An image defogging recovery method and system based on human eye visual non-linear mapping

Abstract

The present application relates to the field of computer vision and image processing technology, and particularly relates to an image defogging recovery method and system based on human eye vision non-linear mapping. The method first performs white balance correction on the foggy image and obtains a transmittance map, and then converts the image to a CIELab color space; in chroma recovery, the hue angle is kept basically unchanged, and only the ab plane chroma vector module is non-linearly recovered according to the transmittance; in brightness recovery, a brightness dark reference determined by the fog medium backscattering is introduced, and the brightness difference component after deducting the dark reference is non-linearly recovered; finally, the recovered brightness component and chroma component are converted back to the RGB space to output the defogging image. The present application has both physical interpretability and visual perception consistency, can effectively improve the problems such as detail loss, color distortion and brightness recovery instability in the thick fog area, and improve the naturalness, clarity and visual comfort of the defogging image.

Description

Technical Field

[0001] This invention relates to the field of computer vision and image processing technology, and in particular to an image dehazing and restoration method and system based on nonlinear mapping of human vision. Background Technology

[0002] Image dehazing is an important research direction in the fields of computer vision and digital image processing. Its main purpose is to reduce or remove image degradation caused by atmospheric scattering media such as fog, haze, water vapor, and smoke, thereby improving image recognizability, color fidelity, contrast, and the reliability of subsequent machine vision tasks. In actual imaging, the light reflected from scene targets is absorbed and scattered by air molecules and suspended particles along its path to the imaging device. This results in the imaging result containing both the direct attenuation component of the target and the atmospheric light component formed by backscattering from the medium. Ultimately, this manifests as decreased contrast of distant targets, an overall grayish-white appearance, obscured details in dark areas, distortion in bright areas, and overall color shift. This type of degradation not only affects the human visual experience but also significantly reduces the accuracy of algorithms in applications such as target detection, semantic segmentation, autonomous driving environmental perception, security monitoring, and remote sensing interpretation. Therefore, high-quality restoration of foggy images has long been a key research focus in this field.

[0003] Existing image dehazing methods can be broadly categorized into three types: The first type is restoration methods based on physical scattering models. These methods typically rely on atmospheric scattering models to restore degraded images by estimating parameters such as atmospheric light, transmittance, or scene depth. The second type is dehazing methods based on image enhancement. These methods focus more on improving visual visibility, using techniques such as histogram equalization, Retinex, multi-scale enhancement, and color correction to improve image appearance. The third type is data-driven methods based on deep learning, which achieve restoration by constructing an end-to-end mapping model between foggy and fog-free images. In general, physical model-based methods offer some interpretability but are highly sensitive to the accuracy of transmittance and atmospheric light estimation. Enhancement-based methods are relatively simple to implement but often lack strict physical constraints, making them prone to color shift, over-enhancement, or detail artifacts. Deep learning-based methods can achieve good results on specific datasets but typically depend on the distribution of training data and have weak interpretability; their generalization stability in real-world, complex foggy scenes remains insufficient.

[0004] For example, Chinese patent document CN104182943B discloses a single-image dehazing method that integrates human visual characteristics. This method is still based on the classic physical restoration framework for single-image dehazing. First, it acquires the dark channel image of the input haze image. Then, it uses dark channel priors to estimate atmospheric illumination values, calculates initial atmospheric transmission values, and refines them using a guided filter. Finally, it combines the estimated atmospheric illumination values ​​with transmittance to achieve clear image restoration. Simultaneously, this patent document introduces a segmentation strategy for saturated and unsaturated regions and a local adjustment concept based on the perceptible threshold of human vision, aiming to reduce halos and noise while dehazing. This patent document demonstrates the practical significance of introducing visual perception constraints into single-image dehazing and also shows that traditional methods relying solely on the dark channel and linear inversion are prone to unstable restoration in areas such as the sky, saturated bright areas, and dense fog. However, further analysis of the patent document's technical approach reveals that its core remains focused on dark channel priors, atmospheric light estimation, transmittance estimation, and linear inversion based on atmospheric scattering models. The so-called introduction of human visual characteristics is mainly reflected in the selection of adjustment parameters for zonal adjustment, threshold setting, and local brightness changes. In other words, this existing technology does not redefine the dehazing recovery variables from the perspective of color perception mechanisms, nor does it migrate the dehazing process from the sensor-response-dominated RGB space to a more visually uniform perceptual color space for modeling. In particular, regarding the phenomenon of asynchronous chromaticity attenuation and brightness changes commonly found in foggy images, the patent document does not establish independent recovery mechanisms for chromaticity and brightness, but instead processes them uniformly within the traditional linear atmospheric scattering inversion framework. Therefore, in areas with strong fog, sky areas, low-saturation areas, and near white or gray targets, problems such as darker images after dehazing, unnatural colors, amplified local noise, and insufficient visual comfort are still likely to occur.

[0005] Chinese patent document CN104537615A recognizes that direct processing in the RGB color space easily leads to color distortion. Therefore, it converts foggy images to the HSV color space for enhancement, arguing that the saturation, hue, and brightness in the HSV space are more consistent with human visual perception. It also performs block-based, multi-scale Retinex enhancement on the V component to improve image clarity and visual effect. This patent document explains that using a perception-dependent color space to separate brightness and color information does indeed help reduce the overall color shift problem caused by directly processing the three RGB channels, demonstrating that existing technology has recognized the significant impact of color space selection on dehazing quality. However, the technical approach of this patent document essentially still falls under the category of image enhancement. Its main basis is local Retinex theory, multi-scale Gaussian convolution, and block enhancement mechanisms, focusing on improving image visual visibility rather than establishing a rigorous restoration model based on the fog propagation mechanism. Although this patent document processes brightness and color separately in the HSV space, it does not introduce a quantitative mapping relationship between transmittance and perceived quantity, nor does it establish a nonlinear attenuation model of backscattered background, dark reference brightness, or fog concentration on the chromaticity modulus. Meanwhile, although the HSV color space is closer to the intuitive description of the human eye than RGB, it is not a strictly visually uniform color space. The same numerical change in different regions may not correspond to the same perceptual change. Therefore, HSV-based enhancement operations may still result in unstable color restoration, distortion of brightness and darkness levels, and over-enhancement in some areas under conditions of dense fog, low contrast, and complex lighting. Especially in applications that require balancing physical consistency and visual naturalness, relying solely on empirical enhancement strategies based on the HSV color space is insufficient to achieve accurate separation and restoration of brightness and chromaticity.

[0006] Based on the two existing technologies mentioned above, it can be seen that although existing dehazing methods have explored both physical models and visual adjustment and perceived color space, as well as enhancement processing, they still have the following common shortcomings: First, most existing technologies still rely mainly on linear inversion in RGB space or empirical enhancement in HSV space, and have not yet established restoration models for the lightness and chromaticity components in the CIELab space, which is more in line with visual uniformity and the comprehensive color difference expression law, respectively, to address the fog degradation mechanism. Second, existing technologies usually use transmittance as the inversion parameter of traditional atmospheric scattering models, without further exploring the nonlinear quantitative relationship between transmittance and human visual perception, especially lacking experimental calibration relationships between transmittance and chromaticity vector magnitude, and between transmittance and lightness difference after subtracting the dark reference. Third, existing technologies generally lack a unified modeling of the differential change laws among the relative constancy of hue, chromaticity amplitude decay, and brightness background enhancement during the color restoration process. Therefore, in typical challenging scenarios such as dense fog areas, bright areas, low saturation areas, and sky areas, it is still difficult to simultaneously achieve detail restoration, color naturalness, and visual comfort.

[0007] Therefore, there is an urgent need to propose a new image dehazing and restoration method: this method should not only preserve the physical interpretability based on transmittance, but also model and restore brightness and chromaticity separately in a more uniform color space from the perspective of human visual perception; furthermore, it should be able to establish a nonlinear mapping relationship between fog concentration and perceived attenuation through experimental means, so that the dehazed image is not only restored numerically, but also more in line with human visual observation habits in terms of color naturalness, brightness comfort and detail level, thereby overcoming the technical defects of existing technologies that make it difficult to balance visual quality and physical consistency. Summary of the Invention

[0008] The technical objective of this invention is to address the common problems in existing image dehazing methods, such as unnatural color restoration, loss of details in foggy areas, unstable brightness inversion, and insufficient visual perception consistency. This invention provides an image dehazing and restoration method based on nonlinear mapping of human vision. By converting the foggy image to the CIELab color space, which better matches the characteristics of human vision, and combining the nonlinear mapping relationship between transmittance and chroma and lightness, the chroma and lightness components are restored separately. This improves the color naturalness, detail clarity, and overall visual comfort of the dehazed image while ensuring physical interpretability.

[0009] Firstly, in order to achieve the above-mentioned objectives, the present invention adopts the following technical solution:

[0010] An image dehazing and restoration method based on nonlinear mapping of human vision includes:

[0011] S1. Obtain the foggy image to be processed, and perform white balance correction on the foggy image to obtain the corrected foggy image. ,in, Represents the pixel position coordinates;

[0012] S2, Based on the corrected foggy image Estimated transmittance map , ;

[0013] S3, will Convert from RGB color space to CIELab color space and separate the lightness component. First chromaticity component Second chromaticity component ;

[0014] S4, according to and Calculate hue angle And the magnitude of the hazy chromaticity vector ;

[0015] S5, while maintaining Under the condition of invariability, based on right Nonlinear recovery is performed to obtain the magnitude of the dehazed chromaticity vector. And further, the first chromaticity component after dehazing is obtained. Second chromaticity component ;

[0016] S6, based on Determine the lightness and darkness reference and based on and right Nonlinear recovery is performed to obtain the lightness component after dehazing. ;

[0017] S7, will , and Convert back to RGB color space to obtain the dehazed image. .

[0018] Preferably, in step S1, the white balance correction adopts any one of the following: gray world white balance method, white balance method based on white area detection, and white balance method based on color temperature estimation.

[0019] And / or, in step S2, the transmittance map Output by a separate transmittance estimation module. The transmittance estimation module employs any one of the following: a transmittance estimation algorithm based on dark channel prior, a transmittance estimation algorithm based on color attenuation prior, a transmittance estimation algorithm based on physical model optimization, or a transmittance estimation algorithm based on deep learning network.

[0020] Preferably, in step S2, when the transmittance estimation module uses a transmittance estimation algorithm based on dark channel priors, the coarse transmittance map... Calculate using the following formula:

[0021] ;

[0022] in, pixel position The coarse transmittance value at that location, To maintain the fog retention coefficient, For color channel index, In pixel position A local window centered on the center. pixel position Located in color channel pixel values ​​on Atmospheric light in color channels The amount on, This is for calculating the minimum value;

[0023] And the coarse transmittance map Perform guided filtering or edge-preserving filtering to obtain the transmittance map. .

[0024] Preferably, in step S3, the corrected foggy image is... When converting to the CIELab color space, use the D65 white point as the reference white point;

[0025] And / or, in step S4, the magnitude of the hazy chromaticity vector Calculate using the following formula:

[0026] ;

[0027] The hue angle Calculate using the following formula:

[0028] .

[0029] Preferably, in step S5, the first chromaticity component after dehazing... Second chromaticity component Calculate according to the following formulas:

[0030] ;

[0031] .

[0032] Calculate using the following formula:

[0033] ;

[0034] in, pixel position Color recovery gain at that location;

[0035] Calculate using the following formula:

[0036] ;

[0037] in, It is a natural exponential function. For reference transmittance, Limiting transmittance for colorimetric restoration;

[0038] Determine using the following formula:

[0039] ;

[0040] in, This is the lower limit threshold for color recovery, and ;

[0041] And / or, in step S6, the brightness component after defogging Calculate using the following formula:

[0042] ;

[0043] in, Calculate using the following formula:

[0044] ;

[0045] For reference transmittance, The limited transmittance used for brightness restoration;

[0046] Determine using the following formula:

[0047] ;

[0048] in, The lower limit threshold for brightness recovery, and .

[0049] Preferably, in step S5, the pre-constructed nonlinear mapping relationship between transmittance and chromaticity perception is an exponential mapping relationship between transmittance and chromaticity vector magnitude, expressed as:

[0050] ;

[0051] in, Let be the magnitude of the chromaticity vector. Here, is the fitting coefficient related to the intrinsic chromaticity of the target, and t is the transmittance. The natural exponential function; the fitting coefficients The results were obtained from experimental data of a standard color chart under controlled fog conditions.

[0052] And / or, in step S6, the pre-constructed nonlinear mapping relationship between transmittance and perceived lightness is an exponential mapping relationship between transmittance and lightness difference component, the expression of which is:

[0053] ;

[0054] in, For the brightness difference component, Here, is the fitting coefficient related to the inherent brightness of the target, and t is the transmittance. The natural exponential function; the fitting coefficients The results were obtained from experimental data of a standard color chart under controlled fog conditions.

[0055] Preferably, in step S6, the brightness / darkness reference... The backscattered radiation is calculated using the zero-reflection target assumption, where the tristimulus values ​​corresponding to the backscattering satisfy the following:

[0056] ;

[0057] in, , , Pixel positions The tristimulus values ​​corresponding to the lightness / darkness baseline. For wavelength, and These are the lower and upper limits of the integration wavelength range, respectively. The spectral power distribution of the incident light source, pixel position Transmittance at that location The backscattering phase function, The backscattering angle, , , This is a CIE standard color matching function;

[0058] Then, the CIELab conversion function fLab(.) is used to obtain... :

[0059] ;

[0060] And / or, in step S7, after the dehazing, the lightness component... The first chromaticity component after defogging and the second chromaticity component after defogging After converting back to the RGB color space, the output pixel values ​​are cropped in color gamut. The cropped output pixel values ​​are... Determine using the following formula:

[0061] ;

[0062] in, pixel position Color channel The output value after cropping. pixel position Color channel The output value before cropping. For color channel index, This represents the maximum value of the channel corresponding to the bit depth of the target image. To perform calculations to obtain the larger value, The operation is performed to select the smaller value.

[0063] Secondly, the present invention also provides an image dehazing and restoration system based on a nonlinear mapping of human vision, the system being configured to perform the method, comprising:

[0064] The image input and white balance correction module is used to acquire the foggy image to be processed and perform white balance correction to obtain the corrected foggy image.

[0065] A transmittance estimation module is used to estimate the transmittance map of the corrected foggy image;

[0066] The color space conversion module is used to convert the corrected hazy image from the RGB color space to the CIELab color space and output the lightness component, the first chromaticity component and the second chromaticity component; the chromaticity recovery module is used to calculate the hue angle, the magnitude of the hazy chromaticity vector and the first and second chromaticity components after dehazing based on the transmittance diagram, the first chromaticity component and the second chromaticity component.

[0067] The lightness restoration module is used to calculate the lightness dark reference, lightness difference component, and lightness component after dehazing based on the transmittance map and lightness component.

[0068] The inverse color space conversion module is used to convert the dehazed lightness component, the dehazed first chromaticity component, and the dehazed second chromaticity component back to the RGB color space to obtain the final dehazed image.

[0069] Thirdly, the present invention also provides a computer-readable storage medium having a computer program or instructions stored thereon, which, when executed by a processor, implement the method.

[0070] Fourthly, the present invention also provides a computer program product, including a computer program or instructions that, when executed by a processor, implement the method.

[0071] This invention shifts the image restoration process for foggy images from the traditional RGB sensor response space to the CIELab color space, which better aligns with human visual perception. It utilizes the nonlinear mapping relationship between transmittance and chromaticity modulus and lightness difference components to decouple and restore chromaticity and lightness information, achieving superior overall technical results. Firstly, this invention maintains a relatively constant hue angle during chromaticity restoration, adaptively amplifying only the modulus of the chromaticity vector in the $ab$ plane. This makes the restored image closer to the real scene in color direction, effectively avoiding common problems in traditional linear inversion such as color cast, graying, distortion in sky areas, and color bleaching in bright areas, significantly improving color naturalness and visual harmony. Secondly, this invention introduces a light difference channel... By subtracting the dark reference and performing nonlinear restoration based on transmittance, this invention can more accurately distinguish between the intrinsic brightness of the target and the background uplift effect caused by fog backscattering. This results in smoother and more stable brightness reconstruction in dense fog areas, distant areas, and low-contrast areas, enhancing details in dark areas and intermediate levels while suppressing noise over-amplification and local over-enhancement caused by low transmittance. Furthermore, the restoration module and transmittance estimation module are decoupled, making it compatible with various existing transmittance acquisition methods. This provides excellent modularity and engineering application flexibility, enabling closed-loop computational restoration without relying on large-scale training samples. Therefore, it also boasts advantages such as high computational efficiency, low implementation cost, clear physical meaning, strong interpretability, and good generalization ability. In summary, this invention achieves a more balanced and significant improvement over existing methods in terms of detail sharpness, brightness comfort, color fidelity, and restoration stability in complex foggy scenes. Attached Figure Description

[0072] Figure 1 This is a schematic diagram of the overall process of an image dehazing and restoration method based on nonlinear mapping of human vision according to the present invention.

[0073] Figure 2 The diagram shows the experimental verification device of this invention, along with a physical image. Figure 2 (a) is a schematic diagram of the experimental verification device of the present invention. Figure 2 (b) is a physical diagram of the experimental verification device of the present invention.

[0074] Figure 3 A statistical graph showing the cosine of the angle between the color block vector and the corresponding fog-free reference vector under different fog concentrations. Figure 3 (a) corresponds to The haze range Figure 3 (b) corresponds to The haze range Figure 3 (c) corresponds to The haze range.

[0075] Figure 4 The measured relationship and fitting curve between the chromaticity vector magnitude and transmittance are shown.

[0076] Figure 5 The figure shows the measured relationship and fitting curve between lightness component and transmittance, with the inset showing the original lightness-transmittance scatter distribution.

[0077] Figure 6 This is a comparison chart showing the performance of the method of this invention and the standard linear inversion method on a public dataset. Detailed Implementation

[0078] The technical solutions in the embodiments of the present invention will be clearly and completely described below. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the protection scope of the present invention.

[0079] I. Terminology Explanation

[0080] To facilitate understanding of the technical solution of this invention, the main terms used herein are briefly explained below. This explanation of terms is for illustrative purposes only and should not be construed as limiting the scope of protection of this invention.

[0081] 1. Transmittance: Transmittance is denoted as... This indicates the position of the reflected light from the scene target at the pixel location. The effective transmittance ratio that can still reach the imaging device after propagating through the fog medium in the corresponding imaging path. The value of satisfies The smaller the value, the more severe the fog attenuation at that location.

[0082] 2. CIELab Color Space: The CIELab color space is a perceptual color space designed with the uniformity of human visual perception as its design goal. Channels represent brightness. The channel represents the chromaticity in the red-green direction. The channels represent chromaticity in the yellow and blue directions. Compared to the RGB color space, CIELab is better suited for analyzing brightness variations and overall color variations relatively independently.

[0083] 3. Chromaticity vector magnitude: The chromaticity vector magnitude is denoted as... This indicates that the pixel is in CIELab space. The overall chromaticity intensity on a plane can be understood as the overall color saturation. This invention compensates for the overall chromaticity attenuation caused by fog by restoring this modulus.

[0084] 4. Hue Angle: The hue angle is denoted as... This indicates that the pixel is in CIELab space. The orientation angle in a plane is used to characterize the overall color direction. This invention is based on experimental findings that fog mainly weakens the chromaticity amplitude, while having a relatively small impact on the hue direction; therefore, the hue angle remains essentially unchanged during restoration.

[0085] 5. Lightness / Darkness Reference: The lightness / darkness reference is denoted as... , indicating at pixel position At a point where the apparent reflectivity of the target approaches zero, the CIELab brightness background is determined solely by backscattering from the fog medium. This quantity is used to separate the effective brightness difference component of the target from the hazy brightness.

[0086] 6. Fog Chamber: A fog chamber is a closed or semi-closed space used in the calibration experiment of this invention to construct a controlled fog environment, which is used to ensure that the fog concentration, illumination conditions, target position and measurement geometry are controllable.

[0087] 7. MOR: MOR stands for Meteorological Optical Range, used to characterize fog concentration. The lower the MOR, the worse the visibility and the higher the fog concentration.

[0088] 8. Recovery Gain: Recovery gain refers to the chromaticity recovery gain or lightness recovery gain generated by transmittance, which is used to map the perceived value under fog conditions back to the perceived value under relatively clear conditions.

[0089] 9. Standard method: The standard method refers to the restoration method that directly performs linear inversion on the image in RGB space based on the traditional atmospheric scattering model. It can usually be expressed as dividing each channel according to transmittance to amplify it and superimposing the atmospheric light term.

[0090] II. System Structure

[0091] Combination Figure 1 This invention can also be understood as an image dehazing and restoration system based on nonlinear mapping of human vision. The system includes at least an image input and white balance correction module, a transmittance estimation module, a color space conversion module, a hue angle and chromaticity analysis module, a brightness and darkness reference estimation module, a chromaticity restoration module, a brightness restoration module, and an inverse color space conversion module.

[0092] The image input and white balance correction module receives the hazy image to be processed and eliminates the color cast caused by the imaging device or environmental spectral distribution, establishing a unified input basis for subsequent transmittance estimation and perceptual restoration. The transmittance estimation module outputs a transmittance map for each pixel. This module can be decoupled from the restoration module, and its output can come from dark channel prior methods, color attenuation priors, deep learning networks, or other mature transmittance estimation methods. The color space conversion module converts the corrected RGB image to the CIELab color space and outputs the luminance component. and chromaticity components , The hue angle and chromaticity analysis module is used in... The hue angle and chromaticity vector magnitude are calculated in the plane. The lightness-dark reference estimation module is used to construct a pixel-level dark reference based on the transmittance and scattering background model. The chromaticity recovery module restores the chromaticity vector while preserving the hue angle. The channel undergoes non-linear restoration. The brightness restoration module uses the dark reference as a base and then performs non-linear restoration on the effective brightness difference components of the target. The inverse color space conversion module then converts the restored color space into a non-linear representation of the target color space. , , The components are remapped back to RGB space to obtain the final dehazed image.

[0093] The system architecture of this invention is not necessarily limited to a standalone hardware module; it can also be implemented on a general-purpose computing platform by a processor calling program instructions from memory. Preferably, it can be implemented on an industrial control computer, server, edge computing unit, or embedded vision terminal equipped with a CPU or GPU. At the software level, it can be implemented using Python, C++, MATLAB, or image processing frameworks based on OpenCV, NumPy, or SciPy.

[0094] III. Core Technical Route and Mechanism of the Invention

[0095] The core technical approach of this invention is not to simply use the linear atmospheric scattering model in the traditional RGB space for inversion, but to break down the dehazing problem into two interconnected sub-paths with different recovery mechanisms: one is the chromaticity recovery path, and the other is the lightness recovery path. Both are driven by the transmittance map, but the recovery variables, physical meanings, and perceptual constraints are different.

[0096] For the chroma path, this invention does not directly address... , Instead of linearly amplifying the two channels separately, we first treat them as a two-dimensional chromaticity vector in the CIELab plane. Experiments show that the cosine of the angle between the chromaticity vector and its haze-free reference vector is close to 1 under different haze concentrations, meaning the overall color direction remains essentially unchanged. The main change is in the vector length being compressed by the haze medium. Therefore, this invention proposes to maintain the hue angle essentially unchanged and only restore the chromaticity vector magnitude as the core constraint, transforming the overall color restoration problem into a one-dimensional magnitude restoration problem. The restored magnitude is then projected back to the original hue direction. This approach has two important effects: first, it avoids the overall hue drift caused by independent amplification of different channels; second, it makes the chromaticity restoration process more consistent with the subjective perception of overall color changes by the human eye.

[0097] Regarding the brightness path, this invention recognizes that the brightness in a hazy image does not entirely originate from target reflection, but is simultaneously superimposed with a significant backscattering background from the hazy medium. If traditional linear inversion is used directly... Transmittance division in the channel can excessively amplify noise and local fluctuations in low transmittance regions, and it makes it difficult to distinguish between components from different sources: low intrinsic brightness of the target and increased fog background. Therefore, this invention first constructs a brightness-darkness benchmark. This benchmark is used to separate the effective brightness difference related to the target from the hazy brightness. Then, transmittance-driven nonlinear restoration is performed on the difference component. Since the object of restoration is no longer the mixed brightness itself, but the effective brightness difference after removing the backscattered background, it can more stably enhance the level details in dense fog areas, while suppressing over-inversion.

[0098] Furthermore, this invention does not write the above-mentioned recovery function as an empirical arbitrary nonlinear transformation, but rather through... Figure 2 The experimental setup shown allows for long-term data collection and calibration of targets with different reflectivities in a controlled fog environment, establishing an exponential mapping relationship between transmittance and chromaticity modulus, and between transmittance and lightness difference.

[0099] IV. Experimental Calibration Basis

[0100] The method of this invention is not based on mere theoretical deduction, but rather on controlled physical experiments. Combined with... Figure 2 In a preferred embodiment, the experimental setup is constructed as follows: a light source, a light-diffusing plate, a transparent fog chamber, a fog generator, a standard color chart, an industrial camera, a fiber optic collimating lens, a spectrometer, and a fan are arranged in a dark room or an experimental space with stable and controllable ambient light. The light source, after passing through the light-diffusing plate, forms an illumination field as uniform as possible, ensuring stable illumination of the standard color chart. The fog generator continuously injects fog into the fog chamber, and the fan improves the uniformity of the medium distribution within the chamber. The industrial camera acquires images of the standard color chart observed through the fog medium from one side, and the spectrometer measures the spectral irradiance distribution from another observation path through the fiber optic collimating lens, in order to invert the corresponding transmittance or establish a correspondence with the transmittance. The experimental process is uniformly data collected, recorded, and processed by a computer.

[0101] This device allows for the continuous acquisition of hazy images and corresponding transmittance measurements of the same color patch under different MOR conditions. After white balance correction, the images are converted to CIELab space. The chromaticity vector, chromaticity modulus, and hue angle of each color patch under different haze concentrations are calculated and then compared with the corresponding values ​​under a haze-free reference state.

[0102] V. Specific Technical Route for Implementing the Method of the Invention

[0103] The following combination Figure 1The following describes the specific implementation process of the present invention in detail with reference to a preferred embodiment. In this embodiment, the transmittance estimation module uses a dark channel prior algorithm as an example, but the present invention is not limited thereto. As long as a transmittance map that is positively correlated with fog concentration and corresponds to the degree of scene attenuation can be output, the recovery module of the present invention can be used compatiblely.

[0104] (a) Step S1: Acquire the foggy image and perform white balance correction

[0105] The input image is denoted as Because camera response curves, sensor spectral sensitivity, lens transmission characteristics, and ambient light color temperature can cause color casts in the original input image, directly estimating transmittance from this image and further reconstructing it in CIELab space will lead to subsequent color distortion. , The combined color shift and haze attenuation effect in the channel disrupt the restoration premise of an approximately constant hue angle. Therefore, this invention preferably performs white balance correction before proceeding with transmittance estimation and visual restoration.

[0106] In a preferred implementation, the gray-world method can be used for automatic white balance correction. Let the average values ​​of the three channels of the input image be... , , The overall color mean is denoted as Then the gain factors of the three channels can be expressed as:

[0107] ;

[0108] in, , , These are the white balance gains for the red, green, and blue channels, respectively. , , These are the average gray values ​​of the input image in the corresponding channels. This is the average value of the overall color.

[0109] Furthermore, the image after white balance correction can be obtained:

[0110] ;

[0111] in, For the white balance corrected image in the channel pixel values ​​on For the original input image in channels pixel values ​​on This represents the white balance gain for the corresponding channel.

[0112] In another preferred implementation, a white area detection method or a white balance method based on known color temperature estimation can also be used. For monitoring images of outdoor roads, power transmission lines, stations, ports, etc., if there are bright but low-saturation areas in the image, these areas can be preferentially used as approximate white references to avoid interference from large areas of colored targets in the whole-image averaging method. Regardless of the white balance method used, the purpose of this invention is to bring the image into a state with smaller overall color bias and more consistent with the reference white point used in experimental calibration.

[0113] Preferably, after white balance correction, normalization processing can be performed, that is, scaling the pixel values ​​of the three channels to... Or within the target depth range, so that subsequent transmittance estimation, color space conversion and exponential recovery formula can be uniformly processed.

[0114] The white balance correction step is mainly used to establish a stable input in subsequent steps.

[0115] (ii) Step S2: Estimate the transmittance map

[0116] The task of step S2 is to provide a fog attenuation intensity representation for each pixel for subsequent perceptual restoration. Unlike traditional dehazing methods, this invention emphasizes the decoupling of the restoration module and the transmittance estimation module. That is, the source of the transmittance map can be flexibly replaced, and this invention focuses on how to use the transmittance map to restore the overall color and brightness.

[0117] In this embodiment, the process of constructing the transmittance map is illustrated using the dark channel prior algorithm as an example. Let the corrected hazy image be... Its color channel The pixel value on The local dark channel plot can then be defined as:

[0118] ;

[0119] in, pixel position The corresponding local dark channel value, In pixel position A local window centered on the center. pixel position within the window In color channels pixel values ​​on This indicates the minimum value operation.

[0120] Subsequently, atmospheric light was estimated. Preferably, a candidate set can be formed by first selecting a certain proportion of the brightest pixels in the dark channel, and then the pixel with the highest overall color brightness can be selected from the candidate set as the atmospheric light position, with its three channel values ​​denoted as follows. , , Based on this, a coarse transmittance map can be obtained:

[0121] ;

[0122] in, This is the coarse transmittance value. The fog-retention coefficient is usually taken as a positive number close to 1. For atmospheric light in the channel The portion on top.

[0123] because Due to the influence of local window minimum value calculation, blocky effects may appear near the edges. Therefore, it is preferable to further refine the image using guided filtering or other edge-preserving filters. The refinement process can be summarized as follows:

[0124] ;

[0125] in, This is the refined transmittance map. This indicates a guided filtering operation. This is a coarse transmittance diagram. For guiding images.

[0126] In engineering implementation, to avoid excessive recovery gain in the extremely low transmittance region during subsequent recovery, it is preferable to set a lower limit threshold for transmittance. The limited transmittance was obtained as follows:

[0127] ;

[0128] in, This refers to the transmittance after amplitude limiting. This is the lower limit threshold for transmittance. and These represent operations to take the larger value and the smaller value, respectively.

[0129] It should be noted that the dark channel prior is only one example of transmittance acquisition. This invention is also fully compatible with color attenuation prior methods, model optimization methods that integrate depth information, neural network-based transmittance regression methods, or other methods that can output transmittance or equivalent fog concentration maps. As long as the transmittance maps output by these methods correspond to the actual attenuation level of the medium, they can be directly integrated into the recovery module of this invention. Therefore, the key to this step lies in the new overall technical solution formed after the transmittance map is introduced into the CIELab perceptual recovery framework.

[0130] (III) Step S3: Convert the image to the CIELab color space

[0131] After obtaining the white balance corrected image and transmittance map, The conversion is performed from RGB space to CIELab space. Preferably, a D65 reference white point is used as the conversion reference. In engineering implementation, the RGB values ​​can be linearized first and then converted to XYZ tristimulus values, and then converted from XYZ to CIELab.

[0132] Its standard expression can be written as:

[0133] ;

[0134] ;

[0135] ;

[0136] in, For hazy brightness components, The first chromaticity component of the hazy state. The second chromaticity component is the hazy state. , , The XYZ tristimulus values ​​of the current pixel. , , To reference the tristimulus values ​​corresponding to the white spots, This is a CIELab standard nonlinear mapping function.

[0137] Through the above transformation, the image is mapped to a space that is closer to the human eye's comprehensive color difference representation, so that the comprehensive color intensity and comprehensive hue direction can be separated in subsequent processing.

[0138] (iv) Step S4: Calculate the hue angle and the magnitude of the hazy chromaticity vector

[0139] In obtaining and Subsequently, both were considered as CIELab spaces. Two-dimensional vector components in the plane. Therefore, the magnitude of the hazy chromaticity vector is defined as:

[0140] ;

[0141] in, pixel position The magnitude of the hazy chromaticity vector at that location. and Pixel positions The first and second chromaticity components of the hazy state at that location.

[0142] The hue angle is defined as:

[0143] ;

[0144] in, pixel position The hue angle at that location, It is the arctangent function in the four quadrants.

[0145] Through this step, the original , The two channels are re-represented as polar coordinates of direction and amplitude. This form highly matches the core experimental conclusion of this invention: that fog primarily compresses chromaticity amplitude while relatively preserving chromaticity direction. Therefore, step S4 is a crucial bridge connecting theoretical understanding and practical implementation. Although its mathematical form is relatively simple, this step represents a significant advancement in this invention compared to simply processing data in CIELab. , Key differences between conventional channel solutions.

[0146] (v) Step S5: Recover the chromaticity components based on transmittance and hue angle

[0147] Step S5 is the key substantive difference between this invention and traditional RGB linear inversion methods, general color space enhancement methods, and methods that directly process channel values ​​only in Lab space. This step is not intended to... , Instead of simply scaling the channels, it employs a comprehensive chromaticity recovery mechanism built upon experimental calibration, perceptual geometry, and stable recovery.

[0148] First, let's start with the experimental conclusions. Figure 3 This indicates that for the vast majority of standard color patches, under fog conditions, their color in the CIELab space... The change in orientation angle in the plane is very small, and the overall hue direction remains basically stable. Therefore, this invention does not change the original hue angle. Only the magnitude of the chromaticity vector is recovered. This aligns with both experimental facts and the subjective experience of the human eye regarding overall color changes. If each color vector were recovered separately... Channels and When two channels have different propagation paths due to noise, quantization error, and transmittance error, the overall hue will be deflected, resulting in a purple sky, yellowish green vegetation, and dirty concrete or equipment surfaces. Keeping the hue angle constant will significantly reduce this problem.

[0149] Secondly, starting from the experimental fit relationship, Figure 4 This indicates that the overall chromaticity modulus and transmittance satisfy an exponential nonlinear relationship, rather than the near-linear proportional relationship implicit in traditional models. Let the transmittance of the sharp reference state be... In a preferred embodiment, it is possible to For the pixel to be restored, the limiting transmittance used for chroma restoration is defined as:

[0150] ;

[0151] in, pixel position The limiting transmittance used for colorimetric restoration The transmittance at that location. The lower limit threshold for chromaticity recovery preferably satisfies .

[0152] Further construct the chromaticity recovery gain:

[0153] ;

[0154] in, pixel position Color recovery gain at that location For reference transmittance, It is a natural exponential function.

[0155] Therefore, the restored composite chromaticity modulus is:

[0156] ;

[0157] in, pixel position The restored composite chromaticity modulus length, To correspond to the overall chromaticity modulus of the hazy state.

[0158] Subsequently, the restored modulus length is reprojected onto the unchanged hue angle. , In the coordinate system:

[0159] ;

[0160] ;

[0161] in, , Pixel positions The recovered first and second chromaticity components, It is a cosine function. It is a sine function.

[0162] The aforementioned recovery link has several significant advantages.

[0163] First, the object to be restored changes from two-dimensional channel values ​​to one-dimensional composite chromaticity modulus, reducing the dimensionality of the problem and shortening the error propagation path. Traditional methods directly address RGB or... , When division or scaling is performed on channels separately, the differences between channels amplify the instability of the overall color. This invention first uses the overall color modulus as a unified recovery object, and then uses the orientation angle to reconstruct it back to a two-dimensional plane, making the overall color recovery more robust.

[0164] Second, the exponential recovery gain is consistent with the experimental calibration. Because the chromaticity modulus attenuation exhibits significant nonlinear compression characteristics in low transmittance regions, linear proportional recovery is prone to insufficient recovery in heavily foggy areas, resulting in a grayish or dull image; conversely, a simpler approach... The gain is high, but it is prone to over-recovery in the low transmittance region, leading to amplification of noise and compression artifacts. This invention utilizes a method derived from... Figure 4 The exponential gain of the fitting relationship ensures that the overall color recovery maintains a sufficient improvement in the medium-high transmittance range, while avoiding uncontrolled surges in the low transmittance range, thus balancing the sense of layering and stability.

[0165] Third, the constant hue angle constraint makes the restored image more consistent with the human eye's overall color perception. Taking the sky as an example, the blue component in foggy images often doesn't completely disappear; instead, the overall chromaticity is compressed and the overall hue is diluted by the fog background. If dehazing is performed directly in the linear RGB space, the sky tends to appear white or have an overall color halo. However, this invention uses the overall hue direction in the CIELab space as a constraint, only amplifying the overall chromaticity modulus, making it easier to restore the sky to a visually natural gradient from light blue to medium blue, rather than a distorted, highly saturated, blocky blue. Similarly, it can maintain a more stable overall color direction for vegetation, road markings, equipment surfaces, soil, and buildings.

[0166] Fourth, the chromaticity restoration and transmittance estimation modules of this invention are decoupled. In other words, this invention does not require the use of a specific transmittance estimation model to be effective. When a more accurate transmittance estimation network becomes available in the future, its output can still be directly connected to the integrated chromaticity restoration module of this invention, thereby continuously improving the final restoration effect.

[0167] In practical engineering, to further improve robustness, an upper limit for chromaticity recovery gain can be introduced. ,Right now:

[0168] ;

[0169] in, This is the gain for chroma recovery after clipping. This is the upper limit of chroma restoration gain.

[0170] When transmittance is extremely low or local textures are mixed with noise, the above upper limit can prevent overall color oversaturation, color spot magnification, or overall color banding. Furthermore, it can be combined with local transmittance variance, adaptive edge masks, or smoothing prior pairs. Post-processing is performed to further enhance the spatial continuity of the overall color restoration.

[0171] In summary, the real innovation of step S5 lies not in using a single formula, but in establishing the following synergistic relationship: Figure 3 Based on this, a constant hue angle constraint is introduced; Figure 4 Based on this, a transmittance-chromaticity modulus index mapping is introduced; and a stable reconstruction of comprehensive color is achieved using polar coordinate geometric relationships. The combination of these three elements forms a comprehensive color restoration mechanism that differs from existing technologies. This mechanism directly determines the core advantages of this invention in terms of comprehensive color naturalness, sky restoration effect, and stability in complex color areas.

[0172] (vi) Step S6: Recover the lightness component based on the lightness-darkness reference

[0173] Corresponding to step S5's restoration of overall chromaticity, step S6 addresses why brightness restoration in foggy images cannot simply rely on linear inversion of transmittance. Traditional methods generally treat brightness or each RGB channel as a similar quantity to be inverted, performing channel-level amplification according to the same transmittance relationship; however, this invention... Figure 5 The experimental results show that the effect of fog on brightness includes both the attenuation of direct light from the target and a significant increase in backscattered background. Therefore, it is necessary to first construct a background baseline and then restore the effective brightness components that are truly relevant to the target.

[0174] For pixel position The lightness component in a foggy state is If the reflectivity of the target at that location is extremely low or even close to zero, the observed brightness mainly comes from backscattering by the fog medium itself. This portion can be considered as the fog background brightness or brightness-darkness reference. This invention denotes it as... .

[0175] In a more complete physical representation, the tristimulus values ​​of the background in the XYZ space can be calculated first:

[0176] ;

[0177] in, , , Pixel positions The tristimulus values ​​of the fog background at that location, For wavelength variables, , These are the lower and upper limits of the integration wavelength range, respectively. The spectral power distribution of the incident light source, Transmittance, The backscattering phase function, The backscattering angle, , , This is a CIE standard color matching function.

[0178] In the experimental calibration setup, if a D65 light source is used and the fog medium is approximated as isotropic scattering, then the following can be taken: Or other calibrated constants or lookup table functions. Based on the above tristimulus values, the lightness and darkness references can be obtained by converting to the CIELab space:

[0179] ;

[0180] in, As a reference for brightness and darkness, The brightness component of the background tristimulus values. For reference white point luminance components, This is a CIELab nonlinear mapping function.

[0181] In practical deployments, to reduce the amount of real-time integration calculations, experimental setups can also be pre-established. The lookup table or fitting function allows for the rapid generation of the corresponding dark reference directly based on transmittance during the runtime phase. In other words, this invention allows for a physical modeling + offline calibration + online table lookup approach, and does not mandate the repeated execution of full-wavelength integration on every image.

[0182] After obtaining the dark reference, calculate the effective lightness difference:

[0183] ;

[0184] in, pixel position The effective lightness difference at that location For hazy brightness components, Used as the benchmark for brightness and darkness.

[0185] Figure 5 This indicates that if we directly put As the object of recovery, its dispersion with transmittance is large, resulting in unstable recovery; however, after subtracting the dark reference, The relationship with transmittance converges significantly and can be fitted using an exponential relationship. Therefore, this invention further defines a limiting transmittance for brightness restoration:

[0186] ;

[0187] in, pixel position Brightness recovery at the limit of transmittance, This is the lower limit threshold for brightness recovery.

[0188] Constructing brightness restoration gain:

[0189] ;

[0190] in, pixel position Brightness recovery gain at that location This is the reference transmittance.

[0191] Therefore, the restored brightness can be expressed as:

[0192] ;

[0193] in, pixel position The restored brightness component.

[0194] This recovery method has a fundamental advantage over traditional linear brightness inversion.

[0195] First, the object to be restored is limited to the effective brightness difference related to the target, rather than the total brightness mixed with the fog background. This difference makes the restoration process more consistent with imaging physics. For low-reflectivity targets, such as dark equipment, shadowed areas, road cracks, and local metal components of distant towers, traditional methods often confuse background scattering with the weak reflection of the target, easily resulting in residual fog or over-brightening during restoration. This invention, however, uses dark reference separation to make the restoration of these areas more focused on bringing back the compressed target brightness difference, rather than blindly increasing the overall brightness.

[0196] Secondly, the exponential recovery gain is consistent with experimental results, exhibiting smoother compensation characteristics in the low transmittance region. Traditional methods based on... Linear inversion in When the gain increases sharply at very low values, it amplifies not only noise but also compression artifacts, JPEG texture blocks, color quantization errors, and local transmittance estimation errors. In contrast, the exponential gain increase of this invention is more controllable, thus enabling more stable recovery of layers in foggy areas without causing the image to appear washed out or producing unnatural halos at the edges.

[0197] Furthermore, the introduction of a dark reference also improves the recovery stability of the sky and large bright background areas. The sky area itself has high background brightness; if simple linear dehazing is used, it is easily over-amplified due to its low transmittance, resulting in unnatural bright spots, color halos, or localized overexposure. By introducing... This invention can first model the brightness increase caused by the fog background in such areas, and then only recover the effective components related to the real sky texture and the overall color difference, so the recovery result is more in line with visual experience.

[0198] From an engineering implementation perspective, step S6 can also set an upper limit for brightness restoration gain. :

[0199] ;

[0200] in, This is the gain for brightness recovery after limiting. This is the upper limit of brightness recovery gain.

[0201] Furthermore, it is also possible to base the analysis on the transmittance gradient or the brightness gradient. Applying local smoothing or edge protection allows for different levels of refinement in flat, low-texture, and high-contrast edge areas of the background. This involves first determining the dark baseline, then calculating the effective lightness difference, and finally performing transmittance-driven nonlinear restoration.

[0202] It is important to emphasize that steps S6 and S5 are independent yet complementary. If only chroma restoration is performed without addressing the lightness background, the restored color will appear dull and lack detail. Conversely, if only lightness restoration is performed without comprehensive chroma restoration, the image will appear brighter but still grayish, with insufficient overall color and unnatural-looking skies and vegetation. Only through the combined action of comprehensive chroma modulus restoration and dark baseline lightness restoration can a clear and natural dehazing result be achieved. This is evident from… Figure 6 The comparison in the text can also be intuitively demonstrated.

[0203] (vii) Step S7: Convert the recovered Lab component back to RGB and output the result.

[0204] After recovery , and Then, it is converted back to XYZ according to the standard CIELab inverse transform procedure, and then converted back from XYZ to the target RGB color space to obtain the output image. To avoid pixel values ​​exceeding the device's color gamut or bit depth range after the inverse transformation, it is preferable to perform color gamut clipping:

[0205] ;

[0206] in, pixel position In the passage The output value after cropping is shown. This is the output value before cropping. This represents the maximum channel value corresponding to the target bit depth.

[0207] In an 8-bit image embodiment, In the floating-point normalization implementation, When there is a risk of overall color overshoot, lightweight local contrast reshaping, global tone mapping, or edge sharpening can be performed on the output results. However, it is preferable to keep these post-processing steps optional to ensure that the present invention mainly comes from the main recovery process of S1-S7, rather than relying on post-processing.

[0208] VI. Specific Application Examples

[0209] The following application examples are used to further illustrate the technical effects of the image dehazing and restoration method based on nonlinear mapping of human vision according to the present invention, but are not intended to limit the scope of protection of the present invention. Experimental conditions not specified are conventional conditions in the art; experimental methods not specified can be implemented using common image acquisition, color space conversion, transmittance estimation, and image quality evaluation procedures in the art.

[0210] (I) Application Example 1: Calibration and Verification of Nonlinear Mapping Laws Based on Controlled Fog Chamber

[0211] 1. Experimental Objective

[0212] The purpose of this application example is twofold: first, to verify whether the overall hue direction of the target in the CIELab space remains basically stable under the influence of fog, thereby supporting the chromaticity recovery strategy of this invention that maintains the hue angle unchanged; and second, to verify whether there is a stable nonlinear mapping relationship between the overall chromaticity modulus and transmittance, and between the lightness difference after subtracting the dark background reference and transmittance, thereby supporting the nonlinear recovery model of this invention.

[0213] 2. Experimental setup

[0214] like Figure 2 As shown, the experimental platform includes: a D65 standard light source, a light-diffusing plate, an industrial camera, a sealed glass fog chamber, a controllable fog generator, a standard color chart, a fiber optic collimator, a spectrometer, a fan, and a computer. Specifically: the D65 standard light source provides stable and repeatable standard illumination; the light-diffusing plate reduces illumination non-uniformity; a 24-color standard color chart is preferred; the industrial camera acquires images of the standard color chart at different fog concentrations; the fog generator continuously injects fog medium into the fog chamber; the spectrometer synchronously acquires transmission spectra to calculate the true transmittance; the fan reduces localized concentration non-uniformity within the chamber; and the computer acquires image sequences, records spectral data, and performs subsequent data analysis.

[0215] 3. Experimental Methods

[0216] In a fog-free state, a reference image of the standard color chart is first taken. Then, a fog generator is activated, causing the fog concentration within the fog chamber to continuously increase from low to high. Simultaneously, an industrial camera acquires a sequence of images from the color chart, and a spectrometer measures the spectral irradiance transmitted through the fog chamber. For each frame, white balance correction is performed, and then the image is converted to CIELab color space to extract the color patches. , , Quantity.

[0217] Transmittance can be calculated using the following formula:

[0218] ;

[0219] in, Transmittance, The spectral irradiance under foggy conditions. The spectral irradiance is under a haze-free reference condition. For wavelength, and These represent the lower and upper limits of the integration wavelength range, respectively.

[0220] For each color patch, calculate the cosine of the angle between the hazy chromaticity vector and the hazy reference chromaticity vector:

[0221] ;

[0222] in, For the first The cosine of the angle between the color blocks, For the first The chromaticity vector of each color block in a foggy state. For the first The chromaticity vector of each color patch in a hazy reference state. For dot product operation, Let be the vector magnitude.

[0223] The overall colorimetric model length is calculated using the following formula:

[0224] ;

[0225] in, To integrate the chromaticity modulus, The first chromaticity component, This is the second chromaticity component.

[0226] The effective difference in lightness is calculated using the following formula:

[0227] ;

[0228] in, The effective brightness difference after subtracting the dark reference. For the luminance component of a foggy image, Used as the benchmark for brightness and darkness.

[0229] 4. Experimental Results and Analysis

[0230] (1) Verification of hue angle constancy

[0231] Figure 3 Statistical results are presented regarding the cosine values ​​of the angle between the hazy chromaticity vector and the non-hazy reference chromaticity vector for each color patch under different fog concentration ranges. Figure 3 It can be seen that:

[0232] exist Within the light to medium fog range, the cosine of the included angle of most color blocks is close to 1, indicating that the overall hue direction hardly changes.

[0233] exist Within the medium fog range, the cosine values ​​of the included angles of most color blocks are still concentrated in the higher range, and the overall hue direction still shows good stability.

[0234] exist Within the dense fog zone, some low-saturation or more sensitive color patches fluctuated more, but the overall hue remained relatively consistent.

[0235] from Figure 3 The overall trend shows that the impact of fog on the overall hue direction is significantly less than its compression effect on the overall chromaticity intensity. This experimental result directly proves that the present invention, which maintains the hue angle unchanged and only restores the overall chromaticity modulus during chromaticity recovery, is consistent with experimental facts and is reasonable.

[0236] (2) Verification of the nonlinear relationship between chromatic modulus and transmittance

[0237] Figure 4 The measured scatter plots and fitted curves of the relationship between the comprehensive chromaticity modulus and transmittance are presented. Figure 4 It can be seen that for targets with different reflectivities under different transmittance conditions, the comprehensive chromaticity modulus exhibits a significant nonlinear increasing trend with increasing transmittance, rather than a simple linear relationship. The following exponential relationship can be preferably used for fitting:

[0238] ;

[0239] in, To integrate the chromaticity modulus, These are the fitting parameters related to the intrinsic chromaticity of the target. Transmittance, It is a natural exponential function.

[0240] Figure 4 This indicates that when the transmittance is low, the overall chromaticity modulus is more significantly compressed by the fog. If traditional linear inversion is used for recovery, insufficient overall color recovery or overshoot in local areas is likely to occur. However, the exponential recovery mechanism of this invention is more consistent with the actual change pattern measured by experiments.

[0241] (3) Verification of the nonlinear relationship between lightness difference and transmittance

[0242] Figure 5 The original scatter plot relationship between lightness and transmittance is given, along with the fitted curve after introducing a dark reference correction. Figure 5 It can be seen that without subtracting the dark reference, the brightness distribution of different targets is relatively dispersed; however, after subtracting the dark reference, a clearer and more stable nonlinear mapping relationship emerges between the effective brightness difference and transmittance. The preferred relationship can be expressed as follows:

[0243] ;

[0244] in, For effective lightness difference, These are fitting parameters related to the intrinsic brightness of the target. Transmittance.

[0245] This indicates that the present invention introduces [a concept / mechanism] into brightness restoration. As a benchmark for fog background brightness, it helps to separate the true brightness difference of the target from the fog scattering background, thereby making brightness recovery more stable, especially beneficial for the layer reconstruction of dense fog areas, distant areas and low contrast areas.

[0246] 5. The technical effects demonstrated by this application example

[0247] This application example demonstrates that:

[0248] 1) The overall hue direction is basically stable under the influence of fog. The strategy of keeping the hue angle unchanged and restoring the overall chromaticity modulus of this invention has experimental support.

[0249] 2) A stable exponential nonlinear relationship exists between the comprehensive chromaticity modulus and transmittance, and the comprehensive chromaticity recovery model of this invention has an objective basis;

[0250] 3) There is also a stable nonlinear relationship between the effective lightness difference after deducting the dark reference and the transmittance. The lightness restoration model of this invention has a physical and experimental basis.

[0251] 4) This invention is not simply experience enhancement, but a recovery method based on controlled experiments and perceptual modeling.

[0252] Application Example 2: Quantitative and Qualitative Validation on Public Fog Image Datasets

[0253] 1. Experimental Objective

[0254] The purpose of this application example is to verify whether the integrated chromaticity-luminance dual-path nonlinear restoration in the CIELab perceptual space proposed in this invention can achieve better restoration quality on real hazy images compared to the standard linear inversion method, when the transmittance map remains consistent.

[0255] 2. Experimental subjects and control scheme

[0256] Sample images from a publicly available real-world fog image dataset were selected as test objects. For each image to be tested, a transmittance map was first obtained using the same transmittance estimation method to ensure fairness in the comparison.

[0257] The comparison methods include:

[0258] Standard method: Performs linear inversion directly on the image in RGB space based on the traditional atmospheric scattering model;

[0259] The method of this invention (Ours) is as follows: First, the image is converted to CIELab space, then a comprehensive chromaticity modulus nonlinear restoration and a lightness nonlinear restoration based on a dark reference are performed, and finally the image is converted back to RGB space to output the result.

[0260] The recovery form of the standard method can be summarized as follows:

[0261] ;

[0262] in, To restore the image in channels pixel values ​​on For foggy images in channels pixel values ​​on For atmospheric light in the channel The amount on, Transmittance.

[0263] The chromaticity recovery gain in the method of this invention is:

[0264] ;

[0265] in, For chroma restoration gain, For reference transmittance, This refers to the limited transmittance.

[0266] The brightness recovery gain in the method of this invention is:

[0267] ;

[0268] in, To restore brightness gain, Limit transmittance to restore brightness.

[0269] 3. Evaluation Indicators

[0270] The following metrics were used to evaluate the recovery results: Peak Signal-to-Noise Ratio (PSNR); Structural Similarity (SSIM); Feature Similarity (FSIM); Visual Signal-to-Noise Ratio (VSI); Fog-Related Feature Similarity (FRFSIM); Vividness Index (VI); Realism Index (RI); and DehIQA.

[0271] Among them, the larger the values ​​of PSNR, SSIM, FSIM, VSI, FRFSIM, VI, and RI, the better the effect; the smaller the value of DehIQA, the better the effect.

[0272] 4. Experimental Data

[0273] Table 1 presents the quantitative comparison results between the method of this invention and the standard method under five different fog concentration conditions.

[0274] Table 1. Performance comparison between the method of the present invention and the standard method under different fog concentrations.

[0275] Furthermore, by averaging the data in Table 1, we can obtain the results shown in Table 2.

[0276] Table 2 Comparison of average performance under five fog concentration conditions

[0277]

[0278] As can be seen from Table 2, the average results for the five fog concentration conditions are as follows:

[0279] The average PSNR is improved by about 6.4051, indicating that the method of the present invention is significantly better than the standard method in terms of overall recovery accuracy;

[0280] The SSIM is improved by an average of approximately 0.1677, indicating that the present invention has a greater advantage in terms of structural preservation.

[0281] FSIM, VSI, FRFSIM, VI, and RI are all improved to varying degrees, indicating that the present invention has better performance in terms of feature preservation, visual perception, dehazing realism, and vividness.

[0282] The DehIQA decreased by an average of approximately 0.0177, indicating that the present invention outperforms the standard method in the overall evaluation of image quality.

[0283] The results of the item-by-item comparison show that, except for FRFSIM which is slightly close at the highest fog concentration, the other major indicators are better than the standard method under most fog concentration conditions, demonstrating good stability and consistency.

[0284] 5. Qualitative Results Analysis

[0285] Figure 6 Visual comparison results of the method of this invention and the standard method are presented. Combined with... Figure 6 It can be seen that for the sky region, see the reference image. Figure 6 In (a1)-(a5), the restoration results of this invention are more natural, and the blue layers are more reasonable. See Figure 6 The results from (b1) to (b5) are often pale, grayish, or have localized halo effects, as seen in the standard method. Figure 6 (c1)-(c5); For distant targets and dense fog areas, see the reference images. Figure 6 In sections (d1)-(d5), this invention enhances details without significantly amplifying quantization noise and texture block effects; for composite color areas such as ground, equipment, and structures, the composite color recovered by this invention is closer to the actual observation effect, with smaller composite color shift. Figure 6 (e1)-(e5);

[0286] Regarding the overall brightness distribution, the restoration result of this invention is smoother and more comfortable, avoiding both localized overbrightness caused by standard methods and overall dullness caused by insufficient restoration; for the standard method, see the reference image. Figure 6 (f1) - (f5).

[0287] Based on the above application examples, it can be concluded that this invention, by constructing comprehensive chromaticity recovery paths and luminance recovery paths in the CIELab perceptual color space and combining them with a transmittance-driven nonlinear mapping relationship, achieves differentiated compensation for the combined color attenuation and luminance enhancement effects in foggy images. Laboratory-controlled fog chamber calibration results demonstrate that the key assumptions and recovery model of this invention have a physical and experimental basis. Comparison results with publicly available datasets show that this invention generally outperforms standard methods in metrics such as PSNR, SSIM, FSIM, VSI, FRFSIM, VI, RI, and DehIQA. Figure 6 The visual results further demonstrate that the present invention has significant technical effects in terms of overall color naturalness, detail restoration, and overall visual comfort.

[0288] The foregoing description of embodiments of the present invention, through which those skilled in the art are able to implement or use the present invention, will be readily apparent to those skilled in the art. Various modifications to these embodiments will be readily apparent to those skilled in the art. The general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles disclosed herein.

[0289] Those skilled in the art will understand that embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-usable storage media (including but not limited to disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0290] This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart... Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0291] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0292] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0293] In a typical configuration, a computing device includes one or more processors (CPU), input / output interfaces, network interfaces, and memory.

[0294] Memory may include non-persistent memory in computer-readable media, such as random access memory (RAM) and / or non-volatile memory, such as read-only memory (ROM) or flash RAM. Memory is an example of computer-readable media.

[0295] Computer-readable media includes both permanent and non-permanent, removable and non-removable media that can store information using any method or technology. Information can be computer-readable instructions, data structures, modules of programs, or other data. Examples of computer storage media include, but are not limited to, phase-change memory (PRAM), static random access memory (SRAM), dynamic random access memory (DRAM), other types of random access memory (RAM), read-only memory (ROM), electrically erasable programmable read-only memory (EEPROM), flash memory or other memory technologies, CD-ROM, digital versatile optical disc (DVD) or other optical storage, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other non-transferable medium that can be used to store information accessible by a computing device. As defined herein, computer-readable media does not include transient computer-readable media, such as modulated data signals and carrier waves.

Claims

1. A method for image dehazing and restoration based on nonlinear mapping of human vision, characterized in that, include: S1. Obtain the foggy image to be processed, and perform white balance correction on the foggy image to obtain the corrected foggy image. ,in, Represents the pixel position coordinates; S2, Based on the corrected foggy image Estimated transmittance map , ; S3, will Convert from RGB color space to CIELab color space and separate the lightness component. First chromaticity component Second chromaticity component ; S4, according to and Calculate hue angle And the magnitude of the hazy chromaticity vector ; S5, while maintaining Under the condition of invariability, based on right Nonlinear recovery is performed to obtain the magnitude of the dehazed chromaticity vector. And further, the first chromaticity component after dehazing is obtained. Second chromaticity component ; S6, based on Determine the lightness and darkness reference and based on and right Nonlinear recovery is performed to obtain the lightness component after dehazing. ; S7, will , and Convert back to RGB color space to obtain the dehazed image. .

2. The method according to claim 1, characterized in that, In step S1, the white balance correction adopts any one of the following methods: gray-world white balance method, white balance method based on white area detection, and white balance method based on color temperature estimation. And / or, in step S2, the transmittance map Output by a separate transmittance estimation module. The transmittance estimation module employs any one of the following: a transmittance estimation algorithm based on dark channel prior, a transmittance estimation algorithm based on color attenuation prior, a transmittance estimation algorithm based on physical model optimization, or a transmittance estimation algorithm based on deep learning network.

3. The method according to claim 2, characterized in that, In step S2, when the transmittance estimation module uses a transmittance estimation algorithm based on dark channel priors, the coarse transmittance map is generated. Calculate using the following formula: ； in, pixel position The coarse transmittance value at that location, To maintain the fog retention coefficient, For color channel index, In pixel position A local window centered on the center. pixel position Located in color channel pixel values ​​on Atmospheric light in color channels The amount on, This is for calculating the minimum value; And the coarse transmittance map Perform guided filtering or edge-preserving filtering to obtain the transmittance map. .

4. The method according to claim 1, characterized in that, In step S3, the corrected foggy image is... When converting to the CIELab color space, use the D65 white point as the reference white point; And / or, in step S4, the magnitude of the hazy chromaticity vector Calculate using the following formula: ； The hue angle Calculate using the following formula: 。 5. The method according to claim 1, characterized in that, In step S5, the first chromaticity component after dehazing Second chromaticity component Calculate according to the following formulas: ； ； Calculate using the following formula: ； in, pixel position Color recovery gain at that location; Calculate using the following formula: ； in, It is a natural exponential function. For reference transmittance, Limiting transmittance for colorimetric restoration; Determine using the following formula: ； in, This is the lower limit threshold for color recovery, and ; And / or, in step S6, the brightness component after defogging Calculate using the following formula: ； in, Calculate using the following formula: ； For reference transmittance, The limited transmittance used for brightness restoration; Determine using the following formula: ； in, The lower limit threshold for brightness recovery, and .

6. The method according to claim 1, characterized in that, In step S5, the pre-constructed nonlinear mapping relationship between transmittance and chromaticity perception is an exponential mapping relationship between transmittance and chromaticity vector magnitude, and its expression is: ； in, Let be the magnitude of the chromaticity vector. Here, is the fitting coefficient related to the intrinsic chromaticity of the target, and t is the transmittance. The natural exponential function; the fitting coefficients Obtained from experimental data of a standard color chart under controlled fog conditions; And / or, in step S6, the pre-constructed nonlinear mapping relationship between transmittance and perceived lightness is an exponential mapping relationship between transmittance and lightness difference component, the expression of which is: ； in, For the brightness difference component, Here, is the fitting coefficient related to the inherent brightness of the target, and t is the transmittance. The natural exponential function; the fitting coefficients The results were obtained from experimental data of a standard color chart under controlled fog conditions.

7. The method according to claim 1, characterized in that, In step S6, the brightness and darkness reference The backscattered radiation is calculated using the zero-reflection target assumption, where the tristimulus values ​​corresponding to the backscattering satisfy the following: ； in, , , Pixel positions The tristimulus values ​​corresponding to the lightness / darkness baseline. For wavelength, and These are the lower and upper limits of the integration wavelength range, respectively. The spectral power distribution of the incident light source, pixel position Transmittance at that location The backscattering phase function, The backscattering angle, , , This is a CIE standard color matching function; Then, the CIELab conversion function fLab(.) is used to obtain... : ； And / or, in step S7, after the dehazing, the lightness component... The first chromaticity component after defogging and the second chromaticity component after defogging After converting back to the RGB color space, the output pixel values ​​are cropped in color gamut. The cropped output pixel values ​​are... Determine using the following formula: ； in, pixel position Color channel The output value after cropping. pixel position Color channel The output value before cropping. For color channel index, This represents the maximum value of the channel corresponding to the bit depth of the target image. To perform calculations to obtain the larger value, The operation is performed to select the smaller value.

8. An image dehazing and restoration system based on nonlinear mapping of human vision, characterized in that, The system is configured to perform the method according to any one of claims 1 to 7, comprising: The image input and white balance correction module is used to acquire the foggy image to be processed and perform white balance correction to obtain the corrected foggy image. A transmittance estimation module is used to estimate the transmittance map of the corrected foggy image; The color space conversion module is used to convert the corrected hazy image from the RGB color space to the CIELab color space and output the lightness component, the first chromaticity component and the second chromaticity component; the chromaticity recovery module is used to calculate the hue angle, the magnitude of the hazy chromaticity vector and the first and second chromaticity components after dehazing based on the transmittance diagram, the first chromaticity component and the second chromaticity component. The lightness restoration module is used to calculate the lightness dark reference, lightness difference component, and lightness component after dehazing based on the transmittance map and lightness component. The inverse color space conversion module is used to convert the dehazed lightness component, the dehazed first chromaticity component, and the dehazed second chromaticity component back to the RGB color space to obtain the final dehazed image.

9. A computer-readable storage medium having a computer program or instructions stored thereon, characterized in that, When the computer program or instructions are executed by a processor, they implement the method of any one of claims 1-7.

10. A computer program product, comprising a computer program or instructions, characterized in that, When the computer program or instructions are executed by a processor, they implement the method of any one of claims 1-7.

Images