040] Step 4.3: Budget adjustment to meet the average bit constraint: Calculate the average number of bits currently allocated: in, The arithmetic mean of the activation bits across all time steps. Let be the activation bit value at time step t.

[0041] like Then, budget adjustments will be implemented: Adjust the quantile threshold (expand / shrink the high-bit interval); or perform "+1 bit / " at several time steps. "1-bit" fine-tuning, prioritizing time step complexity parameters The higher the value, the better the time step complexity parameter. The lower limit is reduced until the budget is met.

[0042] Final output time step precision table: in, This is an optional parameter used to describe the quantization threshold or scaling strategy for this step.

[0043] The embodiments described above address the problems of complex solution processes and heavy deployment chains in existing scheduling methods. Existing methods often require offline sensitivity assessments, complex search or solution processes to obtain time-step accuracy plans, which are not conducive to rapid engineering deployment. This application eliminates the need for dynamic programming or complex optimization solutions; it obtains a time-step plan that meets constraints simply by grading the time-step complexity parameters based on noise schedule parameters and adjusting the budget, making deployment more convenient. This application aims to generate time-step accuracy tables using more direct, implementable, and interpretable rules and budget control methods, simplifying the deployment process.

[0044] This application employs a time step precision table generation and adjustment technique under budget constraints (mapping time step complexity into a finite set of bits and ensuring that the average bit meets the target budget through threshold shifting or local ±1 bit fine-tuning). This technique can balance the quality of key time steps and the overall inference cost under a fixed average bit constraint, solving the technical problem in existing technologies where fixed precision quantization either results in a decrease in image quality or excessive cost, making it difficult to achieve both.

[0045] In some specific embodiments of this application, the quantization bits of the quantizer within the DiT model are configured based on the time-step precision table and the activation saturation rate, and diffusion inference quantization processing is performed on the real degraded low-resolution image using the DiT model with the quantizer configured. This may further include: S51. Input the real degraded low-resolution image into the DiT model. When the diffusion inference enters the current time step t, read the activation bit corresponding to the current time step t from the time step precision table. Use the activation bit as the quantization bit of the quantizer in the DiT model to complete the real-time configuration of the quantizer. S52. Start the forward propagation calculation of the DiT model at the current time step; during the forward propagation process, at the preset quantization point of the DiT model, use the configured quantizer to perform activation quantization calculation. S53. During the activation quantization calculation process, obtain the activation saturation rate corresponding to the current time step, and dynamically correct the quantization bits corresponding to the current time step according to the activation saturation rate and the preset adjustment rules. S54. Based on the corrected quantization bits, complete the forward propagation and sampling update of the DiT model at the current time step.

[0046] Specifically, performing activation quantization calculations using a pre-configured quantizer includes: calling a quantizer that has been configured according to the activation bit value corresponding to the current time step, performing quantization calculations on the activations flowing through the preset quantization points to obtain the quantized activations, and using the quantized activations for subsequent forward propagation calculations at the current time step; During the activation quantization calculation, the activation saturation rate is compared with a preset saturation rate threshold. Based on the comparison result, the quantization bit corresponding to the current time step is dynamically corrected, and the corrected quantization bit is used for activation quantization calculation at the subsequent preset quantization point in the current time step and for subsequent forward propagation calculation.

[0047] For example, the above steps perform diffusion-based super-resolution inference on the input low-resolution image, and at each time step t, activate bits are specified according to the time step precision table. Quantization / dequantization operations are performed on the activation values ​​of preset quantization sites. The activation values ​​of preset quantization sites refer to the intermediate feature output values ​​generated by the DiT model at pre-defined quantization positions during forward propagation, or are calculated using low-bit operators to obtain the output for that time step and advance to the next time step. This step specifically includes: Step 5.1: Loading Time Step-Level Quantization Configuration: Upon entering the current time step t, load the time step precision table. Read the activation bit (and optional parameters) Configure the quantizer's bits and threshold / scaling.

[0048] Step 5.2: Quantization Inference Execution: Perform activation quantization calculations on the preset quantization locations in the DiT model (such as attention projection, feedforward layer output, residual branches, etc.), and complete the model forward and sampling update for this time step.

[0049] In some specific embodiments of this application, the quantization point includes a linear projection layer of a multi-head attention module, an output layer of a feedforward neural network, an input end of a residual connection branch, and an output end of a residual connection branch.

[0050] In some specific embodiments of this application, during the activation quantization calculation process, the activation saturation rate corresponding to the current time step is obtained, and the quantization bits corresponding to the current time step are dynamically corrected according to the activation saturation rate and a preset adjustment rule, including: S531. During the activation quantization calculation process, obtain the current time step. The corresponding activation saturation rate; where the activation saturation rate is used to characterize the proportion of the quantized activation values ​​in the current time step that reach or exceed the preset quantization range boundary; S532, Based on the current time step The corresponding activation saturation rate, for the current time step Correct the corresponding quantization bits: When the activation saturation rate of the current time step is greater than the preset upper threshold of saturation rate, and the quantized bit corresponding to the current time step is less than the maximum bit value in the preset bit budget set, the quantized bit corresponding to the current time step is adjusted to the higher bit value of the next level in the preset bit budget set; when the activation saturation rate of the current time step is less than the preset lower threshold of saturation rate, and the quantized bit corresponding to the current time step is greater than the minimum bit value in the preset bit budget set, the quantized bit corresponding to the current time step is adjusted to the lower bit value of the next level in the preset bit budget set; when the activation saturation rate of the current time step is between the preset lower threshold of saturation rate and the preset upper threshold of saturation rate, the quantized bit corresponding to the current time step remains unchanged. When the quantization bit corresponding to the current time step is increased, the time step with the smallest time step complexity parameter and which has not yet performed quantization bit correction is selected in the subsequent time steps, and its corresponding activation bit is adjusted to the next lower bit value in the preset bit budget set; when the quantization bit corresponding to the current time step is decreased, the time step with the largest time step complexity parameter and which has not yet performed quantization bit correction is selected in the subsequent time steps, and its corresponding activation bit is adjusted to the previous higher bit value in the preset bit budget set, so that the arithmetic mean of the quantization bits corresponding to all time steps satisfies the target average activation bit constraint.

[0051] S533, upon completion of the current time step After the corresponding quantization bit correction, budget compensation adjustment is performed on the quantization bits of the remaining time steps to ensure that the overall average level of the quantization bits across all time steps meets the target average activation bit constraint; specifically: If the current time step If the corresponding quantization bit is increased, then in subsequent time steps, the time step with the smallest time step complexity parameter and which has not yet performed quantization bit correction is selected, and its corresponding activation bit is adjusted to the next lower bit value in the preset bit budget set. If the current time step If the corresponding quantization bit is lowered, then in subsequent time steps, the time step with the largest time step complexity parameter and which has not yet performed quantization bit correction is selected, and its corresponding activation bit is adjusted to the higher bit value of the previous level in the preset bit budget set. Through the above compensation adjustment, the arithmetic mean of the quantized bits corresponding to all time steps is made to satisfy the target average activation bit constraint.

[0052] For example, to improve stability under real degraded inputs or distribution shifts, this application introduces a lightweight feedback correction mechanism for the inference process, namely, closed-loop correction based on activation saturation rate feedback, including the following steps: Statistical activation saturation / truncation rate: The activation saturation rate is statistically calculated during the quantization process at time step t. : in, The set of activated elements to be quantized; This is the quantization threshold for that time step; This represents a function to count the number of elements, where, The number of activation elements whose absolute value exceeds the quantization threshold in the activation values ​​output by the preset quantization point at the current time step t. The total number of activated elements for the preset quantization site output activation values ​​at the current time step t; Triggering rules and corrective actions: like Then at least one corrective action will be performed: Increase the quantized bit at this time step, and update the quantized bit. ; Alternatively, relax the time step threshold / scaling parameter. To reduce truncation; like This can reduce the number of bits to reclaim the budget: Updated quantized bits in, The upper threshold of the saturation rate; The threshold value is the saturation rate. These are the initial quantization bits; Budget preservation: When the number of bits increases due to corrections at certain time steps, the number of bits is reduced or the segmentation threshold is adjusted for subsequent low-sensitivity time steps to ensure that the average number of bits throughout the process still meets the target average number of active bits.

[0053] The embodiments described above address the problems of static precision planning lacking inference process adaptation and being prone to instability under extreme inputs. Most existing solutions use static time-step planning, lacking runtime adaptation. Under real-world degenerate inputs or distribution offsets, phenomena such as activation truncation / overflow ratio increases may occur during inference, making it difficult for static plans to correct in a timely manner. This application aims to introduce a lightweight feedback mechanism, forming a closed loop through activation saturation rate feedback, enabling the inference process to dynamically adjust time-step precision or quantization thresholds based on the runtime state, thereby improving low-bit inference stability and output consistency.

[0054] After completing inference at all time steps, a high-resolution reconstructed image is output as the final super-resolution result.

[0055] This application relates to deep learning model inference deployment, model compression, and low-bit quantization. Specifically, this application addresses the inference process of DiT models in the Real-World Image Super-Resolution task, providing a method for dynamically adjusting the activation quantization precision according to the diffusion time step. This reduces inference computation / storage costs and improves the stability and imaging quality of low-bit inference.

[0056] This application utilizes the DiT model for inference deployment in real-world degraded image super-resolution tasks. The core of this method lies in its direct use of noise schedule parameters inherent in the diffusion sampling process to construct time-step complexity parameters, without relying on activation data statistics, distribution assumptions, or dynamic programming solutions. This generates a time-step accuracy table under a fixed target average activation bit budget, and lightweight closed-loop correction is performed during inference through activation saturation rate feedback, thereby improving the stability and reconstruction quality of low-bit inference.

[0057] This application addresses the challenge of balancing quality and cost in "fixed-precision quantization" due to the varying characteristics of different time steps in diffusion inference. Existing solutions often exhibit two extremes when using a uniform activation bit count across all time steps: uniformly low bit counts lead to excessive errors in critical time steps and degraded reconstruction quality; uniformly high bit counts result in excessive inference overhead and unacceptable deployment costs. This application aims to achieve dynamic adjustment of activation precision at each time step, satisfying the average bit budget while balancing image quality and efficiency. This application achieves adaptive allocation of time-step-level precision under a fixed average bit budget, while also minimizing calibration, facilitating deployment, and ensuring high stability.

[0058] This application employs a closed-loop adaptive correction technique based on saturation rate feedback during the inference phase (which statistically activates the truncation / overflow ratio during runtime and triggers dynamic correction of the threshold or bits while keeping the global budget within limits). This technique can improve the stability and robustness of low-bit inference under real degraded inputs or distribution offsets, and solves the technical problems in existing technologies where static precision plans cannot self-correct and are prone to artifacts and unstable outputs under extreme degradation or abnormal inputs.

[0059] The present application will be further described below with reference to specific embodiments in order to better understand the above technical solutions of the present application. It should be understood that the following are only some examples and are not intended to limit the present application.

[0060] Example 1: Deployment example of DiT inference dynamic quantization for super-resolution of real degraded images 1) Input data and experimental setup Task and Input: Super-resolution of Realistic Degraded Images ( The input is a low-resolution image containing real-world degradations (blurring, noise, compression artifacts, non-ideal downsampling, etc.). The model is a DiT super-resolution model, with a fixed number of inference steps. (like Sampler built-in noise schedule parameters Quantization objects: Several key activations during inference (such as attention projection outputs, intermediate activations in feedforward networks, residual merging points, etc.). Budget and bit set: Target average activation bits. Preset bit budget set Feedback threshold: Saturation threshold : Threshold of saturation rate .

[0061] 2) Implementation steps Step 1: Read the diffused noise schedule and determine the time step index set Load the sampler configuration on the deployment side to obtain the time step index set. and noise schedule parameters for each step (e.g., noise intensity, signal preservation factor, step size parameters, etc.).

[0062] Step 2: Calculate the complexity parameters for each time step. For each time step Construct complexity parameters based on noise schedule parameters The calculation methods are as follows (any one of them can be used): If the noise schedule parameter provides the signal retention strength With noise intensity ,but Or if the noise schedule parameter provides the step intensity With noise intensity ,but Obtain the sequence of complexity parameters Then, the images can be sorted from largest to smallest or segmented by quantile. This step can be completed without any calibration images.

[0063] Step 3: Generate a time-step precision table under the target average activation bit budget. The complexity parameter sequence in descending order It is divided into three levels: high sensitivity, medium sensitivity, and low sensitivity, and mapped to a set of bits: The highly sensitive region is set to 5 bits; The sensitive interval is set to 4 bits; The low-sensitivity interval is set to 3 bits.

[0064] Calculate the current average bits .like Execute budget adjustments: like : Start from the time step with the lowest complexity parameter and reduce the number of bits from 4 to 3; like : Start from the time step with the highest complexity parameter and increase the number of bits from 4 to 5; Until the average bit is satisfied This ultimately forms the time step precision table. .

[0065] Step 4: Perform quantization inference according to the time step precision table For each time step : 1) Read the bit for this step from the time step precision table. Configure the quantizer; 2) Perform quantization and dequantization or low-bit operator calculations on preset activation points; 3) Complete the DiT forward pass and sampling update for this time step, and proceed to the next time step.

[0066] Step 5: Saturation rate feedback correction (closed-loop adaptive) At time step Statistical saturation rate during quantitative execution That is, the proportion of activation that is truncated / overflowed: like This indicates that the quantization step was too "tight," triggering a correction. Increase the quantization bits in this step by one level (e.g., change 4 to 5), or relax the threshold for this step; like This indicates that the quantization margin in this step is relatively large, and the budget can be recovered. Decrease the bit value of this step by one level (for example, change 4 to 3).

[0067] When bits are added at certain time steps, "budget conservation" is achieved by reducing bits in subsequent low-complexity time steps, ensuring that the overall average number of bits remains at 4.

[0068] Step 6: Output the super-resolution result image After completing all time-step inference, a high-resolution result image is output, achieving high-quality reconstruction of the real degraded image.

[0069] 3) Experimental Results To verify the feasibility of this application, 200 images from the publicly available real-world degradation super-resolution test set RealLR200 were selected, with a resolution of [missing information]. A comparison will be made. The comparison methods include: Comparative Example A: Fixed 4-bit activation quantization throughout; Comparative Example B: Fixed 5-bit activation quantization throughout; This application is scheme C: This application (noise schedule driven, budget control, saturation rate feedback) averages 4 bits.

[0070] (1) Image quality indicators Comparative Example A (fixed 4 bits): MUSIQ=67.92, MANIQA=0.4079, ClipIQA=0.5215, LIQE=3.692; Comparative Example B (fixed 5 bits): MUSIQ=69.32, MANIQA=0.4307, ClipIQA=0.5351, LIQE=3.912 (higher cost) Scheme C (in this application, average 4 bits): MUSIQ = 68.99, MANIQA = 0.4268, ClipIQA = 0.5297, LIQE = 3.838 (close to 5-bit quality but cost close to 4 bits). Among them, MUSIQ, MANIQA, ClipIQA, and LIQE are all image quality evaluation metrics.

[0071] Furthermore, for all the above indicators, the larger the indicator value, the higher the image quality. According to the test results of the above image quality evaluation indicators, the image quality of the proposed solution C is similar to that of the comparative example B (fixed 5 bits), while the inference overhead is similar to that of the comparative example A (fixed 4 bits). This shows that the proposed solution C can achieve a reconstruction effect close to that of a higher bit configuration under lower cost conditions.

[0072] (2) Stability performance Figure 2 In RealLQ250:013, it refers to a real-world (REAL) low-resolution image dataset, which contains 250 images, and 013 represents the 13th image. Figure 3 In RealLQ250:155, it refers to a real-world low-resolution image dataset containing 250 images, with 155 representing image number 155. (See reference...) Figure 2 and Figure 3 As can be seen, on an input containing strong noise and compression artifacts, Comparative Example A is prone to local block artifacts and edge breaks; This application improves the accuracy of high-complexity time steps and performs closed-loop correction when the saturation rate is abnormal, resulting in a more stable output and more continuous textures and edges.

[0073] (3) Reasoning overhead On the same hardware: Comparative Example B takes approximately 15% longer to reason compared to Comparative Example A; In this application, Scheme C introduces only minimal feedback statistics overhead (counting / proportional calculation) with an average budget of 4 bits. The overall time consumption is close to that of Comparative Example A (increase of <3%), but the image quality is close to that of Comparative Example B.

[0074] The experimental results above show that this application can automatically generate a time-step accuracy plan based on the diffusion noise schedule without relying on calibration data statistics, and improve low-bit inference stability through a lightweight feedback mechanism, thereby balancing imaging quality and deployment efficiency in real degraded image super-resolution tasks.

[0075] Reference Figure 4 As shown, based on the same inventive concept, another embodiment of this application provides a real-degraded image super-resolution reconstruction system based on noise schedules. The real-degraded image super-resolution reconstruction system 100 includes: The acquisition module 110 is used to acquire the real degraded low-resolution image to be reconstructed and to determine the pre-trained DiT model. The super-resolution parameter extraction module 120 is used to obtain the noise schedule parameters and bit budget constraints corresponding to the DiT model inference process. Complexity metric construction module 130 is used to construct time step complexity parameters based on noise schedule parameters; The precision table generation module 140 is used to generate a time step precision table based on the time step complexity parameter under bit budget constraints. The quantization module 150 is used to configure the quantization bits of the quantizer in the DiT model based on the time step precision table and activation saturation rate, and to perform diffusion inference quantization processing on the real degraded low-resolution image using the DiT model with the quantizer configured. Reconstruction module 160 is used to complete the DiT model diffusion inference at all time steps and output high-resolution reconstructed images.

[0076] It should be noted that the modules in the real degraded image super-resolution reconstruction system based on noise schedule provided in the above embodiments of this application correspond to the steps of the real degraded image super-resolution reconstruction method based on noise schedule in any of the above embodiments. Those skilled in the art can refer to the step features of the real degraded image super-resolution reconstruction method based on noise schedule to implement the corresponding modules in the real degraded image super-resolution reconstruction system based on noise schedule, which will not be repeated here.

[0077] The embodiments described above in this application can realize dynamic quantization deployment of DiffusionTransformer inference for super-resolution of real degraded images. While meeting the requirements of fixed average bit budget and deployment efficiency, it improves the reconstruction quality and overall inference stability of key time steps, and overcomes the problems of strong calibration dependence, complex scheduling solution, difficulty in balancing quality and cost with fixed precision, and lack of runtime adaptability in the prior art.

[0078] In another embodiment of this application, an electronic device is also provided, including a memory and a processor; the memory is used to store program instructions; the processor is used to call the program instructions stored in the memory and execute the steps of the above-described noise-based real-degraded image super-resolution reconstruction method according to the obtained program instructions.

[0079] Optionally, the memory is used to store programs; the memory may include volatile memory, such as random-access memory (RAM), such as static random-access memory (SRAM), double data rate synchronous dynamic random-access memory (DDR SDRAM), etc.; the memory may also include non-volatile memory, such as flash memory. The memory is used to store computer programs (such as application programs and functional modules that implement the above methods), computer instructions, etc., and the aforementioned computer programs and computer instructions can be partitioned and stored in one or more memories. Furthermore, the aforementioned computer programs, computer instructions, data, etc., can be accessed by the processor.

[0080] The aforementioned computer programs, computer instructions, etc., can be stored in partitions within one or more memory locations. Furthermore, the aforementioned computer programs, computer instructions, data, etc., can be accessed by a processor.

[0081] A processor is used to execute a computer program stored in memory to implement the various steps of the methods involved in the above embodiments. For details, please refer to the relevant descriptions in the preceding method embodiments.

[0082] The processor and memory can be separate structures or integrated structures. When the processor and memory are separate structures, they can be coupled together via a bus.

[0083] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0084] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0085] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0086] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0087] The preferred features in the above embodiments can be used individually in any embodiment, or in any combination thereof, provided they do not conflict with each other. Furthermore, parts not described in detail in the embodiments can be implemented using existing technologies.

[0088] The foregoing has described some specific embodiments of this application. It should be understood that this application is not limited to the specific embodiments described above, and those skilled in the art can make various modifications or variations within the scope of the claims, which do not affect the substantive content of this application. The above-described preferred features can be used in any combination without conflict.

Claims

1. A method for super-resolution reconstruction of real-world degraded images based on noisy schedules, characterized in that, include: Obtain the real degraded low-resolution image to be reconstructed and determine the pre-trained DiT model; Obtain the noise schedule parameters and bit budget constraints corresponding to the inference process of the DiT model; Construct time step complexity parameters based on the noise schedule parameters; Under the bit budget constraint, a time step precision table is generated based on the time step complexity parameter; Based on the time step precision table and activation saturation rate, the quantization bits of the quantizer in the DiT model are configured, and the DiT model with the quantizer configured is used to perform diffusion inference quantization processing on the real degraded low-resolution image. Complete the DiT model diffusion inference at all time steps and output a high-resolution reconstructed image.

2. The method for super-resolution reconstruction of real-degraded images based on noise schedules according to claim 1, characterized in that, The process of obtaining the noise schedule parameters and bit budget constraints corresponding to the DiT model inference includes: Define time step index set Where t is the time step number, T This represents the total number of time steps. Load a sampler configuration for the DiT model, the sampler being used to schedule the DiT model to perform noise prediction at each time step, the sampler configuration including noise schedule parameters for each time step; Define the target average activation bits and the preset bit budget set.

3. The method for super-resolution reconstruction of real degraded images based on noise schedules according to claim 2, characterized in that, The construction of time step complexity parameters based on the noise schedule parameters includes: At least two noise schedule features are extracted from the noise schedule parameters, including noise intensity, step intensity, and signal retention intensity. The extracted noise schedule features are used to construct a time step complexity parameter, the expression of which is as follows: ； or, ； or, ； in, Let be the noise intensity at time step t; The signal strength is preserved at time step t; Let be the step strength at time step t; Let t be the time step complexity parameter for time step t; This indicates a preset positive number.

4. The method for super-resolution reconstruction of real-degraded images based on noise schedules according to claim 2, characterized in that, The step of generating a time-step precision table based on the time-step complexity parameter under the bit budget constraint includes: A complexity parameter sequence is constructed using all time step complexity parameters, where each element corresponds to a time step. The elements in the complexity parameter sequence are sorted in descending order of value to obtain a descending complexity parameter sequence. The descending complexity parameter sequence is then divided into high-sensitivity intervals, medium-sensitivity intervals, and low-sensitivity intervals according to the sorting order, where each sensitivity interval is a continuous and non-overlapping numerical interval. Each element within the same sensitive interval is assigned the same activation bit value for its corresponding time step. The activation bit value is the largest for the high-sensitivity interval and the smallest for the low-sensitivity interval. The activation bit values ​​for each sensitive interval are selected from a preset bit budget set. Calculate the arithmetic mean of the activation bit values ​​corresponding to all time steps; when the arithmetic mean is greater than the target average activation bit, select the time step with the smallest time step complexity parameter that has not yet undergone activation bit adjustment, and adjust its corresponding activation bit value to the next lower bit value in the preset bit budget set; when the arithmetic mean is less than the target average activation bit, select the time step with the largest time step complexity parameter that has not yet undergone activation bit adjustment, and adjust its corresponding activation bit value to the previous higher bit value in the preset bit budget set; repeat the above activation bit adjustment process until the arithmetic mean of the activation bit values ​​corresponding to all time steps equals the target average activation bit; A time step precision table is constructed based on the correspondence between each time step and its corresponding activation bit value.

5. The method for super-resolution reconstruction of real-degraded images based on noise schedules according to claim 2, characterized in that, The step involves configuring the quantization bits of the quantizer within the DiT model based on the time-step precision table and activation saturation rate, and then performing diffusion inference quantization processing on the real degraded low-resolution image using the DiT model with the quantizer configured, including: The real degraded low-resolution image is input into the DiT model. When the diffusion inference enters the current time step, the activation bit corresponding to the current time step is read from the time step precision table. The activation bit is used as the quantization bit of the quantizer in the DiT model to complete the real-time configuration of the quantizer. Initiate forward propagation computation of the DiT model at the current time step; during forward propagation, at the preset quantization point of the DiT model, perform activation quantization computation using the configured quantizer; During the activation quantization calculation process, the activation saturation rate corresponding to the current time step is obtained, and the quantization bits corresponding to the current time step are dynamically corrected according to the activation saturation rate and a preset adjustment rule. Based on the corrected quantization bits, complete the forward propagation and sampling update of the DiT model at the current time step.

6. The method for super-resolution reconstruction of real-degraded images based on noise schedules according to claim 5, characterized in that, The quantization point includes a linear projection layer of a multi-head attention module, an output layer of a feedforward neural network, an input end of a residual connection branch, and an output end of a residual connection branch.

7. The method for super-resolution reconstruction of real-degraded images based on noise schedules according to claim 2, characterized in that, The step of dynamically correcting the quantization bits corresponding to the current time step according to the activation saturation rate and a preset adjustment rule includes: When the activation saturation rate of the current time step is greater than the upper threshold of the preset saturation rate, and the quantization bit corresponding to the current time step is less than the maximum bit value in the preset bit budget set, the quantization bit corresponding to the current time step is adjusted to the higher bit value of the next level in the preset bit budget set; when the activation saturation rate of the current time step is less than the lower threshold of the preset saturation rate, and the quantization bit corresponding to the current time step is greater than the minimum bit value in the preset bit budget set, the quantization bit corresponding to the current time step is adjusted to the lower bit value of the next level in the preset bit budget set; when the activation saturation rate of the current time step is between the lower threshold of the preset saturation rate and the upper threshold of the preset saturation rate, the quantization bit corresponding to the current time step remains unchanged. When the quantization bit corresponding to the current time step is increased, the time step with the smallest time step complexity parameter and which has not yet performed quantization bit correction is selected in the subsequent time steps, and its corresponding activation bit is adjusted to the next lower bit value in the preset bit budget set; when the quantization bit corresponding to the current time step is decreased, the time step with the largest time step complexity parameter and which has not yet performed quantization bit correction is selected in the subsequent time steps, and its corresponding activation bit is adjusted to the previous higher bit value in the preset bit budget set, so that the arithmetic mean of the quantization bits corresponding to all time steps is equal to the target average activation bit.

8. A super-resolution reconstruction system for real-world degraded images based on noise schedules, characterized in that, include: The acquisition module is used to acquire the real degraded low-resolution image to be reconstructed and to determine the completed pre-trained DiT model; The super-resolution parameter extraction module is used to obtain the noise schedule parameters and bit budget constraints corresponding to the inference process of the DiT model; A complexity metric construction module is used to construct time step complexity parameters based on the noise schedule parameters. The precision table generation module is used to generate a time step precision table based on the time step complexity parameter under the bit budget constraint. The quantization module is used to configure the quantization bits of the quantizer in the DiT model based on the time step precision table and activation saturation rate, and to perform diffusion inference quantization processing on the real degraded low-resolution image using the DiT model with the quantizer configured. The reconstruction module is used to complete the DiT model diffusion inference at all time steps and output a high-resolution reconstructed image.

9. A non-transitory computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps of the method as described in any one of claims 1-7.

10. An electronic device, characterized in that, include: At least one memory for storing program instructions; At least one processor is configured to invoke program instructions stored in the memory and execute the steps of the method as described in any one of claims 1-7 according to the obtained program instructions.

Images