An intelligent adaptive reconstruction method, system and device for cross-scene image quality enhancement

Abstract

This invention relates to the field of digital image processing technology, and discloses an intelligent adaptive reconstruction method, system, and apparatus for cross-scene image quality enhancement. The method generates an illumination intensity map based on paired supervision of a source domain image and a clear daytime target reference image, maps it to brightness-guided attention, and obtains a scaled illumination attention map after scaling. Brightness is then enhanced and fused, and an image enhancement mask is obtained through differential calculation. The enhanced reconstructed image is obtained by residual superposition reconstruction. A reference domain statistical template and a preset statistical threshold interval are constructed, distribution differences are calculated, and a parameterized mapping output domain alignment enhancement model is iteratively updated. During the inference phase, brightness remapping, contrast and detail gain upper limits, and key regions are automatically set, outputting an enhanced image whose brightness and color distribution are closer to the target reference domain. This invention can output enhanced images close to clear daytime images under various cross-scene conditions.

Description

Technical Field

[0001] This invention relates to the field of digital image processing technology, and more specifically, to an intelligent adaptive reconstruction method, system, and apparatus for cross-scene image quality enhancement. Background Technology

[0002] In applications such as autonomous driving, intelligent transportation, and urban security, images captured by vehicle-mounted or fixed camera equipment are often under complex imaging conditions such as rain, snow, fog, and low light. The imaging results are prone to problems such as indistinguishable details in dark areas, local contrast attenuation, color shift, and bright whitening caused by fog and snow. As a result, the same road structure exhibits obvious differences in appearance distribution under different imaging mechanisms.

[0003] In existing technologies, the first type of method mainly focuses on low-light enhancement, typically involving mapping, equalization, or local contrast enhancement on the luminance channel, supplemented by color restoration. For example, in 2018, Feng Qingzhi published a low-light image enhancement algorithm based on the LIP model and CLAHE in the journal *Optoelectronic Technology Applications*. This algorithm proposes logarithmic image processing of the luminance component in the YUV color space combined with contrast-limited adaptive histogram equalization to improve the brightness and contrast of dark areas, achieving enhancement effects such as... Figure 2 As shown, this low-light image enhancement algorithm still faces typical parameter sensitivity and side effect constraints: when the enhancement-dependent adjustment parameter is too strong, it may cause over-enhancement and noise amplification. At the same time, global histogram equalization essentially sacrifices grayscale levels for contrast enhancement, which can easily lead to loss of output details and unnatural over-enhancement. Furthermore, the design goal of the low-light image enhancement algorithm focuses on "low-light degradation," lacking unified distribution alignment constraints for bright whitening caused by fog and snow scattering, rain streaks and interference, and global color statistical drift caused by weather conditions.

[0004] In existing technologies, the second type of method is represented by dehazing and haze removal, often using dark channel priors, Retinex, or a combination thereof for scattering correction and contrast restoration. Dark channel priors rely on the assumption that the dark channel in non-sky regions is close to zero, thus easily failing in sky regions or bright areas where the target surface color is close to atmospheric light, and potentially producing artifacts such as halos. Simultaneously, Retinex-type methods are prone to information loss in non-uniform dense fog scenes. This indicates that the stability and artifact control of dehazing methods under cross-scene conditions remain challenging. Furthermore, this type of method mainly revolves around scattering models or prior assumptions, lacking a paired domain alignment mechanism that "uses a clear reference domain of the same road segment as a supervision benchmark and simultaneously regresses brightness and color statistics."

[0005] In summary, there is an urgent need for a cross-scene image quality enhancement and intelligent adaptive reconstruction scheme that can establish a paired supervision relationship between the source domain and the clear reference domain at the data level, and output the amplification and enhancement focus automatically under different weather and lighting conditions. Summary of the Invention

[0006] To overcome the aforementioned deficiencies of existing technologies, this invention provides an intelligent adaptive reconstruction method, system, and apparatus for cross-scene image quality enhancement. It calculates an illumination map from a source domain image constructed from cross-scene input images, generates a scaled illumination attention map, achieves pixel-level enhancement intensity scheduling, and uses residual masking for controlled modeling to obtain an enhanced reconstructed image. A domain alignment adaptive module, using the enhanced reconstructed image and the target reference image as supervision, automatically adjusts the amplification and enhancement focus under different weather and lighting mechanisms based on the difference index sequence and distribution difference index output, thus solving the problems mentioned in the background technology.

[0007] To achieve the above objectives, the present invention provides the following technical solution:

[0008] A smart adaptive reconstruction method for cross-scene image quality enhancement includes the following steps:

[0009] Step 1: Obtain cross-scene input images and construct pairing data between source domain images and target reference images.

[0010] Step 2: Calculate the illumination map based on the source domain image to obtain the brightness-guided attention weight map, and then perform scaling and reshaping to generate a scaled illumination attention map.

[0011] Step 3: Overlay the image enhancement mask with the source domain image to obtain the enhanced reconstructed image.

[0012] Step 4: Based on the enhanced reconstructed image and the target reference image, output domain alignment enhancement model is generated, and the amplification and enhancement focus are automatically adjusted under different weather and lighting conditions.

[0013] As a further aspect of the present invention, step 1, acquiring cross-scene input images and constructing pairing data between source domain images and target reference images, includes the following specific content: acquiring cross-scene input images as source domain images; determining a corresponding target reference image for each source domain image, such that the source domain images and target reference images constitute pairing data for training.

[0014] As a further aspect of the present invention, step 2, calculating an illumination map based on the source domain image to obtain a brightness-guided attention weight map, and then performing scaling and reshaping to generate a scaled illumination attention map, includes the following specific steps: calculating the illumination intensity of each pixel in the source domain image to form an illumination intensity map corresponding one-to-one with the pixel position. The values ​​of the illumination intensity map are uniformly normalized to a fixed range, and any abnormal bright or dark spots are cropped. Then, for each pixel position, according to the rule of "lower brightness, higher attention," the pixel's illumination intensity is converted into an attention weight, resulting in an attention weight map with the same resolution as the illumination intensity map. The attention weight map is lightly smoothed to make the weight changes of adjacent pixels more continuous, resulting in a brightness-guided attention weight map. The brightness-guided attention weight map is then subjected to pixel-by-pixel "scaling and reshaping" to generate the scaled illumination attention map.

[0015] As a further aspect of the present invention, step 3, superimposing the image enhancement mask and the source domain image to obtain the enhanced reconstructed image, includes the following specific contents: unifying the source domain image and the scaled illumination attention map to the same spatial resolution and corresponding them pixel by pixel, so that the attention weights in the scaled illumination attention map can directly participate in the enhancement intensity scheduling of each pixel; the source domain image is converted to a color space containing luminance and chrominance components, and the enhancement processing is concentrated on the luminance component to ensure that the color information is mainly maintained by the original chrominance component, reducing the risk of color drift from the source; the scaled illumination attention map is used as a spatially variable control map, dividing pixels into different enhancement types such as "low-light high-attention areas" and "high-brightness occlusion areas that need to be restored": in areas with low brightness and high attention weights, the luminance component undergoes adaptive gamma correction to increase the grayscale of dark areas, and then contrast-limited adaptive histogram equalization is superimposed to improve local visibility, while introducing anti-sharpening masking to compensate for high-frequency details and suppress texture blurring caused by brightening; suppressing texture blurring caused by brightening; utilizing Guided filtering or bilateral filtering decomposes the luminance component into low-frequency illumination and high-frequency details. Local dynamic range compression is applied to the low-frequency illumination to reduce whitening, while mid-frequency contrast is stretched to restore road boundaries and distant target outlines. This enhancement is also constrained by attention weights to avoid overexposure. The luminance enhancement results from different regions are fused at the pixel level according to attention weights to form an enhanced luminance map that balances dark area brightening with highlight occlusion detail recovery. Edge smoothing is applied to the fused result to suppress noise amplification and block artifacts. The enhanced luminance map is recombined with the original chrominance components and mapped back to the source domain image representation to obtain the enhanced result image. A pixel-by-pixel, channel-by-channel differencing method is used to calculate the difference between the enhanced result image and the source domain image at each pixel position, resulting in a difference map with the same size as the source domain image. This difference map is then clipped and lightly smoothed to output an image enhancement mask. Finally, a residual pixel-by-pixel overlay reconstruction method is used to overlay the image enhancement mask with the source domain image to obtain the enhanced reconstructed image.

[0016] As a further aspect of the present invention, step 4, using the enhanced reconstructed image and the target reference image as a supervised basis, outputs a domain alignment enhancement model that automatically adjusts the enhancement magnitude and enhancement focus under different weather and lighting conditions. This includes the following specific content: using a domain alignment adaptive algorithm with the enhanced reconstructed image and the target reference image as a supervised basis, outputs a domain alignment enhancement model that can map the source domain image to a distribution closer to the target reference domain. The domain alignment enhancement model is used to automatically adjust the enhancement magnitude and enhancement focus under different weather and lighting conditions, and outputs a result image that is closer to the clear reference domain.

[0017] A smart adaptive reconstruction system for cross-scene image quality enhancement includes:

[0018] The paired data construction module acquires cross-scene input images as source domain images and determines the corresponding target reference image for each source domain image to form paired data for training.

[0019] The illumination intensity map generation module calculates the brightness component pixel by pixel of the source domain image and performs scaling processing to generate an illumination intensity map with the same size as the source domain image.

[0020] The brightness-guided attention generation module maps pixel illumination intensity to attention weights based on the illumination intensity map and smooths the attention weight map to obtain the brightness-guided attention weight map.

[0021] The scaling and reshaping module applies a preset quadratic scaling transformation to the brightness-guided attention weight map to generate a scaled illumination attention map.

[0022] The enhanced mask generation module takes the source domain image and the scaled illumination attention map as input to perform brightness component enhancement, region weighted fusion and edge smoothing, and obtains the image enhancement mask through pixel-by-pixel channel difference.

[0023] The residual overlay reconstruction module uses a residual pixel-by-pixel overlay reconstruction method to overlay the image enhancement mask with the source domain image to obtain the enhanced and reconstructed image.

[0024] The domain alignment adaptive module uses the enhanced reconstructed image and the target reference image as supervision, constructs a reference domain statistical template and a preset statistical threshold interval, calculates the distribution difference index, and iteratively updates the parameterized mapping structure accordingly to output the domain alignment enhancement model.

[0025] A smart adaptive reconstruction device for cross-scene image quality enhancement includes: a processor, a memory, and program instructions stored in the memory and executed by the processor. When executed, the program instructions enable the smart adaptive reconstruction device to: acquire a cross-scene input image as a source domain image and construct pairing data between the source domain image and a target reference image; calculate and scale the source domain image pixel by pixel to obtain an illumination intensity map, and generate a brightness-guided attention weight map based on the illumination intensity map and perform scaling and reshaping to generate a scaled illumination attention map; generate an image enhancement mask based on the source domain image and the scaled illumination attention map, and reconstruct an enhanced reconstructed image by residual pixel-by-pixel overlay; construct a reference domain statistical template and a preset statistical threshold interval based on the enhanced reconstructed image and the target reference image as a supervised basis, and calculate the distribution difference index; output a domain alignment enhancement model based on an iteratively updated parameterized mapping structure; and calculate scene indicators based on the illumination intensity map and the scaled illumination attention map during the inference stage to automatically set the upper limits of brightness remapping gain, local contrast gain, and detail gain, and determine the enhancement focus area, and output an enhanced result image that is closer to the target reference domain distribution in terms of global brightness distribution and global color distribution.

[0026] The technical effects and advantages of this invention, a smart adaptive reconstruction method, system, and device for cross-scene image quality enhancement, are as follows: Under the paired supervision of a source image and a clear daytime target reference image, this invention first calculates and scales a light intensity map of the same size as the source image pixel by pixel. Then, the light intensity is mapped to brightness-guided attention weights and rescaled to generate a scaled light attention map, enabling the attention weights to participate in pixel-level enhancement intensity scheduling, thereby highlighting the key areas that need to be restored spatially. Subsequently, an image enhancement mask is generated using the source image and the scaled light attention map as input, and residual pixel-by-pixel superposition reconstruction is used to obtain the enhanced image. The image is strongly reconstructed, and the enhancement is injected only in the form of "change" while maintaining geometric consistency. Finally, a reference domain statistical template and a preset statistical threshold interval are constructed based on the enhanced and reconstructed image and the target reference image as supervision. The distribution difference index is calculated and the parameterized mapping structure output domain alignment enhancement model is iteratively updated so that the output is closer to the target reference domain in terms of global brightness distribution and global color distribution. At the same time, in the inference stage, scene indexes are calculated based on the illumination intensity map and the scaled illumination attention map, and the upper limits of brightness remapping gain, local contrast gain and detail gain are automatically set and the enhancement focus area is determined, thereby achieving adaptive control of "amplitude and focus". Attached Figure Description

[0027] Figure 1 This is a flowchart of an intelligent adaptive reconstruction method for cross-scene image quality enhancement according to the present invention.

[0028] Figure 2 This is a schematic diagram illustrating the enhancement effect of existing low-light color images.

[0029] Figure 3 Example diagram of paired samples of cross-scene source domain images and clear reference domain images in this invention.

[0030] Figure 4 A schematic diagram of the structure of an intelligent adaptive reconstruction system for cross-scene image quality enhancement according to the present invention. Detailed Implementation

[0031] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0032] Example 1

[0033] like Figure 1As shown, the present invention provides an intelligent adaptive reconstruction method for cross-scene image quality enhancement, which includes the following steps:

[0034] Step 1: Obtain cross-scene input images and construct pairing data between source domain images and target reference images.

[0035] Step 2: Calculate the illumination map based on the source domain image to obtain the brightness-guided attention weight map, and then perform scaling and reshaping to generate a scaled illumination attention map.

[0036] Step 3: Overlay the image enhancement mask with the source domain image to obtain the enhanced reconstructed image.

[0037] Step 4: Based on the enhanced reconstructed image and the target reference image, output domain alignment enhancement model is generated, and the amplification and enhancement focus are automatically adjusted under different weather and lighting conditions.

[0038] Further, in step 1, during the training data construction phase, cross-scene input images are acquired as source domain images. These cross-scene input images originate from acquisition results under different imaging conditions, including at least one of rain, snow, fog, and low-light scenes. Simultaneously, a corresponding target reference image is determined for each source domain image, forming paired data for training. The target reference image is obtained by acquiring the same scene over multiple time periods under conditions of the same road segment, the same camera pose, and similar viewing angles during a clear daytime period. It is then paired with the corresponding cross-scene source domain image based on geographical location and road structure consistency, forming paired data. The target reference image provides appearance and texture references under clear visibility conditions, providing realistic distribution constraints. Figure 3 The diagram shows a pairing-supervised construction method for the source image and the target reference image.

[0039] Further, in step 2, the source image is kept as a color image, and its color channels are read pixel by pixel. At each pixel location, the channels are synthesized into a luminance component of that pixel according to a preset linear weighting method. The luminance component represents the illumination intensity of that pixel under the current imaging conditions. Subsequently, the luminance components in the entire source image are uniformly scaled. By cropping the effective range of luminance and mapping it to a fixed normalized interval, the luminance dimensions obtained under different weather and exposure conditions are consistent, forming an illumination intensity map that corresponds one-to-one with the pixel location. The illumination intensity map has the same size as the source image.

[0040] The values ​​of the illumination intensity map are uniformly normalized to a fixed range (a continuous range from the darkest to the brightest), and any abnormal bright or dark spots are cropped. Then, for each pixel position, the illumination intensity of the pixel is converted into an attention weight according to the rule of "the lower the brightness, the higher the attention," that is, high illumination intensity is mapped to a smaller attention weight, and low illumination intensity is mapped to a larger attention weight, to obtain an attention weight map with the same resolution as the illumination intensity map. Finally, the attention weight map is lightly smoothed (e.g., using small-scale mean filtering or Gaussian filtering) to suppress attention jitter caused by pixel-level noise, making the weight changes of adjacent pixels more continuous, to obtain a brightness-guided attention weight map.

[0041] The brightness-guided attention weight map is subjected to pixel-by-pixel "scaling and reshaping" to generate a scaled illumination attention map. Specifically, this includes: keeping the value range and resolution of the brightness-guided attention weight map unchanged, and reading the brightness-guided attention weight at each position pixel by pixel (the brightness-guided attention weight is derived from the illumination intensity); then applying a preset quadratic scaling transformation to the brightness-guided attention weight of each pixel, that is, multiplying the brightness-guided attention weight by the result of "subtracting a constant 2" and taking its negative number to obtain the scaled attention weight of the corresponding pixel, forming a new attention map, which is the scaled illumination attention map.

[0042] Further, in step 3, the source image and the scaled illumination attention map are unified to the same spatial resolution and mapped pixel-by-pixel, so that the attention weights in the scaled illumination attention map can directly participate in the enhancement intensity scheduling of each pixel; the source image is converted to a color space containing luminance and chrominance components, and the enhancement processing focuses on the luminance component to ensure that the color information is mainly maintained by the original chrominance component, reducing the risk of color drift from the source; the scaled illumination attention map is used as a spatially variable control map, dividing pixels into different enhancement types such as "low-light high-attention areas" and "high-brightness occlusion areas that need to be restored": in areas with low brightness and high attention weights, the luminance component undergoes adaptive gamma correction to increase the grayscale of dark areas, and then contrast-limited adaptive histogram equalization is superimposed to improve local visibility, while anti-sharpening masking is introduced to compensate for high-frequency details and suppress texture blurring caused by brightening; suppress texture blurring caused by brightening; and utilize guided filtering or Bilateral filtering decomposes the luminance component into low-frequency illumination and high-frequency details. Local dynamic range compression is applied to the low-frequency illumination to reduce whitening, while mid-frequency contrast is stretched to restore road boundaries and distant target outlines. This enhancement is also constrained by attention weights to avoid overexposure. The luminance enhancement results from different regions are fused at the pixel level according to attention weights to form an enhanced luminance map that balances dark area brightening with highlight occlusion detail recovery. Edge smoothing is applied to the fused result to suppress noise amplification and block artifacts. The enhanced luminance map is recombined with the original chrominance components and mapped back to the source domain image representation to obtain the enhanced result image. A pixel-by-pixel, channel-by-channel difference method is used to calculate the difference between the enhanced result image and the source domain image for each pixel position (and each color channel of that pixel), resulting in a difference map with the same size as the source domain image. This difference map is then clipped and lightly smoothed to output as an image enhancement mask.

[0043] A residual pixel-by-pixel overlay reconstruction method is used to overlay the image enhancement mask with the source image to obtain an enhanced reconstructed image. Specifically, this includes: assuming the image enhancement mask and the source image are aligned to the same resolution and coordinates, unifying them to the same color space and numerical representation domain (e.g., maintaining RGB three channels and converting pixel values ​​to floating-point representation to avoid integer overflow and quantization errors); then, performing signed addition on each channel at each pixel location, that is, directly adding the change in the image enhancement mask at that pixel location and channel to the corresponding channel value of the source image to obtain the enhanced output value of that pixel; since the image enhancement mask is a "relative enhancement result..." The pixel-by-pixel change in the source image is used to reconstruct the enhanced result while maintaining the geometric consistency of the input in structure, making only controlled adjustments to brightness, contrast, color, and texture details. The pixel values ​​after superposition are subject to range constraints. When using a 0-1 normalized floating-point domain, the result is truncated to fall within the 0-1 range. When using a 0-255 domain, the result is truncated to 0-255 and then converted to the target data type. In addition, in order to suppress isolated outliers caused by local noise amplification, a light edge smoothing can be applied to the image enhancement mask before or after superposition. However, the smoothing intensity is limited to avoid destroying the edge structure, thereby stabilizing the enhanced reconstructed image.

[0044] Further, step 4, using the enhanced reconstructed image and the target reference image as the supervised basis, employs a domain alignment adaptive module to output a domain alignment enhancement model that can map the source domain image to a distribution close to the target reference domain. Specifically, the domain alignment adaptive module acquires the enhanced reconstructed image set and the target reference image set, establishes a one-to-one pairing relationship between the enhanced reconstructed image set and the target reference image set under coarse alignment conditions of the same road segment, the same geographical location, or the same road structure, performs resolution uniformity and coordinate alignment constraints on each pair of paired images to ensure that pixel coordinates are comparable; and outputs a paired image sample set accordingly. The paired image sample set contains each enhanced reconstructed image and its unique corresponding target reference image, as well as their alignment parameters. The domain alignment adaptive module extracts all target reference images from the paired image sample set, calculates the global luminance and global color statistics of the target reference images within a unified color representation domain, and constructs a reference domain statistical template based on the global luminance and global color statistics. In the reference domain statistical template, segmented statistical calibers are set and statistical bin boundaries are fixed. Based on the statistical distribution of the target reference image set, a preset statistical threshold interval is calculated for each bin. The preset statistical threshold interval is used to define the target distribution boundary of the reference domain. Using the reference domain statistical template and the preset statistical threshold interval as statistical caliber constraints, the domain alignment adaptive module extracts all enhanced reconstructed images from the paired image sample set, calculates the global luminance and global color statistics of each enhanced reconstructed image under the definition of the statistical bin boundaries and the preset statistical threshold interval, and organizes the statistical results (global luminance and global color statistics) corresponding to each enhanced reconstructed image into a source-side statistical vector set. The domain alignment adaptive module, based on the source-side statistical vector set, calculates the statistical differences in brightness and color on an item-by-item basis in terms of dimension, binning, and threshold definition. The difference results are aggregated into a distribution difference index set, which simultaneously outputs an overall difference term and a segmented difference term. The overall difference term is used to constrain global distribution alignment, while the segmented difference term is used to constrain the offset correction of different brightness and color intervals. The distribution difference index set is solidified into a target description, which explicitly lists the target direction and constraint threshold corresponding to each type of difference term. Using the target description as the constraint basis, the domain alignment adaptive module constructs a parameterized mapping form for the domain alignment enhancement model. This parameterized mapping form refers to a family of mapping functions from the source domain image to the target reference domain distribution. An adjustable parameter set is set for this mapping function family. This adjustable parameter set includes at least a subset of parameters for global color correction and a subset of parameters for global brightness remapping, ensuring that the domain alignment enhancement model has adjustable degrees of freedom for both brightness and color difference terms.The domain alignment adaptive module inputs the source domain image from the paired image sample set into the parameterized mapping structure determined by the current iteration parameters, and generates the corresponding output image according to global color correction and global brightness remapping operations, thereby obtaining the mapping result image. In each iteration, the domain alignment adaptive module calculates brightness statistics and color statistics for the mapping result image, and measures the differences with the reference domain statistical template item by item under a unified standard and a preset statistical threshold range, and records them to form a difference index sequence.

[0045] The domain alignment adaptive module outputs a domain alignment enhancement model based on the difference index sequence and the distribution difference index. Specifically, it includes: using the difference index sequence as the convergence trend determination input and the distribution difference index set as the heterogeneous terms constraining the domain statistical caliber input; performing stability analysis on the difference index sequence within a preset iteration window to determine whether each difference term continues to decrease and the decrease magnitude meets a preset threshold condition; comparing the difference index corresponding to the current iteration with the distribution difference index set item by item to confirm that both the brightness difference term and the color difference term have been compressed to within the preset statistical threshold range; selecting the candidate parameter set corresponding to the iteration round with the smallest difference index under the conditions of trend stability and difference term compliance, and binding and solidifying the candidate parameter set with a predetermined parameterized mapping structure to form a domain alignment enhancement model that can be directly used in the inference stage. The domain alignment enhancement model is used to map the source domain image to an output result that is closer to the target reference domain distribution in terms of global brightness distribution and global color distribution.

[0046] The domain alignment enhancement model is used to automatically adjust the enhancement magnitude and focus under different weather and lighting conditions, and output a result image that is closer to the clear reference domain. Specifically, the model converts the input source domain image to a representation domain where luminance and chrominance components are separable, and calculates scene metrics based on the illumination intensity map and scaled illumination attention map of the source domain image. These scene metrics include the proportion of dark areas, the proportion of bright areas occlusion, and the degree of local contrast attenuation. The model automatically sets a set of enhancement magnitude parameters based on the representation domain and scene metrics. These parameters include the overall and piecewise gain of the luminance remapping curve, the upper limit of gain for local contrast enhancement, and the upper limit of high-frequency gain for detail sharpening. Furthermore, based on the scaled illumination attention map… The image identifies key enhancement areas, specifically including road boundaries and distant weak target outlines in low-light areas, as well as areas obscured by fog or snow causing bright whitening. The domain alignment enhancement model increases local contrast and detail gain in key enhancement areas and reduces gain in non-key areas to avoid over-enhancement. The domain alignment enhancement model performs white balance correction and color constancy correction across the entire image to correct color cast, and compresses the output luminance histogram statistics and chromaticity statistics to a preset statistical threshold range within the reference domain through luminance distribution alignment. This makes the output result closer to the sharp reference domain in terms of global luminance and global color distribution. At the same time, it suppresses noise amplification and halo effects through edge preservation artifact suppression, ultimately outputting an enhanced image that is closer to the sharp reference domain.

[0047] In this embodiment, to verify the cross-scene image quality enhancement effect of the enhanced mask generation and residual pixel-by-pixel overlay reconstruction based on the illumination intensity map and scaled illumination attention map described in this application, combined with the domain alignment adaptive output domain alignment enhancement model, peak signal-to-noise ratio (PSNR) is used as an objective evaluation index. On the test sets of rain, snow, fog and low illumination scenes, the original unenhanced image of the cross-scene input image, the enhancement algorithm based on the brightness invariant processing model combined with the limited contrast adaptive histogram equalization, and the two implementation schemes of this application without the domain alignment adaptive module and with the domain alignment adaptive module are compared. The results are shown in Table 1.

[0048] Table 1. PSNR Comparison for Different Scenarios

[0049] method rain Snow fog low light average The input image across scenes is used as the raw, unenhanced image of the source domain image. 15.8 15.2 14.5 16.4 15.5 Based on a brightness-invariant processing model and combined with contrast-limited adaptive histogram equalization 19.3 18.9 18.2 20.1 19.1 Dark Passage Prior (DCP) 18.7 17.9 20.5 17.2 18.6 Retinex 18.1 17.6 19.8 19.0 18.6 This invention (excluding the domain alignment adaptive module) 22.1 21.7 21.2 23.0 22.0 This invention (including a domain alignment adaptive module) 23.0 22.6 22.1 23.8 22.9

[0050] The experimental results in Table 1 show that under four different scene conditions (rain, snow, fog, and low illumination), the PSNR of the original unenhanced image in the source domain is generally low (approximately 14.5–16.4 dB), indicating that adverse imaging conditions significantly introduce noise and structural distortion. After adopting the enhancement algorithm based on a brightness-invariant processing model combined with contrast-constrained adaptive histogram equalization, the PSNR is improved (approximately 18.2–20.1 dB), but the improvement is limited. Without introducing a domain alignment adaptive module, this invention generates an enhancement mask based on the illumination intensity map and scaled illumination attention map, and reconstructs it pixel-by-pixel through residual overlay, further improving the PSNR to approximately 21.2–23.0 dB. After adding the domain alignment adaptive module to the complete scheme, the PSNR in all four scenes continues to increase (approximately +0.8 to +0.9 dB), with an average value of approximately 22.9 dB. This indicates that by iteratively updating the parameterized mapping structure using a reference domain statistical template and threshold interval constraints, the distribution difference between the output and the clear reference domain can be effectively reduced, thereby obtaining more stable and consistent enhanced reconstruction quality in multiple scenes.

[0051] In this embodiment, to further evaluate the improvement effect of the enhanced reconstructed image described in this application in terms of maintaining texture details and overall structural consistency from the perspective of structural similarity, the structural similarity index (SSIM) is used as an objective evaluation index to statistically compare the output results of each method. The results are shown in Table 2.

[0052] Table 2 Comparison of SSIM in various scenarios

[0053] method rain Snow fog low light average The input image across scenes is used as the raw, unenhanced image of the source domain image. 0.55 0.53 0.49 0.57 0.54 Based on a brightness-invariant processing model and combined with contrast-limited adaptive histogram equalization 0.70 0.69 0.66 0.74 0.70 Dark Passage Prior (DCP) 0.68 0.64 0.73 0.60 0.66 Retinex 0.66 0.63 0.71 0.70 0.68 This invention (excluding the domain alignment adaptive module) 0.80 0.79 0.77 0.82 0.80 This invention (including a domain alignment adaptive module) 0.82 0.81 0.80 0.84 0.82

[0054] The experimental results in Table 2 show that, under four scenarios—rain, snow, fog, and low illumination—the SSIM of the original unenhanced image in the source domain is generally low (approximately 0.49–0.57), indicating that the image structure and texture details are significantly damaged under adverse imaging conditions. After adopting the enhancement algorithm based on a brightness-invariant processing model combined with contrast-limited adaptive histogram equalization, the SSIM is improved to approximately 0.66–0.74, indicating some improvement in the recovery of the overall structure, but some detail loss or local distortion still exists. This invention, without introducing a domain alignment adaptive module, uses the illumination intensity map and the scaled illumination attention map... By implementing partitioned enhancement intensity scheduling and leveraging enhancement masks and residual pixel-by-pixel overlay reconstruction, structural consistency is effectively improved, with SSIM further increasing to approximately 0.77–0.82. After adding a domain alignment adaptive module to the complete scheme, SSIM in each scene continues to rise (approximately +0.02 to +0.03), averaging approximately 0.82. This indicates that by iteratively updating the parameterized mapping structure using reference domain statistical templates and threshold interval constraints, the impact of cross-scene distribution offset on structural similarity can be further reduced, thereby achieving a more stable and consistent structural fidelity enhancement effect under different weather and lighting conditions.

[0055] like Figure 4 As shown, an intelligent adaptive reconstruction system for cross-scene image quality enhancement includes:

[0056] The paired data construction module acquires cross-scene input images as source domain images and determines a corresponding target reference image for each source domain image, forming paired data for training.

[0057] The illumination intensity map generation module calculates the brightness component pixel by pixel in the source domain image and performs scaling processing to generate an illumination intensity map with the same size as the source domain image.

[0058] The brightness-guided attention generation module maps pixel illumination intensity to attention weights based on the illumination intensity map and smooths the attention weight map to obtain the brightness-guided attention weight map.

[0059] The scaling and reshaping module applies a preset quadratic scaling transformation to the brightness-guided attention weight map to generate a scaled illumination attention map.

[0060] The enhancement mask generation module takes the source domain image and the scaled illumination attention map as input to perform luminance component enhancement, region weighted fusion and edge smoothing, and obtains the image enhancement mask through pixel-by-pixel channel difference.

[0061] The residual overlay reconstruction module uses a residual pixel-by-pixel overlay reconstruction method to overlay the image enhancement mask with the source domain image to obtain an enhanced and reconstructed image.

[0062] The domain alignment adaptive module uses the enhanced reconstructed image and the target reference image as supervision, constructs a reference domain statistical template and a preset statistical threshold interval, calculates the distribution difference index, and iteratively updates the parameterized mapping structure accordingly to output the domain alignment enhancement model.

[0063] A smart adaptive reconstruction device for cross-scene image quality enhancement includes: a processor, a memory, and program instructions stored in the memory and executed by the processor. When executed, the program instructions enable the smart adaptive reconstruction device to: acquire a cross-scene input image as a source domain image and construct pairing data between the source domain image and a target reference image; calculate and scale the source domain image pixel by pixel to obtain an illumination intensity map, and generate a brightness-guided attention weight map based on the illumination intensity map and perform scaling and reshaping to generate a scaled illumination attention map; generate an image enhancement mask based on the source domain image and the scaled illumination attention map, and reconstruct an enhanced reconstructed image by residual pixel-by-pixel overlay; construct a reference domain statistical template and a preset statistical threshold interval based on the enhanced reconstructed image and the target reference image as a supervised basis, and calculate the distribution difference index; output a domain alignment enhancement model based on an iteratively updated parameterized mapping structure; and calculate scene indicators based on the illumination intensity map and the scaled illumination attention map during the inference stage to automatically set the upper limits of brightness remapping gain, local contrast gain, and detail gain, and determine the enhancement focus area, and output an enhanced result image that is closer to the target reference domain distribution in terms of global brightness distribution and global color distribution.

[0064] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

[0065] In conclusion, 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 smart adaptive reconstruction method for cross-scene image quality enhancement, characterized in that, Includes the following steps: Step 1: Obtain cross-scene input images and construct pairing data between source domain images and target reference images; Step 2: Calculate the illumination map based on the source domain image to obtain the brightness-guided attention weight map, and then perform scaling and reshaping to generate a scaled illumination attention map; Step 3: Overlay the image enhancement mask with the source domain image to obtain the enhanced reconstructed image; Step 4: Using the enhanced reconstructed image and the target reference image as supervision, output domain alignment enhancement model, automatically adjust the amplification and enhancement focus under different weather and lighting mechanisms; In step 4, the domain alignment adaptive module is used to output the domain alignment enhancement model, including: establishing a pairing relationship between the enhanced reconstructed image and the target reference image, and outputting a pairing image sample set; extracting the target reference image, constructing a reference domain statistical template, fixing the statistical bin boundaries in the reference domain statistical template, calculating the preset statistical threshold intervals corresponding to each bin based on the statistical distribution of the target reference image, and then organizing the global brightness statistics and global color statistics corresponding to each enhanced reconstructed image into a source-side statistical vector set; calculating the brightness statistical difference and color statistical difference item by item based on the source-side statistical vector set, merging the difference results into a distribution difference index set, solidifying the distribution difference index set into a target description, using the target description as a constraint, iteratively mapping the paired samples, calculating brightness and color statistics, and measuring the difference using the reference domain statistical template to form a difference index sequence, and outputting the domain alignment enhancement model accordingly.

2. The intelligent adaptive reconstruction method for cross-scene image quality enhancement according to claim 1, characterized in that... The aforementioned output domain alignment enhancement model specifically includes: using the difference index sequence as the convergence trend determination input, using the distribution difference index set as the heterogeneous term domain statistical caliber constraint input, judging the decrease of each difference term in the difference index sequence within a preset iteration window, comparing the difference index corresponding to the current iteration with the distribution difference index set item by item; under the conditions of trend stability and difference term compliance, selecting the candidate parameter set associated with the iteration round with the smallest difference index, and binding and solidifying the candidate parameter set with the predetermined parameterized mapping structure to form the domain alignment enhancement model.

3. The intelligent adaptive reconstruction method for cross-scene image quality enhancement according to claim 2, characterized in that... The domain alignment model automatically adjusts the enhancement magnitude and focus under different weather and lighting conditions, and outputs a result image that is closer to the clear reference domain. This includes: the domain alignment enhancement model converts the input source domain image to a representation domain where luminance and chrominance components are separable, and calculates scene indicators based on the illumination intensity map and scaled illumination attention map of the source domain image; sets a set of enhancement magnitude parameters based on the representation domain and scene indicators; determines the enhancement focus area based on the scaled illumination attention map; performs white balance correction and color constancy correction across the entire image to correct color cast; and compresses the output luminance histogram statistics and chrominance statistics to a preset statistical threshold range in the reference domain through luminance distribution alignment, making the output result closer to the clear reference domain in terms of global luminance and global color distribution, thus outputting an enhanced result image that is closer to the clear reference domain.

4. The intelligent adaptive reconstruction method for cross-scene image quality enhancement according to claim 1, characterized in that, The construction of the pairing data includes: the imaging conditions of the cross-scene input image include at least one of rain, snow, fog and low-light scenes; the target reference image is acquired by multiple time periods of the same scene under the same road segment, the same camera pose and similar viewing angle conditions during clear daytime, and is paired with the corresponding source domain image according to the consistency of geographical location and road structure.

5. The intelligent adaptive reconstruction method for cross-scene image quality enhancement according to claim 1, characterized in that, Obtaining the illumination intensity map includes: keeping the source image as a color image and reading the color channels pixel by pixel; combining each channel into a luminance component at each pixel position according to a preset linear weighting method to characterize the illumination intensity; and cropping and normalizing the luminance component of the entire image to obtain an illumination intensity map with the same size as the source image.

6. The intelligent adaptive reconstruction method for cross-scene image quality enhancement according to claim 1, characterized in that, Obtaining the brightness-guided attention weight map includes: normalizing the illumination intensity map and cropping abnormal bright spots or abnormal dark spots; converting illumination intensity into attention weights according to the rule that lower brightness corresponds to higher attention weights and higher brightness corresponds to lower attention weights; and smoothing the attention weight map using mean filtering to suppress attention jitter caused by pixel-level noise.

7. The intelligent adaptive reconstruction method for cross-scene image quality enhancement according to claim 1, characterized in that, The scaling reshaping includes: while keeping the value range and resolution of the brightness-guided attention weight map unchanged, applying a quadratic scaling transformation to the brightness-guided attention weight at each pixel position, multiplying the brightness-guided attention weight by the result of subtracting a constant 2 and taking the opposite number to obtain the scaling attention weight at the corresponding pixel position in the scaling illumination attention map.

8. An intelligent adaptive reconstruction system for cross-scene image quality enhancement, characterized in that, include: The paired data construction module acquires cross-scene input images as source domain images and determines the corresponding target reference image for each source domain image to form paired data for training. The illumination intensity map generation module calculates the brightness component pixel by pixel of the source domain image and performs scaling processing to generate an illumination intensity map with the same size as the source domain image. The brightness-guided attention generation module maps pixel illumination intensity to attention weights based on the illumination intensity map and smooths the attention weight map to obtain the brightness-guided attention weight map. The scaling and reshaping module applies a preset quadratic scaling transformation to the brightness-guided attention weight map to generate a scaled illumination attention map. The enhanced mask generation module takes the source domain image and the scaled illumination attention map as input to perform brightness component enhancement, region weighted fusion and edge smoothing, and obtains the image enhancement mask through pixel-by-pixel channel difference. The residual overlay reconstruction module uses a residual pixel-by-pixel overlay reconstruction method to overlay the image enhancement mask with the source domain image to obtain the enhanced and reconstructed image. The domain alignment adaptive module uses the enhanced reconstructed image and the target reference image as supervision, constructs a reference domain statistical template and a preset statistical threshold interval, calculates the distribution difference index, and iteratively updates the parameterized mapping structure accordingly to output the domain alignment enhancement model.

9. A smart adaptive reconstruction device for cross-scene image quality enhancement, characterized in that, include: The processor, memory, and program instructions stored in the memory and executed by the processor, wherein when executed, the program instructions enable the intelligent adaptive reconstruction device to: acquire a cross-scene input image as a source domain image and construct pairing data between the source domain image and a target reference image; The source domain image is calculated and scaled pixel by pixel to obtain an illumination intensity map. A brightness-guided attention weight map is generated based on the illumination intensity map and then scaled and reshaped to generate a scaled illumination attention map. An image enhancement mask is generated based on the source domain image and the scaled illumination attention map, and the enhanced reconstructed image is obtained by residual pixel-by-pixel superposition reconstruction. Based on the enhanced reconstructed image and the target reference image, a reference domain statistical template and a preset statistical threshold interval are constructed and the distribution difference index is calculated. Based on the iteratively updated parameterized mapping structure, the output domain alignment enhancement model is generated. During the inference stage, scene indicators are calculated based on the illumination intensity map and the scaled illumination attention map to automatically set the upper limits of brightness remapping gain, local contrast gain and detail gain and determine the key enhancement areas. The resulting enhanced image is output, which is closer to the distribution of the target reference domain in terms of global brightness distribution and global color distribution.

Images