An event camera based calibration-free anti-noise low-light image enhancement method

Abstract

This invention discloses a calibration-free, noise-resistant low-light image enhancement method based on an event camera, belonging to the fields of computer vision and image processing technology. The method first introduces an online physical perturbation synthesis strategy during the training phase, forcing the network to learn adaptive calibration capabilities for uncalibrated data by simulating rigid spatial transformations. During the inference phase, a global coarse alignment network is used to correct large-scale physical parallax at low resolution, followed by sub-pixel-level calibration and denoising at the original resolution using a noise-resistant modulation fine alignment network. The noise-resistant modulation network adaptively shields noise interference through a prediction confidence mask. Next, a spatial attention mechanism is used to achieve weighted fusion of multimodal features. Finally, the output is constrained by edge structure consistency loss. This invention does not rely on expensive precision hardware calibration and can effectively solve the problems of noise interference and non-rigid alignment in extremely low-light environments, achieving clear, ghost-free low-light imaging.

Description

Technical Field

[0001] This invention relates to the field of computer vision and image processing technology, specifically to a low-light image enhancement method that utilizes an event camera and does not require precise hardware calibration during the inference phase. Background Technology

[0002] Image enhancement technology has wide applications in fields such as nighttime autonomous driving, security monitoring, and all-weather robot navigation. Under extremely low light conditions, traditional CMOS image sensors, limited by exposure time and photosensitive principles, typically produce images with severe shot noise, color distortion, and motion blur. Event cameras, as a novel bio-inspired sensor, possess high dynamic range and microsecond-level temporal resolution, enabling asynchronous capture of brightness changes in a scene, unaffected by motion blur. Therefore, utilizing event cameras to assist in low-light enhancement of RGB images has become a current research hotspot.

[0003] However, existing event camera-based enhancement methods face significant challenges in practical applications, primarily including:

[0004] 1. The challenge of physical alignment for multimodal data: Existing mainstream algorithms typically assume perfect pixel-level alignment between the input event stream and the RGB image in both time and space. To meet this assumption, laboratory environments usually rely on expensive beam splitters or high-precision robotic arms for rigorous data acquisition and calibration. However, in consumer-grade binocular devices or automotive scenarios, physical parallax is unavoidable due to assembly tolerances, thermal expansion and contraction, or mechanical vibration. Existing algorithms lack modeling for physical parallax; if the input data is not precisely geometrically corrected, the enhanced image will produce severe ghosting and artifacts, causing the algorithm to fail.

[0005] 2. Noise interference in extremely low light environments: Traditional alignment algorithms, such as optical flow or standard deformable convolutional neural networks (DCN), often rely on the texture gradient of the image to calculate the offset. However, in extremely low light conditions, RGB images are filled with high-frequency shot noise, almost completely obscuring the effective texture. Existing alignment modules are easily misled by noise, predicting random and incorrect offsets, leading to "misaligned" features and ultimately destroying the structural information of the image.

[0006] Therefore, in order to address the problems of low signal-to-noise ratio of traditional CMOS images in low-light environments and the difficulty in eliminating physical parallax in binocular devices, there is an urgent need for an image enhancement method that can adapt to uncalibrated physical parallax and effectively resist interference from extremely low light noise. Summary of the Invention

[0007] The purpose of this invention is to overcome the shortcomings of existing technologies and propose a calibration-free, noise-resistant low-light image enhancement method based on an event camera. This method is applicable to binocular data with overlapping fields of view in the same scene but with installation errors or parallax. This invention does not rely on expensive precision hardware calibration and can effectively solve the problems of noise interference and non-rigid alignment in extremely low-light environments, achieving clear, ghost-free low-light imaging.

[0008] To achieve the above objectives, the present invention adopts the following technical solution: In a first aspect, the present invention provides a calibration-free noise-resistant low-light image enhancement method based on an event camera, the method comprising the following steps: S1. Data Acquisition: Acquire low-light RGB images and event stream data with overlapping fields of view in the same scene; spatial parallax without geometric correction is allowed between the event stream data and the low-light RGB images; S2. Progressive network input: The low-light RGB image and event stream data are input into a pre-trained progressive enhancement network; the network includes a cascaded first-stage coarse processing module and a second-stage fine processing module; S3. First-stage global coarse alignment and reconstruction: In the first-stage coarse processing module, the input data is downsampled, a lightweight offset prediction network is used to predict the global offset, and the event features are deformably resampled based on the global offset to complete the coarse alignment. Then the features are stitched together and reconstructed through the decoder to output the intermediate enhanced image. S4. Second-stage noise-resistant fine alignment: In the second-stage fine processing module, the intermediate enhanced image is used as a priori, and the event features of the original resolution are processed by the noise-resistant modulation fine alignment unit; the unit adaptively filters noise regions by predicting confidence masks to achieve sub-pixel level alignment and denoising. S5. Fine-grained fusion and residual output: The attention mechanism is used to perform weighted fusion of aligned event features and low-light image features. The residual map is generated by the decoder and finally added to the original low-light RGB image to output the final enhanced image.

[0009] Furthermore, the network training process prior to step S2 includes: An online physical perturbation synthesis strategy is introduced at the data input end. During the network training iteration, an affine transformation matrix containing translation and rotation parameters is randomly generated. The affine transformation matrix is ​​used to perform spatial transformation on the aligned event data ground value to generate simulated misalignment data as network input. The aligned clear image is used as a supervision signal to train the network's ability to perform inverse transformation of physical parallax.

[0010] Furthermore, step S3 specifically includes: The low-light RGB image and event stream data are downsampled, and low-resolution image content features and event structure features are extracted through convolutional layers. The image content features and event structure features are concatenated along the channel dimension and input into a lightweight offset prediction network to predict the global offset. The low-resolution event structure features are deformably resampled using the global offset to obtain coarse-aligned event features. The coarse-aligned event features are concatenated with the image content features and input into a coarse decoder to map the features back to the RGB space, outputting the intermediate enhanced image with restored brightness and roughly aligned contours.

[0011] Furthermore, step S4 specifically includes: Features and event stream features of the original resolution low-light RGB image are extracted, and prior features are extracted after upsampling the intermediate enhanced image. These features are then input into the noise-resistant modulation fine alignment unit to predict the position offset. and confidence mask Feature alignment is performed using the following formula:

[0012] in, The current pixel coordinates, The number of sampling points. For convolution weights, Fixed offset; The range of values ​​is It is used to suppress feature aggregation in noisy regions and enhance feature response in textured regions. This indicates the aligned event feature output at the current pixel coordinate p; Indicates the original event features at the sampling location The eigenvalue at that location is calculated using bilinear interpolation.

[0013] Furthermore, step S5 specifically includes: A spatial attention map is generated based on low-light RGB image features to indicate effective texture regions in the scene; the spatial attention map is then used to align the event features generated in step S4. Weighted features are obtained by performing weighted processing; the weighted event features are then concatenated with the original low-light RGB image features along the channel dimension and input into a fine decoder to generate a residual map. The residual map Compared to the original low-light RGB image The formula for pixel-by-pixel addition is: This yields the final enhanced image.

[0014] Furthermore, the training loss function of the network includes reconstruction loss, perceptual loss, and edge structure consistency loss; wherein, the reconstruction loss is used to constrain the pixel-level color reproduction of the enhanced image and the ground truth image; the perceptual loss is used to constrain the consistency of the two in high-level semantic features; the edge structure consistency loss is obtained by calculating the L1 distance between the gradient field of the final enhanced image and the gradient field of the aligned event features, and is used to force the edge structure of the enhanced image to be consistent with the event stream data, thereby eliminating ghosting caused by alignment errors; the calculation formula is:

[0015] in, This indicates the loss of consistency in edge structures. This represents the final enhanced image. Indicates the characteristics of the alignment event. Represents the gradient operator, This represents the L1 norm.

[0016] In a second aspect, the present invention also provides an electronic device, including a processor and a memory, the memory storing machine-executable instructions executable by the processor, the processor executing the machine-executable instructions to implement the above-described method.

[0017] As can be seen from the above technical solution, the present invention proposes a calibration-free noise-resistant low-light image enhancement method based on an event camera. Compared with the prior art, the beneficial effects include: 1. Achieving a balance between large-scale parallax correction and fine texture restoration: This invention adopts a progressive architecture from coarse to fine. Utilizing the physical low-pass filtering characteristics of downsampling operations, noise is effectively suppressed and global large-scale displacement is captured in the first stage; using the original resolution processing in the second stage, sub-pixel-level texture details are accurately restored.

[0018] 2. Reduced deployment costs and improved system robustness: Existing technologies rely on expensive beam splitters or robotic arms for precise calibration. This invention, through an online physical perturbation synthesis strategy, simulates rigid transformation errors in the real physical world, forcing the network to learn inverse transformation capabilities. This allows the method to be directly deployed on low-cost, uncalibrated binocular devices, still outputting high-quality, ghost-free images despite installation errors and parallax.

[0019] 3. Solves the noise interference problem in extremely low light: Existing alignment algorithms are easily misled by shot noise in low light. The noise-resistant modulation fine alignment unit proposed in this invention can adaptively perceive the signal-to-noise ratio of image content by introducing a learnable confidence mask mechanism. It fully utilizes event information in texture-rich areas and automatically cuts off noise propagation in pure noise areas. This "soft gating" mechanism fundamentally ensures the purity of the enhanced image.

[0020] Other features and advantages of the invention will be set forth in the following description, and will be apparent in part from the description, or may be learned by practicing the invention. The objects and other advantages of the invention may be realized and obtained by means of the structures particularly pointed out in the written description and the accompanying drawings.

[0021] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0022] To more clearly illustrate the technical solutions in the embodiments of this application 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 some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0023] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof.

[0024] Figure 1 This is a schematic diagram of the overall process of the calibration-free noise-resistant low-light image enhancement method based on an event camera provided by the present invention.

[0025] Figure 2 This is a schematic diagram of the overall architecture of the calibration-free noise-resistant low-light image enhancement method based on an event camera provided by the present invention.

[0026] Figure 3 This is a schematic diagram of the electronic device structure provided by the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are some embodiments of the present invention, but not all embodiments.

[0028] In the description of this invention, it should be noted that some processes described in this application specification and drawings include multiple operations that appear in a specific order. However, it should be clearly understood that these operations may be performed in any order or in parallel. Furthermore, various numbers are used for descriptive purposes only and should not be construed as indicating or implying relative importance.

[0029] Therefore, the following detailed description of the embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected embodiments of the invention. All other embodiments obtained by those skilled in the art based on the embodiments of the invention without inventive effort are within the scope of protection of the invention.

[0030] See Figure 1 As shown, this invention discloses a calibration-free noise-resistant low-light image enhancement method based on an event camera. The term "calibration-free" specifically refers to the fact that, during model inference or practical application, there is no need to pre-calculate the camera's extrinsic parameter matrix using calibration references (such as a checkerboard pattern), nor is it necessary to pre-correct the input data based on a geometric model. The core process of this method includes: S1. Data Acquisition: Acquire low-light RGB images with overlapping fields of view in the same scene. and event stream data At this point, there is an unknown physical parallax between the two. S2. Progressive network processing: The data is sequentially input into a pre-trained progressive enhancement network, which contains a cascaded first-stage coarse processing module and a second-stage fine processing module.

[0031] The subsequent steps S3 to S5 will be explained in detail in conjunction with the specific modules of each stage.

[0032] For the network training phase, this invention introduces an online physical perturbation synthesis strategy to solve the calibration-free problem. Specifically, in each iteration of network training, rotation parameters are randomly sampled. Translation parameters Construct the affine transformation matrix The truth value of the original aligned event voxels is obtained using this matrix. Perform spatial transformation to generate simulated misalignment data. As network input, and with a clear, aligned image. This serves as a supervisory signal. In this way, the network is forced to learn the inverse transform ability to recover aligned outputs from misaligned inputs.

[0033] Step S3, the first stage of global coarse alignment and reconstruction, aims to capture large-scale physical displacements and restore the basic brightness. Its specific implementation steps are as follows: S31, Multi-scale downsampling: This involves downsampling the input low-light RGB image. and event stream data Input network, where H and W represent height and width, respectively. This represents the number of channels in the event voxel grid. Two consecutive downsampling operations are performed using a convolutional layer with a stride of 2 to extract multi-scale contextual features. This process can be represented as:

[0034] in, This indicates a downsampling operation that includes convolution, normalization, and activation functions; These represent the image content features and event structure features after downsampling, respectively. This represents the number of feature channels.

[0035] S32. Global Migration Prediction: To capture large-scale physical displacements, image features and event features are first concatenated along the channel dimension to obtain a joint feature tensor. This tensor is then input into a lightweight migration prediction network. The calculation formula is as follows:

[0036] in, This indicates a concatenation operation along the channel dimension; It consists of 3 layers of residual blocks and terminal convolutional layers; This is the output global offset, which includes the horizontal offset. and vertical offset .

[0037] S33, Feature Coarse Alignment: Utilizing Predicted Global Offsets For low-resolution event features Perform deformable resampling, i.e., perform spatial transformation operations. This pushes event features to the spatial location of image features:

[0038] in, This is a resampling function based on bilinear interpolation. This is a rough description of the events after alignment.

[0039] S34, Coarse Decoding and Reconstruction: Coarse alignment event features Image content features The concatenated features are then fed into the coarse decoder. The coarse decoder consists of multiple upsampling layers, convolutional layers, and activation functions. Its working principle is as follows: it receives the concatenated features and, through progressive feature mapping and spatial resolution expansion across network layers, decodes the deep features from the low-resolution space and maps them back to the original high-resolution RGB pixel space. The coarse decoder is used to restore basic brightness and reconstruct large-scale structures, outputting an intermediate enhanced image. The image's brightness has been significantly restored, and the overall outline is largely aligned, but there may still be minor subpixel misalignments and blurring, which will be passed on as prior information to the next stage.

[0040] Step S4, the second stage of noise-resistant fine alignment, aims to solve noise interference and sub-pixel parallax residue under extremely low illumination by predicting confidence masks. The specific implementation steps are as follows: S41. Multi-source feature extraction: To capture fine textures and utilize the correction results from the previous stage, the network receives three sets of inputs: a low-light image at the original resolution. event stream data at original resolution and the intermediate enhanced image upsampled to the original size using bilinear interpolation. Features are extracted separately using a feature extraction network, and then concatenated along the channel dimension as input features for the fine alignment unit. :

[0041] in This indicates the convolution feature extraction operation.

[0042] S42, Noise Resistance Modulation Calculation: This involves calculating the features... Input noise-resistant modulation fine alignment unit, such as Figure 2 As shown, this unit contains two parallel convolutional branches, used to predict geometric deformation and quality confidence, respectively: Offset prediction branch: Outputs each pixel Corresponding sampling point position offset ,in This represents the total number of sampling points.

[0043] Mask prediction branch: Outputs the confidence mask for each sample point. The branch ends with a Sigmoid activation function, forcibly constraining the mask value to a certain value. Within the range.

[0044] S43, Weighted Alignment Sampling: Utilizing the predicted... and Characteristics of the original event Perform noise-resistant modulation convolution. The calculation formula is:

[0045] in, The current center pixel coordinates, To fix the convolution grid offset, These are the convolution weights learned by the model. Physical meaning explanation: It acts as an adaptive soft-gating switch. In textured regions with high signal-to-noise ratios, the network predicts... The network fully utilizes high-frequency information from events for sharpening; in pure noise regions with extremely low signal-to-noise ratios, the network predicts... This mechanism automatically blocks feature aggregation at that location. It fundamentally prevents misalignment caused by noise-induced alignment.

[0046] Step S5, fine fusion and residual output, are implemented as follows: S51. Attention Map Generation: To distinguish between effective textured regions and smooth background regions in a scene, a lightweight attention module is used, containing convolutional layers and a sigmoid activation function, to process low-light image features. Generate a spatial attention map:

[0047] in This represents the Sigmoid function. This represents the convolutional feature mapping operation in the lightweight attention module. Regions with high response values ​​indicate areas in the image that are rich in texture and require focused enhancement.

[0048] S52, Weighted Fusion: Utilizing the generated spatial attention map Alignment event features output from step S4 By performing element-wise weighting, we obtain the weighted event characteristics:

[0049] in This indicates an element-wise multiplication operation. This step further enhances the feature representation of the effective texture region while suppressing invalid event noise that may exist in the background region.

[0050] S53, Residual Decoding: Decoding the weighted event features Features of the original low-light image Concatenate along the channel dimension and input to a fine-grained decoder. The decoder reconstructs the image through multiple convolutional layers and outputs a residual map. :

[0051] The residual plot It mainly includes high-frequency detail information missing from the image and the brightness component that needs to be enhanced.

[0052] S54. Final Output: Using global residual join, the residual map is... Compared to the original low light input Pixel-by-pixel addition is performed to obtain the final high-quality enhanced image. :

[0053] Furthermore, in one specific implementation, the training loss function of the network of the present invention is defined as follows: Total loss function It is composed of a weighted average of reconstruction loss, perception loss, and edge structure consistency loss:

[0054] in The hyperparameters are used to balance the weights of each loss term. The specific definitions of each loss term are: Reconstruction losses ( ): Used to constrain the augmented image to match the ground truth image at the pixel level. For consistency, the Charbonnier loss function is used:

[0055] in It is a constant used for numerical stability.

[0056] Perceived loss ( ): Used to enhance the visual appeal of images. Features are extracted using a pre-trained VGG-19 network, and the Euclidean distance between the enhanced image and the ground truth image in the feature space is calculated.

[0057] in This indicates the VGG-19 network's... Feature map of the layer.

[0058] Edge structure consistency loss ( This is the core constraint of this invention. Utilizing the inherent high-frequency edge characteristics of event stream data, the gradient field of the enhanced image is forced to maintain consistency with the aligned event feature gradient field, thereby eliminating ghosting and sharpening edges.

[0059] in Represents the gradient operator, This represents the L1 norm. This loss function ensures that even in areas of extremely low illumination, the enhanced image maintains a clear geometric structure consistent with the event data.

[0060] As described in the above embodiments, those skilled in the art will understand that the present invention specifically provides a calibration-free, noise-resistant low-light image enhancement method based on an event camera. This method addresses the problems of low signal-to-noise ratio in traditional CMOS images and the difficulty in eliminating physical parallax in binocular devices under low-light conditions, proposing a progressive enhancement scheme. Firstly, during the training phase, an online physical perturbation synthesis strategy is introduced, forcing the network to learn adaptive calibration capabilities for uncalibrated data by simulating rigid spatial transformations. During the inference phase, a global coarse alignment network is used to correct large-scale physical parallax at low resolution, followed by sub-pixel-level calibration and denoising at the original resolution using a noise-resistant modulation fine alignment network. The noise-resistant modulation network adaptively shields noise interference through a prediction confidence mask. Next, a spatial attention mechanism is used to achieve weighted fusion of multimodal features. Finally, the output is constrained by edge structure consistency loss. This invention does not rely on expensive precision hardware calibration and can effectively solve the problems of noise interference and non-rigid alignment in extremely low-light environments, achieving clear, ghost-free low-light imaging.

[0061] Additionally, refer to Figure 3 As shown, this embodiment of the invention also provides an electronic device, which may include a processor, a memory, a communication bus and a communication interface, and may also include a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to perform the above-described method.

[0062] In some embodiments, the processor may be composed of integrated circuits, such as a single packaged integrated circuit or multiple integrated circuits packaged with the same or different functions. This includes combinations of one or more central processing units, microprocessors, digital processing chips, graphics processors, and various control chips. The processor is the control core of the electronic device, connecting various components of the device through various interfaces and lines. It executes programs or modules stored in memory and calls data stored in the memory to perform various functions and process data within the electronic device.

[0063] The memory may be, for example, an electrical storage device, a magnetic storage device, an optical storage device, an electromagnetic storage device, a semiconductor storage device, or any suitable combination thereof. More specific examples (a non-exhaustive list) of storage media include: portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), static random access memory (SRAM), portable compact disc read-only memory (CD-ROM), digital multifunction disc (DVD), memory sticks, floppy disks, mechanical encoding devices, and any suitable combination thereof.

[0064] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, electronic devices, or computer program products, etc. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0065] It should be noted that the word "comprising" does not exclude the presence of components or steps not listed in the claims. The words "a" or "an" preceding a component do not exclude the presence of a plurality of such components. This invention can be implemented by means of hardware comprising several different components and by means of a suitably programmed computer.

[0066] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0067] The above description of the disclosed embodiments enables those skilled in the art to make or use the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features disclosed herein.

Claims

1. A calibration-free, noise-resistant low-light image enhancement method based on an event camera, characterized in that, The method includes the following steps: S1. Data Acquisition: Acquire images and event stream data to be enhanced that have overlapping fields of view in the same scene; S2. Network Input: The image to be enhanced and the event stream data are input into a pre-trained progressive enhancement network; the progressive enhancement network includes a cascaded first-stage coarse processing module and a second-stage fine processing module; S3. First-stage global coarse alignment and reconstruction: In the first-stage coarse processing module, the input data is downsampled and the global offset is predicted. Based on the global offset, the event features are deformably resampled to complete the coarse alignment. Then, the features are stitched together and reconstructed through the decoder to output the intermediate enhanced image. S4. Second-stage noise-reducing fine alignment: In the second-stage fine processing module, the intermediate enhanced image is used as a priori to process the event features at the original resolution; the unit adaptively filters noise regions through a prediction confidence mask to achieve sub-pixel level alignment and noise reduction; S5. Fine-grained fusion and residual output: The attention mechanism is used to perform weighted fusion of the aligned event features and the features of the image to be enhanced. The residual map is generated by the decoder and finally added to the original image to be enhanced to output the final enhanced image.

2. The method according to claim 1, characterized in that, In step S1, the image to be enhanced is a low-light RGB image.

3. The method according to claim 1, characterized in that, The progressive augmentation network training process includes: An online physical perturbation synthesis strategy is introduced at the data input end. During the network training iteration, an affine transformation matrix containing translation and rotation parameters is randomly generated. The affine transformation matrix is ​​used to perform spatial transformation on the aligned event data ground value to generate simulated misalignment data as network input. The aligned clear image is used as a supervision signal to train the network's ability to perform inverse transformation of physical parallax.

4. The method according to claim 2, characterized in that, Step S3 specifically includes: S31. Multi-scale downsampling: The low-light RGB image and event stream data are downsampled, and low-resolution image content features and event structure features are extracted through convolutional layers, respectively. S32, Global Offset Prediction: The image content features and event structure features are concatenated along the channel dimension and input into a lightweight offset prediction network to predict the global offset. S33. Feature coarse alignment: The low-resolution event structure features are deformably resampled using the global offset to obtain coarsely aligned event features; S34. Coarse Decoding and Reconstruction: The coarse alignment event features and the image content features are concatenated and input into the coarse decoder. The features are mapped back to the RGB space, and the intermediate enhanced image with restored brightness and roughly aligned contours is output.

5. The method according to claim 4, characterized in that, Step S4 specifically includes: S41. Multi-source feature extraction: Extract the features of the low-light RGB image at the original resolution and the event stream features, and extract the prior features after upsampling the intermediate enhanced image; S42. Noise-resistant modulation calculation: Input the above features into the noise-resistant modulation fine alignment unit, and simultaneously predict the position offset. and confidence mask ; S43. Weighted Alignment Sampling: Perform feature alignment using the following formula: in, Indicates the current pixel coordinates The aligned event characteristics output at the location, The number of sampling points. For convolution weights, For fixed offset, Indicates the original event features at the sampling location The eigenvalue at that location.

6. The method according to claim 5, characterized in that, Step S5 specifically includes: S51. Attention Map Generation: Generate a spatial attention map based on low-light RGB image features to indicate effective texture regions in the scene; S52, Weighted Fusion: The spatial attention map is used to weight the aligned event features generated in step 4 to obtain weighted event features; S53, Residual Decoding: The weighted event features are concatenated with the original low-light RGB image features in the channel dimension and input into the fine decoder to generate a residual map; S54. Final output: The residual image is added pixel by pixel to the original low-light RGB image to obtain the final enhanced image.

7. The method according to claim 3, characterized in that, The network's training loss function includes an edge structure consistency loss, where: the edge structure consistency loss is obtained by calculating the L1 distance between the gradient field of the final enhanced image and the gradient field of the aligned event features, used to constrain the edge structure of the enhanced image to maintain consistency with the event stream data, and the calculation formula is: in, This indicates the loss of consistency in edge structures. This represents the final enhanced image. Indicates the characteristics of the alignment event. Represents the gradient operator, This represents the L1 norm.

8. An electronic device, characterized in that, It includes a processor and a memory, the memory storing machine-executable instructions that can be executed by the processor, the processor executing the machine-executable instructions to perform the method as described in any one of claims 1-7.

Images