A sand and dust image enhancement method based on bright color separation and color cast back protection

Abstract

The application discloses a dust image enhancement method based on bright color separation and color deviation back-off protection, obtains and calculates the global color statistics, dust layer and local activity degree of a dust-degraded image to be processed; respectively learns and enhances the dust-degraded image to obtain a learning-enhanced image and a traditional enhanced image; separates and fuses the brightness and chroma of the learning-enhanced image and the traditional enhanced image to obtain an initial fusion image; jointly recovers the fused brightness and fused chroma of the initial fusion image according to the dust layer of the dust-degraded image to be processed to obtain a joint recovery image; performs limited color feedback and weak red channel balance on the joint recovery image according to neutral pixel RGB statistical deviation to obtain a local chroma correction image; performs intensity rollback or correction rejection of the local chroma correction amount according to the color deviation scores before and after the local chroma correction; and selects a path image with better color deviation score and meeting a preset improvement condition as a final target enhanced image.

Description

Technical Field

[0001] This invention relates to the field of image enhancement technology, and in particular to a method for enhancing sand and dust images based on brightness-color separation and color shift protection. Background Technology

[0002] Suspended particles in dust storms absorb and scatter visible light, causing degradation in acquired images, including reduced contrast, decreased distant information, hazy appearance, overemphasis on red and yellow channels, loss of blue information, and increased local darkness. This degradation exhibits both the atmospheric light superposition characteristics similar to haze imaging and significant wavelength selectivity. Conventional dehazing methods, if they do not distinguish between brightness and chromaticity, easily result in overall darkening, over-enhancing of the sky, or color distortion; simply performing white balance or channel stretching is insufficient to recover structural information obscured by the dust cloud. Existing deep networks can learn the mapping from dusty images to clear images, but larger models typically suffer from large parameter counts and high inference resource requirements; inadequate generalization may also occur when training data differs from the dust concentration, lighting conditions, or camera response of real dusty scenes. Traditional enhancement algorithms have lower computational resource requirements but often suffer from local over-darkness, residual red and yellow color casts, oversaturation, or insufficient removal of haze due to reliance on fixed priors or parameters. Furthermore, existing methods typically perform dehazing, color correction, and contrast enhancement in a simple, sequential manner, lacking decoupling between the effects of brightness and chromaticity, and also lacking a protection mechanism to automatically reduce or reject unreliable color corrections based on changes in color cast before and after processing; when a color correction is not suitable for the current image, the error may be amplified in subsequent processing. Summary of the Invention

[0003] Purpose of the invention: In order to overcome the shortcomings of the existing technology, the present invention provides a sand and dust image enhancement method based on brightness and color separation and color deviation back-off protection, so as to solve the problems of residual red and yellow color deviation, dark area enhancement, hazy appearance, miscorrection of inherent color, and insufficient stability of learning models across scenes in real sand and dust images.

[0004] Technical Solution: To achieve the above objectives, the present invention provides a method for enhancing dust images based on luminance-color separation and color cast protection, comprising the following steps:

[0005] Step 1: Obtain the sand and dust degradation image to be processed, and calculate the global color statistics, sand and dust curtain layer, and local activity of the sand and dust degradation image to be processed.

[0006] Step 2: Enhance the sandstorm-degraded image to be processed using a lightweight learning enhancement model and a traditional image enhancement model, respectively, to obtain the learned enhanced image and the traditional enhanced image.

[0007] Step 3: Separate the luminance and chrominance of the learned enhanced image and the traditional enhanced image, and fuse them with the luminance component of the traditional enhanced image as the main component and the chrominance component of the learned enhanced image as the main component to obtain the initial fused image.

[0008] Step 4: Extract the fused brightness and fused chromaticity from the initial fused image. Estimate the transmittance, atmospheric brightness, and colored atmospheric light based on the dust curtain layer of the sandstorm degradation image to be processed. Construct a brightness recovery path and a chromaticity recovery path based on the estimated transmittance, atmospheric brightness, and colored atmospheric light. Recover the fused brightness and fused chromaticity through the brightness recovery path and the chromaticity recovery path respectively to obtain the target brightness and candidate chromaticity. Perform RGB displayable domain synthesis while preserving brightness based on the obtained target brightness and candidate chromaticity to obtain the joint restored image.

[0009] Step 5: Filter the neutral pixels in the jointly restored image, and perform restricted color feedback and weak red channel balancing on the jointly restored image based on the RGB statistical offset of the neutral pixels to obtain a local chromaticity corrected image.

[0010] Step 6: Calculate the color cast score before and after local color correction, and perform intensity rollback or correction rejection of the local color correction amount based on the color cast score before and after local color correction; and when the triggering condition for final path selection is met, obtain and compare the color cast scores of the main path image and the backup path image, and select the path image with better color cast score and meeting the preset improvement conditions as the final target enhancement image.

[0011] Furthermore, in step 1, the global color statistics, dust curtain layer, and local activity of the degraded dust image to be processed are calculated; the calculation process is as follows:

[0012] ;

[0013] ;

[0014] ;

[0015] ;

[0016] In the formula, This is a global color statistics vector. For local activity, A normalized dust curtain layer; , The mean and standard deviation of the dust degradation image in channel c are respectively. Channel images in dust degradation images The q-th quantile, where Q is the set of quantile levels; ; radius Smooth operation, The input brightness is used for images degraded by sandstorms. This is an absolute value operation; , These are the response values ​​at the x-pixel and y-pixel positions, respectively, within the same dust curtain layer. This is the set of pixels in the degraded image of the sandstorm to be processed; To Take the 97.5th quantile. To prevent the smallest positive number with a denominator of zero, To be The value is truncated to Within the range; Large-scale smooth radius Smooth operation, This represents the positive excess of the red-yellow channel relative to the blue channel in the image of dust degradation. To define the flatness weight, For low saturation weights, This is the brightness weight.

[0017] Furthermore, in step 3, both the learned enhanced image and the traditional enhanced image are separated to obtain the luminance and chrominance components of the learned enhanced image and the traditional enhanced image, respectively. The calculation process is as follows:

[0018] , ; , ;

[0019] In the formula, To learn augmented images, For traditional image enhancement, , These are the luminance and chrominance components of the image, respectively, for learning and enhancing the image. , These are the luminance and chrominance components of a traditional enhanced image, respectively. For brightness extraction operators, This is the chroma extraction operator.

[0020] The fused luminance is obtained by fusing the luminance component of the traditional enhanced image with the luminance component of the learned enhanced image, and the fused chrominance is obtained by fusing the chrominance component of the learned enhanced image with the chrominance component of the traditional enhanced image. The calculation process is as follows:

[0021] ; ;

[0022] In the formula, To blend brightness, To blend colors, The fusion weights are used to determine the fusion brightness. The fusion weights are used to fuse the chroma; an initial fused image is obtained based on the fused luminance and fused chroma components; the calculation process is shown below:

[0023] ;

[0024] In the formula, For the initial fused image, To be The value is truncated to Within the interval, To copy the scalar brightness to the three RGB channels of the image, This is the chromaticity amplitude coefficient.

[0025] Furthermore, in step 4, in the initial fused image Extracting fusion brightness and blending color Based on the dust curtain layer of the degraded dust image to be processed, the transmittance, atmospheric luminosity, and colored atmospheric light are estimated; the calculation process is shown below:

[0026] ;

[0027] ; ;

[0028] ;

[0029] In the formula, Atmospheric luminance, The colored atmospheric light after brightness normalization. Transmittance; This is a collection of pixels from a high-dust curtain layer. To obtain the 95th quantile, To take the 90th quantile; A collection of pixels for a high dust curtain layer A set of pixels with a medium brightness not lower than a preset high percentile; This is the transpose of the RGB luminance weights, where T is the transpose operation; This is the lower limit of transmittance. To restore strength.

[0030] Based on the estimated transmittance, atmospheric luminance, and colored atmospheric light, a luminance recovery path and a chromaticity recovery path are constructed. The luminance recovery path sequentially performs maximum darkening constraint-based recovery, local and mesoscale contrast enhancement, and adaptive hue, shadow enhancement, and specular protection on the fused luminance to obtain the target luminance. The chromaticity recovery path sequentially performs transmittance-based physical chromaticity recovery, joint gating of dust curtain layer, saturation, luminance, and local activity, and removal of the orthographic projection of low-frequency color bias from the source dust to obtain candidate chromaticities.

[0031] Furthermore, based on the obtained target luminance and candidate chrominance, RGB displayable domain synthesis is performed while preserving luminance to obtain a jointly restored image. The candidate chrominance is projected as zero-luminance chrominance, and the maximum allowable chrominance ratio for each channel without crossing the RGB boundaries is calculated using the zero-luminance chrominance and the target luminance. Based on the maximum chrominance ratio of each channel, the minimum feasible ratio for all three RGB channels is calculated. The jointly restored image is then calculated based on the zero-luminance chrominance, the target luminance, and the minimum feasible ratio. The calculation process is shown below:

[0032] ;

[0033] In the formula, To copy the scalar brightness to the three RGB channels of the image, For target brightness, For the minimum feasible proportion, Zero luminance chromaticity To be The value is truncated to Within the range.

[0034] Furthermore, in step 5, neutral pixels in the jointly restored image are selected using pixels with low relative chromaticity and brightness within a preset non-extremely dark, non-extremely bright range. Neutral pixel regions are determined based on these selected neutral pixels. The restricted color gain for each channel is calculated based on the offset between the channel representative value and the channel center value of the neutral pixel region. Spatial mixing of the restricted color gain is performed using an object color preservation map. Restricted color feedback and weak red channel balancing are then applied to the jointly restored image to obtain a local chromaticity corrected image. The calculation for restricted color feedback on the jointly restored image is as follows:

[0035] ;

[0036] ;

[0037] In the formula, This is the intermediate image after spatial weighted correction. To copy the scalar brightness to the three RGB channels of the image, To jointly restore the image, To correct spatial weights, This is channel-by-channel multiplication; Let be the channel gain vector, where These represent the limited color gain for each channel.

[0038] Furthermore, in step 6, the image without local chromaticity correction is used as the reference image, and the obtained locally chromaticity corrected image is used as the candidate image; the color cast score of the reference image and the candidate image are calculated respectively. When the color cast score of the candidate image is less than or equal to the sum of the color cast score and the score tolerance of the reference image, the neutral pixels in the candidate image are selected, and the candidate image is subjected to restricted color feedback and weak red channel balancing based on the RGB statistical offset of the neutral pixels to obtain the main path image.

[0039] When the color cast score of the candidate image is greater than the sum of the color cast score and the score tolerance of the reference image, the local chromaticity correction amount is gradually reduced according to the preset attenuation coefficient, and it is determined whether the current intensity backoff number has reached the maximum number; if the intensity backoff number has not reached the maximum number, the candidate image is corrected by the gradually decreasing local chromaticity correction amount, and a new candidate image is generated; if the intensity backoff number has reached the maximum number, the reference image is rejected after correction; neutral pixels in the reference image are selected, and restricted color feedback and weak red channel balancing are performed on the reference image according to the RGB statistical offset of the neutral pixels to obtain the backup path image.

[0040] Furthermore, based on the obtained main path image and backup path image, the color cast score of the main path image and backup path image is calculated respectively. The color cast score of the main path image is compared with a preset trigger threshold. When the color cast score of the path image is less than the preset trigger threshold, the main path image is output as the final target enhancement image.

[0041] If the color cast score of the main path image is greater than or equal to the preset trigger threshold, the sum of the color cast score and the color cast score improvement amount of the backup path image is compared with the preset minimum improvement threshold. If the sum of the color cast score and the color cast score improvement amount of the backup path image is less than the preset minimum improvement threshold, the backup path image is output as the final target enhancement image. If the sum of the color cast score and the color cast score improvement amount of the backup path image is greater than or equal to the preset minimum improvement threshold, the main path image is output as the final target enhancement image.

[0042] Furthermore, a dust image enhancement system is included to implement a dust image enhancement method based on luminance-color separation and color cast backoff protection. This system comprises: a feature estimation module for calculating global color statistics, dust veil layer, and local activity of the dust-degraded image; a dual-branch enhancement module for enhancing the dust-degraded image using both a lightweight learning enhancement model and a traditional image enhancement model, resulting in a learned enhanced image and a traditional enhanced image; a luminance-color separation and fusion module for separating luminance and chrominance of the learned enhanced image and the traditional enhanced image and fusing them to obtain an initial fused image; a luminance-chrominance joint recovery module for jointly recovering the fused luminance and fused chrominance of the initial fused image using transmittance, atmospheric light intensity, and colored atmospheric light, resulting in a jointly recovered image; a color feedback module for performing restricted color feedback and weak red channel balancing on the jointly recovered image based on the neutral pixel RGB statistical offset, resulting in a locally chrominance corrected image; and a protection decision module for performing intensity backoff or correction rejection of the local chrominance correction based on the color cast score before and after local chrominance correction, and selecting the optimal path image as the final target enhanced image.

[0043] Furthermore, it includes a storage medium storing an executable program, which, when executed by a processor, enables a dust image enhancement method based on luminance-color separation and color shift fallback protection.

[0044] Beneficial Effects: This invention provides a dust image enhancement method based on luminance-color separation and color cast backoff protection. By learning and fusing the luminance-color separation of the image enhancement branch and the traditional image enhancement branch, the learned image enhancement branch mainly contributes chroma, while the traditional image enhancement branch mainly contributes luminance, reducing the risk of color or luminance instability in real-world scenes due to a single branch. Parallel luminance-chroma recovery under the same transmittance coordinates the removal of the dust curtain layer with the suppression of red-yellow color cast. Joint gating of the dust curtain layer, saturation, luminance, and local activity reduces miscorrection of highly saturated objects, dark areas, and high-texture regions. Luminance compensation after RGB displayable domain synthesis and color feedback, while maintaining luminance, reduces luminance drift, oversaturation, and channel truncation caused by color correction. Two-level protection consisting of local correction backoff and final path selection automatically weakens, rejects, or replaces the correction when it is unsuitable for the current image. Collaborative processing of global color transformation and depth-separable local residual networks in the lightweight learning enhancement model reduces the number of model parameters and computational load while maintaining color and detail recovery capabilities. Attached Figure Description

[0045] Figure 1 This is a flowchart of a dust image enhancement method based on brightness-color separation and color shift back protection.

[0046] Figure 2 Network structure diagram for lightweight learning enhancement models.

[0047] Figure 3 This is a flowchart for luminance-chrominance joint recovery.

[0048] Figure 4 A flowchart for local correction rollback and final path selection.

[0049] Figure 5 This is a system block diagram for a dust storm image enhancement system.

[0050] Figure 6 This is a feature distribution map of a sandstorm degradation image.

[0051] Figure 7 Image enhancement result for a dust storm. Detailed Implementation

[0052] The invention will now be further described with reference to the accompanying drawings.

[0053] like Figure 1 As shown, a dust image enhancement method based on brightness-color separation and color cast protection includes the following steps:

[0054] Step 1: Obtain the sand and dust degradation image to be processed, and calculate the global color statistics, sand and dust curtain layer, and local activity of the sand and dust degradation image to be processed;

[0055] Step 2: Enhance the sandstorm-degraded image to be processed using a lightweight learning enhancement model and a traditional image enhancement model, respectively, to obtain the learned enhanced image and the traditional enhanced image.

[0056] Step 3: Separate the luminance and chrominance of the learned enhanced image and the traditional enhanced image, and fuse them with the luminance component of the traditional enhanced image as the main component and the chrominance component of the learned enhanced image as the main component to obtain the initial fused image;

[0057] Step 4: Luminance-Chroma Joint Restoration. Extract the fused luminance and fused chroma from the initial fused image. Estimate the transmittance, atmospheric luminance, and colored atmospheric light based on the dust curtain layer of the degraded dust image to be processed. Construct luminance and chroma restoration paths based on the estimated transmittance, atmospheric luminance, and colored atmospheric light. Restore the fused luminance and fused chroma using the luminance and chroma restoration paths respectively. Restore the fused luminance and fused chroma in parallel under the same transmittance field to obtain the target luminance and candidate chroma respectively. Perform RGB displayable domain synthesis while preserving luminance based on the obtained target luminance and candidate chroma to obtain the jointly restored image.

[0058] Step 5: Filter the neutral pixels in the jointly restored image, and perform restricted color feedback and weak red channel balancing on the jointly restored image based on the RGB statistical offset of the neutral pixels to obtain a local chromaticity corrected image;

[0059] Step 6: Local Correction Rollback and Final Path Selection. Calculate the color cast score before and after local color correction. Based on the color cast score before and after local color correction, perform intensity rollback or correction rejection of the local color correction amount. When the triggering condition for final path selection is met, acquire and compare the color cast scores of the main path image and the backup path image. Select the path image with the better color cast score and that meets the preset improvement conditions as the final target enhancement image, and generate a quality report.

[0060] In step 1, the global color statistics, dust curtain layer, and local activity of the degraded dust image to be processed are calculated; the calculation process is as follows:

[0061] ;

[0062] ;

[0063] ;

[0064] ;

[0065] In the formula, This is a global color statistics vector. For local activity, A normalized dust curtain layer; , The mean and standard deviation of the dust degradation image in channel c are respectively. Channel images in dust degradation images The q-th quantile, where Q is the set of quantile levels; ; radius Smooth operation, The input brightness is used for images degraded by sandstorms. This is an absolute value operation; , These are the response values ​​at the x-pixel and y-pixel positions, respectively, within the same dust curtain layer. The original response function of the dust curtain layer; This is the set of pixels in the degraded image of the sandstorm to be processed; The response values ​​of all pixels in the pixel set of the image to be processed, representing the dust degradation. To Take the 97.5th quantile. To prevent the smallest positive number with a denominator of zero, To be The value is truncated to Within the range; Large-scale smooth radius Smooth operation, This represents the positive excess of the red-yellow channel relative to the blue channel in the image of dust degradation. To define the flatness weight, For low saturation weights, Brightness weight , , , , , All are represented as preset non-negative weight coefficients; square brackets and curly braces in the formula indicate vector concatenation in a given order.

[0066] The input brightness of the dust degradation image The positive excess of red and yellow relative to the blue channel in the dust degradation image Define flatness weights Low saturation weight Brightness weight The calculation is as follows:

[0067] ;

[0068] ;

[0069] ;

[0070] ;

[0071] ;

[0072] ;

[0073] In the formula, For brightness extraction operators, The image shows the degraded dust image to be processed; , , These are the R channel value, G channel value, and B channel value of the x-th pixel, respectively. To be The value is truncated to Within the interval; D represents the local activity. S represents the channel span; Y represents the input brightness of the dust degradation image. The x-th pixel in the c-th channel of the image in the dust degradation image; The normalized upper limit for local activity can be 0.045; This is a normalized parameter for the channel span, used to measure pixel saturation, and can be 0.32; The initial threshold for brightness can be 0.18; This represents the width of the brightness normalization interval.

[0074] like Figure 2 As shown, in step 2, the lightweight learning enhancement model is a lightweight global-local co-learning enhancement model, which includes a global color transformation branch and a depth-separable local residual recovery branch. The global color transformation branch predicts the third-order color transformation matrix and the three-dimensional color bias based on the global color statistics composed of the mean, standard deviation and multiple quantiles of each color channel of the sandstorm degradation image to be processed. The depth-separable local residual recovery branch receives the stitching result of the sandstorm degradation image to be processed and the image after global color transformation, and outputs the restricted local residual through a coding structure of at least two scales, a depth-separable residual structure with multiple holes, a decoding structure and skip connections.

[0075] The specific global color transformation branch is calculated based on the sand and dust degradation image to be processed. The mean, standard deviation, and multiple quantiles of each channel are used to obtain a 21-dimensional global color statistical vector z. This 21-dimensional global color statistical vector z is then passed sequentially through fully connected layers of dimensions 21 to 64, a GELU activation function, and a fully connected layer of dimensions 64 to 12, outputting nine matrix parameters and three bias parameters, which are a 3×3 color matrix and a 3-dimensional color bias. Subsequently, a global color transformation is performed on the 3×3 color matrix and the 3-dimensional color bias to obtain the globally color-transformed image. The calculation process is shown below:

[0076] ;

[0077] ;

[0078] ;

[0079] ;

[0080] In the formula, This is the 12-dimensional parameter vector output by the fully connected network. From Dimension mapping to A fully connected layer of dimension For GELU activation function, It is a 21-dimensional global color statistics vector. It is a 3×3 color matrix. 3D color bias. The first 9 terms of a 12-dimensional parameter vector. The last three terms of the 12-dimensional parameter vector; It is a 3-order identity matrix; , All are amplitude coefficients; It is the hyperbolic tangent function. For matrix rearrangement, This is the image after global color transformation; global color transformation is achieved through matrix multiplication and vector addition.

[0081] The specific depth-separable local residual recovery branch can be used to process the dust-degraded image. Image after global color transformation The input is concatenated along the channel dimension to obtain a six-channel input. This six-channel input is then transformed into a 24-channel input via a 3×3 convolution. The 24-channel input is then sequentially passed through two lightweight residual blocks of the first 24-channel input, a 48-channel downsampling block with a stride of 2, two lightweight residual blocks of the first 48-channel input, a 96-channel downsampling block with a stride of 2, a 96-channel dilated lightweight residual block with dilatancy rates of 1, 2, 4, and 2, a bilinear upsampling block skipping a connection with the 48-channel input, two lightweight residual blocks of the second 48-channel input, a bilinear upsampling block skipping a connection with the 24-channel input, two lightweight residual blocks of the second 24-channel input, and a three-channel convolution to obtain a restricted local residual. Specifically, the two lightweight residual blocks of the first 24-channel input are skipped between the bilinear upsampling block and the 48-channel input, and the two lightweight residual blocks of the first 48-channel input are skipped between the 96-channel dilated lightweight residual blocks with dilatancy rates of 1, 2, 4, and 2. The calculation process is shown below:

[0082] ;

[0083] In the formula, The output of the second 24-channel lightweight residual block and the three-channel convolution. For constrained local residuals, This is the residual amplitude coefficient; This is the hyperbolic tangent function, used to limit the range of residuals.

[0084] The globally color-transformed image is then added to and fused with the restricted local residuals to obtain the learned augmented image output by the lightweight learning augmentation model; the calculation process is shown below:

[0085] ;

[0086] In the formula, For constrained local residuals, This is the image after global color transformation. To be The value is truncated to Within the interval, Enhance images for learning.

[0087] The lightweight residual block includes a 3×3 depthwise convolution, a pointwise 1×1 expanded convolution, a GELU activation function, a pointwise 1×1 compressed convolution, learnable scaling, and residual summation. Let the input feature of the lightweight residual block be F, then the output of the lightweight residual block is calculated as follows:

[0088] ;

[0089] In the formula, F represents the input features of the lightweight residual block. The output characteristics of lightweight residual blocks, For learnable channel scaling parameters, To compress pointwise convolution, To extend pointwise convolution, For channel-by-channel multiplication, Let be a depthwise convolution with a dilation rate of d; in the above formula, the terminal addition indicates that the scaled convolution branch is added to the input feature as a residual, where 24, 48, and 96 channels and the number of each residual block are related to... Figure 2 Consistent; other channel widths or number of layers can be used while maintaining global color transformation, depth-separable local multi-scale residual recovery, and skip connection relationships.

[0090] Traditional image enhancement models perform spatially weighted white balance, yellowing removal, contrast enhancement, and detail enhancement based on a normalized dust curtain layer V and local activity D, resulting in a traditionally enhanced image. .

[0091] In step 3, both the learned enhanced image and the traditional enhanced image are separated to obtain the luminance and chrominance components of the learned enhanced image and the traditional enhanced image, respectively. The calculation process is as follows:

[0092] , ; , ;

[0093] In the formula, To learn augmented images, For traditional image enhancement, , These are the luminance and chrominance components of the image, respectively, for learning and enhancing the image. , These are the luminance and chrominance components of a traditional enhanced image, respectively. For brightness extraction operators, This is the chroma extraction operator.

[0094] Set the pixel position to x, and set the normalized RGB image to x. Brightness weight vector And remember The luminance component is then extracted as follows:

[0095] ;

[0096] In the formula, x is the pixel position. To perform a weighted summation of the three channels at pixel position x, The luminance extraction operator is used. The chrominance component extraction is shown below:

[0097] ;

[0098] In the formula, For the chromaticity extraction operator, the subtraction operation in the above formula is used to remove the luminance component common to the three channels from the RGB vector. The above equation satisfies That is, chromaticity with zero brightness.

[0099] The fused luminance is obtained by fusing the luminance component of the traditional enhanced image as the primary component with the luminance component of the learned enhanced image. The fused chrominance is obtained by fusing the chrominance component of the learned enhanced image as the primary component with the chrominance component of the traditional enhanced image. The calculation process is as follows:

[0100] ; ;

[0101] In the formula, To blend brightness, To blend colors, The fusion weights are used to determine the fusion brightness. The blending weights are the blending weights for the chromaticity; where ,and , A value of 0.22 can be taken; where ,and , The value can be 1.00; in the above formula, addition and coefficient multiplication represent pixel-wise linear weighting.

[0102] An initial fused image is obtained based on the fused luminance and fused chrominance components; the calculation process is shown below:

[0103] ;

[0104] In the formula, For the initial fused image, To be The value is truncated to Within the range, it is used to limit the output channel value; To copy the scalar brightness to the three RGB channels of the image, is the chromaticity amplitude coefficient, where , 0.90 is acceptable.

[0105] like Figure 3 As shown, in step 4, in the initial fused image Extracting fusion brightness and blending color Based on the dust curtain layer of the degraded dust image to be processed, the transmittance, atmospheric luminosity, and colored atmospheric light are estimated; the calculation process is shown below:

[0106] ;

[0107] ; ;

[0108] ;

[0109] In the formula, Atmospheric luminance, The colored atmospheric light after brightness normalization. The initial colored atmospheric light is obtained based on the channel quantile; Transmittance; This is a collection of pixels from a high-dust-coverage layer. To obtain the 95th quantile, To take the 90th quantile; A collection of pixels for a high dust curtain layer A set of pixels with a medium brightness not lower than a preset high percentile; This is the transpose of the RGB luminance weights, where T is the transpose operation; This is the lower limit of transmittance. To restore strength; To prevent the smallest positive number where the denominator is zero; the vector fraction is calculated separately for each RGB channel, where division is used to make the weighted brightness of colored atmospheric light equal to... Consistent.

[0110] A set of pixels with high dust cover was selected based on the dust cover layer, and atmospheric luminance was calculated based on this set. Zero-luminance chromaticity was then calculated based on the normalized colored atmospheric light. The calculations are shown below:

[0111] ;

[0112] ;

[0113] In the formula, To obtain the 85th quantile, For normalized dust curtain layer, The zero-brightness chromaticity of colored atmospheric light. This is the transpose of the RGB brightness weights. To copy scalar brightness to the three RGB channels of the image.

[0114] The recovery strength The calculation is as follows:

[0115] ;

[0116] ;

[0117] In the formula, The preset base strength; , All are preset weighting coefficients; This is expressed as the median; , These are the severity baseline and the scale, respectively. The severity of the dust storm is derived from the median value of the dust curtain layer. To be The value is truncated to Within the range.

[0118] Based on the estimated transmittance, atmospheric luminance, and colored atmospheric light, luminance recovery paths and chromaticity recovery paths are constructed. The luminance recovery paths are used to sequentially perform maximum darkening-constrained recovery, local and mesoscale contrast enhancement, adaptive hue adjustment, shadow enhancement, and specular protection on the fused luminance to obtain the target luminance. The fused luminance is then subjected to maximum darkening-constrained recovery; first, the physically recovered luminance is calculated, and then the maximum decrease relative to the fused luminance is limited. The calculation process is shown below:

[0119] ;

[0120] ;

[0121] In the formula, To physically restore brightness, To blend brightness, Atmospheric luminance, Transmittance, The blended brightness after applying the maximum darkening constraint. The maximum allowable decrease in brightness. , All are coefficients for the blending brightness adjustment darkening limit; max is the maximum value, and min is the minimum value. max and min together limit the minimum allowable value for restored brightness; where physical restored brightness is calculated. The fractional expression represents the inversion based on transmittance.

[0122] Local and mesoscale contrast enhancement and adaptive hue were performed on the fused brightness. First, the detail gain at both scales was adjusted using local activity and the dust curtain layer, and then local and mesoscale contrast enhancement and adaptive hue were applied. The calculation process is shown below:

[0123] ; ;

[0124] ; ;

[0125] ;

[0126] In the formula, For small-scale brightness details. This represents mid-scale brightness detail. for smooth radius, for smooth radius, The blended brightness after applying the maximum darkening constraint; This is a local activity gating, where D is the local activity. V represents the dust curtain layer gate; , , All are gated normalized parameters; This is the normalized upper limit for local activity, which can be 0.018; The activation threshold for the dust curtain layer gating can be 0.3; The normalized interval width for the transition from 0 to 1 for the dust curtain layer gate can be 0.65; when but ,when but ,when but . , All are overall detail gain coefficients. , All are basic gain weights. , All are gating incremental weights; To be The value is truncated to Within the interval; in the above formula, subtraction is used to extract residuals at different scales, and weighted summation is used to enhance local and mesoscale contrast.

[0127] Adaptive hue adjustment, shadow enhancement, and specular protection are applied to the blended brightness, with the median brightness before and after restoration set as follows: and The hue index is determined by the target median and then boosted by the shading operator. With highlight shoulder compression operator The target brightness is calculated; the calculation process is shown below:

[0128] ;

[0129] ;

[0130] ;

[0131] In the formula, , , These represent the median of the blended luminance, the median of the recovered luminance, and the median of the target luminance, respectively. The allowable decrease in median, For adaptive hue index, This is the lower limit of the adaptive hue index. For the highlight shoulder compression operator, Enhance the shadow operator. The operation is performed pixel-wise exponentiation, where V represents the dust curtain layer and D represents the local activity. Target brightness; To prevent the smallest positive number with a denominator of zero; To be The value is truncated to Within the interval; the shadow lifting operator The specular shoulder compression operator applies constrained compensation to non-black pixels below the shadow upper limit, and adjusts for dust curtain layer and local activity. Apply smooth shoulder compression to brightness levels near the upper limit.

[0132] The fused chromaticity is sequentially subjected to physical chromaticity restoration based on transmittance through the chromaticity restoration path, joint gating of dust curtain layer, saturation, brightness, and local activity, and removal of the orthographic projection of low-frequency color shift direction of the source dust to obtain candidate chromaticities. Physical chromaticity restoration based on transmittance is then performed on the fused chromaticity, and the physical correction amount relative to the fused chromaticity is calculated using the transmittance of the brightness restoration path. The calculation process is shown below:

[0133] ;

[0134] ;

[0135] In the formula, For chromaticity recovered based on a physical model, To blend colors, The zero-brightness chromaticity of colored atmospheric light. Transmittance, This is the physical correction amount relative to the blended chromaticity. This is the upper limit of the chromaticity correction amplitude. To be The value is truncated to Within the range.

[0136] A joint gating system for dust curtain layer, saturation, brightness, and local activity was applied to the fused chromaticity. Separate gating systems were calculated for dust curtain layer chromaticity correction, saturation chromaticity correction, brightness chromaticity correction, and local activity chromaticity correction. The calculation process is shown below:

[0137] ;

[0138] ;

[0139] ;

[0140] ;

[0141] ;

[0142] In the formula, For saturation and chromaticity correction gating, For luminance and chromaticity correction gating, For localized activity chromaticity correction gating, For color correction gating of the dust curtain layer, For final physical chromaticity correction gating; The difference between the maximum and minimum values ​​of the three RGB channels of the initial fused image. For the blended brightness, D represents the local activity. It is a dust curtain layer; The channel span threshold at which enhanced protection for highly saturated objects begins is preset to 0.1; This is the transition width, used to describe the channel span range where saturation protection changes from weak to strong; the default value is 0.28. The luminance threshold at which luminance gating begins to allow chroma correction is set to 0.04 by default. The brightness threshold ranges from 0 to 1, with a default value of 0.4. The maximum attenuation ratio for chromaticity correction in high-activity areas is preset to 0.45; This is the normalized upper bound for local activity, used to map local activity to [0, 1], with a default value of 0.05. This is the overall intensity, the overall intensity coefficient used for physical colorimetric correction, preset to 0.95; , These are the base weights and incremental weights for modulating the severity of the sandstorm. The default value is 0.35. The default value is 0.65; The dust severity of the entire image is obtained from the median of the dust curtain layer; it is a total image quantity, ranging from [0, 1]. The multiplication of each gate indicates that a larger correction intensity is retained only when all relevant conditions are met simultaneously.

[0143] To remove the low-frequency color shift of the source dust through orthogonal projection, the source colored atmospheric light is first estimated based on the brighter pixels of the high dust curtain layer in the original dust-degraded image. According to the source of colored atmospheric light The direction of unit low-frequency color shift is calculated, and then the low-frequency orthographic projection and low-frequency inverse color shift correction amounts are calculated sequentially; the calculation process is shown below:

[0144] ;

[0145] ;

[0146] ;

[0147] In the formula, For the direction of low-frequency color shift, To blend the low-frequency orthographic projection of chromaticity in this direction, The chromaticity smoothing radius is... To blend color transpose; To obtain the maximum value, only the forward projection is retained; This is the upper limit of the chromaticity correction amplitude; This is the low-frequency reverse color shift correction amount. It is a 2-norm; To prevent the smallest positive number with a denominator of zero; To be The value is truncated to Within the range.

[0148] Low-frequency reverse color shift correction amount Apply saturation and chromaticity correction gating Brightness and chromaticity correction gating Local activity colorimetric correction gating Dust curtain layer color correction gating , recorded as joint gating Specifically, saturation and chromaticity correction gating is employed. Brightness and chromaticity correction gating Local activity colorimetric correction gating Dust curtain layer color correction gating Performing independent calculations for the same pixel location and then multiplying the results of each gating is called joint gating; this is equivalent to gating saturation and chroma correction. Brightness and chromaticity correction gating Local activity colorimetric correction gating Dust curtain layer color correction gating Correction amount for low-frequency reverse color shift Perform independent calculations to obtain four gating calculation results; multiply the four gating calculation results together to obtain... .

[0149] Based on fused color Low-frequency reverse color shift correction amount Physical correction amount relative to blended chromaticity and final physical colorimetric correction gating The candidate chromaticities were calculated as follows:

[0150] ;

[0151] In the formula, To blend colors; For joint gating, it is represented as the amount of low-frequency inverse color shift correction. Apply saturation and chromaticity correction gating Brightness and chromaticity correction gating Local activity colorimetric correction gating Dust curtain layer color correction gating ; This is the low-frequency reverse color shift correction amount. This is the physical correction amount relative to the blended chromaticity. For final physical chromaticity correction gating; Candidate chromaticity; This represents the amount of physical correction relative to the blended chroma after gating. This represents the low-frequency inverse color shift correction amount after gating; the addition in the above formula represents the superposition of two types of correction amounts on the fused chromaticity.

[0152] Based on the obtained target luminance and candidate chrominance, RGB displayable domain synthesis is performed while preserving luminance to obtain a jointly restored image; the candidate chrominance is projected as a zero-luminance chrominance, and the calculation process is as follows:

[0153] ;

[0154] In the formula, As candidate chromaticity, To copy scalar brightness to the three RGB channels of the image.

[0155] The maximum permissible chromaticity ratio for each channel without crossing the RGB boundaries is calculated using zero-brightness chromaticity and target brightness. Based on the maximum chromaticity ratio of each channel, the minimum feasible ratio for all three RGB channels is calculated. The calculation process is as follows:

[0156] ;

[0157] ;

[0158] In the formula, The maximum allowable chroma ratio for the c-th channel. For target brightness, The zero-luminance chromaticity of the c-th channel, For the minimum feasible proportion, This indicates that the zero chromaticity component has no scale restriction; the piecewise operation selects the upper bound based on the positive, negative, or zero of the chromaticity component, and nested min is used to obtain the strictest channel constraint.

[0159] The jointly restored image is calculated based on zero-luminance chromaticity, target luminance, and minimum feasible ratio; the calculation process is shown below:

[0160] ;

[0161] In the formula, To copy the scalar brightness to the three RGB channels of the image, For target brightness, For the minimum feasible proportion, Zero luminance chromaticity To be The value is truncated to Within the range; For joint image restoration; all three channels use the same minimum feasible scale. This allows the chromaticity direction to be maintained while restricting channel crossings.

[0162] In step 5, neutral pixels in the image are jointly recovered by filtering pixels with low relative chromaticity and brightness within a preset non-extremely dark and non-extremely bright range. The neutral pixel region is determined based on the filtered neutral pixels; the calculation is as follows:

[0163] ;

[0164] ;

[0165] In the formula, Relative chromaticity per pixel For a set of neutral pixels, To prevent the smallest positive number with a denominator of zero; This is the 35th quantile of relative chromaticity, which is low relative chromaticity; , These are the 15th and 92nd percentiles of the brightness, respectively, indicating that the brightness falls within the preset non-extremely dark and non-extremely bright range. , These are the effective lower limit and effective upper limit values ​​for the c-th channel, respectively; the multiple inequalities in the set of neutral pixels represent the conditions that candidate pixels must simultaneously satisfy low relative chromaticity, medium brightness, and non-saturation.

[0166] The constrained color gain for each channel is calculated based on the offset between the channel representative value and the channel center value of the neutral pixel region. First, the RGB median vector of the neutral pixel and the center values ​​of the three channels are calculated. Then, the color cast score of the jointly reconstructed image is calculated based on the RGB median vector of the neutral pixel and the center values ​​of the three channels, as shown below:

[0167] ;

[0168] ;

[0169] ;

[0170] In the formula, This is the RGB median vector of the neutral pixel, which is the channel representative value or neutral anchor point of the neutral pixel region; The center value of the three channels. To take the median of each channel within the set of neutral pixels, To score the color cast of the jointly restored image, , , These are the R channel components, G channel components, and B channel components, respectively. The L2 norm represents the Euclidean distance from the neutral anchor point to the grayscale axis; To copy scalar brightness to the three RGB channels of the image.

[0171] The feedback strength is determined based on the consistency of the color cast score, neutral anchor point offset, and overall image mean offset of the jointly restored image. Specifically, the color cast score of the input image is linearly normalized between the normal threshold of 0.04 and the full intensity threshold of 0.18. The severity of color cast was obtained. Calculate the cosine similarity between the input image and the current image based on the neutral anchor point offset direction, mapping values ​​from 0.05 to 0.70 to... , obtain the direction confidence ; among them when At this time, a conservative correction component of no more than 0.35 can be introduced to avoid severe color casts from being completely lost due to unstable orientation estimation. Calculate the cosine similarity between the current neutral anchor point offset and the RGB mean offset of the entire image, and then... Mapping to 1 Statistical directional consistency is obtained Based on the severity of color cast directional confidence Consistency with statistical direction Calculate the feedback strength; feedback strength The calculation process is as follows:

[0172] ;

[0173] In the formula, The severity of residual color deviation, , representing the confidence level for the color bias direction. To ensure statistical consistency, 0.9 is the overall feedback coefficient, and 0.45 and 0.55 are the weights of the fixed basic term and the statistical consistency term, respectively. If the magnitude of any directional vector is too small, a default consistency is used to avoid division by zero and unstable feedback.

[0174] The initial limited gain of each channel is calculated based on the center value of each channel and the center values ​​of the three channels. The initial limited gain of each channel and the feedback strength are then used as the basis for the calculation. The constrained color gain for each channel is calculated as follows:

[0175] ;

[0176] ;

[0177] In the formula, Let be the initial limited gain of the c-th channel. The limited color gain for the c-th channel; For feedback strength, Maximum allowable gain; To prevent the smallest positive number with a denominator of zero, To be The value is truncated to Within the interval; the exponentiation operation in the above formula is used to continuously shrink the gain to near unity gain.

[0178] Simultaneously, in order to perform spatial blending calculations of the protection map for highly saturated dark objects, corrected spatial weights are calculated based on the protection map for highly saturated dark objects; the calculation process is shown below:

[0179] ;

[0180] ;

[0181] In the formula, This is a color protection image for highly saturated dark objects, also known as an object color protection image. To preset the protection weight, To be The value is truncated to Within the range; To jointly restore the span of the three RGB channels of the image; The channel span threshold at which high saturation protection begins to function is preset to 0.1; This is the transition width, used to describe the channel span range from 0 to 1 for high saturation protection; the default value is 0.3. For target brightness, The upper limit of brightness for dark protection is set to 0.82 by default; This is the transition width, used to describe the brightness range as dark protection decreases, with a default value of 0.55. When a pixel simultaneously has a large channel span and low brightness... Approaching 1 When the value approaches 0.28, the color feedback weakens significantly; if any condition is not met, Approaching 0 Approximately 1. The multiplication of the two truncation results in the above formula is used to simultaneously detect high saturation and dark areas.

[0182] Spatial blending of the restricted color gain is performed using the object color preservation map. Restricted color feedback and weak red channel balancing are then applied to the jointly restored image to obtain a locally chromaticity corrected image. The calculation for restricted color feedback on the jointly restored image is shown below:

[0183] ;

[0184] ;

[0185] In the formula, For the joint reconstruction image after limited color feedback, This is the intermediate image after spatial weighted correction. To copy the scalar brightness to the three RGB channels of the image, To jointly restore the image, To correct spatial weights, This is channel-by-channel multiplication; Let be the channel gain vector, where These represent the limited color gain for each channel; To be The value is truncated to Within the interval; the square brackets in the formula for calculating the local chromaticity corrected image multiplied by the difference between the brightness of the jointly reconstructed image and the brightness of the intermediate image. Three channels are then added back to compensate for the brightness changes before and after correction, which is to compensate for the brightness difference before and after correction.

[0186] Jointly restored image after limited color feedback After balancing the weak red channel, a locally chromaticity corrected image is obtained. That is The weak red channel balancing method determines a red channel gain of less than or equal to 1 based on the excess of the red channel relative to the green and blue channels in the neutral anchor point and the overall image mean, and uses the same object color protection map and the brightness difference before and after compensation correction. Specifically, the excess of the red channel relative to the green and blue channel mean in the neutral anchor point and the excess of the red channel mean relative to the green and blue channel mean in the overall image are calculated separately; the excess of the red channel relative to the green and blue channel mean in the neutral anchor point is subtracted by an allowable threshold of 0.012, retaining only positive values; the excess of the red channel mean relative to the green and blue channel mean in the overall image is subtracted by an allowable threshold of 0.02, retaining only positive values; then, the excess of the neutral anchor point red channel relative to the green and blue channel mean and the excess of the red channel mean relative to the green and blue channel mean in the overall image are weighted and fused using coefficients of 0.72 and 0.28 to obtain the comprehensive red excess. Using 0.75 as the red channel balance strength, the red channel gain is calculated based on the overall red excess; the neutral anchor red channel value in the denominator is no less than 0.15, and the gain is limited to... The green and blue channel gains are kept at 1, with 0.15 used to prevent the denominator in dark areas from being too small, and 0.9 serving as the lower limit for the red channel gain. A spatial weighting correction is applied based on the red channel gain. Protecting images of highly saturated dark objects The process involves attenuating and correcting highly saturated dark regions, and compensating for the brightness difference before and after channel correction to maintain consistent output brightness. The calculation formula is the same as that for restricted color feedback, except that the restricted color gain is replaced with the red channel gain. The final result is a locally chromaticity corrected image. .

[0187] like Figure 4 As shown, in step 6, the image without local chromaticity correction is used as the reference image. ,in To maintain the brightness composition calculation, For target brightness, To fuse chromaticity; the resulting locally chromaticity corrected image is... As candidate images, the color cast scores of the reference image and the candidate image are calculated separately. When the color cast score of the candidate image is less than or equal to the sum of the color cast score and the score tolerance of the reference image, neutral pixels in the candidate image are selected. Based on the RGB statistical offset of the neutral pixels, restricted color feedback and weak red channel balancing are performed on the candidate image to obtain the main path image. The formula for determining whether the color cast score of the candidate image is less than or equal to the sum of the color cast score and the score tolerance of the reference image is as follows:

[0188] ;

[0189] In the formula, The candidate image for the k-th intensity regression; The initial candidate image is the calculated local chromaticity corrected image. As the reference image, For scoring tolerance; For color cast scoring calculation; Score the color cast of the candidate images. The color cast score is given to the baseline image.

[0190] When the color cast score of the candidate image is greater than the sum of the color cast score and the score tolerance of the reference image, the local chromaticity correction amount is successively reduced according to the preset attenuation coefficient, and it is determined whether the current intensity backoff count has reached the maximum count; the formula for the color cast score of the candidate image being greater than the sum of the color cast score and the score tolerance of the reference image is as follows:

[0191] ;

[0192] In the formula, Score the color cast of the candidate images. The color cast score is given to the baseline image.

[0193] The calculation process for successively reducing the local chromaticity correction amount according to the preset attenuation coefficient is shown below:

[0194] , ;

[0195] In the formula, This represents the local chromaticity correction amount for the k-th intensity regression. This is the initial local chromaticity correction amount. This is the attenuation coefficient, typically taken as 0.5; To perform an exponential operation on the attenuation coefficient, representing a gradual decrease with each number of intensity retreats; simultaneously Update the value of k.

[0196] If the maximum number of intensity downtime steps is not reached, the candidate image is corrected using a progressively decreasing local chromaticity correction amount, and a new candidate image is generated; the calculation process is shown below:

[0197] ;

[0198] In the formula, To maintain the brightness composition calculation; For local chromaticity correction images The brightness is the same as the target brightness. Same; For local chromaticity correction images chromaticity, The local chromaticity correction amount for the kth intensity regression This represents the total amount of local colorimetric correction.

[0199] If the number of intensity rollbacks reaches the maximum, then If the reference image is not found, correction is performed and the reference image is rejected. Neutral pixels in the reference image are selected, and restricted color feedback and weak red channel balancing are performed on the reference image based on the RGB statistical offset of the neutral pixels to obtain the alternative path image.

[0200] Based on the obtained main path image and backup path image, the color cast score of the main path image and backup path image is calculated respectively. The color cast score of the main path image is compared with a preset trigger threshold. When the color cast score of the path image is less than the preset trigger threshold, the main path image is output as the final target enhancement image.

[0201] When the color cast score of the main path image is greater than or equal to a preset trigger threshold, the sum of the color cast score and the improvement amount of the color cast score of the backup path image is compared with a preset minimum improvement threshold, where the minimum improvement threshold is the color cast score of the main path image. When the sum of the color cast score and the improvement amount of the color cast score of the backup path image is less than the preset minimum improvement threshold, the backup path image is output as the final target enhancement image. When the sum of the color cast score and the improvement amount of the color cast score of the backup path image is greater than or equal to the preset minimum improvement threshold, the main path image is output as the final target enhancement image. The calculation process is as follows:

[0202] ;

[0203] In the formula, Enhance the image to achieve the final goal. Image for alternative paths, Main path image; The color cast score of the main path image. Score the color cast of the alternative path image. The improvement amount for color cast rating; The preset trigger threshold for the main path; For logical AND operation, This is a logical NOT operation.

[0204] The backup path image does not need to be generated in parallel with the main path image; the main path image can be calculated first; only if the color cast score of the main path image is... Reaching the trigger threshold Then, calculate the alternative path image and its color deviation score. Only when the color cast improvement of the backup path image relative to the main path image exceeds the minimum improvement amount. Only when the condition is met will the alternative path image be output; otherwise, the main path image will still be output. In the current implementation, , If the first-level protection has set the local correction ratio to 0, the main path image and the backup path image may be the same or approximately the same. and The color cast scores are for the main path image and the backup path image, respectively. The preset trigger threshold for the main path to be calculated to initiate the backup path image. The minimum color cast score improvement required to select an alternative path image.

[0205] like Figure 5 As shown, a dust image enhancement system is included to implement a dust image enhancement method based on luminance-color separation and color cast backoff protection. It comprises: a feature estimation module for calculating global color statistics, dust curtain layer, and local activity of the dust degradation image; a dual-branch enhancement module for enhancing the dust degradation image using a lightweight learning enhancement model and a traditional image enhancement model, respectively, to obtain a learned enhanced image and a traditional enhanced image; a luminance-color separation and fusion module for separating luminance and chrominance of the learned enhanced image and the traditional enhanced image and fusing them to obtain an initial fused image; a luminance-chrominance joint recovery module for jointly recovering the fused luminance and fused chrominance in the initial fused image using transmittance, atmospheric light intensity, and colored atmospheric light, to obtain a joint recovered image; a color feedback module for performing restricted color feedback and weak red channel balancing on the joint recovered image based on the neutral pixel RGB statistical offset, to obtain a locally chrominance corrected image; and a protection decision module for performing intensity backoff or correction rejection of the local chrominance correction based on the color cast score before and after local chrominance correction, and selecting the optimal path image as the final target enhanced image.

[0206] When outputting the enhanced target image, the system can simultaneously generate a quality report. The quality report may include at least one of the following: input and output file identifiers, color cast scores before and after correction, overall RGB mean, neutral anchor RGB value, mean offset, orientation consistency, color feedback intensity, number of color feedback iterations, cumulative channel gain, red channel gain, median of the dust curtain layer, mean and minimum transmittance, chromaticity correction backoff ratio, luminance statistics, and alternate path selection status. This quality report is used to record the processing status and is not a necessary limitation for the formation of the enhanced target image.

[0207] The system includes a storage medium containing an executable program, which, when executed by a processor, enables a dust image enhancement method based on luminance-color separation and color shift fallback protection.

[0208] Example: First, feature analysis is performed on the selected dust degradation image to obtain the feature distribution map of the dust degradation image, such as... Figure 6 As shown; Figure 6 The algorithm employs a four-row, multi-column layout, with each column corresponding to a sandstorm scene. The first row represents the input sandstorm degraded image, the second row represents the normalized histograms of the red, green, and blue RGB channels of the input sandstorm degraded image, the third row represents the estimated sandstorm curtain layer of the input sandstorm degraded image, with grayscale intensity representing the degree of sandstorm occlusion or colored air light influence at each pixel location, and brighter locations corresponding to higher estimated sandstorm curtain layer values. The fourth row represents the estimated local activity of the input sandstorm degraded image, with grayscale intensity representing the degree of brightness change relative to a local smooth base, and edge and texture regions typically exhibiting higher activity. The sandstorm curtain layer and local activity are used for transmittance estimation, brightness restoration, chromaticity joint gating, and local contrast enhancement, respectively.

[0209] The following degradation characteristics can be observed from the normalized histograms of the red, green, and blue RGB channels of the second-row input image of a degraded sandstorm: The peak positions, distribution centers, or main regions of the normalized histograms of the red, green, and blue RGB channels exhibit varying degrees of lateral misalignment; as the separation distance between the three channels increases, the channel energy imbalance intensifies, and the overall tone shift of the image is usually more pronounced; when comparing the channel mean or histogram distribution centers of typical sandstorm images, the three channels typically show an intensity order of red channel first, followed by green channel, and then blue channel, i.e., μR>μG>μB; this order reflects the dominance of red and green components relative to blue components, which is an important statistical manifestation of the image's yellowish and reddish appearance; most pixels in each channel are concentrated in a relatively limited grayscale range, with fewer pixels at the low and high ends, resulting in a reduced usable dynamic range; this distribution pattern corresponds to the overall compression of brightness levels, insufficient difference in brightness, and decreased contrast in sandstorm images.

[0210] A dust image enhancement method based on brightness and color separation and color shift protection is used to enhance the degraded dust image to be processed, resulting in the enhanced dust image image, as shown below. Figure 7 As shown; Figure 7 The method employs a four-row, multi-column layout, with each column corresponding to the same sandstorm scene: the first row is the input sandstorm degradation image, showing the original red-yellow color cast, low contrast, and hazy degradation; the second row is the learned enhancement image, showing the result of the lightweight learning model after global color transformation and local residual recovery; the third row is the traditional enhancement image, showing the result of the traditional branch after white balance, yellow removal, brightness, and detail enhancement; the fourth row is the final target enhancement image, showing the output after brightness-color separation and fusion, brightness-chroma joint recovery, color feedback, and two-level backoff protection. By comparing each row and column, the differences between the learned branch and the traditional branch in color recovery and brightness structure recovery can be qualitatively explained, as well as the fusion effect of the final target enhancement image on the two types of information. The low-frequency red-yellow shift in the final target enhancement image is suppressed, scene hierarchy and local structure are restored, and dark areas, highlights, and areas with inherent colors are protected; it can be seen that the method of this invention can balance sandstorm layer removal, color recovery, inherent color protection, and stability in real-world scene processing. Figure 7 This is a qualitative description only, and does not include accuracy rates, method rankings, or other quantitative experimental conclusions.

[0211] Meanwhile, in this invention, both the learned enhanced image and the traditional enhanced image undergo luminance and chrominance separation operations, which are not limited to the RGB weighted luminance form, but can also be implemented in YCbCr, Lab, IPT, or other color spaces that can separate luminance and chrominance information; the lightweight learning enhancement model can be replaced by other learnable models that can generate color transformations and local residuals based on global color statistics and local image features; traditional image enhancement can use white balance, color attenuation, local contrast, guided filtering, or a combination thereof; dust curtain layer estimation can be achieved by color attenuation, local extrema, dark channel, guided filtering, or a learning model; color cast scoring can use neutral anchor point channel offset, chrominance center distance, grayscale world deviation, or a combination thereof, as long as it can compare the degree of color cast before and after correction.

[0212] The above description is merely a preferred embodiment of the present invention. Those skilled in the art can make several modifications and optimizations based on the above disclosure without departing from the basic principles described above. These modifications and optimizations should be considered within the scope of protection as understood by the present invention.

Claims

1. A method for enhancing dust images based on luminance-color separation and color cast protection, characterized in that: Includes the following steps: Step 1: Obtain the sand and dust degradation image to be processed, and calculate the global color statistics, sand and dust curtain layer, and local activity of the sand and dust degradation image to be processed; Step 2: Enhance the sandstorm-degraded image to be processed using a lightweight learning enhancement model and a traditional image enhancement model, respectively, to obtain the learned enhanced image and the traditional enhanced image. Step 3: Separate the luminance and chrominance of the learned enhanced image and the traditional enhanced image, and fuse them with the luminance component of the traditional enhanced image as the main component and the chrominance component of the learned enhanced image as the main component to obtain the initial fused image; Step 4: Extract the fused brightness and fused chromaticity from the initial fused image. Estimate the transmittance, atmospheric brightness, and colored atmospheric light based on the dust curtain layer of the dust degradation image to be processed. Construct brightness recovery path and chromaticity recovery path based on the estimated transmittance, atmospheric brightness, and colored atmospheric light. The fused brightness and fused chromaticity are recovered by using the brightness recovery path and the chromaticity recovery path respectively, and the target brightness and candidate chromaticity are obtained respectively. Based on the obtained target luminance and candidate chrominance, perform RGB displayable domain synthesis while preserving luminance to obtain a jointly restored image; Step 5: Filter the neutral pixels in the jointly restored image, and perform restricted color feedback and weak red channel balancing on the jointly restored image based on the RGB statistical offset of the neutral pixels to obtain a local chromaticity corrected image; Step 6: Calculate the color cast score before and after local color correction, and perform intensity rollback or correction rejection of the local color correction amount based on the color cast score before and after local color correction; and when the triggering condition for final path selection is met, obtain and compare the color cast scores of the main path image and the backup path image, and select the path image with better color cast score and meeting the preset improvement conditions as the final target enhancement image.

2. The sand and dust image enhancement method based on luminance-color separation and color shift protection according to claim 1, characterized in that: In step 1, the global color statistics, dust curtain layer, and local activity of the degraded dust image to be processed are calculated; the calculation process is as follows: ； ； ； ； In the formula, This is a global color statistics vector. For local activity, A normalized dust curtain layer; , The mean and standard deviation of the dust degradation image in channel c are respectively. Channel images in dust degradation images The q-th quantile, where Q is the set of quantile levels; ; radius Smooth operation, The input brightness is used for images degraded by sandstorms. This is an absolute value operation; , These are the response values ​​at the x-pixel and y-pixel positions, respectively, within the same dust curtain layer. This is the set of pixels in the degraded image of the sandstorm to be processed; To Take the 97.5th quantile. To prevent the smallest positive number with a denominator of zero, To be The value is truncated to Within the range; Large-scale smooth radius Smooth operation, This represents the positive excess of the red-yellow channel relative to the blue channel in the image of dust degradation. To define the flatness weight, For low saturation weights, This is the brightness weight.

3. The dust image enhancement method based on luminance-color separation and color shift protection according to claim 1, characterized in that: In step 3, both the learned enhanced image and the traditional enhanced image are separated to obtain the luminance and chrominance components of the learned enhanced image and the traditional enhanced image, respectively. The calculation process is as follows: ， ； ， ； In the formula, To learn augmented images, For traditional image enhancement, , These are the luminance and chrominance components of the image, respectively, for learning and enhancing the image. , These are the luminance and chrominance components of a traditional enhanced image, respectively. For brightness extraction operators, For chroma extraction operators; The fused luminance is obtained by fusing the luminance component of the traditional enhanced image with the luminance component of the learned enhanced image, and the fused chrominance is obtained by fusing the chrominance component of the learned enhanced image with the chrominance component of the traditional enhanced image. The calculation process is as follows: ； ； In the formula, To blend brightness, To blend colors, The fusion weights are used to determine the fusion brightness. The fusion weights are used to fuse the chroma; an initial fused image is obtained based on the fused luminance and fused chroma components; the calculation process is shown below: ； In the formula, For the initial fused image, To be The value is truncated to Within the interval, To copy the scalar brightness to the three RGB channels of the image, This is the chromaticity amplitude coefficient.

4. The dust image enhancement method based on luminance-color separation and color shift protection according to claim 1, characterized in that: In step 4, in the initial fused image Extracting fusion brightness and blending color Based on the dust curtain layer of the degraded dust image to be processed, the transmittance, atmospheric luminosity, and colored atmospheric light are estimated; the calculation process is shown below: ； ； ； ； In the formula, Atmospheric luminance, The colored atmospheric light after brightness normalization. Transmittance; This is a collection of pixels from a high-dust-coverage layer. To obtain the 95th quantile, To take the 90th quantile; A collection of pixels for a high dust curtain layer A set of pixels with a medium brightness not lower than a preset high percentile; This is the transpose of the RGB luminance weights, where T is the transpose operation; This is the lower limit of transmittance. To restore strength; Based on the estimated transmittance, atmospheric luminance, and colored atmospheric light, a luminance recovery path and a chromaticity recovery path are constructed. The target brightness is obtained by sequentially performing maximum darkening constraint restoration, local and mesoscale contrast enhancement, adaptive hue, shadow enhancement, and highlight protection on the fused brightness through the brightness restoration path. Candidate chromaticities are obtained by sequentially performing physical chromaticity restoration based on transmittance, joint gating of dust curtain layer, saturation, brightness and local activity, and removal of the orthogonal projection of low-frequency color bias of source dust through the chromaticity restoration path.

5. The dust image enhancement method based on luminance-color separation and color shift protection according to claim 1, characterized in that: Based on the obtained target luminance and candidate chrominance, RGB displayable domain synthesis is performed while preserving luminance to obtain a jointly restored image. The candidate chrominance is projected as zero-luminance chrominance, and the maximum allowable chrominance ratio for each channel without crossing the RGB boundaries is calculated using the zero-luminance chrominance and target luminance. Based on the maximum chrominance ratio of each channel, the minimum feasible ratio for all three RGB channels is calculated. The jointly restored image is then calculated based on the zero-luminance chrominance, target luminance, and minimum feasible ratio. The calculation process is shown below: ； In the formula, To copy the scalar brightness to the three RGB channels of the image, For target brightness, For the minimum feasible proportion, Zero luminance chromaticity To be The value is truncated to Within the range.

6. The sand and dust image enhancement method based on luminance-color separation and color shift protection according to claim 1, characterized in that: In step 5, neutral pixels in the image are jointly recovered by screening pixels with low relative chromaticity and brightness in a preset non-dark and non-bright range, and the neutral pixel region is determined based on the screened neutral pixels. The restricted color gain for each channel is calculated based on the offset between the channel representative value and the channel center value in the neutral pixel region. Spatial blending of the restricted color gain is performed using the object color preservation map. Restricted color feedback and weak red channel balancing are then applied to the jointly restored image to obtain a locally chromaticity corrected image. The calculation of restricted color feedback for the jointly restored image is shown below: ； ； In the formula, This is the intermediate image after spatial weighted correction. To copy the scalar brightness to the three RGB channels of the image, To jointly restore the image, To correct spatial weights, This is channel-by-channel multiplication; Let be the channel gain vector, where These represent the limited color gain for each channel.

7. The dust image enhancement method based on luminance-color separation and color shift protection according to claim 1, characterized in that: In step 6, the image without local color correction is used as the reference image, and the obtained locally color-corrected image is used as the candidate image. The color cast score of the reference image and the candidate image are calculated respectively. When the color cast score of the candidate image is less than or equal to the sum of the color cast score and the score tolerance of the reference image, the neutral pixels in the candidate image are selected. Based on the RGB statistical offset of the neutral pixels, the candidate image is subjected to restricted color feedback and weak red channel balancing to obtain the main path image. When the color cast score of the candidate image is greater than the sum of the color cast score and the score tolerance of the reference image, the local chromaticity correction amount is gradually reduced according to the preset attenuation coefficient, and it is determined whether the current intensity backoff number has reached the maximum number; if the intensity backoff number has not reached the maximum number, the candidate image is corrected by the gradually decreasing local chromaticity correction amount, and a new candidate image is generated; if the intensity backoff number has reached the maximum number, the reference image is rejected after correction; neutral pixels in the reference image are selected, and restricted color feedback and weak red channel balancing are performed on the reference image according to the RGB statistical offset of the neutral pixels to obtain the backup path image.

8. The sand and dust image enhancement method based on luminance-color separation and color shift protection according to claim 7, characterized in that: Based on the obtained main path image and backup path image, the color deviation score of the main path image and backup path image is calculated respectively. The color deviation score of the main path image is compared with the preset trigger threshold. When the color deviation score of the path image is less than the preset trigger threshold, the main path image is output as the final target enhancement image. When the color cast score of the main path image is greater than or equal to the preset trigger threshold, the sum of the color cast score of the backup path image and the improvement amount of the color cast score is compared with the preset minimum improvement threshold. When the sum of the color cast score and the color cast improvement amount of the backup path image is less than the preset minimum improvement threshold, the backup path image is output as the final target enhancement image; when the sum of the color cast score and the color cast improvement amount of the backup path image is greater than or equal to the preset minimum improvement threshold, the main path image is output as the final target enhancement image.

9. The sand and dust image enhancement method based on luminance-color separation and color shift protection according to claim 1, characterized in that: Including a dust image enhancement system for implementing a dust image enhancement method based on brightness-color separation and color shift back protection; It includes: a feature estimation module for calculating global color statistics, dust curtain layer, and local activity of the dust degradation image; a dual-branch enhancement module for enhancing the dust degradation image using a lightweight learning enhancement model and a traditional image enhancement model, respectively, to obtain a learned enhanced image and a traditional enhanced image; a luminance-color separation and fusion module for separating luminance and chrominance of the learned enhanced image and the traditional enhanced image and fusing them to obtain an initial fused image; a luminance-chrominance joint recovery module for jointly recovering the fused luminance and fused chrominance in the initial fused image using transmittance, atmospheric light luminance, and colored atmospheric light, to obtain a joint recovered image; a color feedback module for performing restricted color feedback and weak red channel balancing on the joint recovered image based on the neutral pixel RGB statistical offset, to obtain a locally chrominance corrected image; and a protection decision module for performing intensity backoff or correction rejection of the local chrominance correction based on the color shift score before and after local chrominance correction, and selecting the optimal path image as the final target enhanced image.

10. The dust image enhancement method based on luminance-color separation and color shift protection according to claim 1, characterized in that: The system includes a storage medium containing an executable program, which, when executed by a processor, enables a dust image enhancement method based on luminance-color separation and color shift fallback protection.

Images