Super-resolution low-light image enhancement method and system based on block overlap reasoning

Abstract

This invention discloses a method and system for enhancing ultra-high resolution low-light images based on block overlap inference, belonging to the field of image processing technology. The method segments the ultra-high resolution image into sub-blocks with overlapping regions, and uses a deep learning model based on the UNet architecture to independently enhance each sub-block. The model incorporates a Feature Linear Modulation (FiLM) module to adaptively modulate the underlying features using target brightness parameters. It calculates the ratio of the local average brightness of the sub-block to the global average brightness, and uses logarithmic scaling to determine the local target brightness, ensuring brightness continuity. The overlapping regions are progressively fused using a cosine window function to eliminate block artifacts. This invention solves the problems of memory limitations, block artifacts, and brightness discontinuities in low-light enhancement of ultra-high resolution images, achieving high-quality real-time enhancement, and is suitable for scenarios such as wide-area monitoring and astronomical observation.

Description

Technical Field

[0001] This invention relates to a method and system for enhancing ultra-high resolution low-light images based on block overlap inference, belonging to the field of image processing technology. Background Technology

[0002] With the rapid development of image sensor technology, ultra-high resolution imaging equipment has been widely used in wide-area surveillance, astronomical observation, medical imaging, remote sensing and mapping, and other fields. However, ultra-high resolution images generally suffer from low signal-to-noise ratio, blurred details, and color distortion under low-light conditions, which seriously affect the subsequent visual analysis and application effects.

[0003] Traditional low-light image enhancement algorithms mainly include histogram equalization-based methods, Retinex-based methods, and gamma correction-based methods. While these methods are computationally efficient, they have significant limitations when processing ultra-high-resolution images: First, traditional methods typically use globally uniform parameter settings, making it difficult to adaptively adjust according to the lighting conditions of different regions; second, in extremely low-light environments, traditional algorithms are prone to noise amplification, detail loss, and color distortion; third, traditional methods lack an understanding of the image's semantic information, failing to effectively recover details obscured by noise.

[0004] In recent years, deep learning-based low-light enhancement algorithms have made significant progress. Representative works such as EnlightenGAN and ZeroDCE learn the mapping relationship from low-light images to normal-light images through end-to-end neural networks, and their enhancement effects are significantly better than traditional methods. However, these deep learning methods face serious challenges when processing ultra-high-resolution images: First, there's the issue of video memory limitations. Ultra-high resolution images have a massive number of pixels; if stored using 32-bit floating-point numbers, the memory usage per image is substantial. Furthermore, deep learning models need to store intermediate feature maps, gradients, and other information during inference, causing video memory requirements to grow exponentially. Currently, the video memory capacity of mainstream GPUs is insufficient to support end-to-end processing of such large-scale images.

[0005] Second, there's the issue of computational efficiency. Ultra-high-resolution images have a massive number of pixels, making direct full-image inference computationally extremely complex and unsuitable for real-time processing. Using conventional convolutional neural networks, inference times could reach tens of seconds or even minutes, failing to meet the requirements of real-time applications such as monitoring and navigation.

[0006] Third, the block artifact problem. To address the issues of video memory and computational efficiency, existing technologies typically employ a simple image segmentation strategy, dividing a large image into multiple smaller blocks, processing them separately, and then stitching them together. However, this simple segmentation method ignores the continuity of contextual information between blocks, resulting in obvious block artifacts such as brightness discontinuities and color banding at block boundaries in the stitched image, severely impacting image quality.

[0007] Fourth, the issue of brightness consistency. Local lighting conditions may differ across different image blocks. Applying the same enhancement parameters independently to each block can lead to uneven brightness distribution across the entire image, resulting in localized overexposure or underexposure. Existing technologies lack effective mechanisms for maintaining brightness consistency, making it difficult to achieve globally coordinated enhancement effects.

[0008] To address the aforementioned issues, there is an urgent need in this field for a low-light enhancement method that can effectively process ultra-high resolution images, balance enhanced quality and computational efficiency, and eliminate block artifacts and brightness discontinuities. Summary of the Invention

[0009] The technical problem to be solved by this invention is that existing low-light enhancement technologies have defects such as limited video memory, low computational efficiency, obvious block effect, and discontinuous brightness when processing ultra-high resolution images, making it difficult to achieve high-quality, low-latency real-time image enhancement.

[0010] To address the aforementioned technical problems, this invention provides a method for enhancing ultra-high resolution low-light images based on block overlap inference, comprising the following steps: Step S1: Perform image block processing on the ultra-high resolution original image to obtain multiple overlapping sub-blocks; Step S2: Calculate the global average brightness based on the ultra-high resolution original image; Step S3: Calculate the local target brightness of each sub-block obtained in step S1 based on the brightness consistency control strategy; Step S4: Use a deep learning model based on an encoder-decoder structure to perform low-light enhancement processing on each sub-block; Step S5: A progressive fusion strategy based on a cosine window function is used to progressively fuse the overlapping areas of each sub-block; Step S6: Output the final ultra-high resolution enhanced image.

[0011] Optionally, step S3 includes: Step S3.1: Calculate the local average brightness of each sub-block obtained in step S1. ; Step S3.2, based on local average brightness Compared with global average brightness Calculate the luminance ratio: ,in ; Step S3.3: Apply logarithmic scaling to the brightness ratio to obtain the scaled brightness ratio. : ; Step S3.4, based on local average brightness Compared with global average brightness The relationship is used to calculate the adaptive factor corresponding to each sub-block. ; Step S3.5: Calculate the target brightness of each sub-block, i.e., the local target brightness. This is so that the data can be input into a deep learning model to perform low-light enhancement processing on each sub-block.

[0012] Optionally, step S3.4 includes: For each sub-block: like At that time, ; like At that time, ; like At that time, ; in These are adaptive adjustment parameters used to adjust the overall target brightness. The preset first threshold is used to control the threshold for judging dark areas; and These are adaptive adjustment parameters used for target brightness adjustment in medium brightness and bright brightness areas, respectively.

[0013] Optionally, in step S4, the encoder part of the deep learning model based on the encoder-decoder structure includes multiple downsampling layers, each layer employing a two-layer residual block ResBlock structure; the decoder part includes multiple upsampling layers, each layer also employing a two-layer residual block ResBlock structure; multi-scale feature information is transmitted between the encoder and decoder through skip connections; a feature linear modulation module FiLM is introduced at the bottom layer to receive the target brightness parameters of each sub-block output in step S3. The feature map is mapped to a feature vector through a fully connected layer, decomposed into scale parameters and offset parameters, and then an affine transformation is performed on the low-light feature map. The final output is the image of each sub-block after low-light enhancement processing.

[0014] Optionally, step S5 includes: For any pixel within the overlapping area of ​​adjacent sub-blocks, the final brightness output value is calculated according to the following formula: ; in and These are the output values ​​of the deep learning models for the left and right sub-blocks, respectively; and Let be the weighting coefficients of the left and right sub-blocks at this position, respectively, satisfying: ; ; ; in This represents the normalized distance of a pixel from the boundary of the right sub-block. Indicates the width of the overlapping area.

[0015] Optionally, when performing image block processing on the ultra-high resolution original image in step S1, the width of the overlapping region of adjacent sub-blocks is determined according to the receptive field of the subsequent deep learning network.

[0016] Optionally, step S2 includes: Convert ultra-high resolution raw images to grayscale images; Calculate the average grayscale value of all pixels in the ultra-high resolution original image.

[0017] This invention also provides an ultra-high resolution low-light image enhancement system based on block overlap inference. The system is used to implement the above method and includes an image segmentation module, a global brightness calculation module, a low-light enhancement processing module, a brightness consistency control module, and an image fusion module. The brightness consistency control module adaptively calculates the target brightness based on the local brightness characteristics of each sub-block. The low-light enhancement processing module performs low-light enhancement processing on each sub-block using a deep learning model based on an encoder-decoder structure, based on the target brightness of each sub-block calculated by the brightness consistency control module. The image fusion module uses a progressive fusion strategy based on a cosine window function to progressively fuse the overlapping areas of each sub-block to obtain the final ultra-high resolution enhanced image.

[0018] Optionally, in the low-light enhancement processing module, the encoder part of the deep learning model based on the encoder-decoder structure includes multiple downsampling layers, each layer employing a two-layer residual block ResBlock structure; the decoder part includes multiple upsampling layers, each layer also employing a two-layer residual block ResBlock structure; multi-scale feature information is transmitted between the encoder and decoder through skip connections; a feature linear modulation module FiLM is introduced at the bottom layer to receive the target brightness parameters of each sub-block. The feature map is mapped to a feature vector through a fully connected layer, decomposed into scale parameters and offset parameters, and then an affine transformation is performed on the low-light feature map. The final output is the image of each sub-block after low-light enhancement processing.

[0019] The present invention also provides applications of the above method in the fields of wide-area surveillance, astronomical observation, medical imaging, and remote sensing mapping.

[0020] The beneficial effects of this invention are: 1. Effectively solves the memory limitation problem in low-light enhancement of ultra-high resolution images. By using a block processing strategy, the ultra-high resolution image is divided into multiple manageable sub-blocks, and each sub-block is inferred independently, which significantly reduces the memory requirements for a single inference, enabling real-time enhancement processing of ultra-high resolution images to be achieved on ordinary GPUs.

[0021] 2. Block artifacts and boundary discontinuities are eliminated. By introducing an overlapping region reasoning strategy and a progressive fusion mechanism, a smooth transition between adjacent sub-blocks at the boundary is ensured, effectively avoiding block artifacts such as brightness discontinuities and color banding common in traditional block partitioning methods.

[0022] 3. Global brightness consistency was achieved. Through a brightness consistency maintenance mechanism, the target brightness is adaptively calculated based on the local brightness characteristics of each sub-block, ensuring the coordination and naturalness of the overall brightness distribution of the enhanced image and avoiding local overexposure or underexposure. 4. Provides precise control over the statistical characteristics of the output image. By integrating the user-defined target brightness parameters into the underlying feature modulation of the network through the feature modulation module, the controllability and predictability of the enhancement results are achieved, meeting the personalized needs of different application scenarios.

[0023] 5. It balances enhancement quality with computational efficiency. While ensuring high-quality enhancement, it fully utilizes the parallel computing capabilities of modern GPUs by processing multiple sub-blocks in parallel, achieving low-latency real-time processing. This makes it suitable for applications with high real-time requirements, such as wide-area monitoring and astronomical observation.

[0024] 6. It exhibits good generalization ability and adaptability. Through a zero-reference-based self-supervised fine-tuning strategy, the model can be effectively adapted to specific sensors without the need for paired real low-light-normal-light image data, reducing data acquisition and annotation costs and improving the practicality and generalizability of the method. Attached Figure Description

[0025] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying 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] Figure 1A flowchart of a low-light enhancement method for ultra-high resolution images based on block overlap inference provided in an embodiment of the present invention; Figure 2 A schematic diagram of the network structure of a deep learning model provided in an embodiment of the present invention; Figure 3 This is a schematic diagram illustrating the working principle of the feature modulation module in the deep learning model provided in this embodiment of the invention; Figure 4 This is a schematic diagram of progressive fusion of overlapping regions in a deep learning model provided in an embodiment of the present invention; Figure 5 A flowchart illustrating the brightness consistency preservation mechanism in a deep learning model provided in this embodiment of the invention; Figure 6 A schematic diagram illustrating the training process of a deep learning model provided in an embodiment of the present invention; Figure 7 This is a schematic diagram of the deep learning model fine-tuning process provided in an embodiment of the present invention; Figure 8 The image shows a comparison of low-light enhancement effects provided in an embodiment of the present invention. Detailed Implementation

[0027] To make the objectives, technical solutions, and advantages of the present invention clearer, the embodiments of the present invention will be described in further detail below with reference to the accompanying drawings.

[0028] Example 1 This embodiment provides a method for enhancing ultra-high resolution low-light images based on block overlap inference. See [link to relevant documentation]. Figure 1 The method includes: Step 101: Perform image block processing on the ultra-high resolution original image; The ultra-high resolution original image is segmented into multiple overlapping sub-blocks. The size of each sub-block can be set according to actual needs, and the size of the overlapping region is determined by the receptive field of the subsequent deep learning network (the size of the overlapping region must meet the requirement of being at least larger than the receptive field), usually set to 32 pixels. For example, for an image with a resolution of 3840×2160, if the sub-block size is set to 640×640 pixels and the overlapping region is 32 pixels, it can be segmented into 6×4 sub-blocks.

[0029] Step 102: Calculate the global average brightness based on the ultra-high resolution original image; Calculate the global average brightness of the ultra-high resolution original image .

[0030] Specifically, after converting the image to a grayscale image, the average grayscale value of all pixels is calculated. , A grayscale image representing the original ultra-high resolution image.

[0031] Step 103: Brightness consistency control; When performing low-light enhancement processing on each sub-block, it is necessary to first calculate the local average brightness of the sub-block based on the grayscale image. Combined with global average brightness Calculate local target brightness To ensure that the processing results of each sub-block maintain continuity in the overall brightness level, i , j () indicates the sub-block number. The specific calculation process is as follows: Figure 5 As shown: (1) Calculate the luminance ratio based on the local average luminance and the global average luminance: ,in ; (2) Apply logarithmic scaling to process the brightness ratio to obtain the scaled brightness ratio. : ; (3) Based on local average brightness Compared with global average brightness Based on the relationship, calculate the adaptive factor corresponding to each sub-block: like At that time, ; like At that time, ; like At that time, ; in These are adaptive adjustment parameters used to adjust the overall target brightness. The preset first threshold is used to control the threshold for judging dark areas; and These are adaptive adjustment parameters used for target brightness adjustment in medium brightness and bright brightness areas, respectively. (4) Calculate the target brightness of each sub-block, i.e., the local target brightness. : .

[0032] To address the issue of varying local lighting conditions across different image blocks, applying the same enhancement parameters independently to each block can lead to uneven brightness distribution in the overall image, resulting in localized overexposure or underexposure. Existing technologies lack effective mechanisms for maintaining brightness consistency, making it difficult to achieve globally coordinated enhancement. This application's solution specifically designs the aforementioned low-light enhancement processing scheme. First, the brightness ratio of each sub-block to the global image is calculated. This ratio is then scaled, and corresponding adaptive parameters and adaptive factors are set for dark, medium-brightness, and bright areas. Based on the scaled brightness ratio and the adaptive adjustment parameters, the final target brightness for each sub-block is determined, ensuring continuity in the overall brightness level across all sub-blocks and guaranteeing brightness consistency in the stitched image.

[0033] Step 104: Perform low-light enhancement processing on each sub-block; Based on the local target brightness of each sub-block output in step 103, independent low-light enhancement processing is performed on each sub-block using a deep learning model based on an encoder-decoder structure. For example... Figure 2 As shown, the encoder of this model contains multiple downsampling layers, each employing a two-layer residual block (ResBlock) structure; the decoder contains multiple upsampling layers, each also employing a two-layer residual block (ResBlock) structure; multi-scale feature information is transmitted between the encoder and decoder via skip connections; a Feature-wise Linear Modulation (FiLM) module is introduced at the bottom layer to receive the target brightness parameters. (e.g., 0.5), it is mapped to a feature vector through a fully connected layer, decomposed into scale parameters and offset parameters, and an affine transformation is performed on the low-light feature map. The final output is the image of each sub-block after low-light enhancement processing.

[0034] Step 105: Merge overlapping regions; After obtaining the sub-block images output by the deep learning model based on the encoder-decoder structure, the final brightness of each pixel in the non-overlapping region of each sub-block is the brightness of each sub-block image output by the deep learning model. For each pixel in the overlapping region of adjacent sub-blocks, progressive fusion is performed to eliminate the inter-block boundary effect.

[0035] For any pixel within the overlapping region, the formula for calculating its final brightness output value is: ; in and These are the output values ​​of the deep learning models for the left and right sub-blocks, respectively; and Let be the weighting coefficients of the left and right sub-blocks at this position, respectively, satisfying: ; ; ; in This represents the normalized distance of a pixel from the boundary of the right sub-block. Indicates the width of the overlapping area.

[0036] Unlike current target recognition and tracking applications that simply discard one side of the pixel value or take the average of the two values ​​in overlapping areas, this application proposes a progressive fusion strategy based on a cosine window function, which solves the block effect caused by block processing, such as obvious brightness discontinuity and color banding at the block boundaries of the stitched image.

[0037] Step 106: Output the enhanced image; The enhancement results of all sub-blocks are stitched together and fused to obtain the final ultra-high resolution enhanced image.

[0038] like Figure 6 As shown, the present invention provides a training method for deep learning models based on encoder-decoder structures, which specifically includes the following steps: Step 201: Dataset Construction.

[0039] Multiple low-light image datasets were used for mixed training to increase sample diversity and scene coverage. These datasets are publicly available: the LOL (Low-Light Image Enhancement) dataset, the VE-LOL (Virtual Enhanced LOL) dataset, and the LSRW dataset. The LOL dataset is a widely used foundational dataset in the field of low-light image enhancement, containing 485 pairs of low-light and normal-exposure images, covering a variety of indoor and outdoor scenes.

[0040] The VE-LOL dataset is an extension of the LOL dataset. It generates more diverse low-light scenes through virtual augmentation technology, making up for the limited number of samples in the LOL dataset.

[0041] The LSRW dataset is a larger-scale low-light image enhancement benchmark dataset containing over 1000 pairs of low-light and normal-light images, covering a wider range of scene types and lighting conditions.

[0042] Step 202: Data Preprocessing. All image pairs are cropped into fixed-size patches to increase the number of training samples. To prevent the model from over-focusing on denoising and causing blurred image details, Gaussian noise can be randomly added to the label images. All data is divided into training, validation, and test sets in a 7:2:1 ratio.

[0043] Step 203: Model Training. A multi-component loss function is used for model optimization, including pixel-level reconstruction loss, perceptual loss, color fidelity loss, and edge structure loss. The Adam optimizer is used for training.

[0044] The multi-component loss function is:

[0045] in, This represents the pixel-level reconstruction loss, used to ensure that the output image is as close as possible to the target image at the pixel level:

[0046] in It is a small positive number (usually set to) ), used to avoid gradient discontinuities; It is the first image data output by the low-light enhancement module. pixel, It is the first of the target output image (Ground Truth) pixel, Indicates the total number of pixels.

[0047] Represents perceptual loss, used to preserve the semantic information and visual perceptual quality of an image:

[0048] in This represents the feature mapping of the j-th layer of the pre-trained network. , and These represent the number of channels, height, and width of the layer, respectively. The pre-trained network uses the VGG network. This indicates that the enhancement module outputs image data. This represents the target output image (Ground Truth).

[0049] This indicates color fidelity loss, used to maintain the consistency of image colors:

[0050] in This represents the mean value of image I in channel p. This represents the mean value of image I in channel q.

[0051] This represents edge structure loss, used to preserve the edge and texture details of an image:

[0052] in To represent the image gradient operator (typically using the Sobel or Scharr operator), this application uses the Sobel operator. The first image output by the enhancement module represents the... pixel, Represents the target output image (Ground Truth) of the th Pixel.

[0053] In actual training, the loss weights are set as follows: , , , This configuration optimizes training efficiency while maintaining image quality.

[0054] Training termination condition: The training rounds are set to 1000 rounds, meaning the training ends after 1000 rounds.

[0055] like Figure 7 As shown, in order to achieve better low-light image enhancement, the method of the present invention also includes a method for fine-tuning the model based on the correspondence of the sensor-acquired real low-light images, specifically including the following steps: Step 301: Pseudo-label generation. For the set of real low-light images acquired by the target sensor... The pre-trained zero-reference low-light image enhancement model Zero-DCE can be used to generate corresponding pseudo-labels. Zero-Reference Deep Curve Estimation (Zero-DCE) is a deep learning model for low-light image enhancement (Guo C, Li C, Guo J, et al. Zero-reference deep curve estimation for low-light image enhancement[C] / / Proceedings of the IEEE / CVF Conference on Computer Vision and Pattern Recognition. 2020: 1780-1789). It employs a lightweight convolutional neural network to predict a set of dynamic curve parameters that define a non-linear brightness adjustment curve for adaptively adjusting the brightness and contrast of the image. This model requires no paired training data and achieves low-light enhancement solely through a non-reference loss function. However, it cannot add a modulation module for output mean control.

[0056] Step 302: Self-supervised fine-tuning. The network model proposed in this invention is then applied... The data is then fine-tuned. To reduce computational costs, the input image and pseudo-labels are first downsampled, and then the optimization objective is the mean squared error loss. A small learning rate (e.g., between 1e-6 and 5e-6) is used during the fine-tuning process.

[0057] To verify the effectiveness of this invention, this embodiment also conducted a comprehensive quantitative and qualitative evaluation using a real dataset. The experimental settings included: an input image resolution of 3840×2160, a block size of 640×640 pixels, an overlapping region of 32 pixels, and global target brightness parameters. Set to 0.5, adaptively adjust parameter A value of 0.2 is used to adjust the brightness in the medium-brightness area. A value of 0.5 is used to adjust the highlight area. The value is 0.3.

[0058] like Figure 8 As shown, from left to right, the results are: (a) the original low-light image; (b) the result using only the block inference strategy (without using brightness consistency control and overlap fusion); and (c) the enhanced result using the complete scheme of this invention (including brightness consistency control and overlap fusion).

[0059] from Figure 8It is evident that image (b) using only block-based reasoning exhibits severe discontinuities at block boundaries, manifested as significant brightness discontinuities and color banding, resulting in poor image quality. In contrast, the result (c) employing the complete solution of this invention effectively eliminates brightness differences and boundary effects between blocks, resulting in a more natural and continuous overall image with more complete detail preservation.

[0060] Example 2 This embodiment provides an ultra-high resolution image low-light enhancement system based on block processing, including the following modules: Image segmentation module: Used to segment ultra-high resolution raw images into multiple sub-blocks with overlapping regions, supporting custom sub-block size and overlapping region size.

[0061] Global brightness calculation module: used to calculate the global average brightness of the ultra-high resolution original image, providing a reference for subsequent brightness consistency control.

[0062] Low-light enhancement processing module: Used to perform independent low-light enhancement processing on each sub-block, integrating a deep learning model based on an encoder-decoder structure, and supporting the parameter modulation function of the feature modulation module.

[0063] Brightness consistency control module: used to calculate the local target brightness when performing low-light enhancement processing on each sub-block, ensuring that the processing results of each sub-block maintain continuity in the overall brightness level.

[0064] Image fusion module: used to progressively fuse overlapping areas of adjacent sub-blocks to eliminate inter-block boundary effects.

[0065] The ultra-high resolution image input image segmentation module segments the image into low-light image sub-blocks. Simultaneously, the image data is sent to the global brightness calculation module to calculate the global brightness average. The low-light image sub-blocks output by the image segmentation module are sent to the brightness consistency control module to obtain the local target brightness. The image brightness sub-blocks and low-light enhancement sub-blocks are sent to the low-light enhancement processing module to obtain the enhanced image sub-blocks. The enhanced sub-blocks are then sent to the image fusion module to obtain the final fused image.

[0066] This system can be deployed on servers or embedded devices equipped with GPUs and supports both batch processing and real-time streaming processing modes. In real-time streaming processing mode, the system can be integrated with an ISP pipeline to achieve real-time low-light enhancement of raw image data output from sensors.

[0067] Some steps in the embodiments of the present invention can be implemented using software, and the corresponding software program can be stored in a readable storage medium, such as an optical disc or a hard disk.

[0068] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. 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 method for enhancing ultra-high resolution low-light images based on block overlap inference, characterized in that, The method includes: Step S1: Perform image block processing on the ultra-high resolution original image to obtain multiple overlapping sub-blocks; Step S2: Calculate the global average brightness based on the ultra-high resolution original image; Step S3: Calculate the local target brightness of each sub-block obtained in step S1 based on the brightness consistency control strategy; Step S4: Use a deep learning model based on an encoder-decoder structure to perform low-light enhancement processing on each sub-block; Step S5: A progressive fusion strategy based on a cosine window function is used to progressively fuse the overlapping areas of each sub-block; Step S6: Output the final ultra-high resolution enhanced image.

2. The method according to claim 1, characterized in that, Step S3 includes: Step S3.1: Calculate the local average brightness of each sub-block obtained in step S1. ; Step S3.2, based on local average brightness Compared with global average brightness Calculate the luminance ratio: ,in ; Step S3.3: Apply logarithmic scaling to the brightness ratio to obtain the scaled brightness ratio. : ; Step S3.4, based on local average brightness Compared with global average brightness The relationship is used to calculate the adaptive factor corresponding to each sub-block. ; Step S3.5: Calculate the target brightness of each sub-block, i.e., the local target brightness. This is so that the data can be input into a deep learning model to perform low-light enhancement processing on each sub-block.

3. The method according to claim 2, characterized in that, Step S3.4 includes: For each sub-block: like At that time, ; like At that time, ; like At that time, ; in These are adaptive adjustment parameters used to adjust the overall target brightness. The preset first threshold is used to control the threshold for judging dark areas; and These are adaptive adjustment parameters used for target brightness adjustment in medium brightness and bright brightness areas, respectively.

4. The method according to claim 3, characterized in that, In step S4, the encoder part of the deep learning model based on the encoder-decoder structure includes multiple downsampling layers, each layer employing a two-layer ResBlock structure; the decoder part includes multiple upsampling layers, each layer also employing a two-layer ResBlock structure; the encoder and decoder transmit multi-scale feature information through skip connections; a feature linear modulation module (FiLM) is introduced at the bottom layer to receive the target brightness parameters of each sub-block output in step S3. The feature map is mapped to a feature vector through a fully connected layer, decomposed into scale parameters and offset parameters, and then an affine transformation is performed on the low-light feature map. The final output is the image of each sub-block after low-light enhancement processing.

5. The method according to claim 4, characterized in that, Step S5 includes: For any pixel within the overlapping area of ​​adjacent sub-blocks, the final brightness output value is calculated according to the following formula: ； in and These are the output values ​​of the deep learning models for the left and right sub-blocks, respectively; and Let be the weighting coefficients of the left and right sub-blocks at this position, respectively, satisfying: ； ； ； in This represents the normalized distance of a pixel from the boundary of the right sub-block. Indicates the width of the overlapping area.

6. The method according to claim 5, characterized in that, When performing image block processing on the ultra-high resolution original image in step S1, the width of the overlapping region of adjacent sub-blocks is determined according to the receptive field of the subsequent deep learning network.

7. The method according to claim 6, characterized in that, Step S2 includes: Convert ultra-high resolution raw images to grayscale images; Calculate the average grayscale value of all pixels in the ultra-high resolution original image.

8. A super-resolution low-light image enhancement system based on block overlap inference, characterized in that, The system is used to implement the method according to any one of claims 1-7, and the system includes an image segmentation module, a global brightness calculation module, a low light enhancement processing module, a brightness consistency control module, and an image fusion module; The brightness consistency control module is used to adaptively calculate the target brightness based on the local brightness characteristics of each sub-block. The low-light enhancement processing module uses a deep learning model based on an encoder-decoder structure to perform low-light enhancement processing on each sub-block based on the target brightness of each sub-block calculated by the brightness consistency control module. The image fusion module uses a progressive fusion strategy based on a cosine window function to progressively fuse the overlapping areas of each sub-block to obtain the final ultra-high resolution enhanced image.

9. The system according to claim 8, characterized in that, The low-light enhancement processing module uses a deep learning model based on an encoder-decoder structure. The encoder part comprises multiple downsampling layers, each using a two-layer ResBlock structure. The decoder part comprises multiple upsampling layers, each also using a two-layer ResBlock structure. Multi-scale feature information is transmitted between the encoder and decoder via skip connections. A Feature Linear Modulation (FiLM) module is introduced at the bottom layer to receive the target brightness parameters of each sub-block. The feature map is mapped to a feature vector through a fully connected layer, decomposed into scale parameters and offset parameters, and then an affine transformation is performed on the low-light feature map. The final output is the image of each sub-block after low-light enhancement processing.

10. The application of the method according to any one of claims 1-7 in the fields of wide-area surveillance, astronomical observation, medical imaging, and remote sensing mapping.

Images