Low-illumination image enhancement method based on illumination perception modeling

Abstract

The invention discloses a low-illumination image enhancement method based on illumination perception modeling, belongs to the technical field of image processing and computer vision, and aims to solve the problems of insufficient image brightness, fuzzy details and structural distortion in a low-illumination scene. The method comprises the following steps: extracting illumination prior information of an input image, and averagely constructing an illumination guide map through a channel; performing feature coding by combining the original image and the illumination image, generating illumination modulation mapping and executing enhancement operation; performing context modeling based on a multi-head attention mechanism guided by illumination, and capturing a global dependency relationship between image areas; reconstructing an image structure and details by adopting a multi-scale coding and decoding network, and improving local texture expression in a residual error correction mode; and finally, the reconstructed residual signal and the enhanced image are fused, and a high-quality image with a clear structure and natural colors is output. According to the method, the structure consistency and texture definition are kept while the image brightness is improved, and the method has the advantages of low complexity and high adaptability and is suitable for low-illumination perception scenes such as night monitoring, intelligent transportation and mobile terminal vision.

Description

TECHNICAL FIELD

[0001] The present application relates to a low-light image enhancement method based on illumination perception modeling. The method addresses the problems of insufficient image brightness and missing structural information under complex lighting conditions, and achieves structure preservation and detail enhancement of low-light images through illumination prior guidance, feature fusion, and multi-scale reconstruction mechanism. The system has good universality and deployability, and is suitable for image quality improvement tasks in various low-light scenes such as night video analysis, intelligent driving perception, mobile device image optimization, industrial detection, remote sensing processing, and medical image enhancement. BACKGROUND

[0002] In recent years, with the widespread application of artificial intelligence and image processing technology, low-light image enhancement has become a key pre-processing link in many computer vision tasks. In the fields of intelligent monitoring, autonomous driving, mobile terminal photography, industrial detection, remote sensing mapping, and medical imaging, images are often acquired in weak light environments, resulting in low overall brightness, blurred details, increased noise, and color deviation. If such images are not effectively enhanced, it will seriously affect the recognition accuracy, system stability, and subjective visual experience of downstream tasks.

[0003] For example, in a night security system, image enhancement can significantly improve the robustness of person recognition and anomaly detection; in an autonomous driving environment, clear recognition of traffic signs, pedestrians, and road conditions in low-light images relies on stable image enhancement mechanisms; in an industrial quality inspection process, image clarity and texture expression directly affect the reliability of defect detection and classification judgment; and in a medical scenario, low-contrast images that cannot restore structural and texture details will weaken the accuracy of medical analysis and diagnosis support.

[0004] Although existing low-light image enhancement methods have made some progress in brightness enhancement and basic detail restoration, there are still many deficiencies. Traditional image enhancement methods such as histogram equalization and Gamma correction ignore the context information of the image and only adjust the pixel intensity, which can easily introduce problems such as overexposure, artifacts, or structural distortion. Although deep learning-based image enhancement models improve overall image quality, they generally have the defects of complex models, high computational resource consumption, and lack of illumination prior perception ability, making it difficult to efficiently run on resource-constrained terminal devices. In addition, most current methods fail to effectively model the internal illumination distribution of the image, resulting in a lack of relevance and semantic consistency in the enhancement process; structural information and texture details are often mixed and processed, which can easily cause over-enhancement in some areas of the image and insufficient preservation in other areas, affecting the overall perceptual quality.

[0005] Therefore, there is an urgent need for a low-illumination image enhancement method and system based on illumination perception modeling. This scheme can fully utilize the potential illumination distribution characteristics in the image, combine global semantic modeling mechanism and detail reconstruction strategy, and perform adaptive enhancement processing on low-light images. While maintaining the naturalness of image structure and color, it has good deployability, generalization ability and efficient running ability, and meets the image enhancement needs in multiple scenarios. SUMMARY

[0006] In view of the technical problems of uneven brightness distribution, weak image structure expression ability, serious detail loss and low running efficiency of existing models in low-illumination images in actual application, the present application provides a low-illumination image enhancement method and system based on illumination perception modeling. Based on the illumination prior guidance, the image brightness features and deep semantic relationship are fused to realize high-quality reconstruction and enhancement of dark light images.

[0007] Unlike the enhancement methods relying on manual features or simple convolution processing in the past, the present application first obtains an illumination map through channel statistics to guide the feature extraction process to perceive the illumination distribution change. Then, a light modulation mapping is constructed to dynamically adjust the brightness of the original image, and the intermediate illumination features are retained for guiding the subsequent modeling. By introducing a context modeling module based on illumination guidance, the non-local correlation between image regions is established, effectively improving the global consistency and background restoration ability. Then, a multi-scale reconstruction architecture is used to optimize the image at the structure level and texture level, and the residual signal is used to further fine repair the local detail area. Finally, the multi-source information is fused to generate an output image with higher perception quality and natural reality.

[0008] The present application has the advantages of simple structure, high computational efficiency and flexible deployment, and is particularly suitable for image processing tasks with high requirements for brightness, structure and texture in weak light monitoring, night vehicle imaging, edge vision devices and image perception terminals.

[0009] In a first aspect, the present application provides a low-illumination image enhancement method and system based on illumination perception modeling, comprising the following steps:

[0010] S1, receiving a low-illumination image input and calculating the channel average value to form an illumination prior map; jointly encoding the input image and the illumination prior map to generate illumination guided feature representation;

[0011] S2, estimating the light modulation mapping based on the above features, performing light adjustment operation on the original image to generate a light enhanced image, and retaining the intermediate light features to assist subsequent restoration;

[0012] S3, send the light enhanced image and its light features into the long-range dependency modeling network, enhance the semantic association between different regions of the image through the illumination guided multi-head attention mechanism, and obtain the image features with global context awareness;

[0013] S4, the image features are reconstructed by a multi-scale encoding-decoding structure, and the details of the light enhanced image are corrected by a residual method, to obtain a residual signal with strong structural consistency and clear texture;

[0014] S5, the above residual signal and the light enhanced image are fused to generate a final enhanced image output, which has natural color, strong noise suppression ability and structure preserving low-light image performance.

[0015] In a second aspect, the present application provides a low-light image enhancement system based on illumination perception modeling, characterized in that it comprises:

[0016] The illumination prior generation module is configured to receive the original low-light image, obtain an illumination map by channel dimension statistical calculation, reflect the light intensity distribution of each region in the image, and provide physical guidance information for subsequent enhancement modeling;

[0017] The light modulation estimation module is configured to combine the original image and the illumination map for joint encoding to generate a light modulation mapping image, and then perform adaptive brightness adjustment on the original image to output a preliminary enhanced image and intermediate illumination features;

[0018] The context-dependent modeling module is configured to receive the light enhanced image and the intermediate illumination features, use the multi-head attention mechanism based on illumination guidance to construct long-range dependencies, and enhance the structural consistency and semantic expression ability between regions of the image;

[0019] The multi-scale feature reconstruction module is configured to use a multi-scale encoding-decoding structure to reconstruct the structure and texture of the image at different levels based on the context enhancement, and correct the local details through a residual path to generate more realistic high-frequency information;

[0020] The image fusion module is configured to fuse the reconstructed residual signal with the light enhanced image to generate a final output image. The fusion process supports multiple strategies such as weighted summation, feature splicing, etc. to achieve higher image naturalness, clarity and color restoration.

[0021] In a third aspect, the present application further provides an image processing device, characterized in that it comprises one or more processors and a memory connected in communication with the processors, and the memory stores computer program instructions for performing an image enhancement task. When the program is executed by the processor, the processor is configured to perform the low-light image enhancement method based on illumination perception modeling according to the first aspect of the present application, specifically including but not limited to the following processing flow: illumination prior generation, illumination modulation mapping estimation, context-dependent modeling, multi-scale feature reconstruction, and fusion output. The device can be deployed in a single computer, server or embedded system, has the ability to run independently, and is suitable for various intelligent devices that need to perform image enhancement in low-light environments.

[0022] In a fourth aspect, the present application further provides a computer-readable storage medium having stored thereon a computer program for implementing low-light image enhancement. When the program is loaded and executed by one or more processors, it can complete all steps of the method according to the first aspect of the present application, including illumination modeling, feature guidance, context semantic enhancement, residual signal restoration and image output, supports the loading and execution of the complete model, and meets the deployment requirements in multiple scenarios.

[0023] The present application has the following technical advantages:

[0024] (1) Illumination-guided image understanding capability: by introducing an illumination prior graph to construct a physically meaningful illumination modeling pathway, the image enhancement has the ability to perceive real illumination changes, reducing over-enhancement or color cast problems from the source;

[0025] (2) Context-dependent modeling mechanism: using a multi-head attention mechanism for cross-region semantic modeling, effectively improving the mutual consistency between local regions and overall structures in dark-light images, and enhancing the collaborative restoration capability of background and edges;

[0026] (3) Multi-scale detail reconstruction and residual correction: using an encoding-decoding structure to hierarchically reconstruct image main stems and texture information, and combining a residual branch to achieve fine repair of detail areas, improving image sharpness and texture realism;

[0027] (4) Decoupled modeling and high adaptation deployment capability: through independent module division and intermediate feature representation connection mechanism, the combinability and flexibility of the system structure are enhanced, supporting rapid deployment and efficient execution on multiple types of computing platforms, adapting to image enhancement tasks in resource-constrained scenarios such as edge terminals, embedded chips and mobile devices.

[0028] (5) The application is widely applicable: the application is not only applicable to night monitoring image improvement, but also applicable to automatic driving perception optimization, industrial image preprocessing, mobile image sharpening, remote sensing image enhancement and other low-light environment image processing tasks, and has good expansibility and practicality. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 It is a low-illumination image enhancement method based on illumination perception modeling. DETAILED DESCRIPTION

[0030] In order to make the purpose, technical scheme and advantages of the application clearer, the low-illumination image enhancement method and system based on illumination perception modeling are described in detail below with specific examples.

[0031] Example 1

[0032] The embodiment proposes a low-illumination image enhancement method based on illumination perception modeling, the core of which is to introduce an illumination perception mechanism, to realize the coordinated optimization of image brightness, structure and texture through illumination feature guidance, context-dependent modeling and hierarchical detail reconstruction, etc. The main process of the method includes the following five stages:

[0033] S1: Illumination prior map generation and feature guided coding; S2: Illumination modulation mapping estimation and brightness enhancement; S3: Context-dependent modeling and global consistency optimization; S4: Multi-scale structure reconstruction and residual detail correction; S5: Final image fusion and enhancement output.

[0034] Further, the specific method of S1 is:

[0035] S11: Obtain the original low-illumination image input, which can be collected from night shooting devices, low-light video monitoring systems, industrial quality inspection cameras or other low-light environment imaging devices.

[0036] S12: Perform statistical processing on the channel information of the input image, calculate the overall brightness distribution trend, and generate a single-channel illumination prior map to describe the illumination intensity of different regions in the image.

[0037] S13: Joint coding of the illumination prior map and the original image as the input of subsequent feature extraction, guiding the network to pay attention to the brightness insufficient area, and realizing the collaborative modeling of illumination information and image structure.

[0038] Further, the specific method of S2 is:

[0039] S21: Use the illumination guided feature to predict the illumination modulation mapping, which is used to reflect the brightness adjustment requirement of each image region and has spatial adaptability.

[0040] S22: Perform brightness enhancement processing on the original image according to the modulation mapping, realize the brightness improvement of the dark part region, and maintain the naturalness of the bright part region, avoiding over-enhancement or uneven brightness.

[0041] Further, the specific method of S3 is:

[0042] S31: Send the brightness-enhanced image and its illumination features into the context modeling module, model the semantic association between different regions in the image through a multi-scale attention mechanism, and improve the consistent expression of the global structure.

[0043] S32: The output context-enhanced features integrate illumination information and semantic structure, which helps to enhance the coordination between target regions and backgrounds in the image, and provides structured prior for subsequent reconstruction modules.

[0044] Further, the specific method of S4 is:

[0045] S41: Construct a multi-scale reconstruction module to decode the context features layer by layer, and gradually restore the main structural information in the image, such as edges, contours, and region boundaries.

[0046] S42: At the same time, construct a residual repair branch to learn the differences between the structure reconstruction results and the enhanced image, capture high-frequency detail features, including texture, reflection, and fine lines.

[0047] S43: Collaboratively reconstruct the main structure and local details of the image to improve the overall image's level of detail and clarity.

[0048] Further, the specific method of S5 is:

[0049] S51: Fuse the residual repair results with the brightness-enhanced image to integrate structure, illumination, and detail information, and generate the final enhanced image.

[0050] S52: The output image has good brightness distribution, structural integrity, and texture performance, and is suitable for night vision enhancement, low-light image quality inspection, mobile image optimization, and other low-illumination image processing scenarios.

[0051] Embodiment 2:

[0052] To further illustrate the low-illumination image enhancement method and system based on illumination perception modeling proposed by the present application, the following combines mathematical expressions and Figure 1 The steps of embodiment 1 are described in detail.

[0053] Step 201:

[0054] Obtain the original low-illumination image I∈R H×W×3, where H and W represent the height and width of the image, respectively, and the three channels are red, green, and blue. To depict the overall brightness distribution of the image, the channel average method is used to generate the illumination prior map L e R H×W , whose calculation expression is:

[0055]

[0056] where I c (x, y) represents the value of the pixel at position (x, y) in the original image in channel c e {R, G, B}. This illumination map reflects the relative brightness distribution of the entire image in the spatial position, which can be used to guide the network to focus on the insufficiently illuminated area. Then, the original image I and the illumination map L are spliced in the channel dimension to form an enhanced input:

[0057] I in = Concat(I, L)

[0058] The resulting input tensor I in e R H×W×3 contains three-channel original images and single-channel illumination information. This spliced feature is sent to the guided feature encoding module e (·) to extract the illumination guided feature F1 e R h×w×d , which is represented as follows:

[0059] F1 = e (I in , q e )

[0060] where e (·) is a shallow encoding network containing several convolution layers, activation functions, and down-sampling operations, and q e is its set of learnable parameters. Through this operation, the model not only extracts the low-level texture features of the image, but also captures the spatial structure information related to brightness. The illumination guided feature F1 output by this step provides perceptual support for subsequent illumination modulation mapping estimation (step 202), making the subsequent image enhancement more locally adaptive and physically reasonable.

[0061] Step 202:

[0062] The core goal of this step is to estimate the pixel-level illumination modulation mapping based on the illumination guided feature, and use this mapping for brightness enhancement of the original image to achieve local adaptation, structure preservation, and illumination enhancement. First, the illumination guided feature F1 obtained in the previous step is input into the illumination modulation estimation module q (·), which is composed of multiple convolution layers and normalization activation functions, to generate a two-dimensional illumination modulation map M, which represents the illumination compensation coefficient that should be applied at each spatial position, with the calculation expression as follows:

[0063]

[0064] where qm The learnable parameters of the modulation network are denoted as. The value range of the modulation map M is constrained by a Sigmoid function in the interval [0, 1], ensuring the physical reasonableness of the illumination adjustment amplitude in the enhancement process. Subsequently, the original image I is processed at the pixel level using the modulation map to generate an image I e ∈R H×W×3 , and the enhancement operation is expressed as:

[0065] I e (x,y,c)=I(x,y,c)·(1+M(x,y))

[0066] where (x, y) represents the spatial position in the image, and c is the color channel index. This enhancement method is essentially an adaptive gain mechanism, allowing the model to perform differentiated brightness adjustment in different regions based on the illumination estimation results, thereby avoiding over-enhancement or local underexposure. In addition, to preserve the intermediate feature information in brightness adjustment, the modulation map M will continue to be used as guiding information in the subsequent context modeling step (step 203), ensuring the entire enhancement process has consistent and physically interpretable consistency.

[0067] Step 203: This step aims to enhance the structural association between different regions in the image through a multi-head attention mechanism, improving the semantic consistency and regional collaborative restoration ability of low-light images in the global range. First, the brightness-enhanced image I e ∈R H ×W×3 is concatenated with the modulation map M in the channel dimension to obtain a fused image input I c ∈R H×W×4 :

[0068] I c =Concat(I e ,M)

[0069] The fused image retains the texture information of the enhanced image and the illumination guidance information of the modulation map, which helps the subsequent network to perceive the illumination differences and structural relationships between regions. The fused image retains the texture information of the enhanced image and the illumination guidance information of the modulation map, which helps the subsequent network to perceive the illumination differences and structural relationships between regions. Then, I c is input into the context modeling module, which is constructed based on a multi-head attention mechanism and specifically includes two branches: channel attention (CA) and spatial attention (SA):

[0070] F2=SA(CA(I C ))+I C

[0071] where CA(·) is used to model the importance relationship between channels, and to enhance the channel dimension feature expression ability; SA(·) models the long-range dependence between spatial positions, so that the model can identify the structure alignment relationship between distant regions; the residual connection is to retain the original fusion features to prevent overfitting of the attention module. The context feature map F2∈R H×W×d While maintaining the expression of details, cross-region structure modeling under illumination guidance is achieved, especially in the edge weakening and texture blurred area, which has stronger semantic reconstruction ability. The feature map serves as the input for structure reconstruction and detail restoration, providing structural level context support for multi-scale reconstruction and residual correction in subsequent steps.

[0072] Step 204: This step aims to hierarchically reconstruct image structure and texture based on context-aware features F2, and supplement local details through a residual branch, thereby obtaining a structure-complete and texture-clear image representation. First, input the context feature map F2 into the multi-scale decoder module D(·), which consists of multiple encoding-decoding layers (including upsampling and convolution) for restoring the main structure content of the image. The obtained structure reconstruction image is denoted as:

[0073] I s = D(F2; θ d )

[0074] where θ d are the trainable parameters of the structure reconstruction network. The structure image I s ∈R H×W×3 captures the general outline, region boundary and basic brightness information of the image, but may be missing or blurred in detail areas. To this end, an independent residual detail correction module R(·) is designed to model the difference between the structure image and the brightness enhancement image, and output a detail compensation image det:

[0075] det = R(I e ,I s ; θ r )

[0076] where I e is the brightness enhancement image in step 202, and θ r are the parameters of the residual module. The module can use multiple shallow convolution + residual connection structures to emphasize the reconstruction of high-frequency texture and detail edges.

[0077] Finally, the structure image and the residual image are added pixel by pixel to form the image representation I r after detail correction:

[0078] I r = I s + R

[0079] The residual enhancement mechanism not only improves the texture expressiveness of the image, but also maintains the stability of the overall structure, significantly alleviating the common edge blur and detail loss problems in low-light images.

[0080] Step 205: This step aims to integrate the multi-layer information of the structure reconstruction image and the brightness enhancement image, and output the final image I through a feature-guided fusion mechanism out , to achieve the enhancement goal of brightness balance, structure fidelity and detail clarity.

[0081] First, the residual correction image I r ∈R H×W×3 obtained in step 204 is combined with the brightness enhancement image I e ∈R H×W×3 in step 202 to input the fusion module for integration. In order to fully exert the complementarity of each branch feature, an adaptive weighted fusion strategy is adopted:

[0082] I out (x,y,c)=α(x,y)·I r (x,y,c)+(1―α(x,y))·I e (x,y,c)

[0083] Where α(x,y)∈[0,1] is the fusion weight map, representing the weight coefficient of the residual image at spatial position (x,y), which is generated based on the context-guided feature and the illumination modulation map, reflecting the dependence of different regions on structure information and brightness information.

[0084] The fusion module internally combines the global guide features generated by the aforementioned context modeling, and refers to the illumination modulation map, using a lightweight attention mechanism to dynamically adjust the fusion proportion of different regions and channels, thereby achieving unified enhancement of structure and detail in the entire image range. This method is suitable for image enhancement tasks in various low-light environments, including night monitoring, vehicle vision systems, industrial defect detection, terminal image optimization, and other application scenarios, and has wide practicality and good deployment performance.

[0085] Embodiment 3:

[0086] The embodiment discloses a low-light image enhancement device based on illumination perception modeling, which is designed for image quality improvement in weak light environment and is suitable for scenes such as night monitoring, industrial detection, security image preprocessing, etc. The device includes at least one processor and a memory in communication with the processor, the memory having pre-stored computer program instructions executable by the processor, for driving the following functional modules:

[0087] The illuminance prior generation module is configured to receive the original low-illumination image, obtain an illuminance map by channel dimension statistics calculation, reflect the light intensity distribution of each region in the image, and provide physical guidance information for subsequent enhancement modeling.

[0088] The illumination modulation estimation module is configured to jointly encode the original image and the illuminance map to generate an illumination modulation mapping image, and perform adaptive brightness adjustment on the original image to output a preliminary enhanced image and an intermediate illuminance feature.

[0089] The context-dependent modeling module is configured to receive the illumination enhanced image and the intermediate illuminance feature, use a multi-head attention mechanism based on illuminance guidance to construct long-distance dependency, and enhance the structural consistency and semantic expression ability between regions of the image.

[0090] The multi-scale feature reconstruction module is configured to perform hierarchical reconstruction of structure and texture of the image based on context enhancement using a multi-scale encoding-decoding structure, and correct local details through a residual path to generate more realistic high-frequency information.

[0091] The image fusion module is configured to fuse the reconstructed residual signal and the illumination enhanced image to generate a final output image. The fusion process supports multiple strategies such as weighted summation, feature splicing, etc. to achieve higher image naturalness, clarity and color restoration.

[0092] Embodiment 4:

[0093] The fourth embodiment provided in the present application is a terminal processing device supporting illuminance-aware image enhancement, which can be flexibly deployed in a single server, an industrial edge computing platform or a local image processing terminal, and is suitable for image enhancement processing tasks in low-illumination environments. The multiple functional modules built-in the device support on-demand combination operation, and can also be packaged as a unified image enhancement framework, with good module decoupling, system lightweight and application expansion.

[0094] The device is suitable for various weak light scenes such as night industrial quality inspection, low-illumination visual analysis, security monitoring image preprocessing, etc. and can realize stable and fast image enhancement operation in different deployment environments. The system as a whole can effectively improve the image brightness, contrast and visual clarity while maintaining the structural consistency and texture integrity of the image, thereby enhancing the environmental perception and image understanding ability under complex lighting conditions.

Claims

1. A low-light image enhancement method based on illumination perception modeling, characterized in that, The method comprises the following steps: S1, receiving a low-illumination image input and calculating the channel average value to form an illumination prior map; the input image and the illumination prior map are jointly encoded to generate a light guide feature representation; S2, based on the above feature, the illumination modulation mapping is estimated, the illumination adjustment operation is performed on the original image to generate an illumination enhanced image, and the intermediate illumination feature is retained to assist subsequent restoration; S3, the illumination enhanced image and its illumination feature are sent to a long-range dependence modeling network, and the semantic association between different regions of the image is enhanced through the illumination guide multi-head attention mechanism to obtain image features with global context awareness; S4, a multi-scale encoding-decoding structure is used to reconstruct the image features in layers, and the details of the illumination enhanced image are corrected through the residual method to obtain a residual signal with strong structural consistency and clear texture; S5, the above residual signal and the illumination enhanced image are fused to generate a final enhanced image output, which has a natural color, strong noise suppression ability and a low-illumination image performance effect with structural fidelity.

2. The low-light image enhancement method based on illumination perception modeling according to claim 1, characterized in that, In step S1, the channels of the input image are averaged to generate an illumination prior map; the illumination prior map and the input image are spliced in the channel dimension and then sent to an encoding network to extract illumination guide features that fuse illumination information, which are used to strengthen the dark area perception ability. 3.The low-light image enhancement method based on illumination perception modeling of claim 1, wherein, In step S2, the illumination modulation mapping is estimated by a lightweight regression network and is used to perform illumination enhancement on the original image; the enhanced image and the intermediate modulation feature are used as inputs for the subsequent reconstruction process.

4. The low-light image enhancement method based on illumination perception modeling according to claim 1, characterized in that, In step S3, the multi-head attention module with illumination perception is introduced to model the context of the illumination enhanced image, capture the long-range dependence relationship between image regions, and improve the global structural consistency of the image.

5. The low-light image enhancement method based on illumination perception modeling according to claim 1, characterized in that, In step S4, the encoding-decoding structure is used to perform multi-scale reconstruction on the image features to extract texture details at different resolutions; meanwhile, a residual correction module is introduced to compensate the details of the enhanced image to generate a residual signal with structural consistency.

6. The low-light image enhancement method based on illumination perception modeling according to claim 1, characterized in that, In step S5, the final enhanced image is generated by fusing the residual signal and the illumination enhanced image; The fusion process integrates structural information and detail information by using weighted superposition or feature splicing.

7. A low-light image enhancement system based on illumination perception modeling, characterized in that, The system comprises a plurality of functional modules for performing the steps of claims 1 to 6, the modules comprising: an illumination prior generation module, an illumination modulation estimation module, a context dependence modeling module, a multi-scale feature reconstruction module, and an image fusion module; the system is suitable for low-illumination scenes such as night video enhancement, industrial quality inspection, and medical image preprocessing.

8. An image processing apparatus for low-light image enhancement, characterized by, The system comprises one or more processors and a memory, the memory storing instructions that, when executed by the processor, cause the processor to perform the method of any one of claims 1 to 6, comprising: obtaining a low-illumination image; generating an illumination prior map; estimating an illumination modulation map and performing enhancement; modeling context information using a multi-head attention mechanism; performing multi-scale residual reconstruction; and fusing the reconstructed residual signal and the enhanced image to output the final result.

Images