Method for evaluating isp based on serialized multi-agent

Abstract

The present application relates to the field of computer vision and artificial intelligence, in particular to an ISP parameter tuning and evaluation method based on serialized multi-agent. The method comprises the following steps: obtaining original images of different environments to obtain a RAW-RGB data set; constructing a serialized multi-agent reinforcement learning model, training the serialized multi-agent reinforcement learning model through the RAW-RGB data set to obtain a trained serialized multi-agent reinforcement learning model; and performing parameter prediction on an input RAW image according to the trained serialized multi-agent reinforcement learning model to obtain ISP parameters. In this way, the serialized multi-agent architecture can be used to map the ISP pipeline module, the multi-modal fusion input of image features, gradient features and ISP parameter features can be combined, and the serialized process parameter optimization can be performed module by module, and the mechanism of combining local and global rewards can be used to guide the agent to generate high-quality RGB images and realize task-driven optimization.

Description

Technical Field

[0001] This invention generally relates to the fields of computer vision and artificial intelligence, and more specifically, to an evaluation method for ISP coordination based on serialized multi-agent systems. Background Technology

[0002] Image signal processors (ISPs) are image signal processing pipelines that convert images from RAW format to RGB format for human viewing. Hardware ISPs often produce better image quality and faster processing times than software-generated ones. ISPs have a wide range of applications, including mobile phones, cameras, and autonomous driving. Furthermore, ISPs serve as a crucial bridge between image sensors and downstream applications. They consist of a series of modules responsible for different image processing tasks, such as denoising, white balance, and sharpening. Each module contains a set of adjustable hyperparameters, and these hyperparameters are directly related to the final image quality. Currently, the industry largely relies on imaging experts to manually fine-tune the parameters of each module within the ISP to adapt it to different scenarios. Typically, experts evaluate the image quality generated by the ISP based on subjective perception and, based on objective indicators such as structural similarity or peak signal-to-noise ratio, comprehensively evaluate the effectiveness of a set of parameters, ultimately finding the relatively optimal set of parameters among a large number of image signal processing parameter combinations. However, since experts can only adjust a limited number of parameters based on a single scene diagram each time, the entire process is time-consuming and labor-intensive, and inevitably introduces human subjective error.

[0003] In recent years, with the in-depth research of academia and industry on the parameter tuning problem of image signal processors, many methods for image signal processing proxies or image signal processing parameter prediction have emerged. Some evolutionary strategy-based methods continuously search and iterate in the parameter space to maximize parameter fitness, and finally find a set of parameters with the highest fitness on the current image as the optimal parameter combination. Some convolutional neural network-based methods directly learn a model for predicting parameters to generate a set of image signal processing parameters for different images. However, evolutionary strategy-based methods often lack targeted prediction capabilities based on new images and can only find the corresponding parameters by searching again. Convolutional neural network-based methods require large-scale RAW-RGB data for model training, but currently, there is a lack of relevant large-scale datasets in the industry.

[0004] Therefore, parameter tuning in image signal processors still faces many challenges, such as the large scale of prediction parameters, poor tuning results, and poor correlation and interpretability of parameters inside and outside modules. Previous work has hardly considered the strong correlation and dependency between parameters of different modules, often treating them as independent problems, which limits the global optimization effect. At the same time, current methods in reward design mostly use a single global index, lacking fine-grained guidance on the effects of local modules, which limits the model convergence efficiency and the final image quality. Summary of the Invention

[0005] This invention provides an ISP parameter tuning and evaluation scheme based on serialized multi-agent systems. This scheme addresses the problems of existing ISP parameter tuning methods, such as long manual processing time, difficulty in modeling strong parameter dependencies, and challenges in global quality assessment.

[0006] This invention provides an ISP coordination evaluation method based on serialized multi-agent systems. The method includes: acquiring raw images from different environments to obtain a RAW-RGB dataset;

[0007] A sequential multi-agent reinforcement learning model is constructed, and the sequential multi-agent reinforcement learning model is trained using the RAW-RGB dataset to obtain the trained sequential multi-agent reinforcement learning model.

[0008] The ISP parameters are obtained by predicting the parameters of the input RAW image using a trained sequential multi-agent reinforcement learning model.

[0009] Further, training the sequential multi-agent reinforcement learning model using the RAW-RGB dataset includes:

[0010] S201. Input the RAW images in the training set and the parameters corresponding to each module in the ISP to the ISP to obtain the corresponding RGB images;

[0011] S202. The RGB image is processed by downstream tasks to obtain the environment's score and reward value for the current image, as well as the fused features of the RGB image after feature extraction and fusion; the fused features are used as the state information of the current environment to be input to the current agent.

[0012] S203. The policy network in the current agent outputs the corresponding action based on the agent's state information.

[0013] S204. After adding noise to the output action, fuse it with the parameters corresponding to each module in the ISP to obtain the updated ISP parameters;

[0014] S205. Input the updated ISP parameters and RAW image back into the ISP to obtain the RGB image after the agent's parameter tuning.

[0015] S206. Calculate the local reward of the current agent based on the current task and the RGB image, and feed it back to the current agent;

[0016] S207. Iterate the next agent from S202 to S206 until all agents have been traversed, complete the agent serialization, and output the updated RGB image.

[0017] Furthermore, the fusion features are obtained from RGB images through feature extraction and fusion; the feature extraction yields semantic features, gradient features, and third-type features;

[0018] The semantic features are used to characterize global color and scene structure information, and the semantic features are obtained through the ResNet50 feature extraction network;

[0019] The gradient feature is used to capture edge details and texture changes. It is obtained by calculating the horizontal and vertical gradients using the Sobel operator after grayscale processing of the RGB image, and then using the gradient magnitude map as the gradient feature.

[0020] The third type of feature is obtained by normalizing and expanding the parameter vectors of all current ISP modules to a dimension of the same size as the semantic features and gradient features.

[0021] The semantic features, gradient features, and third-type features are concatenated and nonlinearly transformed to obtain fused features.

[0022] Furthermore, after obtaining the updated RGB image, the global reward of the agent is calculated based on the updated RGB image, and a reward function is constructed using the agent's local reward and global reward.

[0023] Furthermore, the construction of the reward function using the agent's local and global rewards includes:

[0024]

[0025]

[0026] in, Let be the total reward of the i-th agent at time step t; For hyperparameters; Let be the local reward value obtained by the i-th agent at time step t; Let $\frac{ ... The score obtained in the downstream task for the final RGB image generated for the j-th round of serialization of the multi-agent.

[0027] Furthermore, the method for calculating the local reward of the current agent based on the current task and the RGB image includes:

[0028]

[0029] in, Generate the score for the i-th agent in the downstream task for the RGB image generated at time step t; Let be the local reward value obtained by the i-th agent at time step t.

[0030] Furthermore, each iteration generates a tuple, which includes the observations, current environment state, action, current image score, and reward value generated by each agent during the training of the serialized multi-agent reinforcement learning model; the tuple is stored in the buffer of the corresponding agent as a training sample.

[0031] Furthermore, it also includes: calculating the value network loss and policy network loss based on the training samples in each buffer, performing backpropagation, and updating the weight parameters of the policy network.

[0032] Furthermore, the value network loss is calculated using the loss function of the value network, including:

[0033] L(φ) =

[0034] Where L(φ) is the loss function of the value network; φ is the network parameter in the value network; For the first A single agent; For time steps; For cumulative discount rewards; This refers to the batch size. The number of single agents; This represents the state corresponding to the i-th single agent; express Value in a given state; This is a hyperparameter that controls the boundaries in the clip function, and its value is (0,1). Indicates the previous round Value in a given state; For the summation variable; These are the weighting coefficients for the cross-agent consistency loss; Indicates the first The agent in The value outcome under a given state.

[0035] Furthermore, the policy network loss is calculated using the loss function of the policy network, including:

[0036]

[0037] in, Let be the loss function of the policy network; These are network parameters in the policy network; This refers to the batch size. For the first A single agent; The number of single agents; This represents the current advantage function value for a single agent. The entropy of the strategy; To control the hyperparameters of the entropy coefficient; The importance ratio of the current strategy to the previous strategy; The action output by the i-th single agent; For the observation value of the i-th single agent; For the summation variable; This is a hyperparameter that controls the boundaries in the clip function, and its value is (0,1). Weights for perceived loss; To perceive loss; Indicates the current policy probability; This represents the probability of the strategy in the previous round;

[0038] The perceptual loss is calculated using a perceptual loss function, including:

[0039]

[0040] in, A predicted RGB image generated by ISP pipeline optimization parameters; A high-quality target RGB image for reference; This is the feature extraction function for the l-th layer of the pre-trained network (ResNet50); Here are the number of channels, height, and width of the feature map at layer l; Weighting coefficients for features at different levels

[0041] Compared with the prior art, the present invention has the following beneficial technical effects:

[0042] This invention utilizes a serialized multi-agent architecture to map ISP pipeline modules, combines multimodal fusion inputs of image features, gradient features, and ISP parameter features, performs serialized process parameter optimization module by module, and employs a mechanism that combines local and global rewards to guide the agent to achieve high-quality RGB image generation and task-driven optimization.

[0043] The technical solution provided in this application involves the collection, processing, and use of user personal information and related data in strict accordance with the provisions of relevant laws and regulations, and has obtained the authorization and consent of the relevant users. During the implementation process, all personal information involved is protected by encryption technology, and necessary security measures are taken to ensure data privacy and user rights. The collection, storage, and use of all data follow the principles of transparency, legality, and compliance in order to protect users' personal privacy and information security to the greatest extent.

[0044] It should be understood that the description in the Summary of the Invention is not intended to limit the key or essential features of the embodiments of the present invention, nor is it intended to restrict the scope of the invention. Other features of the invention will become readily apparent from the following description. Attached Figure Description

[0045] The above and other features, advantages, and aspects of the various embodiments of the present invention will become more apparent from the accompanying drawings and the following detailed description. In the drawings, the same or similar reference numerals denote the same or similar elements, wherein:

[0046] Figure 1 A flowchart of an ISP coordination evaluation method based on serialized multi-agents according to an embodiment of the present invention is shown;

[0047] Figure 2 A schematic diagram of the structure of a serialized multi-agent reinforcement learning model according to an embodiment of the present invention is shown;

[0048] Figure 3 A schematic diagram of an ISP coordination evaluation method based on serialized multi-agent systems according to an embodiment of the present invention is shown. Detailed Implementation

[0049] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, 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 scope of protection of the present invention.

[0050] Furthermore, the term "and / or" in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, or B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0051] Figure 1 A flowchart of the ISP coordination evaluation method based on serialized multi-agent system according to an embodiment of the present invention is shown.

[0052] The method includes:

[0053] S101. Obtain the original images of different environments to obtain the RAW-RGB dataset.

[0054] In this embodiment, the original images of different environments can be indoor and outdoor scene images constructed for human subjective image quality assessment tasks, or existing datasets such as the MSCOCO dataset.

[0055] In this embodiment, the MSCOCO dataset is selected, and the image inversion method is used to process the MSCOCO dataset to obtain the RGB dataset. Then, the RGB dataset is preprocessed to obtain the RAW dataset.

[0056] As one implementation method of this embodiment, preprocessing may include image cropping and image resizing. The preprocessed RAW dataset is divided into a training set and a test set, with a data ratio of 8:2.

[0057] S102. Construct a serialized multi-agent reinforcement learning model, and train the serialized multi-agent reinforcement learning model using the RAW-RGB dataset to obtain the trained serialized multi-agent reinforcement learning model.

[0058] like Figure 2 As shown in this embodiment, in the serialized multi-agent reinforcement learning model, the RGB image and ISP parameters are used as inputs to the feature fusion module. After passing through the ResNet50 feature extraction network, Sobel operator, and parameter normalization, the corresponding fused features are output and used as the state information of the current environment to input to the agent network.

[0059] In this embodiment, as Figure 3 As shown, the feature fusion module mainly consists of three types of feature extraction pathways:

[0060] 1. Semantic feature extraction pathway ( The input RGB image is first processed by a ResNet50 deep convolutional neural network to extract high-level semantic features. These features can characterize the texture, structure, and global scene information of objects in the image, providing contextual reference for optimizing ISP parameters.

[0061] 2. Gradient feature extraction pathway ( To better characterize the edge and detail features of an image, the Sobel operator is used to calculate the gradients in the horizontal and vertical directions of the input image, resulting in gradient feature maps. These features can highlight edge, texture, and local structural changes, and have direct reference value for parameter adjustment of ISP functional modules such as sharpening, noise reduction, and detail enhancement.

[0062] 3. Parameter feature extraction pathway ( The ISP parameters in the current iteration step are first normalized to eliminate the influence of different dimensions on model convergence. Simultaneously, a dimensionality expansion operation is used to embed the parameters into a representation space consistent with image features, resulting in a parameter feature representation. This processing method enables the fusion of image-perceptual information and numerical parameter information within the same feature space, thereby achieving cross-modal information interaction.

[0063] After the above three types of features are extracted, they are fused in the channel dimension through feature concatenation. First, they are compressed by a module (1×1 convolution) to obtain a one-dimensional feature vector, which is then used as the input of the subsequent agent network.

[0064] The intelligent agent network model consists of an Actor network and a Critic network.

[0065] Actor Network:

[0066] Composed of a multilayer perceptron (MLP), it performs a nonlinear mapping on the fused state features to obtain an intermediate representation of the local observations. Subsequently, an adaptive average pooling layer is used to compress and summarize the features, ultimately outputting a single agent action, i.e., the adjustment amount of the ISP parameters. The dimension of the output action is consistent with the number of ISP parameters to be optimized in the target functional module. The new parameter settings are obtained by superimposing this action with the previous ISP parameters.

[0067] Critic Network:

[0068] Also composed of a multilayer perceptron, it is used to perform global representation modeling of fused state features to obtain intermediate state values. Then, an adaptive average pooling layer outputs the state value, which characterizes the overall merits of executing the policy in the current state and provides an evaluation signal for policy updates.

[0069] Through the above design, this embodiment can fully combine image semantic information, low-level gradient information and current ISP parameter state to achieve cross-modal multi-source information fusion and intelligent decision-making, ensuring that the output ISP parameter adjustment amount can not only improve the image visual quality, but also better serve the downstream perception task.

[0070] In one specific implementation, the Actor network model includes:

[0071] The first multilayer perceptron (MLP) includes several fully connected layers, which take the current state of the environment as input and output a single result of local observations.

[0072] The first adaptive average pooling layer takes a single result from a local observation as input and outputs an action for a single agent. This action represents the adjustment amount of the ISP parameters. The dimension of the agent's action is consistent with the number of ISP parameters in the functional module the agent is predicting. The action is added to the previous ISP parameters to obtain the current ISP parameters.

[0073] The Critic network model includes:

[0074] The second layer perceptron, which includes several fully connected layers, is used to take the current state of the environment as input and output a single result of the state.

[0075] The second adaptive average pooling layer is used to take a single state outcome as input and output the state value of a single agent.

[0076] In this embodiment, training the serialized multi-agent reinforcement learning model using the RAW-RGB dataset includes:

[0077] S201. Input the RAW images in the training set and the parameters corresponding to each module in the ISP into the ISP to obtain the corresponding RGB images.

[0078] S202. The RGB image is processed by downstream tasks to obtain the environment's score and reward value for the current image, as well as the fused features of the RGB image after feature extraction and fusion; the fused features are used as the state information of the current environment to be input to the current agent.

[0079] In this embodiment, the fusion features are obtained from RGB images through feature extraction and fusion; the feature extraction yields semantic features, gradient features, and third-type features.

[0080] Specifically, the semantic features are used to characterize global color and scene structure information, and the semantic features are obtained through the ResNet50 feature extraction network.

[0081] The gradient feature is used to capture edge details and texture changes. It is obtained by calculating the horizontal and vertical gradients using the Sobel operator after grayscale processing of the RGB image, and then using the gradient magnitude map as the gradient feature.

[0082] Normalize all current ISP module parameter vectors and expand them to the same dimension as the two features mentioned above, and use them as the third type of feature.

[0083] The three types of features are concatenated and nonlinearly transformed in the feature fusion module to generate fused features, which serve as the current state information of the agent.

[0084] S203. The policy network in the current agent outputs corresponding actions based on the agent's state information. The action output by each agent is the change in the parameters of the corresponding ISP functional module. The optimized ISP parameters of the current agent will be inherited by the corresponding module parameters in all subsequent agents' ISP parameters.

[0085] S204. After adding noise to the output action, it is fused with the parameters corresponding to each module in the ISP to obtain the updated ISP parameters. That is, only the ISP parameters of the module corresponding to the current agent are updated, and the parameters of other modules are updated and optimized in sequence.

[0086] S205. Input the updated ISP parameters and RAW image back into the ISP to obtain the RGB image after the agent's parameter tuning.

[0087] S206. Calculate the local reward of the current agent based on the current task and the RGB image, and feed it back to the current agent.

[0088] In this embodiment, the local reward function for each agent is as follows:

[0089]

[0090] in, It is the score obtained in the downstream task of the RGB image generated by the i-th agent at time step t. This is the local reward value obtained by the i-th agent at time step t. Meanwhile, the next agent will perform feature extraction and fusion operations based on the current RGB image to obtain its state information.

[0091] S207. Iterate the next agent from S202 to S206 until all agents have been traversed, complete the agent serialization, and output the updated RGB image.

[0092] In this embodiment, after all agents have completed the above S202~S206 processes, that is, after all ISP parameters have been updated once, the corresponding updated RGB image will be obtained. The model calculates the global reward based on the final output image and uses it as part of the reward function for each agent.

[0093] The global reward is calculated as follows:

[0094]

[0095] The score obtained in the downstream task for the final RGB image generated for the j-th round of serialization of the multi-agent; Let be the global reward value of the multi-agent sequence in the j-th round; and the final reward function of each agent is designed as a fusion of global and local rewards, as follows:

[0096]

[0097] in, Let be the total reward of the i-th agent at time step t; This is a hyperparameter used to control the ratio between local and global rewards.

[0098] In this embodiment, each iteration generates a tuple, which includes the observation value, current environment state value, action value, current image score and reward value generated by each agent during the training of the serialized multi-agent reinforcement learning model; the tuple is stored in the buffer of the corresponding agent as a training sample.

[0099] In some embodiments, the method further includes calculating the value network loss and policy network loss based on the training samples in each buffer, performing backpropagation, updating the weight parameters of the policy network, optimizing the ISP parameters, and improving the final image quality.

[0100] The loss function consists of the value network (Critic) loss function and the policy network (Actor) loss function.

[0101] Specifically, a cross-agent prediction difference regularization term is introduced into the value network (Critic) loss function to ensure that the value estimates of neighboring agents remain consistent, reflecting the continuity between ISP modules. After calculating the value network (Critic) loss function, the value network weights φ are updated through directional propagation.

[0102] In the policy network (Actor) loss function, a perceptual loss is introduced to improve image quality through semantic and detail feature constraints. After calculating the policy network (Actor) loss function, the policy network weights θ are updated through backpropagation.

[0103] The cumulative discounted return must be calculated before calculating the value function loss. The definition is as follows:

[0104]

[0105] Where T is the expected step size; γ is the decay factor, with a value range of (0,1]; t is the time step; and R is the reward value in the tuple. For cumulative discount rewards; i represents the i-th single agent; j represents the index of a time step.

[0106] In this embodiment, to encourage consistency in value estimation among neighboring agents and reflect the continuity of the ISP pipeline, a cross-agent prediction difference regularization term is added to the value network loss function.

[0107] Value network loss functions include:

[0108] L(φ) =

[0109] Where L(φ) is the loss function of the value network; φ is the network parameter in the value network; For the first A single agent; For time steps; For cumulative discount rewards; This refers to the batch size. The number of single agents; This represents the state corresponding to the i-th single agent; express Value in a given state; This is a hyperparameter that controls the boundaries in the clip function, and its value is (0,1). Indicates the previous round Value in a given state; For the summation variable; The weighting coefficient for cross-agent consistency loss controls the proportion of this constraint in the total loss; Indicates the first The agent in The value outcome under a given state.

[0110] In this embodiment, in order to obtain high-quality images, an additional image dimension feature constraint is introduced into the policy network loss function to improve detail representation.

[0111] The policy network loss function includes:

[0112]

[0113] in, Let be the loss function of the policy network; These are network parameters in the policy network; This refers to the batch size. For the first A single agent; The number of single agents; This represents the current advantage function value for a single agent. The entropy of the strategy; To control the hyperparameters of the entropy coefficient; The importance ratio of the current strategy to the previous strategy; The action output by the i-th single agent; For the observation value of the i-th single agent; For the summation variable; This is a hyperparameter that controls the boundaries in the clip function, and its value is (0,1). Weights for perceived loss; To calculate the perceptual loss, the difference between the features extracted based on RESNET50 and the ground truth image is calculated. Indicates the current policy probability; This represents the probability of the strategy in the previous round.

[0114] The dominance function is used to measure the quality of a particular action relative to the average level. Specifically, the dominance function... Let be the difference between the cumulative discounted reward and the state value of the i-th agent:

[0115]

[0116] The perceptual loss is calculated using a perceptual loss function, including:

[0117]

[0118] in, A predicted RGB image generated by ISP pipeline optimization parameters; A high-quality target RGB image for reference; This is the feature extraction function for the l-th layer of the pre-trained network (ResNet50); The number of channels, height, and width of the feature map at layer l are used for normalization; These are weighting coefficients for features at different layers, used to balance the contributions of different semantic levels.

[0119] Through the above embodiments, since the estimation of the advantage function depends on the value output of the value network, and the policy network needs to use the advantage function, the consistency constraint of the value network can improve the stability of state value estimation, thereby making the advantage function estimation more reliable. Therefore, the value network not only optimizes itself independently, but also indirectly enhances the effectiveness of policy network updates. It is evident that the value network can guarantee the stability of value estimation and continuity across modules, while the policy network can learn robust parameter tuning strategies based on the value network and generate better images under perceptual loss constraints. The two networks work together to achieve sequential multi-agent tuning of ISP parameters.

[0120] In this embodiment, the training of the sequential multi-agent reinforcement learning model is also included, and the process specifically includes:

[0121] S301. Input the RAW images in the test set and the parameters corresponding to each module in the ISP to the ISP to obtain the corresponding RGB images.

[0122] S302. The RGB image is processed by downstream tasks to obtain the environment's score and reward value for the current image, as well as the fused features of the RGB image after feature extraction and fusion; the fused features are used as the state information of the current environment to be input to the current agent.

[0123] In this embodiment, the fusion features are obtained from RGB images through feature extraction and fusion; the feature extraction yields semantic features, gradient features, and third-type features.

[0124] Specifically, the semantic features are used to characterize global color and scene structure information, and the semantic features are obtained through the ResNet50 feature extraction network.

[0125] The gradient feature is used to capture edge details and texture changes. It is obtained by calculating the horizontal and vertical gradients using the Sobel operator after grayscale processing of the RGB image, and then using the gradient magnitude map as the gradient feature.

[0126] Normalize all current ISP module parameter vectors and expand them to the same dimension as the two features mentioned above, and use them as the third type of feature.

[0127] The three types of features are concatenated and nonlinearly transformed in the feature fusion module to generate fused features, which serve as the current state information of the agent.

[0128] S303. The policy network in the current agent outputs corresponding actions based on the agent's state information. The action output by each agent is the change in the parameters of the corresponding ISP functional module. The optimized ISP parameters of the current agent will be inherited by the corresponding module parameters in all subsequent agents' ISP parameters.

[0129] S304. After adding noise to the output action, it is fused with the parameters corresponding to each module in the ISP to obtain the updated ISP parameters. That is, only the ISP parameters of the module corresponding to the current agent are updated, and the parameters of other modules are updated and optimized in sequence.

[0130] S305. Input the updated ISP parameters and RAW image back into the ISP to obtain the RGB image after the agent's parameter tuning.

[0131] S306. Calculate the local reward of the current agent based on the current task and the RGB image, and feed it back to the current agent.

[0132] In this embodiment, the local reward function for each agent is as follows:

[0133]

[0134] in, It is the score obtained in the downstream task of the RGB image generated by the i-th agent at time step t. This is the local reward value obtained by the i-th agent at time step t. Meanwhile, the next agent will perform feature extraction and fusion operations based on the current RGB image to obtain its state information.

[0135] S307. Iterate the next agent through S302~S306 until all agents have been traversed. After multiple rounds of the above process, the parameter prediction work for a RAW graph is completed. Repeat the above operation for multiple RAW graphs in the test set to complete the entire test process.

[0136] S103. Based on the trained sequential multi-agent reinforcement learning model, perform parameter prediction on the input RAW image to obtain the ISP parameters.

[0137] According to embodiments of the present invention, by mapping the functional modules of the ISP pipeline to multiple agents and optimizing them sequentially according to the actual hardware order, a parameter tuning strategy highly consistent with the actual ISP process is achieved, exhibiting good engineering adaptability and scalability. The method innovatively introduces a multi-feature fusion mechanism, comprehensively utilizing semantic features, gradient features, and parameter features for state description, preserving global color and structural information while enhancing detailed modeling of edge textures and parameter states. Furthermore, the present invention proposes a novel value function and policy function design: based on traditional reinforcement learning loss, a cross-agent prediction difference regularization term is added to constrain the prediction consistency between different agents, improving the module collaborative optimization capability; simultaneously, an image dimensional feature constraint loss is introduced, combined with perceptual feature constraints on the network output, improving the tuning process in both global and detailed aspects. This combined loss not only effectively enhances detail representation and edge fidelity but also improves the stability and generalization of multi-agent decision-making. The overall method demonstrates significant advantages in visual quality and downstream task adaptability, possessing broad application value and practical promotion potential.

[0138] It should be noted that, for the sake of simplicity, the foregoing method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all optional embodiments, and the actions and modules involved are not necessarily essential to the present invention.

[0139] The specific embodiments described above do not constitute a limitation on the scope of protection of this invention. Those skilled in the art should understand that various modifications, combinations, sub-combinations, and substitutions can be made according to design requirements and other factors. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of this invention should be included within the scope of protection of this invention.

Claims

1. A method for evaluating ISP coordination based on serialized multi-agent systems, characterized in that, include: Obtain raw images from different environments to obtain the RAW-RGB dataset; A sequential multi-agent reinforcement learning model is constructed, and the sequential multi-agent reinforcement learning model is trained using the RAW-RGB dataset to obtain the trained sequential multi-agent reinforcement learning model. The ISP parameters are obtained by predicting the parameters of the input RAW image based on the trained sequential multi-agent reinforcement learning model. The step of training the serialized multi-agent reinforcement learning model using the RAW-RGB dataset includes: S201. Input the RAW images in the training set and the parameters corresponding to each module in the ISP to the ISP to obtain the corresponding RGB images; S202. The RGB image is processed by downstream tasks to obtain the environment's score and reward value for the current image, as well as the fused features of the RGB image after feature extraction and fusion; the fused features are used as the state information of the current environment to be input to the current agent. S203. The policy network in the current agent outputs the corresponding action based on the agent's state information. S204. Add noise to the output action and fuse it with the parameters corresponding to each module in the ISP to obtain the updated ISP parameters; S205. Input the updated ISP parameters and RAW image back into the ISP to obtain the RGB image after the agent's parameter tuning. S206. Calculate the local reward of the current agent based on the current task and the RGB image, and feed it back to the current agent; S207. Iterate the next agent from S202 to S206 until all agents have been traversed, complete the agent serialization, and output the updated RGB image. After obtaining the updated RGB image, the global reward of the agent is calculated based on the updated RGB image, and a reward function is constructed using the agent's local reward and global reward. The construction of the reward function using the agent's local and global rewards includes: in, Let be the total reward of the i-th agent at time step t; For hyperparameters; Let be the local reward value obtained by the i-th agent at time step t; Let $\frac{ ... The score obtained in the downstream task for the final RGB image generated for the j-th round of serialization of the multi-agent.

2. The method according to claim 1, characterized in that, The fusion features are obtained from RGB images through feature extraction and fusion; the feature extraction yields semantic features, gradient features, and third-type features. The semantic features are used to characterize global color and scene structure information, and the semantic features are obtained through the ResNet50 feature extraction network; The gradient feature is used to capture edge details and texture changes. It is obtained by calculating the horizontal and vertical gradients using the Sobel operator after grayscale processing of the RGB image, and then using the gradient magnitude map as the gradient feature. The third type of feature is obtained by normalizing and expanding the parameter vectors of all current ISP modules to a dimension of the same size as the semantic features and gradient features. The semantic features, gradient features, and third-type features are concatenated and nonlinearly transformed to obtain fused features.

3. The method according to claim 1, characterized in that, The calculation of the local reward of the current agent based on the current task and the RGB image includes: in, Generate the score for the i-th agent in the downstream task for the RGB image generated at time step t; Let be the local reward value obtained by the i-th agent at time step t.

4. The method according to claim 1, characterized in that, Each iteration generates a tuple, which includes the observations, current environment state, action, current image score, and reward value generated by each agent during the training of the serialized multi-agent reinforcement learning model; the tuple is stored in the buffer of the corresponding agent as a training sample.

5. The method according to claim 4, characterized in that, Also includes: Based on the training samples in each buffer, calculate the value network loss and policy network loss, perform backpropagation, and update the weight parameters of the policy network.

6. The method according to claim 5, characterized in that, The value network loss is calculated using the value network's loss function, including: L(φ) = Where L(φ) is the loss function of the value network; φ is the network parameter in the value network; For the first A single agent; For time steps; Let $t$ be the cumulative discounted reward for the i-th agent at time step $t$. This refers to the batch size; The number of single agents; This represents the state corresponding to the i-th single agent; express Value in a given state; This is a hyperparameter that controls the boundaries in the clip function, and its value is (0,1). Indicates the previous round Value in a given state; For the summation variable; These are the weighting coefficients for the cross-agent consistency loss; Indicates the first The agent in The value outcome under a given state.

7. The method according to claim 5, characterized in that, The policy network loss is calculated using the policy network's loss function, including: in, Let be the loss function of the policy network; These are network parameters in the policy network; This refers to the batch size; For the first A single agent; The number of single agents; This represents the current advantage function value for a single agent. The entropy of the strategy; To control the hyperparameters of the entropy coefficient; The importance ratio of the current strategy to the previous strategy; The action output by the i-th single agent; For the observation value of the i-th single agent; For the summation variable; This is a hyperparameter that controls the boundaries in the clip function, and its value is (0,1). Weights for perceived loss; To perceive loss; Indicates the current policy probability; This represents the probability of the strategy in the previous round; The perceptual loss is calculated using a perceptual loss function, including: in, A predicted RGB image generated by ISP pipeline optimization parameters; A high-quality target RGB image for reference; This is the feature extraction function for the l-th layer of the pre-trained network (ResNet50); Here are the number of channels, height, and width of the feature map at layer l; These are the weighting coefficients for features at different layers.

Images