A distributed image restoration method, system, electronic device, and storage medium

By using a native two-dimensional matrix structure and a dynamic event triggering mechanism in a distributed system, the image restoration method solves the problems of computational dimensionality expansion and network congestion, and achieves high-precision, low-latency image restoration results.

CN123071197APending Publication Date: 2026-11-10SHENZHEN POLYTECHNIC +1
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611560655.3
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-10-10
Publication Date
2026-11-10

AI Technical Summary

Technical Problem

Existing distributed image restoration methods suffer from computational dimensionality expansion, memory overflow, and network congestion when processing high-resolution images due to vectorization. Furthermore, the communication control mechanism exhibits Zeno behavior, making it difficult to deploy effectively in practical hardware.

Method used

In distributed computing systems, local state variables with native two-dimensional matrix structures are used, combined with alternating optimization iteration and event triggering mechanisms based on dynamic thresholds to avoid vectorized operations, and image restoration is achieved through sparse communication.

Benefits of technology

It effectively reduces computational overhead, preserves the topological correlation between image pixels, improves image restoration accuracy and texture detail, while reducing network bandwidth consumption and improving system stability.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123071197A_ABST
    Figure CN123071197A_ABST
Patent Text Reader

Abstract

This application proposes a distributed image restoration method, system, electronic device, and storage medium. The method includes: constructing local state variables at computing nodes that maintain the original two-dimensional matrix structure, including a local image estimation matrix and first and second local fuzzy matrices; acquiring degraded image data and loading it into the estimation matrix, and initializing the two fuzzy matrices; performing alternating optimization iterations on the state variables, sequentially fixing any two terms and independently updating the remaining term; when the state change combined with the communication weight reaches a dynamic threshold that decays exponentially over time, sending updated variables to neighboring nodes and resetting the change; performing consistency constraint corrections based on the received variables, and outputting the restored image if the convergence condition is met. This scheme, based on original matrix iteration and a time lower bound communication mechanism, avoids memory overload and network congestion, achieving globally accurate convergence.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This invention relates to a distributed image restoration method, system, electronic device, and storage medium, belonging to the field of computer vision technology. Background Technology

[0002] In recent years, with the significant increase in data scale and system complexity, optimization problems in the fields of intelligent computing and network communication have become increasingly complex. Large-scale optimization problems often exceed the hardware processing capabilities of standard centralized algorithms. Therefore, distributed algorithms based on multi-agent systems have emerged and are widely used in machine learning, interdisciplinary decision-making, and distributed mobile robot networks. In practical engineering scenarios, such as multi-UAV collaborative environmental perception, high-definition remote sensing image processing from satellite constellations, and distributed joint analysis of medical images, multiple agents or computing nodes often need to collaboratively process blind image restoration tasks.

[0003] In current distributed optimization problems, decision variables are typically processed as vectors. However, in practical engineering challenges such as image restoration, there are large amounts of high-dimensional data stored in native two-dimensional matrix format. Traditional distributed continuous-time subgradient algorithms, which must flatten the image matrix into a one-dimensional vector, cause a dramatic increase in computational dimensionality. For high-resolution image matrices, this forced vectorization produces massive high-dimensional vectors, with computational complexity increasing exponentially. This makes individual computing nodes highly susceptible to memory overflows and even system crashes when processing high-resolution images. Furthermore, forced vectorization severs the native spatial correlation between pixels, failing to effectively preserve the structure and features of the original problem. This directly results in restored images often having blurred edges and distortion, severely losing the original two-dimensional texture details.

[0004] Meanwhile, existing multi-agent systems have limited network transmission capabilities and are frequently subject to external environmental interference, all of which affect the uninterrupted data exchange and processing between nodes, thereby impairing the performance of distributed optimization algorithms. Research shows that introducing event-triggered mechanisms can significantly reduce communication load and enhance communication stability; however, existing communication control mechanisms often suffer from Zeno behavior defects in their underlying logic. Specifically, when the system state approaches the optimal solution (i.e., local measurement errors tend to zero), the triggering interval of the continuous-time synchronization mechanism rapidly shrinks, causing nodes to engage in an unlimited number of high-frequency communications within a finite timeframe. This pathological triggering mechanism instantly exhausts network bandwidth and consumes a large amount of node computing power, leading to severe network congestion, a surge in distributed network latency, and even system unavailability, making it difficult to deploy such distributed algorithms in actual physical hardware.

[0005] Therefore, the existing technology has significant shortcomings and urgently needs further improvement and refinement. Summary of the Invention

[0006] In view of the shortcomings of the prior art, the purpose of this invention is to provide a distributed image restoration method, system, electronic device and storage medium.

[0007] According to an embodiment of the present invention, a first solution is provided as follows: 1. A distributed image restoration method, applied to a target computing node in a distributed computing system, the method being executed by the target computing node, comprising the steps of: S1. Allocate memory space in the target computing node to construct local state variables that maintain the original two-dimensional matrix structure. The local state variables include a local image estimation matrix that represents the target restored image, and a first local fuzzy matrix and a second local fuzzy matrix that represent the two spatial degradation dimensions, respectively. S2. Acquire two-dimensional degraded image data and load the local image estimation matrix as the initial estimation value, while initializing the first local blur matrix and the second local blur matrix; S3. Perform alternating optimization iteration on the local state variables. In a single iteration, fix any two two-dimensional matrices in sequence, and independently update the remaining two-dimensional matrix under the condition of satisfying the image physical constraints. S4. Calculate the change in local state variables relative to the previous communication trigger state. When the change in local state variables, combined with the communication weight parameters of neighboring nodes, reaches a preset dynamic threshold that decays exponentially over time, send the updated local state variables to the neighboring nodes and reset the change in local state variables. S5. Based on the received local state variables of the neighboring nodes, perform consistency constraint correction. If the preset convergence condition is met, output the restored target image based on the synchronized local image estimation matrix.

[0008] Furthermore, the step of allocating memory space in the target computing node to construct local state variables that maintain the original two-dimensional matrix structure includes: allocating local private two-dimensional contiguous memory space for the local state variables in the local random access memory of the target computing node, wherein the local image estimation matrix, the first local fuzzy matrix, and the second local fuzzy matrix are respectively mapped to non-overlapping memory address ranges, so that the processing unit of the target computing node can lock the corresponding memory address ranges for data reading and updating in a single iteration cycle when performing subsequent alternating optimization iterations.

[0009] Furthermore, the two-dimensional spatial dimensions of the local image estimation matrix, the first local blur matrix, and the second local blur matrix satisfy the following mapping relationship: the local image estimation matrix is ​​a two-dimensional matrix with dimension m×n, which is consistent with the original resolution of the two-dimensional degraded image data; the first local blur matrix is ​​a cyclic matrix with dimension m×m, used to characterize the spatial degradation characteristics of the imaging system in the row direction; the second local blur matrix is ​​a cyclic matrix with dimension n×n, used to characterize the spatial degradation characteristics of the imaging system in the column direction; where m and n are the number of pixels of the two-dimensional degraded image data in the first spatial dimension and the second spatial dimension, respectively.

[0010] Furthermore, the step of loading the two-dimensional degraded image data into the local image estimation matrix as the initial estimate includes: assigning the original pixel intensity values ​​of the two-dimensional degraded image data to the matrix elements corresponding to the local image estimation matrix to avoid the fragmentation of spatial correlation, and constraining the matrix element values ​​in the local image estimation matrix within a preset boundary interval representing the image luminance or color channel, so as to construct a closed set feasible region for subsequent local gradient update calculation.

[0011] Furthermore, the step of initializing the first local fuzzy matrix and the second local fuzzy matrix includes: extracting principal component features from the two-dimensional degraded image data using a preset singular value decomposition algorithm, and configuring the extracted principal component features as the initial states of the first local fuzzy matrix and the second local fuzzy matrix, respectively.

[0012] Furthermore, the condition of satisfying the physical constraints of the image specifically refers to: combining a preset regularization penalty term for preserving the spatial features of the two-dimensional image, and within the feasible region that satisfies the non-negative bounded constraints of the image pixels, using the proximal gradient descent rule to perform local gradient update calculation.

[0013] Furthermore, the step of calculating the local state change of the current local state variable relative to the previous communication trigger state includes: calculating the local state variable error matrix of the target computing node and the Frobenius norm of the Lagrange dual variable error matrix introduced based on the image physical constraints, and using the sum of the norms as the local state change.

[0014] Furthermore, the communication weight parameters are determined by a Laplace matrix or adjacency matrix that characterizes the communication topology of the distributed computing system.

[0015] Furthermore, the formula for determining whether the preset dynamic threshold of exponential decay over time has been reached is as follows: ; in, These are preset constant parameters. Let i represent the set of adjacent nodes of the i-th target computation node. This represents the communication weight parameter between the target computation node i and its neighboring node j. This represents the amount of change in the local state. and Both represent preset constants greater than zero, and t represents the continuous running time of the system. This represents the preset dynamic threshold that decays exponentially over time.

[0016] Furthermore, the step of sending the updated local state variables to the adjacent nodes and resetting the local state change amount includes: at the triggering time when the dynamic threshold is determined to be reached, extracting the latest local state variables in the local memory of the target computing node and broadcasting them to the corresponding adjacent nodes through the communication channel of the distributed network; synchronously clearing and resetting the accumulated local state change amount in the target computing node to zero as the starting point for error calculation of the next adjacent triggering cycle, thereby converting continuous time information interaction into discrete event-driven interaction.

[0017] Furthermore, the step of correcting the consistency constraint based on the received local state variables of the neighboring nodes includes: calculating the state difference between the local state variables of the local node and the received local state variables of the neighboring node, generating a consistency error term in combination with the communication weight parameters, and substituting the consistency error term as a penalty constraint feedback into the gradient update calculation of the alternating optimization iteration, so as to drive the local optimization state of the target computing node to move closer to the state of its neighboring nodes.

[0018] Furthermore, the step of outputting the restored target image based on the synchronized local image estimation matrix if the preset convergence condition is met includes: when the state difference between the target computing node and each of the neighboring nodes tends to zero, determining that the local state variables of the local node and the local state variables of the global network have reached a consensus convergence; extracting the finally converged local image estimation matrix from the local memory of the target computing node and outputting it as the globally optimal restored target image.

[0019] According to an embodiment of the present invention, a second solution is provided as follows: A distributed image restoration method is applied to a distributed computing system, the distributed computing system including multiple computing nodes connected by communication, the method comprising the following steps: Memory space is allocated in each of the computing nodes to construct local state variables that maintain the original two-dimensional matrix structure. The local state variables include a local image estimation matrix representing the target restored image, and a first local blur matrix and a second local blur matrix representing the two spatial degradation dimensions, respectively. Two-dimensional degraded image data is acquired, and each computing node loads the two-dimensional degraded image data into its own local image estimation matrix as an initial estimation value. At the same time, the first local blur matrix and the second local blur matrix of each computing node are initialized respectively. Each computing node is controlled to perform alternating optimization iterations in parallel. In a single iteration, each computing node fixes any two two-dimensional matrices in sequence and independently updates the remaining two-dimensional matrix under the condition of satisfying the physical constraints of the image. An event-triggered asynchronous communication step based on a dynamic threshold is executed between the multiple computing nodes: when the change in the local state of any computing node, combined with the communication weight parameter of the system communication topology, reaches a preset dynamic threshold that decays exponentially over time, the computing node is triggered to send the updated local state variables to its neighboring nodes and reset the change in the local state of the computing node. Each computing node performs consistency constraint correction based on the local state variables of its neighboring nodes. When it is determined that the local state variables of each computing node in the distributed computing system meet the preset global consistency convergence condition, the synchronized local image estimation matrix is ​​extracted to output the restored target image.

[0020] According to an embodiment of the present invention, a third solution is provided as follows: A distributed image restoration apparatus, applied to any of the above-mentioned distributed image restoration methods, the apparatus comprising: The memory construction module is used to allocate memory space in the target computing node to construct local state variables that maintain the original two-dimensional matrix structure. The local state variables include a local image estimation matrix representing the target restored image, and a first local fuzzy matrix and a second local fuzzy matrix representing two spatial degradation dimensions, respectively. An initialization module is used to acquire two-dimensional degraded image data and load the local image estimation matrix as the initial estimation value, while simultaneously initializing the first local blur matrix and the second local blur matrix. The optimization iteration module is used to perform alternating optimization iteration on the local state variables. In a single iteration, any two two-dimensional matrices are fixed in sequence, and the remaining two-dimensional matrix is ​​updated independently under the condition of satisfying the image physical constraints. The event interaction module is used to calculate the change in local state variables relative to the previous communication trigger state. When the change in local state variables, combined with the communication weight parameters of neighboring nodes, reaches a preset dynamic threshold that decays exponentially over time, the module sends the updated local state variables to the neighboring nodes and resets the change in local state variables. The output module is used to perform consistency constraint correction based on the received local state variables of neighboring nodes. If the preset convergence condition is met, the restored target image is output based on the synchronized local image estimation matrix.

[0021] A computer device includes a memory and a processor, the memory storing a computer program that, when executed by the processor, causes the processor to perform the following steps: Memory space is allocated in the target computing node to construct local state variables that maintain the original two-dimensional matrix structure. The local state variables include a local image estimation matrix representing the target restored image, and a first local blur matrix and a second local blur matrix representing the two spatial degradation dimensions, respectively. Two-dimensional degraded image data is acquired and loaded into the local image estimation matrix as the initial estimation value, while the first local blur matrix and the second local blur matrix are initialized. Alternating optimization iteration is performed on the local state variables. In a single iteration, any two two-dimensional matrices are fixed in sequence, and the remaining two-dimensional matrix is ​​independently updated under the condition of satisfying the image physical constraints. The local state change of the current local state variable relative to the previous communication trigger state is calculated. When the local state change, combined with the communication weight parameters of the neighboring nodes, reaches a preset dynamic threshold that decays exponentially over time, the updated local state variables are sent to the neighboring nodes and the local state change is reset. Consistency constraint correction is performed based on the received local state variables of the neighboring nodes. If the preset convergence condition is met, the restored target image is output based on the synchronized local image estimation matrix.

[0022] A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the following steps: Memory space is allocated in the target computing node to construct local state variables that maintain the original two-dimensional matrix structure. The local state variables include a local image estimation matrix representing the target restored image, and a first local blur matrix and a second local blur matrix representing the two spatial degradation dimensions, respectively. Two-dimensional degraded image data is acquired and loaded into the local image estimation matrix as the initial estimation value, while the first local blur matrix and the second local blur matrix are initialized. Alternating optimization iteration is performed on the local state variables. In a single iteration, any two two-dimensional matrices are fixed in sequence, and the remaining two-dimensional matrix is ​​independently updated under the condition of satisfying the image physical constraints. The local state change of the current local state variable relative to the previous communication trigger state is calculated. When the local state change, combined with the communication weight parameters of the neighboring nodes, reaches a preset dynamic threshold that decays exponentially over time, the updated local state variables are sent to the neighboring nodes and the local state change is reset. Consistency constraint correction is performed based on the received local state variables of the neighboring nodes. If the preset convergence condition is met, the restored target image is output based on the synchronized local image estimation matrix.

[0023] Compared with the prior art, the beneficial effects of the technical solution provided in this application are as follows: This invention processes degraded image data within the native two-dimensional matrix space, avoiding the use of traditional vectorized dimensionality reduction operations. Under a distributed framework, it effectively overcomes the problems of model state space expansion and memory overflow of computing nodes, significantly reducing computational overhead. At the same time, this mechanism fully preserves the two-dimensional physical topological correlation between image pixels, effectively protecting the features and structure of the original problem, thereby significantly improving the accuracy of image restoration and the representation of texture details.

[0024] In addition, this solution introduces a dynamic event triggering mechanism based on exponential decay over time to replace continuous time synchronization. By ensuring that there is a strictly positive lower bound on the time interval between triggering events, the potential Zeno behavior during the system convergence period is eliminated from the underlying logic, unnecessary data flow is reduced, high-frequency invalid communication is effectively suppressed, network bandwidth consumption is reduced, and the system's operational stability in constrained hardware environments is improved. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. 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.

[0026] in: Figure 1This is a flowchart of a distributed image restoration method applied to a target computing node, provided in an embodiment of the present invention. Figure 2 This is a schematic diagram of the communication topology and internal node structure of a distributed computing system provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the evolution curves of local state changes and dynamic thresholds in an embodiment of the present invention; Figure 4 This is a scatter plot showing the distribution of event trigger times for each computing node in this embodiment of the invention. Figure 5 This is a visual comparison diagram of the blind image restoration effect between the embodiments of the present invention and the prior art; Figure 6 This is a flowchart of a distributed image restoration method applied to a distributed computing system, provided in an embodiment of the present invention. Figure 7 A virtual structure module diagram of the distributed image restoration device provided in the embodiments of the present invention; Figure 8 This is a structural block diagram of a computer device in one embodiment. Detailed Implementation

[0027] To enable those skilled in the art to better understand the technical solutions in this application, the technical solutions in the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0028] Example 1 This embodiment provides a distributed image restoration method, which is applied to a distributed computing system.

[0029] Conventional non-distributed computing systems rely on a single central server for centralized image restoration processing. Centralized processing requires the central node to have extremely high computing power and memory. However, when faced with massive amounts of high-definition remote sensing images or complex medical images, the computing power bottleneck of a single device can easily lead to memory overflow or even system crashes. In contrast, distributed computing systems consist of multiple computing nodes with relatively limited computing power but interconnected with each other. This method chooses to execute based on the target computing node precisely because distributed systems have decentralized characteristics. There is no scheduling center with a global view in the network; each node only has local information. This embodiment enables the entire network to spontaneously compute the globally optimal solution through local cooperation between nodes by giving a single target computing node the ability to compute autonomously and interact on demand.

[0030] This method processes two-dimensional degraded image data. In practical engineering applications such as multi-UAV collaboration and satellite remote sensing, the initial images captured by the terminal devices are inevitably blurred and noisy due to objective factors such as the physical limitations of optical sensors, atmospheric turbulence interference, and relative motion of equipment. These objective limitations determine that the acquired images are usually degraded. Image restoration is a common problem that needs to be solved in the field of image processing, and blind image restoration algorithms must be used as the underlying support. Since traditional centralized algorithms cannot be directly transferred to distributed nodes with limited computing power and bandwidth, the underlying computing architecture and communication mechanism must be reconstructed.

[0031] Based on the above scenario, the core technical problem solved by this embodiment is: how to avoid the collapse of node computing power caused by image matrix vectorization in a distributed network without central scheduling, and overcome the high-frequency network congestion caused by continuous state synchronization between nodes, so as to achieve high-precision image restoration with extremely low system resource overhead.

[0032] To solve the above technical problems, such as Figure 1 , Figure 2 As shown, the method provided in this embodiment is executed by the target computing node, and specifically includes the following steps: S1. Allocate memory space in the target computing node to construct local state variables that maintain the original two-dimensional matrix structure. The local state variables include a local image estimation matrix that represents the target restored image, and a first local fuzzy matrix and a second local fuzzy matrix that represent the two spatial degradation dimensions, respectively. S2. Acquire two-dimensional degraded image data and load the local image estimation matrix as the initial estimation value, while initializing the first local blur matrix and the second local blur matrix; S3. Perform alternating optimization iteration on the local state variables. In a single iteration, fix any two two-dimensional matrices in sequence, and independently update the remaining two-dimensional matrix under the condition of satisfying the image physical constraints. S4. Calculate the change in local state variables relative to the previous communication trigger state. When the change in local state variables, combined with the communication weight parameters of neighboring nodes, reaches a preset dynamic threshold that decays exponentially over time, send the updated local state variables to the neighboring nodes and reset the change in local state variables. S5. Based on the received local state variables of the neighboring nodes, perform consistency constraint correction. If the preset convergence condition is met, output the restored target image based on the synchronized local image estimation matrix.

[0033] The technical advantages of this solution lie in its innovative underlying design of both data structures and communication mechanisms. Through steps S1 and S2, this method avoids the conventional approach of forcibly flattening images into one-dimensional vectors using traditional distributed optimization algorithms. Instead, it directly allocates and maintains the original two-dimensional matrix form of the degraded image in the node's memory. This design not only avoids an exponential increase in model complexity and significantly reduces the memory load on edge computing nodes, but more importantly, it preserves the two-dimensional spatial topology between pixels.

[0034] In step S3, the core challenge of blind image restoration lies in the fact that the original image content and spatial degradation features influence and constrain each other in the imaging model. In this embodiment, the target computing node adopts an alternating locking method for optimization iteration. In a single calculation, the node updates independently and gradually approaches the optimal state of the remaining variables by fixing some of the variables.

[0035] In step S4, by introducing an event-triggered mechanism based on a dynamic threshold, the target computing node only broadcasts data to neighboring nodes when its own state change reaches the decay threshold. This determination mechanism transforms continuous communication actions into discrete event-driven processes, eliminating invalid and frequent interactions. Specifically, the local state change represents the error between the actual state and the estimated state.

[0036] Finally, through step S5, each node completes the consistency error correction with extremely low-frequency information exchange, and eventually achieves global consensus convergence spontaneously.

[0037] The method in this embodiment successfully balances local computing power limitations and network-wide communication bottlenecks, achieving high-precision, low-latency, and extremely stable distributed blind image restoration.

[0038] The core features of this embodiment include an alternating optimization strategy implemented for interleaved variables within the native two-dimensional matrix space, which forms a close underlying collaborative mechanism with the dynamic event-triggered communication mechanism that eliminates Zeno behavior.

[0039] Specifically, in real-world distributed matrix optimization scenarios, the alternating updates of the two highly intertwined variables—original image content and spatial degradation features—can lead to drastic and nonlinear fluctuations in local measurement errors during iteration. If only conventional static thresholds or triggers based on simple error proportions are used, the network will inevitably fall into Zeno's behavior—where local errors tend to zero and trigger time intervals shrink instantaneously—as the system state approaches the global optimum.

[0040] The dynamic threshold constructed in this invention, which decays exponentially over time, is a low-level constraint mechanism that works in deep collaboration with the local matrix alternation optimization algorithm. By introducing a decay term that evolves over time, the dynamic threshold formula establishes a safety boundary at both the theoretical and execution levels: no matter how rapidly the local error approaches zero due to the matrix alternation iteration within the computing node, there is always and necessarily a strictly positive lower bound for the time interval between two state broadcast events of the computing node.

[0041] This design, which combines the alternating iterative characteristics of the native matrix space with a communication mechanism that imposes a mandatory lower bound on time, eliminates the conflict between the triggering mechanism and the rapid convergence of the matrix at the hardware and algorithm execution levels. It overcomes the technical bias in existing technologies that require sacrificing algorithm accuracy or real-time communication to maintain system stability. The scheme achieves stable global accurate convergence of high-dimensional matrices with an extremely sparse discrete communication network, resulting in significant and beneficial technical effects.

[0042] Example 2 This embodiment elaborates on the processing method of Embodiment 1.

[0043] For step S1, in order to accurately depict the image degradation process and guide the subsequent restoration process, this embodiment introduces a matrix degradation model:

[0044] In this model, parameter G is the known observed degraded image matrix, parameter X is the target image matrix to be restored, parameters H1 and H2 are the fuzzy cyclic matrices of the two spatial degradation dimensions, and parameter W represents the additional noise matrix. Based on this degradation model, the target computing node strictly maintains the two-dimensional matrix characteristics of each parameter when constructing local state variables.

[0045] Specifically, the steps of allocating memory space in the target computing node to construct local state variables that maintain the original two-dimensional matrix structure include: In the local random access memory of the target computing node, a local private two-dimensional contiguous memory space is allocated for the local state variables. The local image estimation matrix, the first local fuzzy matrix, and the second local fuzzy matrix are respectively mapped to non-overlapping memory address ranges, so that the processing unit of the target computing node can lock the corresponding memory address ranges for data reading and updating in a single iteration cycle when performing subsequent alternating optimization iterations.

[0046] The advantage of this design lies in its ability to map each matrix variable to non-overlapping memory address ranges in local random access memory, thereby ensuring data security and memory read efficiency at the hardware level during subsequent alternating optimization and update processes. This underlying memory isolation and periodic locking mechanism provides underlying device support for alternating optimization iterations in a distributed collaborative environment.

[0047] Furthermore, the two-dimensional spatial dimensions of the local image estimation matrix, the first local blur matrix, and the second local blur matrix satisfy the following mapping relationship: The local image estimation matrix corresponds to the degradation model described above. , is a two-dimensional matrix with dimensions m×n, and maintains the same resolution as the original resolution of the two-dimensional degraded image data; The first local fuzzy matrix corresponds to the degradation model described above. , is a cyclic matrix of dimension m×m, used to characterize the spatial degradation features of the imaging system in the row direction; The second local fuzzy matrix corresponds to the degradation model described above. , is a cyclic matrix of dimension n×n, used to characterize the spatial degradation features of the imaging system in the column direction; Where m and n are the number of pixels in the two-dimensional degraded image data in the first and second spatial dimensions, respectively.

[0048] With the above configuration, local state variables such as the local image estimation matrix and the fuzzy cyclic matrix are directly stored in the contiguous memory allocated to the nodes in the form of two-dimensional matrices. Compared with conventional distributed subgradient algorithms, which must forcibly convert the m×n image into a one-dimensional long vector, thus causing a sharp increase in the computational dimension and making the computing nodes prone to memory overflow, this embodiment retains the two-dimensional structure of each data item in the underlying memory allocation stage, effectively controlling the model complexity of the algorithm and fundamentally avoiding the memory overload problem.

[0049] For step S2, the target computing node needs to complete the initialization of each local state variable.

[0050] The steps of acquiring two-dimensional degraded image data and loading it into a local image estimation matrix specifically include: assigning the original pixel intensity values ​​of the two-dimensional degraded image data to the matrix elements corresponding to the local image estimation matrix to avoid the fragmentation of spatial correlation; and constraining the matrix element values ​​in the local image estimation matrix within a preset boundary range representing the image luminance or color channel to construct a closed set feasible region for subsequent local gradient update calculation.

[0051] In blind image restoration projects, the feasible region of this closed set is specifically defined as:

[0052] Wherein, parameter L is a zero matrix and parameter U is a matrix with all elements being 255, corresponding to the lower and upper bounds of the grayscale or color channels of the digital image, respectively.

[0053] During the algorithm startup phase, the target computing node uses the degraded observation image matrix G as the initial image state, that is... The initial estimate is loaded and constrained by the aforementioned boundary interval to ensure that it strictly falls within the feasible region Ω of the closed set, thus providing a basis for continuous-time convergence.

[0054] The steps of initializing the first local fuzzy matrix and the second local fuzzy matrix specifically include: using a preset singular value decomposition algorithm to extract principal component features from the two-dimensional degraded image data, and configuring the extracted principal component features as the initial states of the first local fuzzy matrix and the second local fuzzy matrix, respectively.

[0055] Since the initial blur kernel for blind image restoration is a completely unknown blind state, during the distributed collaborative startup phase, each computing node directly calls the singular value decomposition algorithm to process the degraded image data, extracting the principal component matrix and using it as the initial state. and The input value is used to extract the principal direction of spatial degradation from the observation data itself, providing a physically meaningful starting point for the subsequent accurate update of the fuzzy matrix.

[0056] For step S3, the target computing node needs to perform alternating optimization iterations on the constructed local state variables. In the physical scenario of blind image restoration, the local image estimation matrix representing the image is highly intertwined with the first and second local fuzzy matrices representing degradation features. To achieve an efficient solution, this embodiment adopts an alternating iteration strategy, decomposing the overall optimization problem into three sub-problems that are solved alternately in sequence.

[0057] The conditions for satisfying the physical constraints of the image specifically refer to: combining a pre-defined regularization penalty term to preserve the spatial features of the two-dimensional image, and using the proximal gradient descent rule to perform local gradient update calculations within the feasible region that satisfies the non-negative bounded constraints of the image pixels.

[0058] Specifically, the target computation node constructs the following optimization objective function, which includes a regularization penalty term, during the solution process: ; In this objective function, parameters β, β1, and β2 are regularization parameters greater than zero, and parameter D... x and D y These represent the gradient operator matrices along the horizontal and vertical directions, respectively. The penalty term in the formula... This is a pre-defined regularization penalty term used to preserve the spatial features of a two-dimensional image, suppressing image noise and maintaining edge texture.

[0059] Within a single iteration cycle, the processing unit of the target computing node sequentially performs the following three stages of optimization and update: In the first stage, the current state of the first local fuzzy matrix H1 and the second local fuzzy matrix H2 is fixed, and the local optimal value of the local image estimation matrix X is solved using the proximal gradient descent rule. At this time, the update calculation is strictly constrained within the feasible region of the closed set containing the upper and lower bounds of the pixels. In the second stage, the updated local image estimation matrix X and the second local fuzzy matrix H2 from the previous state are fixed, and the first local fuzzy matrix H1 is updated independently, while the non-negativity constraint of matrix elements is satisfied during the calculation process. In the third stage, the updated local image estimation matrix X and the first local fuzzy matrix H1 are fixed, and the second local fuzzy matrix H2 is updated independently.

[0060] Through the above alternating optimization iteration, the target computing node, in a state of no central scheduling, relies on local memory resources to transform the complex multivariate optimization problem into a continuous update of a single variable in the original matrix space, which greatly reduces the node computing power consumption of a single calculation.

[0061] Regarding step S4, to address the high-frequency network congestion issue caused by continuous-time state synchronization, the target computing node independently executes event triggering judgment based on dynamic thresholds internally. The step of calculating the local state change of the current local state variable relative to the previous communication trigger state specifically includes: calculating the error matrix of the local state variable of the target computing node, and the Frobenius norm of the error matrix of the Lagrange dual variable introduced based on the image physical constraints, and using the sum of these norms as the local state change. In the underlying continuous-time optimization algorithm, the Lagrange dual variable is specifically used to encode and adjust the system's physical constraints.

[0062] The determination process requires the use of the communication weight parameters of adjacent nodes. These communication weight parameters are determined by the Laplace matrix or adjacency matrix that characterizes the communication topology of the distributed computing system. These parameters truly reflect the connection status and influence strength between the target computing node and its adjacent nodes in the distributed network.

[0063] The formula for determining whether a target computing node has reached the preset dynamic threshold of exponential decay over time is as follows: ; in, These are preset constant parameters. Let i represent the set of adjacent nodes of the i-th target computation node. This represents the communication weight parameter between the target computation node i and its neighboring node j. This represents the amount of change in the local state. and Both represent preset constants greater than zero, and t represents the continuous running time of the system. This represents the preset dynamic threshold that decays exponentially over time.

[0064] Combination Figure 3 As shown in the evolution curve diagram, during the algorithm's running cycle, the measurement error curve, representing the change in local state, fluctuates continuously over time, but its trigger boundary is always subject to smooth decay. Suppression and strict constraint of the envelope.

[0065] The steps of sending the updated local state variables to adjacent nodes and resetting the local state changes specifically include: At the trigger moment when the dynamic threshold is reached, the latest local state variables in the local memory of the target computing node are extracted and broadcast to the corresponding neighboring nodes through the communication channel of the distributed network; at the same time, the accumulated local state changes in the target computing node are cleared and reset to serve as the starting point for error calculation in the next adjacent triggering cycle, thereby converting continuous time information interaction into discrete event-driven interaction.

[0066] like Figure 4 As shown, the triggering times of each computing node exhibit a clear discrete distribution. Even when the system runtime approaches the latter half and the state is close to the optimal solution, the triggering actions on each node do not exhibit an infinitely dense, pathological clustering. This performance demonstrates that the decision mechanism of this embodiment ensures a strictly positive lower bound on the time interval between consecutive triggering events, thus eliminating Zeno-like behavior that could lead to network paralysis at the underlying execution level.

[0067] For step S5, after the target computing node interacts with the discrete event trigger, it needs to perform consistency correction on the optimization iteration that is currently in progress locally in order to ensure the consistency of the computing direction of the entire network.

[0068] The steps for correcting consistency constraints based on the received local state variables of neighboring nodes specifically include: The local state variables of the target computing node are calculated to be different from the local state variables of the neighboring nodes. A consistency error term is generated by combining the communication weight parameters. The consistency error term is used as a penalty constraint feedback and substituted into the gradient update calculation of the alternating optimization iteration to drive the local optimization state of the target computing node to move closer to the state of its neighboring nodes.

[0069] In the underlying algorithm model, this consistency error term is represented as a sum of node state differences incorporating adjacency weight parameters. Once the target computation node receives the latest state broadcast from its neighboring nodes, it immediately incorporates this sum of differences as a penalty factor into the current proximal gradient descent calculation. This mechanism ensures that each autonomous computation node is constrained and guided by the overall distributed network state on a macroscopic level, thereby guaranteeing the stable evolution of the decentralized network towards a unified optimal goal.

[0070] If the preset convergence condition is met, the specific steps for outputting the restored target image based on the local image estimation matrix after synchronization include: When the state difference between the target computing node and each of its neighboring nodes approaches zero, it is determined that the local state variables of the local node and the local state variables of the global network have reached a consensus convergence; the finally converged local image estimation matrix is ​​extracted from the local memory of the target computing node and output as the globally optimal restored target image.

[0071] like Figure 5 As shown, after processing by the method of this invention, each node can spontaneously achieve globally consistent convergence with extremely low communication overhead. Compared with traditional computation methods that employ forced vectorization operations, this embodiment, because it performs state allocation and iterative calculations within the native matrix space from beginning to end, maintains extremely high clarity and spatial fidelity in the edges and subtle textures of the target object in the final output restored target image. This fully demonstrates that this scheme, while overcoming the bottlenecks of local computing power and overall network communication bandwidth, comprehensively improves the execution efficiency and accuracy of blind image restoration tasks.

[0072] Example 3 This embodiment targets the visual perception scenario of a drone swarm. Each computing node is often in a high-speed moving state when it is working. This mobility characteristic causes the distributed communication topology to change dynamically over time. At the same time, each edge node also needs to face strict battery energy constraints. In scenarios where the topology changes drastically or the node's power drops to the point of exhaustion, the existing execution scheme may lead to premature energy depletion or the inability to close the consistency error term due to the node suddenly moving out of the communication range, thereby causing local network collapse and interruption of the overall recovery task.

[0073] To address the aforementioned technical problems, this embodiment proposes an image restoration method, comprising the following steps: S301. Real-time detection of wireless link status parameters between the target computing node and each neighboring node; conversion of the link status parameters into dynamic communication weights according to a preset signal attenuation model; and updating the adjacency matrix in local memory accordingly to generate a time-varying communication topology matrix. S302. Periodically obtain the residual power ratio of the target computing node, and use a preset nonlinear inverse proportional function to map the residual power ratio to an adaptive energy penalty coefficient. Set the adaptive energy penalty coefficient to increase exponentially as the residual power ratio decreases. S303. Perform alternating optimization iteration on the local state variables within the original two-dimensional matrix structure, fix any two two-dimensional matrices in turn, and independently update the remaining two-dimensional matrix. S304. Calculate the Frobenius norm of the local state variable error matrix and the dual variable error matrix of the target computing node respectively, and use the sum of the two norms as the local state change at the current timestamp. S305. The adaptive energy penalty coefficient is introduced into the preset dynamic threshold calculation formula that decays exponentially with time for product fusion, and combined with the real-time weight extracted from the time-varying communication topology matrix to generate a composite event trigger threshold. S306. Compare the local state change amount with the composite event triggering threshold. When the local state change amount exceeds the composite event triggering threshold, send the updated local state variable to the currently active neighboring node in the time-varying communication topology matrix and reset the local state change amount. S307. Receive the local state variables sent by the currently active neighboring nodes, extract the timestamp difference during the data packet transmission process to generate delay compensation weights, combine the delay compensation weights with the time-varying communication topology matrix, and perform consistency constraint correction with delay compensation on the local alternating optimization iteration.

[0074] In the above implementation scheme, the triggering boundary inside the computing node is no longer solely controlled by the runtime, but is bound to the node's own endurance and the connectivity of the external network. The composite event triggering threshold generated in step S305 exhibits a decay characteristic similar to the basic scheme when the node has sufficient power and the network is stable, thereby prompting the network to converge quickly. However, when the node's power is critical, the adaptive energy penalty coefficient will raise the triggering threshold to suppress the node's data transmission frequency, causing it to use its remaining computing power for the iterative update of the local native matrix.

[0075] The event-triggered mechanism in this embodiment possesses a high degree of environmental adaptability: in harsh mobile edge networks, the system can proactively extend the lifespan of low-power nodes, avoiding the avalanche effect caused by single-point failures. Simultaneously, by employing delay compensation constraints, it ensures that even under extreme conditions of frequent topology disconnections and reconnections, it can still smoothly complete the convergence calculation of high-dimensional matrix variables using asymmetric asynchronous information, significantly enhancing the stability of the image restoration algorithm under real-world complex conditions.

[0076] Example 4 In actual deployed distributed sensor networks, abnormal communication of individual computing nodes is inevitable during long-term operation. These anomalies will broadcast abnormal local state variables that deviate significantly from the normal computing trajectory of the scheme in this embodiment. In the basic distributed coordination scheme, these abnormal variables will spread through the event-triggered network and pollute the iteration direction of normal nodes.

[0077] To address this technical problem, this embodiment provides a distributed image restoration method, including the following steps: S401. During the event-triggered asynchronous communication process, when the target computing node receives the latest local state variable sent by any neighboring node, it extracts the local state variables participating in the current iteration in local memory in real time, and calculates the state difference matrix between the two by subtracting them element by element. S402. Perform eigenvalue decomposition on the state difference matrix, extract the maximum singular value of the state difference matrix, and use it as the absolute deviation index to measure the current state change of the adjacent node. S403. Synchronously obtain the gradient descent step size of the objective function of the target computing node in the local alternating optimization iteration process, and generate an adaptive convergence expectation benchmark value for the current computing stage in combination with the current global iteration progress of the system. S404. The absolute deviation index is compared with the adaptive convergence expected benchmark value. When it is determined that the absolute deviation index exceeds the adaptive convergence expected benchmark value, the difference of the excess part is input into the preset smoothing saturation function to calculate the dynamic trust decay coefficient for the adjacent node. S405. The communication weight parameters initially allocated to the adjacent node by the system are multiplied and scaled using the dynamic trust decay coefficient to generate adaptive suppression weights. S406. The latest local state variables sent by each neighboring node are weighted and aggregated with the corresponding adaptive suppression weights to generate a consistency error term with anomaly isolation capability. S407. The consistency error term is fed back as a penalty factor into the alternating optimization iteration calculation of the target computing node, and when the adaptive suppression weight for a certain neighboring node is detected to drop to near zero, the neighboring node is determined to be invalid and the state synchronization interaction with the neighboring node is directly disconnected.

[0078] In the implementation of this embodiment, the solution does not rely on historical state storage. The target computing node uses its current state and gradient descent expectation as a review benchmark. Through singular value calculation, the huge state differences are directly quantified into single-value deviation indicators. The system constructs a dynamic trust decay mechanism through a smoothing saturation function, so that the node with more drastic state changes has a lower weight coefficient in the consensus network.

[0079] This embodiment does not require additional memory allocation to store historical states. It can achieve soft isolation of abnormal data simply through real-time state comparison and adaptive weight decay. When some nodes in the network experience hardware failure or data mutation, the target computing node can actively shield abnormal interference and maintain stable convergence of local iterative computing. This ensures the final accuracy of image restoration and the overall stability of the system in real environments with high single-point failure risk.

[0080] Example 5 When drone swarms perform large-scale remote sensing and mapping tasks, the onboard sensor arrays typically capture both broadband intensity data containing rich texture details and sparsely distributed narrowband multispectral data simultaneously. To reconstruct full-resolution, high-definition multispectral images, the conventional approach is to use the broadband data as a guiding layer for intensive interpolation and cross-channel filtering. However, in drone swarm networks, individual nodes rely on battery power and have limited computing power, and the wireless communication bandwidth between swarms is susceptible to interference. Requiring a single drone to independently perform high-dimensional matrix intensive operations at the tens of millions of pixels level can easily lead to system crashes, while centrally transmitting massive amounts of raw data back can cause link congestion. These are the practical technical problems that urgently need to be solved when deploying multispectral sensors in edge scenarios such as drone swarms.

[0081] To address the above problems, this embodiment provides an image restoration method, including: S501. In the local random access memory of the target UAV node, allocate non-overlapping two-dimensional continuous memory space for the wide spectrum intensity matrix acquired by the airborne sensor and the multiple mutually orthogonal narrowband sparse sampling matrices. S502. Extract the local neighborhood gradient features of the broadband intensity matrix and generate a structure-guided weight matrix by combining it with a preset attenuation mapping function, which serves as a reference benchmark for the reconstruction of multispectral missing elements. S503. Construct a distributed joint optimization objective function adapted to UAV swarm collaboration, introduce the structure-guided weight matrix as a data fidelity term into the objective function, and transform the cross-channel edge smoothing constraint into a spatial penalty term; S504. In the native two-dimensional matrix space, perform alternating optimization iteration on the distributed joint optimization objective function. In a single iteration cycle, the target UAV node fixes the wide spectrum intensity matrix and the structure guidance weight matrix, and updates the state of each of the narrowband sparse sampling matrices independently in sequence. S505. Calculate the Frobenius norm of the error matrix of each narrowband sparse sampling matrix of the target UAV node, and sum them as the joint local state change. S506. Compare the joint local state change with the trigger boundary that incorporates the wireless communication weights of neighboring UAV nodes. When the joint local state change exceeds a preset dynamic threshold that decays exponentially over time, send the updated narrowband sparse sampling matrices to the neighboring UAV nodes through the airborne wireless channel and reset the joint local state change. S507. Receive matrix data broadcast by neighboring UAV nodes to generate a consistency error term, and feed it back into the gradient descent step of the local alternating optimization iteration. S508. When the cluster's global uniformity convergence condition is met, output a full-resolution multispectral target image that incorporates wide-spectrum high-frequency details.

[0082] This embodiment effectively solves the deployment challenge of high-order multispectral sensors on computationally limited flight nodes by transforming the wide-spectral structure guidance and cross-channel constraint mechanism into mathematical penalty terms and integrating them into a distributed alternating iterative architecture. The cluster network can collaboratively output full-resolution restored images with both extremely high resolution and signal-to-noise ratio with extremely low system resource overhead, significantly improving the environmental perception quality of UAV formations in complex working conditions.

[0083] Example 6 This embodiment provides a distributed image restoration method, which is applied to a distributed computing system containing multiple computing nodes with communication connections.

[0084] like Figure 6 As shown, the distributed image restoration method provided in this embodiment includes the following steps: P1. Allocate memory space in each of the computing nodes to construct local state variables that maintain the original two-dimensional matrix structure. The local state variables include a local image estimation matrix representing the target restored image, and a first local fuzzy matrix and a second local fuzzy matrix representing the two spatial degradation dimensions, respectively. P2. Acquire two-dimensional degraded image data, and make each computing node load the two-dimensional degraded image data into its own local image estimation matrix as the initial estimation value, while simultaneously initializing the first local blur matrix and the second local blur matrix of each computing node respectively. P3. Control each computing node to perform alternating optimization iteration in parallel. In a single iteration, each computing node fixes any two two-dimensional matrices in sequence and independently updates the remaining two-dimensional matrix under the condition of satisfying the image constraints. P4. Execute an event-triggered asynchronous communication step based on a dynamic threshold between the multiple computing nodes: When the change in the local state of any computing node, combined with the communication weight parameter of the system communication topology, reaches a preset dynamic threshold that decays exponentially over time, the computing node is triggered to send the updated local state variables to its neighboring nodes and reset the change in the local state of the computing node. P5. Each computing node performs consistency constraint correction based on the local state variables of the neighboring nodes received. When it is determined that the local state variables of each computing node in the distributed computing system meet the preset global consistency convergence condition, the synchronized local image estimation matrix is ​​extracted to output the restored target image.

[0085] Example 7 like Figure 7 As shown, this embodiment provides a distributed image restoration device, applied to the distributed image restoration method described in any one of embodiments one to five above. The device includes: The memory construction module 100 is used to allocate memory space in the target computing node to construct local state variables that maintain the original two-dimensional matrix structure. The local state variables include a local image estimation matrix representing the target restored image, and a first local blur matrix and a second local blur matrix representing two spatial degradation dimensions, respectively. The initialization module 200 is used to acquire two-dimensional degraded image data and load the local image estimation matrix as the initial estimation value, and at the same time initialize the first local blur matrix and the second local blur matrix. The optimization iteration module 300 is used to perform alternating optimization iteration on the local state variables. In a single iteration, any two two-dimensional matrices are fixed in sequence, and the remaining two-dimensional matrix is ​​updated independently under the condition of satisfying the image constraints. The event interaction module 400 is used to calculate the local state change of the current local state variable relative to the previous communication trigger state. When the local state change, combined with the communication weight parameter of the neighboring node, reaches a preset dynamic threshold that decays exponentially over time, the updated local state variable is sent to the neighboring node and the local state change is reset. The output module 500 is used to perform consistency constraint correction based on the local state variables of the received neighboring nodes. If the preset convergence condition is met, the restored target image is output based on the local image estimation matrix after synchronization.

[0086] The distributed image restoration device provided in this embodiment has the same execution mechanism for each virtual functional module as the corresponding method embodiment described above, and can stably achieve global accurate convergence of high-dimensional matrices.

[0087] Example 8 This embodiment provides a computer-readable storage medium storing computer instructions for causing a computer to execute the distributed image restoration method described in any one of the preceding embodiments one to six.

[0088] like Figure 8 As shown, this embodiment also provides a computer device, including at least one processor and a memory communicatively connected to the at least one processor. The memory stores a computer program executable by the at least one processor, which, when executed by the at least one processor, enables the computer device to perform the distributed image restoration method described in any one of embodiments one through six above.

[0089] Specifically, this computer device can be a centralized server, a personal computer, or a dedicated edge computing node. Internally, it primarily uses a system bus to connect the processor, internal memory, non-volatile memory, and network interface to achieve efficient data interaction and processing. The network interface is used to establish communication channels with other adjacent nodes in the distributed network to perform event-triggered asynchronous communication.

[0090] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), RAMbus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and RAMbus dynamic RAM (RDRAM), etc.

[0091] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0092] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are specific and detailed, they should not be construed as limiting the scope of this patent application. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A distributed image restoration method, applied to a target computing node in a distributed computing system, characterized in that, The method is executed by the target computing node and includes the following steps: S1. Allocate memory space in the target computing node to construct local state variables that maintain the original two-dimensional matrix structure. The local state variables include a local image estimation matrix that represents the target restored image, and a first local fuzzy matrix and a second local fuzzy matrix that represent the two spatial degradation dimensions, respectively. S2. Acquire two-dimensional degraded image data and load the local image estimation matrix as the initial estimation value, while initializing the first local blur matrix and the second local blur matrix; S3. Perform alternating optimization iteration on the local state variables. In a single iteration, fix any two two-dimensional matrices in sequence, and independently update the remaining two-dimensional matrix under the condition of satisfying the image physical constraints. S4. Calculate the change in local state variables relative to the previous communication trigger state. When the change in local state variables, combined with the communication weight parameters of neighboring nodes, reaches a preset dynamic threshold that decays exponentially over time, send the updated local state variables to the neighboring nodes and reset the change in local state variables. S5. Based on the received local state variables of the neighboring nodes, perform consistency constraint correction. If the preset convergence condition is met, output the restored target image based on the synchronized local image estimation matrix.

2. The distributed image restoration method according to claim 1, characterized in that, The step of allocating memory space in the target computing node to construct local state variables that maintain the original two-dimensional matrix structure includes: In the local random access memory of the target computing node, a local private two-dimensional contiguous memory space is allocated for the local state variables. The local image estimation matrix, the first local fuzzy matrix, and the second local fuzzy matrix are respectively mapped to non-overlapping memory address ranges, so that the processing unit of the target computing node can lock the corresponding memory address ranges for data reading and updating in a single iteration cycle when performing subsequent alternating optimization iterations.

3. The distributed image restoration method according to claim 1, characterized in that, The two-dimensional spatial dimensions of the local image estimation matrix, the first local blur matrix, and the second local blur matrix satisfy the following mapping relationship: The local image estimation matrix is ​​a two-dimensional matrix with dimensions m×n, and its resolution is consistent with the original resolution of the two-dimensional degraded image data. The first local blur matrix is ​​a cyclic matrix of dimension m×m, used to characterize the spatial degradation features of the imaging system in the row direction; The second local blur matrix is ​​a cyclic matrix of dimension n×n, used to characterize the spatial degradation features of the imaging system in the column direction; Where m and n are the number of pixels in the two-dimensional degraded image data in the first and second spatial dimensions, respectively.

4. The distributed image restoration method according to claim 1, characterized in that, The steps of loading two-dimensional degraded image data into the local image estimation matrix as initial estimates include: The original pixel intensity values ​​of the two-dimensional degraded image data are assigned to the matrix elements corresponding to the local image estimation matrix to avoid the fragmentation of spatial correlation. The matrix element values ​​in the local image estimation matrix are constrained within a preset boundary range representing the image luminance or color channel to construct a closed set feasible region for subsequent local gradient update calculation.

5. The distributed image restoration method according to claim 1, characterized in that, The steps of initializing the first local fuzzy matrix and the second local fuzzy matrix include: Principal component features are extracted from the two-dimensional degraded image data using a preset singular value decomposition algorithm, and the extracted principal component features are configured as the initial states of the first local fuzzy matrix and the second local fuzzy matrix, respectively.

6. The distributed image restoration method according to claim 1, characterized in that, The conditions for satisfying the physical constraints of the image specifically refer to: combining a preset regularization penalty term for preserving the spatial features of the two-dimensional image, and using the proximal gradient descent rule to perform local gradient update calculations within the feasible region that satisfies the non-negative bounded constraints of the image pixels.

7. The distributed image restoration method according to claim 1, characterized in that, The step of calculating the change in the local state variable of the current local state variable relative to the previous communication trigger state includes: Calculate the local state variable error matrix of the target computing node and the Frobenius norm of the Lagrange dual variable error matrix introduced based on the image physical constraints, and use the sum of the norms as the local state change.

8. The distributed image restoration method according to claim 7, characterized in that, The communication weight parameters are determined by the Laplace matrix, which characterizes the communication topology of the distributed computing system.

9. The distributed image restoration method according to claim 8, characterized in that, The formula for determining whether the preset dynamic threshold of exponential decay over time has been reached is as follows: ; in, These are preset constant parameters. Let i represent the set of adjacent nodes of the i-th target computation node. This represents the communication weight parameter between the target computation node i and its neighboring node j. This represents the amount of change in the local state. and Both represent preset constants greater than zero, and t represents the continuous running time of the system. This represents the preset dynamic threshold that decays exponentially over time.

10. The distributed image restoration method according to claim 1, characterized in that, The step of sending the updated local state variables to the neighboring nodes and resetting the local state change includes: At the trigger moment when the dynamic threshold is determined to be reached, the latest local state variables in the local memory of the target computing node are extracted and broadcast to the corresponding neighboring nodes through the communication channel of the distributed network; The accumulated local state changes within the target computing node are simultaneously cleared and reset to zero, serving as the starting point for error calculation in the next adjacent triggering cycle, thereby converting continuous-time information interaction into discrete event-driven interaction.

11. The distributed image restoration method according to claim 1, characterized in that, The step of correcting the consistency constraints based on the received local state variables of neighboring nodes includes: The local state variables of the target computing node are calculated to be different from the local state variables of the neighboring nodes. A consistency error term is generated by combining the communication weight parameters. The consistency error term is used as a penalty constraint feedback and substituted into the gradient update calculation of the alternating optimization iteration to drive the local optimization state of the target computing node to move closer to the state of its neighboring nodes.

12. The distributed image restoration method according to claim 11, characterized in that, The step of outputting the restored target image based on the synchronized local image estimation matrix if the preset convergence condition is met includes: When the state difference between the target computing node and each of the neighboring nodes approaches zero, it is determined that the local state variables of the local node and the local state variables of the global network have reached a consensus and converged. The finally converged local image estimation matrix is ​​extracted from the local memory of the target computing node and output as the globally optimal restored target image.

13. A distributed image restoration method, applied to a distributed computing system, the distributed computing system comprising multiple computing nodes connected in communication, characterized in that, The method includes the following steps: Memory space is allocated in each of the computing nodes to construct local state variables that maintain the original two-dimensional matrix structure. The local state variables include a local image estimation matrix representing the target restored image, and a first local blur matrix and a second local blur matrix representing the two spatial degradation dimensions, respectively. Two-dimensional degraded image data is acquired, and each computing node loads the two-dimensional degraded image data into its own local image estimation matrix as an initial estimation value. At the same time, the first local blur matrix and the second local blur matrix of each computing node are initialized respectively. Each computing node is controlled to perform alternating optimization iterations in parallel. In a single iteration, each computing node fixes any two two-dimensional matrices in sequence and independently updates the remaining two-dimensional matrix under the condition of satisfying the physical constraints of the image. An event-triggered asynchronous communication step based on a dynamic threshold is executed between the multiple computing nodes: when the change in the local state of any computing node, combined with the communication weight parameter of the system communication topology, reaches a preset dynamic threshold that decays exponentially over time, the computing node is triggered to send the updated local state variables to its neighboring nodes and reset the change in the local state of the computing node. Each computing node performs consistency constraint correction based on the local state variables of its neighboring nodes. When it is determined that the local state variables of each computing node in the distributed computing system meet the preset global consistency convergence condition, the synchronized local image estimation matrix is ​​extracted to output the restored target image.

14. A distributed image restoration device, characterized in that, The apparatus used in the distributed image restoration method according to any one of claims 1 to 12 includes: The memory construction module is used to allocate memory space in the target computing node to construct local state variables that maintain the original two-dimensional matrix structure. The local state variables include a local image estimation matrix representing the target restored image, and a first local fuzzy matrix and a second local fuzzy matrix representing two spatial degradation dimensions, respectively. An initialization module is used to acquire two-dimensional degraded image data and load the local image estimation matrix as the initial estimation value, while simultaneously initializing the first local blur matrix and the second local blur matrix. The optimization iteration module is used to perform alternating optimization iteration on the local state variables. In a single iteration, any two two-dimensional matrices are fixed in sequence, and the remaining two-dimensional matrix is ​​updated independently under the condition of satisfying the image physical constraints. The event interaction module is used to calculate the change in local state variables relative to the previous communication trigger state. When the change in local state variables, combined with the communication weight parameters of neighboring nodes, reaches a preset dynamic threshold that decays exponentially over time, the module sends the updated local state variables to the neighboring nodes and resets the change in local state variables. The output module is used to perform consistency constraint correction based on the received local state variables of neighboring nodes. If the preset convergence condition is met, the restored target image is output based on the synchronized local image estimation matrix.

15. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores computer instructions for causing a computer to perform the distributed image restoration method as described in any one of claims 1 to 12.

16. A computer device, characterized in that, include: At least one processor, and a memory communicatively connected to the at least one processor; wherein the memory stores a computer program executable by the at least one processor, the computer program being executed by the at least one processor to enable the computer device to perform the distributed image restoration method as described in any one of claims 1 to 12.