Single image dehazing method based on block-wise nonlinear brightness prior

Abstract

The application discloses a single image defogging method based on block-by-block nonlinear brightness prior, and belongs to the technical field of image processing. The fog-containing image is divided into local blocks, and the block average brightness is calculated. A prior block-by-block monotone increasing nonlinear mapping is constructed to represent the corresponding relationship between the fog-containing and clear block brightness. The atmospheric scattering model is combined, and the atmospheric light vector is modeled as a vector. The parameterized recovery model is composed of the vector and the PPWF. Three scalar parameters are used as the core. Through multi-objective joint optimization, the optimal parameters are obtained by using the alternating optimization and golden section search. Finally, the defogging image is generated. The application has the advantages of few parameters, low complexity and no need of training data. The application can significantly improve the clarity and global contrast of the far and near scenes while maintaining the image structure, effectively suppresses the halo, over-enhancement and color deviation, has good robustness to different fog densities and light conditions, and is suitable for real-time and embedded defogging applications of single-channel or multi-channel images.

Description

Technical Field

[0001] This invention relates to the fields of image processing and computer vision technology, and more specifically, to a single-image dehazing method that combines an atmospheric scattering model, block-by-block monotonic nonlinear brightness mapping, and multi-objective optimization. Background Technology

[0002] Atmospheric impurities in hazy weather degrade image quality, leading to decreased contrast, blurred details, and color distortion. This not only affects the user's visual experience but also hinders intelligent vision tasks requiring high-quality input. Existing dehazing methods still face many challenges: traditional methods based on dark channel priors rely too heavily on prior assumptions, easily producing halos and color casts in the sky or bright areas; deep learning-based methods, while performing well on specific datasets, typically rely on large-scale labeled samples, resulting in insufficient generalization ability and high computational resource consumption; and physical model-based methods, despite their sound theoretical foundation, are often susceptible to noise and texture interference due to unstable estimates of atmospheric light values ​​and transmittance, leading to block artifacts in practical applications. Furthermore, existing image dehazing methods have a fundamental drawback: they consistently ignore the brightness consistency between the blurred image and its dehazing result, resulting in widespread over-enhancement and significant color casts. These shortcomings are particularly pronounced in real-world engineering applications such as drone aerial photography and security monitoring, directly impacting the stability and practical usability of dehazing results. Summary of the Invention

[0003] To address the aforementioned problems in existing technologies, this invention, in its overall design, starts from the spatial characteristics of the impact of fog on image brightness and constructs a constrained parametric dehazing model. Specifically, this invention recognizes that statistical modeling of brightness at a local scale helps to suppress noise and texture interference while maintaining sensitivity to changes in fog distribution. Based on this, by introducing a block-by-block monotonically increasing nonlinear brightness prior, the changing trend of brightness recovery during the dehazing process is explicitly constrained, thereby limiting the solution space of the dehazing result and avoiding over-enhancement and brightness imbalance. Furthermore, by combining the above prior with an atmospheric scattering model and using a small number of global parameters for unified optimization, efficient and stable image dehazing can be achieved while ensuring structural consistency and natural brightness.

[0004] To solve the above problems, the technical solution adopted by the present invention is as follows:

[0005] A single-image dehazing method based on block-by-block nonlinear brightness prior, the method comprising the following steps:

[0006] Step S1: Image acquisition and normalization. Acquire the foggy image I (single-channel or multi-channel) and normalize the pixel intensity to [0,1].

[0007] Step S2: Block division and block mean calculation. Divide I into preset blocks, traverse all local blocks of the hazy image, and calculate the average brightness of each block. To address the issues of image brightness modeling stability and spatial non-uniformity, local blocks use fixed-size non-overlapping or partially overlapping windows. The average brightness of each local block is calculated. It is obtained by the arithmetic mean of the pixels within the window.

[0008] Step S3: Constructing the prior map.

[0009] To address the issues of over-enhancement and brightness imbalance in traditional dehazing methods, a prior block-by-block monotonically increasing nonlinear function (PPWF) is constructed to characterize the mapping relationship between the average brightness of local blocks in a foggy image and the average brightness of local blocks in the corresponding clear image.

[0010] (12)

[0011] in, The coordinates of the local block center are... To correspond to the average brightness of the sharp image within this block, a and b are parameters used to control the amplitude and shape of the fitted curve, respectively. Their derivatives are:

[0012] (13)

[0013] in, for To suppress excessive enhancement and ensure numerical stability, the derivative of PPWF is constrained as follows:

[0014] Step S4: Atmospheric light estimation and parameterization model.

[0015] Traverse the local block and select The largest block is the atmospheric light region; atmospheric light A is modeled as

[0016] (14)

[0017] in, The intensity coefficient of atmospheric light A; It is the color vector of atmospheric light, which consists of three components: red, green, and blue channels; , and These represent the components of the color vector in the red, green, and blue channels, respectively, and their values ​​are the input image values ​​in the corresponding channels. The average brightness value within the local block centered on the luminance.

[0018] Atmospheric scattering model:

[0019] (15)

[0020] in, The pixel values ​​of the hazy image. The pixel values ​​of the dehazed image. Transmittance.

[0021] Combining formulas (12), (14), and (15), and through formula transformation, the coarse transmission map is obtained. and the reflected value of the restored dehazed image They are respectively:

[0022] (16)

[0023] (17)

[0024] Step S5: Multi-objective joint optimization,

[0025] To resolve the conflict between enhancing sharpness and achieving natural brightness, a multi-objective joint optimization strategy is employed to solve for the optimal parameter set. :

[0026] (18)

[0027] Where α, β, and γ are the weight coefficients of each item and are non-negative constants;

[0028] Structural consistency item for:

[0029] (19)

[0030] in, It is the KL divergence, used to measure the consistency of the local average brightness distribution between a hazy and a dehazed image. and These are the probability distributions of the local average brightness values ​​of the foggy and defoggy images, respectively, with the number of bins N ranging from 64 to 128.

[0031] Contrast Enhancement for:

[0032] (20)

[0033] in, It is the Shannon entropy, used to reflect the degree of dispersion of the brightness distribution in a dehazed image;

[0034] Brightness balance item for:

[0035] (twenty one)

[0036] in, This represents the average brightness of the dehazed image.

[0037] Alternating optimization and golden section search are used for a, b, By performing a one-dimensional rotation solution, the optimal parameter set is obtained. .

[0038] The parameter search range is a∈[0,1], b∈[0,1], ∈[0.8,1.2], when iterative updates cause the target to decrease and When crossing the boundary, The parameter remains within this interval; the termination condition is if the parameter change is less than 0.01 (before and after one iteration, (a,b, The maximum absolute change is less than 0.01).

[0039] Step S6: Transmission image calculation and restoration

[0040] Optimal parameter set Substitute the values ​​into formula (16) in step S4 to calculate the coarse transmission map, and then use formula (17) to calculate the reflectance of the dehazed image. Finally, the reflectivity of atmospheric light is used to reconstruct the image, generating the final dehazed image.

[0041] An electronic device includes a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the program, implements the single-image dehazing method based on block-by-block nonlinear brightness prior.

[0042] A computer-readable storage medium having computer instructions stored thereon, which, when executed by a processor, implement the single-image dehazing method based on block-by-block nonlinear brightness prior.

[0043] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0044] This invention solves the problems of over-enhancement and color cast caused by traditional dehazing techniques: existing technologies often ignore the brightness consistency between the hazy and dehazed images, leading to over-enhancement and color cast in the dehazed image. This invention innovatively proposes a block-by-block nonlinear brightness prior, explicitly modeling the monotonically increasing nonlinear mapping relationship between hazy image blocks and clear image blocks. Through this brightness consistency constraint, this invention can adaptively preserve the lighting characteristics and color fidelity of the original scene while achieving image dehazing, fundamentally suppressing over-enhancement and color cast, and significantly improving visual quality (this conclusion is supported by data in comparative experiments, see Table 1).

[0045] This invention significantly reduces model complexity and computational overhead: it reconstructs the complex single-image dehazing problem into a model containing only three scalar parameters. The parameterized recovery model of this invention significantly reduces the solution space of the dehazing problem, eliminating the need for the algorithm to rely on massive deep neural networks or paired training datasets. Combined with the "filter optimization" method proposed in this invention, the algorithm achieves high efficiency while ensuring high-quality output, making it highly suitable for deployment on edge devices with limited computing power, such as drones and security monitoring systems.

[0046] This invention constructs a multi-objective joint optimization function that includes information gain, exposure control, and pixel histogram preservation. This function dynamically balances image detail recovery and noise suppression. It maximizes the extraction of texture details hidden under fog by utilizing information entropy, prevents overexposure in highlights or complete blacking in shadows by using the exposure term, and maintains the consistency of pixel histogram distribution between input and output images by using KL divergence constraints. Furthermore, the guided filtering introduced in the final step not only smooths the transmission image but also effectively eliminates the block effect and halo artifacts caused by block processing, resulting in good performance under different fog concentrations and lighting conditions. Attached Figure Description

[0047] Figure 1 This is a flowchart of a single-image dehazing method based on block-by-block nonlinear brightness prior. Detailed Implementation

[0048] To enhance understanding of the present invention, the embodiments will be described in detail below with reference to the accompanying drawings.

[0049] Example: A single-image dehazing method based on block-by-block nonlinear brightness prior is described in detail below. Figure 1 As shown, the specific steps are as follows:

[0050] Step S1: Image acquisition and normalization.

[0051] Input: A hazy color image I with pixel intensity range [0, 255].

[0052] Preprocessing: Channel normalization from I to I∈ [0,1]; Calculation of luminance components for statistical and guided filtering. :

[0053] (twenty two)

[0054] in, , and These represent the pixel values ​​at the corresponding pixel positions in the red, green, and blue channels of the normalized input image, respectively; subsequent calculations are performed in the floating-point numerical domain [0,1].

[0055] Parameter configuration: Window size w*w = 16×16, step size s = 8; histogram binning N = 64; target weights. =5, =500, =0.01; Guided filter radius r=40, regularization Parameter search range: a∈[0,1], b∈[0,1], ∈[0.8,1.2];

[0056] Step S2: Block partitioning and calculation of block mean.

[0057] Local blocks use fixed-size non-overlapping or partially overlapping windows, calculated by arithmetic mean. Center of each block corresponding The block mean is calculated and interpolated to form a block mean map M of the same size as the original image (for subsequent reuse). The pixel values ​​of M are then statistically analyzed using N bins within the interval [0,1] and normalized to obtain the mean brightness distribution p of the foggy blocks.

[0058] Step S3: Constructing the prior map.

[0059] Based on the principle of minimizing fitting error, PPWF is defined as:

[0060] (twenty three)

[0061] Derivative constraints are used for numerical stability and to suppress over-enhancement:

[0062] (twenty four)

[0063] Step S4: Atmospheric light estimation and parameterization model.

[0064] Select from all blocks Largest block center The atmospheric light vector is estimated based on its location as follows:

[0065] (25)

[0066] Step S5: Multi-objective joint optimization (solving a, b, ),

[0067] Objective function:

[0068] (26)

[0069] Three complementary constraints:

[0070] (27)

[0071] (28)

[0072] (29)

[0073] Atmospheric scattering model:

[0074] (30)

[0075] The parametric image restoration model is obtained by combining and transforming equations (23), (25), and (30):

[0076] (31)

[0077] (32)

[0078] Solution strategy: Alternating one-dimensional golden section search (optimizing a→b→ sequentially) (Keep the rest of the parameters fixed).

[0079] Single candidate (a, b, The search process for ) is as follows:

[0080] 1) Calculate the coarse transmission map: Substitute the current candidate parameters into formula (31) to calculate the coarse transmission map. ;

[0081] 2) Generate coarse dehazing map: Based on formula (30), perform inverse transformation on the atmospheric scattering model to generate the coarse transmission map. Inversion is performed to calculate a coarse estimate of the dehazed image. ;

[0082] 3) Statistical indicators: coarse estimation results for dehazed images The brightness channel calculation block mean map is obtained by performing histogram statistics on N equal-width bins in the interval [0,1], and then normalizing the results to obtain the brightness distribution. The average brightness distribution of the hazy block (Original patch mean distribution) and Calculate the structural consistency term using formulas (27) and (28). Entropy term ;Depend on Calculate the average brightness of the dehazed image The brightness constraint term is calculated using formula (29). The objective function value is obtained by weighting the three terms using formula (26). Comparing different candidates , choose to The largest candidate will be the next direction for updates.

[0083] 4) Termination and Boundary: If the parameter change is less than 0.01 (before and after one iteration, (a,b, If the maximum absolute change is less than 0.01, the search stops; if the search results indicate an out-of-bounds error... And after crossing the boundary If it's lower, then... Fixing to the nearest boundary value (0.8 or 1.2) is equivalent to the optimal solution in that direction falling on the boundary;

[0084] Step S6: Final transmission image and restoration.

[0085] The optimal parameter set output in step S5 Substituting into formula (31) yields the final coarse transmission map. At this point, guided filtering is used for refinement (guide filter takes...). ), to obtain a fine transmission map with well-preserved edges .

[0086] Finally, transmittance ( Substituting into formula (32), we obtain the reflectance of a high-quality dehazed image. The image was reconstructed based on atmospheric light to generate the final dehazed image. The evaluation results are shown in Table 1.

[0087] Table 1 provides a quantitative comparison of dehazing methods for different datasets.

[0088]

[0089] It should be noted that the above embodiments are not intended to limit the scope of protection of the present invention. Equivalent transformations or substitutions made based on the above technical solutions all fall within the scope of protection of the claims of the present invention.

Claims

1. A single-image dehazing method based on block-by-block nonlinear brightness prior, characterized in that, Includes the following steps: Step S1: Image acquisition and normalization. Acquire the foggy image I, where I is a single-channel or multi-channel image, and normalize the pixel intensity to [0,1]. Step S2: Block division and calculation of block mean. Divide I into preset blocks, traverse all local blocks of the foggy image, and calculate the average brightness of each block. Step S3: Prior mapping construction. Construct a prior block-by-block monotonically increasing nonlinear function PPWF to characterize the mapping relationship between the average brightness of local blocks in a foggy image and the average brightness of corresponding clear local blocks in the image. Step S4: Atmospheric light estimation and parameterization modeling. Model atmospheric light and, in conjunction with the atmospheric scattering model (ASM), establish a model based on a, b, ... A parametric image restoration model with three scalars as core optimization variables, where a and b are parameters used to control the amplitude and shape of the fitted curve, respectively. The intensity coefficient of atmospheric light A; Step S5: Multi-objective joint optimization. The optimal parameter set is solved by adopting a multi-objective joint optimization strategy, while considering three constraints: structural consistency, contrast enhancement and brightness balance. Step S6: Transmission image calculation and restoration. Substitute the optimal parameter set into the restoration model from step S4 to calculate the coarse transmission image, and optimize it using guided filtering (GF) to generate the final dehazed image. In step S2, the local blocks use non-overlapping or partially overlapping windows of a fixed size, and the average brightness of each local block... Obtained by the arithmetic mean of the pixels within the window; Based on the curves of individual local patches in clear yet hazy images from the SOTS dataset fitted by MATLAB and 10 representative patch-by-patch fitting curves sampled from the I-Haze dataset, the prior patch-by-patch monotonically increasing nonlinear function PPWF in step S3 is defined as follows, based on the principle of minimizing fitting error: (1) in, The coordinates of the local block center are... The average brightness of the corresponding sharp image within this block; a and b are parameters used to control the amplitude and shape of the fitted curve, respectively, and the derivative of PPWF is: (2) in, for To suppress excessive enhancement and ensure numerical stability, the derivative of PPWF is constrained as follows: .

2. The single-image dehazing method based on block-by-block nonlinear brightness prior as described in claim 1, characterized in that, The atmospheric light A estimation in step S4 includes: Step S41: Select average brightness The largest local block, representing the pure atmospheric light region, has the following coordinates: (3) Step S42: Model atmospheric light A as follows: (4) in, The intensity coefficient of atmospheric light A; It is the color vector of atmospheric light, which consists of three components: red, green, and blue channels; , and These represent the components of the color vector in the red, green, and blue channels, respectively.

3. The single-image dehazing method based on block-by-block nonlinear brightness prior as described in claim 2, characterized in that, The parameterized image restoration model in step S4 is based on the atmospheric scattering model: (5) The coarse transmission pattern is obtained by transforming formulas (1) and (4). and the reflected value of the restored dehazed image The calculations are as follows: (6) (7) in, The pixel values ​​of the hazy image. The pixel values ​​of the dehazed image. For transmittance, CR(·) represents the guided filter operator to optimize the smoothness and edge preservation of the coarse transmission map.

4. The single-image dehazing method based on block-by-block nonlinear brightness prior as described in claim 3, characterized in that, The multi-objective joint optimization in step S5 aims to maximize the following expression: (8) in, The optimal parameter set is defined by α, β, and γ, which are non-negative constants and represent the weighting coefficients of each term. The structural consistency term is also included. for: (9) in, It is the KL divergence, used to measure the consistency of the local average brightness distribution between a hazy and a dehazed image. and These are the probability distributions of the local average brightness values ​​of the foggy and defoggy images, respectively, with the number of bins N ranging from 64 to 128. Contrast Enhancement for: (10) in, It is the Shannon entropy, used to reflect the degree of dispersion of the brightness distribution in a dehazed image; Brightness balance item for: (11) in This represents the average brightness of the dehazed image.

5. The single-image dehazing method based on block-by-block nonlinear brightness prior as described in claim 4, characterized in that, Step S5 is solved using a strategy combining alternating optimization and golden section search, including: Step S51: Initialize the parameter search range a∈[0,1], b∈[0,1], ∈[0.8,1.2], and the initial value of b is set to the midpoint of its search range; Step S52: Fix any two parameters, optimize the third parameter using the golden ratio search, iterate through the three parameters, and search... If the value exceeds [0.8, 1.2] and causes the objective function to decrease, maintain... Within that interval; Step S53: Terminate the iteration when the parameter change is less than 0.01, and output the optimal parameter set. .

6. An electronic device comprising a memory, a processor, and a computer program stored in the memory and executable on the processor, characterized in that, When the processor executes the program, it implements the single-image dehazing method based on block-by-block nonlinear brightness prior as described in any one of claims 1 to 5.

7. A computer-readable storage medium storing computer instructions thereon, characterized in that, When executed by the processor, the computer instructions implement the single-image dehazing method based on block-by-block nonlinear brightness prior as described in any one of claims 1-5.

Images