Event camera real-time deblurring method based on dynamic edge guidance

Abstract

The invention provides an event camera real-time deblurring method based on dynamic edge guidance. The event camera real-time deblurring method comprises the following steps: acquiring a to-be-recovered blurred image and event stream data in the same field of view; extracting an image modal feature and an event modal feature by using a double-flow visual encoder; constructing a dynamic edge guiding module, predicting pixel-level multi-level fuzzy probability distribution, and generating a dynamic space edge priori graph; removing redundant background events by using an edge prior graph, and screening out a key event Token rich in motion information; and constructing a cross-modal fusion module based on bidirectional RWKV, carrying out long-distance space-time modeling on the screened event Token and image features with linear calculation complexity, and finally decoding and outputting a clear image. According to the method, the sensor noise is effectively suppressed and the calculation redundancy is reduced under the condition of not depending on the traditional hard gating through a'guide-selection 'normal form, and the image deblurring which is high in signal-to-noise ratio and high in structural similarity and meets the real-time requirement is realized.

Description

Technical Field

[0001] This invention relates to the fields of computer vision and computational photography, specifically to a real-time deblurring method for event cameras based on dynamic edge guidance. Background Technology

[0002] Image deblurring, a core task in computer vision and computational photography, aims to restore sharp images by inversely modeling the motion degradation process, providing crucial underlying perceptual support for applications such as autonomous driving, drone navigation, and mobile smart terminal imaging. In recent years, deep learning-based deblurring methods have made significant progress in handling complex dynamic scenes by constructing encoder-decoder architectures to learn the mapping relationship between blurred and sharp images. However, traditional frame-based cameras are limited by their inherent photon integration mechanism and fixed sampling frequency. If there is relative motion between scene objects or the camera within the exposure time, it leads to the loss of high-frequency details, texture boundary degradation, and visual artifacts. This performance limitation stems from the lack of physical motion cues in a single blurred frame. To overcome the perceptual limitations of traditional cameras, event cameras have been introduced into the deblurring task. These sensors possess microsecond-level temporal resolution and extremely high dynamic range, enabling asynchronous recording of brightness change events, thus providing the possibility of capturing high-frequency motion trajectories. Current mainstream methods typically encode event streams into voxel grids and employ cross-modal attention mechanisms to fuse event information with image features to achieve restoration.

[0003] Despite some progress, existing methods still face three key challenges. First, event data exhibits extremely high spatial sparsity and is susceptible to sensor thermal noise. The massive amount of redundant events generated, especially in complex texture regions, not only increases computational overhead but also introduces non-stationary noise during fusion, leading to artifacts in the restored results. Second, existing solutions largely rely on computationally complex global self-attention or cross-attention mechanisms (such as Transformer-based architectures), whose computational complexity increases quadratically with the amount of input data, making it difficult to meet the real-time processing requirements of embedded devices when handling high-resolution or long-sequence data. Third, existing methods generally lack effective structure guidance mechanisms, often treating all events indiscriminately and ignoring the spatial non-uniformity of image blur intensity. This prevents precise semantic focusing on motion boundary regions crucial for deblurring, resulting in structural blurring at object contours in the restored image. In short, existing methods are significantly inadequate in selecting spatial sparsity for modeling event flows and in their real-time spatiotemporal modeling capabilities, making it difficult to cope with massive redundant data interference and real-time bottlenecks. The aforementioned limitations reveal a core issue: how to construct a real-time deblurring framework that can adaptively filter key motion cues and achieve long-range spatiotemporal modeling with linear computational complexity, so as to achieve a balance between high fidelity and high efficiency. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a real-time deblurring method for event cameras based on dynamic edge guidance. By constructing a spatial structure prior based on blur probability, the method guides the model to adaptively filter key motion information in the spatiotemporal dimensions. Combined with a cross-modal linear recursive scanning mechanism, it achieves efficient feature fusion, thereby significantly improving the real-time performance and reconstruction accuracy of event-assisted deblurring in complex high-speed motion scenes, thus solving the problems mentioned in the background. The technical solution provided by this invention is as follows:

[0005] A real-time deblurring method for event cameras based on dynamic edge guidance includes the following steps:

[0006] Step 1: Collect blurred image data and corresponding event data in dynamic scenes, obtain clear images synchronized with the time as real labels, and construct an event deblurring dataset for training and evaluation.

[0007] Step 2: Spatiotemporally encode the original event data to convert it into an event feature tensor; use a two-stream visual encoder to extract deep semantic features from the blurred image data and event data respectively, as input for subsequent cross-modal interaction;

[0008] Step 3: Construct a dynamic edge-guided deblurring network based on the RWKV architecture. This network includes a dynamic edge-guided module, an edge-guided token selection module, and a bidirectional RWKV cross-modal fusion module.

[0009] Step 4: The preprocessed dataset is divided into training, validation, and test sets. The training set is used to train the model through end-to-end backpropagation. The dynamic edge guidance module is responsible for predicting the pixel-level blur probability distribution and synthesizing the dynamic edge map. The edge guidance token selection module removes redundant background events based on the edge map. The bidirectional RWKV cross-modal fusion module performs long-distance spatiotemporal modeling of the selected key event tokens and image features with linear complexity. The pixel-level reconstruction loss function is minimized through gradient descent algorithm to optimize the model parameters.

[0010] Step 5: After training is completed, select the model weights with the best performance on the validation set, and use them to infer the real-time blurred image and event stream to generate a high-fidelity clear image.

[0011] Preferably, step 2 specifically involves: the event feature tensor is constructed by dividing the event data within the exposure time into T equally divided time sub-intervals and embedding them linearly through interpolation; this tensor is then divided into fixed-size patches and flattened into token sequences; dual-stream Vision-RWKV is used as the backbone network, and the linear attention mechanism in the RWKV blocks is utilized to capture long-distance dependencies, outputting the downsampled latent image features respectively. and potential characteristics of events Subsequently, the two token sequences—blurred image data and event data—are reconstructed into spatial feature maps, represented as follows: and ,in For the number of channels, and This represents the resolution after downsampling.

[0012] Preferably, the dynamic edge guidance module is constructed as follows: firstly, image features are... Event characteristics The data is concatenated along the channel dimension, and cross-modal information is integrated through a 1×1 convolutional layer to obtain fused features. The calculation formula is as follows:

[0013]

[0014] in Indicates channel splicing. This represents a 1×1 convolutional layer;

[0015] Next, edge estimation is modeled as a probabilistic fuzzy ranking prediction task: utilizing a projection layer. Fusion features Perform mapping to predict Logit response values ​​for five different fuzzy levels:

[0016]

[0017] For each pixel in the feature map Calculate its belonging to the th order through the Softmax operation. The probability of each fuzzy level :

[0018]

[0019] This yields a pixel-level probability distribution map reflecting different degrees of blur.

[0020] Finally, the predicted fuzzy probabilities at each level are weighted and summed to generate the final dynamic spatial edge map. :

[0021]

[0022] in, The predefined fuzzy intensity weighting coefficients are used; the resulting As a spatial probability prior: in Areas with lower values ​​correspond to static backgrounds or slight blurring. Areas with higher values ​​correspond to severe motion blur and object boundaries.

[0023] Preferably, the edge-guided Token selection module has the following specific structure:

[0024] First, calculate the event token sequence. self-similarity matrix Row average Quantify internal redundancy;

[0025] Then the dynamic space edge prior map is used. Flattened into token-level priors and modulate similarity scores ,in As a spatial suppression factor, it reduces the redundancy score in regions with high fuzziness probability.

[0026] Then select the redundancy score. Lowest front Each location retains the corresponding event feature vector:

[0027]

[0028] in, , To retain the proportion, The number of event tokens; the filtered event feature vector is denoted as... .

[0029] Preferably, the specific structure of the bidirectional RWKV cross-modal fusion module is as follows: the input is image features. and key event features after filtering First, a unified sequence is formed by concatenating along the sequence dimensions. ,in For the number of image tokens, To preserve the number of event tokens; then process via cross-modal RWKV blocks:

[0030]

[0031]

[0032] in For query offset operation, bidirectional Operator, For activation function, and It is the weight matrix; the final output feature map is:

[0033]

[0034] High-quality image reconstruction for subsequent decoders.

[0035] Preferably, an optimization algorithm based on gradient descent is used to perform end-to-end backpropagation training on the network model. The training process updates the network parameters by minimizing the pixel-level reconstruction loss function, and its specific form is as follows:

[0036]

[0037]

[0038] In the formula, This represents the deblurred image obtained from network prediction. This indicates the corresponding real, clear image label. This represents the L1 distance, used to constrain the consistency between the predicted image and the real sharp image in pixel space.

[0039] Compared with existing technologies, the beneficial effects achieved by this invention are as follows: On the one hand, this invention introduces a Dynamic Edge Guidance Module (DEGM), which uses deep semantic priors to replace traditional hard gradient operators. This enables precise localization of motion degradation regions and guides the Token Selection Module to dynamically filter approximately 20% of redundant event information, significantly suppressing the interference of sensor noise on restoration quality. On the other hand, this invention replaces the quadratic complexity attention mechanism of the traditional Transformer with a linear complexity RWKV fusion architecture, greatly reducing computational overhead while maintaining the global receptive field, thus reducing the single-frame inference time to 0.0084s. This invention fully utilizes the high temporal domain characteristics of event cameras and the texture details of images, achieving ultimate real-time performance while effectively solving the problems of structural artifacts and detail loss in complex dynamic scenes, and has extremely high industrial application value. Attached Figure Description

[0040] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0041] Figure 1 This is the overall flowchart of the method of the present invention;

[0042] Figure 2 This is a diagram of the overall network architecture of the present invention;

[0043] Figure 3 This is a network structure diagram of the dynamic edge guidance module of the present invention. Detailed Implementation

[0044] 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.

[0045] To make the above-mentioned objectives, features and effects of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0046] Example 1: A real-time deblurring method for event cameras based on dynamic edge guidance, such as... Figure 1 As shown, the specific steps are as follows.

[0047] Step 1, Dataset Preparation. Blurred image data and corresponding event data from dynamic scenes are collected, and clear images synchronized with these events are obtained as ground truth labels to construct an event-assisted deblurring dataset for training and evaluation. In this embodiment, the proposed deblurring method is evaluated on two datasets: GoPro and REBlur. Both datasets are collected from real or simulated scenes, contain frame images and event data, and are widely used in event-driven deblurring tasks. For GoPro, the model is trained on the training set and evaluated on its test set; GoPro provides 22 training sequences and 11 test sequences with an image resolution of 640×360, covering various dynamic scenes. For REBlur, a real-world dataset is used for generalization evaluation, containing 983 pairs of blurred and sharp images with an image resolution of 260×360, covering linear and non-linear motion modes.

[0048] Step 2: Spatiotemporally encode the original event data to convert it into an event feature tensor; use a two-stream visual encoder to extract deep semantic features from the blurred image data and the event feature tensor respectively, as input for subsequent cross-modal interaction.

[0049] Specifically, the event feature tensor is constructed by dividing the event data within the exposure time into T equally divided time sub-intervals and embedding them linearly through interpolation. This tensor is then divided into fixed-size patches and flattened into a token sequence. A dual-stream Vision-RWKV network is used as the backbone, leveraging the linear attention mechanism within the RWKV blocks to capture long-range dependencies, and outputting the downsampled latent image features. and potential characteristics of events The two token sequences were then reconstructed into a spatial feature map, represented as follows: and ,in For the number of channels, and This represents the resolution after downsampling. In this embodiment, the time window T=6.

[0050] Step 3: Construct a dynamic edge-guided deblurring network based on a parallel visual RWKV architecture. This network includes a dynamic edge guidance module, an edge guidance token selection module, and a bidirectional RWKV cross-modal fusion module, such as... Figure 2 As shown, the network adopts a "guided selection" paradigm, using structural priors to guide feature sparsification.

[0051] The dynamic edge guidance module aims to synthesize cross-modal structural priors to bridge modal differences, such as Figure 3 The construction method shown is as follows: First, image features are... Event characteristics The data is concatenated along the channel dimension, and cross-modal information is integrated through a 1×1 convolutional layer to obtain fused features. The calculation formula is as follows:

[0052]

[0053] in Indicates channel splicing. This represents a 1×1 convolutional layer; this fusion feature is used to identify localized degradation regions caused by object or camera motion.

[0054] Next, edge estimation is modeled as a probabilistic fuzzy ranking prediction task: utilizing a projection layer. Fusion features Perform mapping to predict Logit response values ​​for five different fuzzy levels:

[0055]

[0056] For each pixel in the feature map Calculate its belonging to the th order through the Softmax operation. The probability of each fuzzy level :

[0057]

[0058] This yields a pixel-level probability distribution map reflecting different degrees of blur.

[0059] Finally, the predicted fuzzy probabilities at each level are weighted and summed to generate the final dynamic spatial edge map. :

[0060]

[0061] in, The predefined fuzzy intensity weighting coefficients are used; the resulting As a spatial probability prior: in Areas with lower values ​​correspond to static backgrounds or slight blurring. Regions with higher values ​​correspond to severe motion blur and object boundaries; this prior map is used to instruct subsequent modules to focus on motion-rich regions in the event sequence. This dynamic edge guidance mechanism, specifically designed for event data, explicitly integrates motion information from images and events, effectively converting high-density temporal cues into spatially aligned prior representations. It demonstrates significant advantages in complex scenarios with blur, noise, and event redundancy, while also providing a robust and refined deblurring coding foundation for downstream event selection modules.

[0062] This invention introduces an edge-guided token selection module. Previous event deblurring methods typically rely on full event processing, but this ignores spatial redundancy, leading to high computational overhead. This invention establishes a hybrid selector combining self-similarity calculation and edge-guided filtering. Unlike Transformer-based selection, this module captures sequence order without explicit positional encoding, thus exhibiting stronger robustness when processing asynchronous event data. Specifically, the event token sequence is first calculated. self-similarity matrix Row average Quantify internal redundancy; then use the dynamic space edge prior map. Flattened into token-level priors and modulate similarity scores ,in As a spatial suppression factor, the redundancy score is reduced in high fuzziness probability regions; then the redundancy score is selected. Lowest front Each location retains its corresponding event feature vector, serving as a key motion clue for subsequent image detail recovery. , To retain the percentage (e.g., 0.8). The number of event tokens is denoted as . For smooth backgrounds or static regions with high scores, their corresponding event features are suppressed to filter sensor background noise and significantly reduce computational redundancy in subsequent fusion modules. The filtered sparse event features are denoted as . .

[0063] The bidirectional RWKV cross-modal fusion module employs an attention mechanism with linear computational complexity. Its specific structure is as follows: the input is image features. and key event features after filtering First, a unified sequence is formed by concatenating along the sequence dimensions. ,in For the number of image tokens, To preserve the number of event tokens, the module includes two recursive paths: forward and backward. It captures the spatiotemporal relationships between tokens using a linear recursive formula, thus preserving the original... The complexity is reduced to ,in The length of the token sequence; the specific processing procedure is as follows:

[0064] calculate

[0065]

[0066] Then merge

[0067]

[0068] in For query offset operation, bidirectional Operator, For activation function, and It is a weight matrix; through bidirectional linear interaction, high temporal resolution event motion information is accurately compensated into the feature representation of the blurred image, realizing cross-modal global dependency modeling; the final output is a feature map.

[0069]

[0070] High-quality image reconstruction is achieved through subsequent decoding. The decoding stage integrates an edge-prior-based selection mechanism, cross-modal long-range modeling capabilities, and linear complex fusion. Unlike previous methods that typically process images and events independently, the bidirectional RWKV cross-modal fusion module explicitly and jointly models deblurred spatial patterns and motion cues, significantly improving representation consistency and detail completeness in complex scenes with severe blur, low texture, and rapid motion.

[0071] Step 4: The preprocessed dataset is divided into training, validation, and test sets. The training set is used to train the model through end-to-end backpropagation. The dynamic edge guidance module is responsible for predicting the pixel-level blur probability distribution and synthesizing the dynamic edge map. The edge guidance token selection module removes redundant background events based on the edge map. The bidirectional RWKV cross-modal fusion module performs long-distance spatiotemporal modeling of the selected key event tokens and image features with linear complexity. The pixel-level reconstruction loss function is minimized through the gradient descent algorithm to optimize the model parameters.

[0072] This embodiment uses the Adam optimizer with a learning rate set to 1×10⁻⁶. -4The model was trained on a GoPro with a batch size of 8. An end-to-end backpropagation training algorithm based on gradient descent was used to train the network model. The training process updates the network parameters by minimizing the pixel-level reconstruction loss function, as follows:

[0073]

[0074]

[0075] In the formula, This represents the deblurred image obtained from network prediction. This indicates the corresponding real, clear image label. This represents the L1 distance, used to constrain the consistency between the predicted image and the real sharp image in pixel space.

[0076] Step 5: After training is completed, select the model weights with the best performance on the validation set, and use them to infer the real-time blurred image and event stream to generate a high-fidelity clear image.

[0077] To verify the effectiveness of the proposed method, it was compared with several existing methods, including MIMO-UNet (S.-J. Cho, et al., ICCV, 2021), NAFNet (L. Chen, et al., ECCV, 2022), XYDeblur (Seo-Won Ji, et al., CVPR, 2022), E2VID (H. Rebecq, et al., IEEE TPAMI, 2021), EVSNN (L. Zhu, et al., CVPR, 2022), eSL-Net (B. Wang, et al., ECCV, 2020), EDI (L. Pan, et al., CVPR, 2019), RED (F. Xu, et al., ICCV, 2021), EVDI (X. Zhang, et al., CVPR, 2022), and LEDVDI (S. Lin, et al.). ECCV (2020), EFNet (L. Sun, et al., ECCV, 2022), and TRMD (K. Chen, et al., IEEE TMM, 2024) were compared. This embodiment uses Peak Signal-to-Noise Ratio (PSNR), Structural Similarity (SSIM), and Learned Perceptual Image Patch Similarity (LPIPS) as the main evaluation metrics on GoPro and REBlur. In addition, inference time (TIME), number of parameters (Params), and floating-point operations (FLOPs) are also introduced. Table 1 shows the performance comparison of different methods on the GoPro dataset. Frame-based and event-based methods generally perform worse than event-assisted methods. Frame-based methods such as MIMO-UNet and NAFNet rely on pure image data, resulting in high LPIPS errors. Event-based methods such as E2VID and EVSNN also perform poorly, reflecting their difficulty in capturing dense and fine-grained details in the absence of image texture. In contrast, event-assisted methods generally achieve stronger performance, but still have shortcomings in fusion efficiency. EDI and RED aim for simple fusion, but their temporal modeling is relatively shallow, making them ill-suited for complex and blurred scenes, and their performance remains insufficient. LEDVDI introduces a cyclic structure to adapt to event sparsity, while EFNet enhances fusion capabilities through cross-modal attention; however, both remain weak in redundancy suppression. TRMD utilizes residual learning to guide deblurring, exhibiting strong robustness in structured scenes, but struggles to adapt to complex, fine-grained dynamic changes. Notably, existing methods generally do not explicitly utilize dynamic edge priors or construct efficient selection mechanisms.The model proposed in this invention achieved a PSNR of 38.67 on GoPro, an improvement of approximately 2.0 dB over TRMD. Compared with existing state-of-the-art methods, the method of this invention achieves a good balance between accuracy and efficiency, even outperforming EFNet, fully demonstrating the effective modeling capability of the proposed model for redundant structures in event data.

[0078] Table 1: Performance Comparison of the Method of the Present Invention and Other Methods

[0079]

[0080] Example 2: The computer-readable storage medium of this example stores a computer program that, when executed by a processor, implements the steps of the real-time deblurring method for an event camera based on dynamic edge guidance in Example 1.

[0081] The computer-readable storage medium in this embodiment can be an internal storage unit of the terminal, such as the terminal's hard disk or memory; the computer-readable storage medium in this embodiment can also be an external storage device of the terminal, such as a plug-in hard disk, smart memory card, secure digital card, flash memory card, etc. equipped on the terminal; furthermore, the computer-readable storage medium can include both the terminal's internal storage unit and external storage devices.

[0082] The computer-readable storage medium of this embodiment is used to store computer programs and other programs and data required by the terminal. The computer-readable storage medium can also be used to temporarily store data that has been output or will be output.

[0083] Example 3: The computer device of this example includes a processor, a memory, and a computer program stored in the memory and executable on the processor. When the processor executes the program, it implements the steps in the real-time deblurring method for event cameras based on dynamic edge guidance in Example 1.

[0084] In this embodiment, the processor can be a central processing unit, or other general-purpose processors, digital signal processors, application-specific integrated circuits, off-the-shelf programmable gate arrays or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor or any conventional processor, etc. The memory can include read-only memory and random access memory, and provides instructions and data to the processor. A portion of the memory can also include non-volatile random access memory. For example, the memory can also store device type information.

[0085] Those skilled in the art will clearly understand that each implementation can be achieved using software plus the necessary general-purpose hardware platform, or of course, hardware. Based on this understanding, the above technical solutions, in essence or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product can be stored in a computer-readable storage medium, such as ROM / RAM, magnetic disk, optical disk, etc., and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute the methods described in the various embodiments or some parts of the embodiments.

[0086] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A real-time deblurring method for event cameras based on dynamic edge guidance, characterized in that, Includes the following steps: Step 1: Collect blurred image data and corresponding event data in dynamic scenes, obtain clear images synchronized with the time as real labels, and construct an event deblurring dataset for training and evaluation. Step 2: Spatiotemporally encode the original event data to convert it into an event feature tensor; use a two-stream visual encoder to extract deep semantic features from the blurred image data and event data respectively, as input for subsequent cross-modal interaction; Step 3: Construct a dynamic edge-guided deblurring network based on the RWKV architecture. This network includes a dynamic edge-guided module, an edge-guided token selection module, and a bidirectional RWKV cross-modal fusion module. Step 4: Divide the preprocessed dataset into a training set, a validation set, and a test set, and use the training set to train the model through end-to-end backpropagation. The dynamic edge guidance module is responsible for predicting the pixel-level blur probability distribution and synthesizing a dynamic edge map; the edge guidance token selection module removes redundant background events based on the edge map; the bidirectional RWKV cross-modal fusion module performs long-distance spatiotemporal modeling of the selected key event tokens and image features with linear complexity; and the pixel-level reconstruction loss function is minimized through the gradient descent algorithm to optimize the model parameters. Step 5: After training is completed, select the model weights with the best performance on the validation set, and use them to infer the real-time blurred image and event stream to generate a high-fidelity clear image.

2. The real-time deblurring method for event cameras based on dynamic edge guidance according to claim 1, characterized in that, Step 2 specifically involves: the event feature tensor is constructed by dividing the event data within the exposure time into T equally divided time sub-intervals and embedding them linearly through interpolation. This tensor is then divided into fixed-size patches and flattened into token sequences. A dual-stream Vision-RWKV network is used as the backbone network, leveraging the linear attention mechanism within the RWKV blocks to capture long-range dependencies, and outputting the downsampled latent image features. and potential characteristics of events ; The two token sequences, namely the blurred image data and the event data, were then reconstructed into spatial feature maps, represented as follows: and ,in For the number of channels, and This represents the resolution after downsampling.

3. The real-time deblurring method for event cameras based on dynamic edge guidance according to claim 2, characterized in that, The dynamic edge guidance module is constructed as follows: first, image features are... Event characteristics The data is concatenated along the channel dimension, and cross-modal information is integrated through a 1×1 convolutional layer to obtain fused features. The calculation formula is as follows: ； in Indicates channel splicing. This represents a 1×1 convolutional layer; Next, edge estimation is modeled as a probabilistic fuzzy ranking prediction task: utilizing a projection layer. Fusion features Perform mapping to predict Logit response values ​​for five different fuzzy levels: ； For each pixel in the feature map Calculate its belonging to the th order through the Softmax operation. The probability of each fuzzy level : ； This yields a pixel-level probability distribution map reflecting different degrees of blur. Finally, the predicted fuzzy probabilities at each level are weighted and summed to generate the final dynamic spatial edge map. : ； in, The predefined fuzzy intensity weighting coefficients are used; the resulting As a spatial probability prior: in Areas with lower values ​​correspond to static backgrounds or slight blurring. Areas with higher values ​​correspond to severe motion blur and object boundaries.

4. The real-time deblurring method for event cameras based on dynamic edge guidance according to claim 3, characterized in that, The specific structure of the edge-guided Token selection module is as follows: First, calculate the event token sequence. self-similarity matrix Row average Quantify internal redundancy; Then the dynamic space edge prior map is used. Flattened into token-level priors and modulate similarity scores ,in As a spatial suppression factor, it reduces the redundancy score in regions with high fuzziness probability. Then select the redundancy score. The lowest front Each location retains the corresponding event feature vector: ； in, , To retain the proportion, The number of event tokens; the filtered event feature vector is denoted as... .

5. A real-time deblurring method for event cameras based on dynamic edge guidance according to claim 4, characterized in that, The specific structure of the bidirectional RWKV cross-modal fusion module is as follows: the input is image features. and key event features after filtering First, a unified sequence is formed by concatenating along the sequence dimensions. ,in For the number of image tokens, To preserve the number of event tokens; then process via cross-modal RWKV blocks: ； ； in For query offset operation, bidirectional Operator, For activation function, and It is the weight matrix; the final output feature map is: ； High-quality image reconstruction for subsequent decoders.

6. The real-time deblurring method for event cameras based on dynamic edge guidance according to claim 5, characterized in that, An optimization algorithm based on gradient descent is used to train the network model through end-to-end backpropagation. The training process updates the network parameters by minimizing the pixel-level reconstruction loss function, as follows: ； ； In the formula, This represents the deblurred image obtained from network prediction. This indicates the corresponding real, clear image label. This represents the L1 distance, used to constrain the consistency between the predicted image and the real sharp image in pixel space.

7. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the steps in the real-time deblurring method for event cameras based on dynamic edge guidance as described in any one of claims 1-6.

8. A computer device comprising a processor, a memory, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the steps in the real-time deblurring method for event cameras based on dynamic edge guidance as described in any one of claims 1-6.

Images