Multi-objective optimization image processing method, system and device and storage medium

Abstract

The invention discloses a multi-objective optimization image processing method, system and device and a storage medium, which are corresponding schemes, and in the scheme, through multi-dimensional evaluation, an image processing model realizes effective tradeoff among different optimization objectives; a preference sample pair is constructed through an evaluation result, and a preference optimization process is executed, so that a single-target optimization gradient corresponding to each evaluation target is calculated, and the optimization direction of a target processing model is guided to flexibly adapt to different application scenes and task requirements; a universal generative recovery model and a multi-objective optimization framework are adopted, so that repeated design of the model for a single task or index is avoided; according to the method, preference optimization is carried out on different evaluation targets, single-target optimization gradients are calculated, the single-target optimization gradients are further adjusted and aligned, a comprehensive optimization gradient is generated, gradient conflicts possibly generated in the multi-target optimization process are effectively reduced, the stability and the implementability of the model optimization process are improved, and the model optimization efficiency is improved. And engineering realization and deployment are facilitated.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a multi-objective optimized image processing method, system, device, and storage medium. Background Technology

[0002] Traditional image restoration methods based on convolutional neural networks or transformers typically prioritize minimizing pixel differences, optimizing the model by reducing objective metrics such as peak signal-to-noise ratio (PSNR). While these methods generally ensure numerical consistency, they often result in overly smoothed images, making it difficult to recover realistic textures and details. To improve the subjective visual experience, researchers have further introduced perceptual loss and adversarial training to make the restored results appear more natural. However, these methods still primarily optimize around a single or limited numerical objective, easily introducing unrealistic details or color deviations, and the contradiction between objective metrics and subjective perceptual metrics remains difficult to reconcile effectively.

[0003] As application scenarios become increasingly complex, image processing tasks often involve multiple evaluation criteria simultaneously, including pixel fidelity, structural consistency, subjective visual quality, and downstream task performance. These objectives are not entirely consistent, and some are difficult to accurately characterize using explicit and reliable loss functions. To address the challenge of explicitly modeling complex evaluation objectives, preference optimization methods have gained increasing attention. These methods do not rely directly on a single numerical indicator but instead transform subjective judgments into learnable supervisory signals based on human relative preferences for generated results. Direct Preference Optimization (DPO), proposed in the field of language models, directly optimizes model parameters using pairwise preference data, achieving efficient preference learning without the need for an explicit reward model. However, when this optimization objective is directly transferred to diffusion models, the differences between the model training mechanism and language models can easily introduce estimation errors, thus limiting the preference alignment effect. The direct preference optimization scheme mentioned above can be found in reference 1: R. Rafailov, A. Sharma, E. Mitchell, S. Ermon, CD Manning, and C. Finn, “Direct Preference Optimization: Your Language Model is Secretly a Reward Model,” July 29, 2024, arXiv: arXiv:2305.18290. doi: 10.48550 / arXiv.2305.18290.

[0004] To address the challenge of simultaneously balancing pixel consistency and human subjective preferences in image generation and processing using diffusion models, existing technologies have proposed improvement schemes based on preference optimization. Direct Semantic Preference Optimization (DSPO) introduces preference constraints into the training process of the diffusion model. It incorporates a semantic instance-level alignment mechanism during preference modeling and combines textual semantic descriptions to constrain key instance regions, thereby suppressing local artifacts and semantic inconsistencies in the generated results and improving the matching degree between the generated image and human subjective preferences. On the other hand, D2PO-SR addresses the high cost of manual preference annotation by fusing quality assessment signals from full-reference and no-reference images to construct preference supervision. It then forms multiple preference sample pairs based on multiple random generation results from the same model and employs a hierarchical preference optimization strategy to weight and update different preference samples, improving the model's perceptual quality and stability without relying on extensive manual annotation. Similarly, the corresponding solutions for the above two types can be found in reference 2: M. Cai et al., “DSPO: Direct Semantic Preference Optimization for Real-World Image Super-Resolution,” Apr. 21, 2025, arXiv: arXiv:2504.15176. doi: 10.48550 / arXiv.2504.15176; and reference 3: R. Wu et al., “DP2O-SR: Direct Perceptual Preference Optimization for Real-World Image Super-Resolution,” Oct. 21, 2025, arXiv:arXiv:2510.18851. doi: 10.48550 / arXiv.2510.18851.

[0005] However, most existing preference optimization methods still focus on a single task or a single generative model, typically aligning only to a single preference objective. When multiple evaluation objectives exist simultaneously, such as considering customer perception quality, subjective perception, and downstream task performance, existing methods lack a unified coordination mechanism. Optimization directions corresponding to different objectives are prone to conflict, thus affecting training stability and the final result.

[0006] Against this backdrop, multi-objective optimization offers a new approach to addressing the aforementioned problems in a unified manner. By coordinating multiple evaluation objectives and seeking trade-offs among different objectives under Pareto optimality, more realistic details can be recovered while maintaining structural consistency and taking into account the needs of downstream tasks. However, even with the introduction of preference optimization or multi-index loss, existing research still generally suffers from the following shortcomings: (1) There is a lack of a unified multi-objective preference optimization image processing system. Existing methods are mostly designed for specific models or tasks, and lack an integrated architecture that covers preference sample generation, screening and optimization, making it difficult to extend to different tasks or application scenarios.

[0007] (2) The ability to adjust the target is limited. Most methods still rely on fixed loss weights or simple weighting strategies, which are difficult to adjust flexibly according to different application needs or user preferences. There is a lack of effective coordination between the target gradients, which can easily lead to directional conflicts.

[0008] (3) Training efficiency is limited. Multi-objective independent optimization brings additional computational overhead. If there is no reasonable gradient accumulation and correction mechanism, the system is difficult to balance engineering efficiency and training stability while ensuring performance.

[0009] Therefore, there is an urgent need for a general image processing system that can uniformly handle the needs of multiple objective preferences and has adaptive trade-off capabilities.

[0010] In view of this, the present invention is hereby proposed. Summary of the Invention

[0011] The purpose of this invention is to provide a multi-objective optimized image processing method, system, device, and storage medium that can balance and take into account the multi-dimensional image optimization needs, thereby improving the visual quality and applicability of images.

[0012] The objective of this invention is achieved through the following technical solution: A multi-objective optimization image processing method, comprising: The input image is processed by multiple different generative restoration models to generate multiple candidate image samples, forming a set of processed image samples. Based on multi-dimensional evaluation objectives, candidate image samples in the processed image sample set are evaluated respectively, and the preference sample pairs corresponding to each evaluation objective are selected by comparing the evaluation results. Using the preference sample pairs corresponding to each evaluation objective, the preferences of the corresponding evaluation objectives are optimized respectively, and the optimization direction is aligned to obtain the comprehensive gradient. Then, the comprehensive gradient is used to optimize the objective processing model.

[0013] A multi-objective optimized image processing system, comprising: The preference sample generation module processes the input image using multiple different generative recovery models to generate multiple candidate image samples, forming a set of processed image samples. The multi-objective preference sample screening module evaluates candidate image samples in the processed image sample set according to multi-dimensional evaluation objectives, and selects the preference sample pairs corresponding to each evaluation objective by comparing the evaluation results. The multi-objective image processing model optimization module is used to optimize the preferences of each evaluation target by using the preference sample pairs corresponding to each evaluation target, and to align the optimization directions to obtain a comprehensive gradient, and then use the comprehensive gradient to optimize the target processing model.

[0014] A processing device includes: one or more processors; and a memory for storing one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method.

[0015] A readable storage medium storing a computer program that, when executed by a processor, implements the aforementioned method.

[0016] As can be seen from the technical solutions provided by the present invention above: (1) Multi-dimensional evaluation objectives can be considered, so that the image processing results can achieve a balance between various optimization objectives; (2) Preference sample pairs are generated through the evaluation results, and the relative information of preference samples under different optimization objectives guides the optimization direction of the target network, thereby flexibly adapting to different task requirements; (3) A general framework suitable for various image processing tasks is constructed, supporting a variety of generative recovery models and evaluation strategies, which facilitates the expansion of new indicators or preference optimization branches; (4) By calculating the preference optimization gradient under different objectives separately, and then uniformly aligning the optimization direction, the multi-objective optimization process is stable and efficient, which is convenient for engineering implementation and deployment. Attached Figure Description

[0017] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0018] Figure 1 This is a flowchart of a multi-objective optimization image processing method provided in an embodiment of the present invention.

[0019] Figure 2 This is a schematic diagram of the overall architecture of a multi-objective optimization image processing method provided in an embodiment of the present invention.

[0020] Figure 3 This is a schematic diagram of a multi-objective optimized image processing system provided in an embodiment of the present invention.

[0021] Figure 4 This is a schematic diagram of a processing device provided in an embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0023] First, the following explanations are provided for the terms that may be used in this article: The terms "comprising," "including," "containing," "having," or other similar semantic descriptions should be interpreted as non-exclusive inclusion. For example, including a technical feature element (such as raw material, component, ingredient, carrier, dosage form, material, size, part, component, mechanism, device, step, process, method, reaction conditions, processing conditions, parameter, algorithm, signal, data, product or article of manufacture, etc.) should be interpreted as including not only the expressly listed technical feature element, but also other technical feature elements that are not expressly listed and are well-known in the art.

[0024] The term "composed of" excludes any technical features not expressly listed. When used in a claim, it closes the claim to exclude all technical features other than those expressly listed, except for associated conventional impurities. If the term appears only in a clause of a claim, it limits the claim to the elements expressly listed in that clause; elements recited in other clauses are not excluded from the overall claim.

[0025] The following provides a detailed description of a multi-objective optimized image processing method, system, device, and storage medium provided by the present invention. Contents not described in detail in the embodiments of the present invention are prior art known to those skilled in the art. Where specific conditions are not specified in the embodiments of the present invention, conventional conditions in the art or conditions recommended by the manufacturer shall apply. Where the manufacturers of reagents or instruments used in the embodiments of the present invention are not specified, they are all conventional products that can be purchased commercially.

[0026] Example 1 like Figure 1 The diagram shown is a flowchart of a multi-objective optimization image processing method provided by an embodiment of the present invention, which mainly includes the following steps: Step 1: The input image is processed by multiple different generative processing models selected by the user to generate multiple candidate image samples, which are then summarized to form a set of processed image samples.

[0027] In this embodiment of the invention, the processed image sample set includes multiple sets of images obtained after processing the input image, which differ in terms of structural restoration characteristics, detail rendering degree, and color reproduction ability.

[0028] Step 2: Evaluate the candidate image samples in the processed image sample set according to the multi-dimensional evaluation objectives selected by the user, and select the preference sample pairs corresponding to each evaluation objective by comparing the evaluation results.

[0029] In this embodiment of the invention, for each evaluation target, candidate image samples are evaluated separately, and ranked according to the evaluation results. Based on the ranking of the evaluation results, multiple (e.g., two or more) candidate samples are selected from the ranking results to construct corresponding preference sample pairs. For example, in a machine task evaluation scenario, the performance index of each candidate image sample in the corresponding task can be calculated separately. Based on the ranking of the performance index, the images with the best and worst performance are selected as a pair of preference samples. Of course, considering that the best and worst performing candidate image samples may not be suitable for subsequent preference optimization, multiple best and worst performing candidate image samples can be selected based on the ranking of the evaluation results to construct corresponding preference sample pairs. Alternatively, a portion of candidate image samples can be selected from the middle of the ranking to construct corresponding preference sample pairs. The specific choice can be determined by the user based on the actual situation.

[0030] Step 3: Using the preference sample pairs corresponding to each evaluation objective, optimize the preferences of the corresponding evaluation objective respectively, and align the optimization directions to obtain the comprehensive gradient. Then, use the comprehensive gradient to optimize the objective processing model.

[0031] In this embodiment of the invention, based on the preference sample pair corresponding to each evaluation objective, a preference optimization process is performed for the corresponding evaluation objective, and the corresponding optimization gradient is calculated. Subsequently, the optimization gradient generated for each evaluation objective is adjusted and aligned to generate a comprehensive gradient for multi-objective optimization.

[0032] In this embodiment of the invention, the target processing model can be one of the generative processing models or other newly introduced image processing models. After optimization through the comprehensive gradient, a multi-objective optimized image processing model is obtained, which can be used for multi-objective optimized image processing.

[0033] The above-mentioned solutions provided by the embodiments of the present invention have the following main advantages: (1) This invention constructs a complete system from preference sample generation, multi-objective preference sample screening to multi-objective image processing model optimization, realizing the collaborative optimization of multiple evaluation objectives such as objective quality, downstream tasks and subjective preferences under the same framework, avoiding the repeated development of independent models for a single objective, and significantly improving consistency and deployment efficiency at the system level.

[0034] (2) This invention obtains real preference information under different optimization objectives by screening preference samples, and performs optimization direction alignment in the multi-objective image processing model optimization stage. It can achieve flexible optimization direction selection according to application scenarios and user preferences, so that the image processing results can achieve a balance between structural fidelity, texture details, subjective naturalness and downstream task performance, and comprehensively improve the image processing quality and application value.

[0035] (3) This invention has good versatility and continuous expansion capability, and can handle different image processing tasks. The system supports extending new generative restoration models to obtain candidate samples with more dimensions, or adding quality evaluation branches of other dimensions to adapt to more optimization objectives, without having to significantly modify the existing structure, and has broad prospects for engineering applications.

[0036] To more clearly demonstrate the technical solution and its effects provided by the present invention, the method provided by the embodiments of the present invention will be described in detail below with reference to specific examples.

[0037] I. Overall Overview of the Plan

[0038] This invention provides a multi-objective optimized image processing method that can balance and accommodate multi-dimensional image optimization needs, thereby improving the visual quality and applicability of images. Figure 2 As shown, the system mainly comprises three stages: preference sample generation, multi-objective preference sample screening, and multi-objective image processing model optimization. First, the system receives the input image, and the user selects N generative processing models. The input image is processed by each of the N generative processing models, resulting in a set of candidate image samples. Then, the user selects M evaluation targets from multiple evaluation dimensions, including subjective quality evaluation indicators and machine task evaluation indicators, and optionally incorporates user subjective rating information. In the multi-objective preference sample screening stage, candidate image samples are compared and screened under the quality evaluation branches corresponding to each evaluation dimension, constructing preference sample pairs for different optimization targets. Finally, in the multi-objective image processing model optimization stage, a preference optimization process is performed based on the preference sample pairs corresponding to each single target. The corresponding optimization gradient directions are calculated, aligned, and fused to form a unified multi-objective optimization direction. This multi-objective optimization direction is fed back to the image processing model to be optimized to update parameters, thereby obtaining an image processing result with balanced comprehensive performance under multi-objective constraints.

[0039] II. Detailed introduction of the plan.

[0040] The following section provides a detailed introduction to the three stages involved in this invention: preference sample generation, multi-objective preference sample screening, and multi-objective preference optimization.

[0041] 1. Preference sample generation stage.

[0042] The preference sample generation stage generates a set of candidate image samples for each input image. This stage begins with the user selecting N generative processing models to construct diverse generation paths. The input image is then processed by generative processing models 1 through N, resulting in multiple sets of processed images that differ in structural recovery characteristics, detail rendering, and color reproduction capabilities. These processed images are then aggregated to form a set of sample images. The generated sets of results complement each other in different optimization dimensions, such as texture representation and overall visual quality, providing a sufficient and effective sample basis for subsequent multi-objective preference sample selection.

[0043] 2. Multi-objective preference sample screening stage.

[0044] The multi-objective preference sample screening stage compares and filters candidate processing results based on different evaluation objectives. This stage allows users to select M evaluation dimensions from multiple dimensions, including subjective quality evaluation indicators and machine task evaluation indicators, and sets corresponding screening evaluation branches for each evaluation objective. In subjective quality optimization scenarios, user subjective scoring or human preference feedback can be introduced as auxiliary evaluation branches. For example, in high-fidelity tasks, candidate results can be compared based on objective quality evaluation indicators such as pixel error, structural similarity, and texture fidelity to construct relative superiority relationships between images. In various downstream task scenarios, candidate results can be evaluated based on their performance scores in object detection, recognition, or other machine vision tasks to obtain corresponding image ranking information. In subjective quality optimization scenarios, human preference annotation, user subjective scoring, or user feedback can be combined to determine more favored candidate results based on visual naturalness and overall impression, thus forming subjective preference relationships. Each of the above evaluation branches evaluates the candidate results and constructs a set of preference relationship information for different evaluation objectives. Each evaluation branch outputs a set of preference information reflecting the relative merits of candidate results, providing a basis for determining the optimization direction of a single target and aligning different directions during multi-target optimization in the subsequent multi-target image processing model optimization process.

[0045] In this embodiment of the invention, the multi-objective preference sample screening stage supports adding preference screening branches and expanding the evaluation objective to a new optimization objective.

[0046] 3. Multi-objective image processing model optimization stage.

[0047] In the multi-objective image processing model optimization stage, based on the preference sample pairs obtained in the aforementioned multi-objective preference sample screening stage, a single-objective preference optimization process is first performed independently for each evaluation objective, thereby obtaining the optimization gradient direction corresponding to each objective. Given that the optimization directions generated by different objectives may be inconsistent or conflicting in direction and magnitude, the system further introduces an optimization direction alignment mechanism to align and fuse the optimization gradient directions of each objective. The contribution ratio of different objectives can be adjusted as needed through methods such as weighted fusion, thus forming a more stable and representative unified multi-objective optimization direction. Finally, this unified optimization direction is used to update the parameters of the optimization objective processing model, enabling the model to achieve an effective trade-off between multi-objective constraints.

[0048] In this embodiment of the invention, for a single evaluation objective, the corresponding optimization gradient is obtained by calculating the loss of preference optimization; the subsequent comprehensive optimization gradient can be achieved by introducing gradient alignment and fusion mechanisms commonly used in meta-learning. For example, the gradients generated by different evaluation objectives can be projected to eliminate conflicting components between gradients; or a weighted fusion method can be used to assign corresponding weights according to the importance of different evaluation objectives, and the gradients of each objective can be weighted and summed to obtain a stable and consistent comprehensive optimization gradient for subsequent unified updates of model parameters.

[0049] Through the above description of the embodiments, those skilled in the art can clearly understand that the above embodiments can be implemented by software, or by using software plus necessary general-purpose hardware platforms. Based on this understanding, the technical solutions of the above embodiments can be embodied in the form of a software product, which can be stored in a non-volatile storage medium (such as a CD-ROM, USB flash drive, mobile hard drive, etc.), including several instructions to cause a computer device (such as a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0050] Example 2 This invention also provides a multi-objective optimized image processing system, which is mainly used to implement the methods provided in the foregoing embodiments, such as... Figure 3 As shown, the system mainly includes: The preference sample generation module processes the input image using multiple different generative recovery models to generate multiple candidate image samples, forming a set of processed image samples. The multi-objective preference sample screening module evaluates candidate image samples in the processed image sample set according to multi-dimensional evaluation objectives, and selects the preference sample pairs corresponding to each evaluation objective by comparing the evaluation results. The multi-objective image processing model optimization module is used to optimize the preferences of each evaluation target by using the preference sample pairs corresponding to each evaluation target, and to align the optimization directions to obtain a comprehensive gradient, and then use the comprehensive gradient to optimize the target processing model.

[0051] In this embodiment of the invention, the processed image sample set includes multiple sets of images obtained after processing the input image, which differ in terms of structural restoration characteristics, detail rendering degree, and color reproduction capability.

[0052] In this embodiment of the invention, the step of evaluating candidate image samples in the processed image sample set according to multi-dimensional evaluation objectives and selecting preference sample pairs corresponding to each evaluation objective by comparing the evaluation results includes: evaluating candidate image samples for each evaluation objective, sorting candidate image samples according to the evaluation results, and selecting multiple candidate samples from the sorting results to construct corresponding preference sample pairs according to the different evaluation results.

[0053] In this embodiment of the invention, the step of using the preference sample pairs corresponding to each evaluation target to perform preference optimization on the corresponding evaluation targets and aligning the optimization directions to obtain a comprehensive gradient includes: performing a preference optimization process for a single evaluation target based on the preference sample pairs corresponding to each evaluation target, and calculating the corresponding optimization gradient; subsequently, adjusting and aligning the optimization gradient generated for each evaluation target to generate a comprehensive gradient for multi-objective optimization.

[0054] Since the main technical details of this system have been described in detail in previous embodiments, they will not be repeated here.

[0055] Those skilled in the art will understand that, for the sake of convenience and brevity, the above-described division of functional modules is used as an example. In practical applications, the above functions can be assigned to different functional modules as needed, that is, the internal structure of the system can be divided into different functional modules to complete all or part of the functions described above.

[0056] Example 3 The present invention also provides a processing device, such as Figure 4 As shown, it mainly includes: one or more processors; a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the method provided in the foregoing embodiments.

[0057] Furthermore, the processing device also includes at least one input device and at least one output device; in the processing device, the processor, memory, input device, and output device are connected via a bus.

[0058] In this embodiment of the invention, the specific types of the memory, input device, and output device are not limited; for example: Input devices can be touchscreens, image acquisition devices, physical buttons, or mice, etc. The output device can be a display terminal; The memory can be random access memory (RAM) or non-volatile memory, such as disk storage.

[0059] Example 4 The present invention also provides a readable storage medium storing a computer program that, when executed by a processor, implements the method provided in the foregoing embodiments.

[0060] In this embodiment of the invention, the readable storage medium is a computer-readable storage medium and can be disposed in the aforementioned processing device, for example, as a memory in the processing device. Furthermore, the readable storage medium can also be any medium capable of storing program code, such as a USB flash drive, portable hard drive, read-only memory (ROM), magnetic disk, or optical disk.

[0061] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the scope of the claims. The information disclosed in the background section is intended only to enhance the understanding of the overall background technology of the present invention and should not be construed as an admission or implication in any way that such information constitutes prior art known to those skilled in the art.

Claims

1. A multi-objective optimization image processing method, characterized in that, include: The input image is processed by multiple different generative restoration models to generate multiple candidate image samples, forming a set of processed image samples. Based on multi-dimensional evaluation objectives, candidate image samples in the processed image sample set are evaluated respectively, and the preference sample pairs corresponding to each evaluation objective are selected by comparing the evaluation results. Using the preference sample pairs corresponding to each evaluation objective, the preferences of the corresponding evaluation objectives are optimized respectively, and the optimization direction is aligned to obtain the comprehensive gradient. Then, the comprehensive gradient is used to optimize the objective processing model.

2. The multi-objective optimization image processing method according to claim 1, characterized in that, The processed image sample set includes multiple sets of images obtained after processing the input image, which differ in terms of structural restoration characteristics, detail rendering level, and color reproduction ability.

3. The multi-objective optimization image processing method according to claim 1, characterized in that, The process of evaluating candidate image samples in the processed image sample set based on multi-dimensional evaluation objectives, and selecting preference sample pairs corresponding to each evaluation objective by comparing the evaluation results, includes: For each evaluation target, candidate image samples are evaluated separately, and the candidate image samples are ranked according to the evaluation results. Based on the different evaluation results, multiple candidate samples are selected from the ranking results to construct corresponding preference sample pairs.

4. The multi-objective optimization image processing method according to claim 1, characterized in that, The process of using preference sample pairs corresponding to each evaluation objective to optimize the preferences for the corresponding evaluation objective and aligning the optimization directions to obtain the comprehensive gradient includes: Based on the preference sample pairs corresponding to each evaluation objective, a preference optimization process is performed for each single evaluation objective, and the corresponding optimization gradient is calculated. Subsequently, the optimization gradients generated for each evaluation objective are adjusted and aligned to generate a comprehensive gradient for multi-objective optimization.

5. A multi-objective optimized image processing system, characterized in that, include: The preference sample generation module processes the input image using multiple different generative recovery models to generate multiple candidate image samples, forming a set of processed image samples. The multi-objective preference sample screening module evaluates candidate image samples in the processed image sample set according to multi-dimensional evaluation objectives, and selects the preference sample pairs corresponding to each evaluation objective by comparing the evaluation results. The multi-objective image processing model optimization module is used to optimize the preferences of each evaluation target by using the preference sample pairs corresponding to each evaluation target, and to align the optimization directions to obtain a comprehensive gradient, and then use the comprehensive gradient to optimize the target processing model.

6. The multi-objective optimized image processing system according to claim 5, characterized in that, The processed image sample set includes multiple sets of images obtained after processing the input image, which differ in terms of structural restoration characteristics, detail rendering level, and color reproduction ability.

7. The multi-objective optimized image processing system according to claim 5, characterized in that, The process of evaluating candidate image samples in the processed image sample set based on multi-dimensional evaluation objectives, and selecting preference sample pairs corresponding to each evaluation objective by comparing the evaluation results, includes: For each evaluation target, candidate image samples are evaluated separately, and the candidate image samples are ranked according to the evaluation results. Based on the different evaluation results, multiple candidate samples are selected from the ranking results to construct corresponding preference sample pairs.

8. The multi-objective optimized image processing system according to claim 5, characterized in that, The process of using preference sample pairs corresponding to each evaluation objective to optimize the preferences for the corresponding evaluation objective and aligning the optimization directions to obtain the comprehensive gradient includes: Based on the preference sample pairs corresponding to each evaluation objective, a preference optimization process is performed for each single evaluation objective, and the corresponding optimization gradient is calculated. Subsequently, the optimization gradients generated for each evaluation objective are adjusted and aligned to generate a comprehensive gradient for multi-objective optimization.

9. A processing device, characterized in that, include: One or more processors; Memory, used to store one or more programs; Wherein, when the one or more programs are executed by the one or more processors, the one or more processors cause the one or more processors to implement the method as described in any one of claims 1 to 4.

10. A readable storage medium storing a computer program, characterized in that, When a computer program is executed by a processor, it implements the method as described in any one of claims 1 to 4.

Images