A fire-fighting smoke image sharpening method based on image enhancement

Abstract

This invention discloses a method for sharpening fire smoke images based on image enhancement, belonging to the field of fire image processing technology. Addressing the problems of flame highlights interfering with atmospheric light estimation and the distortion of transmittance estimation edges due to spatial non-uniform variations in smoke density in fire scenes, this invention introduces a partitioned adaptive atmospheric light estimation method based on collaborative perception of the flame-smoke region within an atmospheric scattering model framework. The image is divided into flame, dense smoke, transition, and sparse smoke regions using color-brightness joint features. After eliminating flame highlights, atmospheric light is independently estimated for each region and then weighted and fused to generate a spatially continuously varying atmospheric light distribution map. Simultaneously, a multi-scale adaptive transmittance refinement method driven by smoke density gradients is introduced. Based on the dark channel gradient, the guided filter parameters are adaptively adjusted to maintain edge sharpness at the dense smoke-sparse smoke boundary and ensure smoothness and consistency in uniform regions, suppressing halo artifacts and blocky effects.

Description

Technical Field

[0001] This invention relates to the field of fire protection image processing technology, and in particular to a method for clarifying fire smoke images based on image enhancement. Background Technology

[0002] In fire rescue scenarios, images acquired by on-site video surveillance and image acquisition systems are often severely obscured by heavy smoke, resulting in a sharp decrease in image contrast, color shifts, and significant loss of detail. This significantly hinders the efficiency of fire command and dispatch and personnel search and rescue. Extensive research has been conducted on image restoration under smoke or haze conditions, based on atmospheric scattering models. Among these, the dark channel prior method has achieved relatively ideal restoration results in natural haze scenarios. However, fire scenes and natural haze scenes differ fundamentally—fire scenes simultaneously contain bright flame areas and dense, dark smoke areas, which have diametrically opposed brightness characteristics. Classical dark channel prior methods typically use the brightest pixel globally as the estimation basis when estimating atmospheric light, implicitly assuming that atmospheric light is uniform and constant across the entire image. When this method is directly applied to fire smoke images, the bright pixels in the flame area severely interfere with the estimation results of atmospheric light, causing the restored image to be overexposed and blown out in the flame area while remaining blurry in the smoke area, creating an unbalanced phenomenon that is difficult to balance. Meanwhile, traditional transmittance estimation methods often employ fixed-size local windows when performing minimum filtering or mean calculations, based on the assumption that smoke density is approximately uniform within the local neighborhood. Fire smoke is extremely concentrated near the fire source but rapidly thins out at a distance, and is irregularly turbulently distributed due to thermal disturbances, exhibiting significant spatial non-uniformity. Fixed-window transmittance estimation produces halo artifacts in the transition region between dense and sparse smoke, and blocky effects appear in areas with drastic changes in smoke density gradients, leading to blurred edges in the reconstructed image. These two issues are coupled within the solution chain of the atmospheric scattering model. The accuracy of atmospheric light estimation directly affects the accuracy of transmittance calculation, while the local adaptability of transmittance determines the final reconstruction quality. Therefore, it is urgent to collaboratively improve the atmospheric light estimation and transmittance refinement processes within the framework of the atmospheric scattering model to achieve effective clarity enhancement of fire smoke images. Summary of the Invention

[0003] The purpose of this invention is to provide a method for clarifying fire smoke images based on image enhancement, in order to solve the technical problems of inaccurate atmospheric light estimation caused by the coexistence of flame highlights and dense smoke in fire scenes, and halo artifacts and loss of edge structure caused by non-uniform spatial variation of smoke density in fixed window transmittance estimation.

[0004] To achieve the above objectives, the present invention provides a method for clarifying fire smoke images based on image enhancement, comprising the following steps:

[0005] Preprocessing is performed on the input raw fire smoke image to convert the image from BGR color space to HSV color space to obtain the hue channel, saturation channel and luminance channel. Bilateral filtering is performed on the luminance channel to obtain the denoised luminance channel. Then, the three RGB channels are normalized to obtain the normalized image.

[0006] Flame region detection is performed using color-luminance joint features in HSV space. The detection results are morphologically refined to obtain a flame mask. Dark channel values ​​are calculated for non-flame regions and the grading threshold is adaptively determined based on the cumulative distribution function of the dark channel values. The image is divided into four types of regions: flame region, dense smoke region, transition region, and sparse smoke region, and a region label map is generated.

[0007] When estimating atmospheric light, pixels in the flame region are excluded. In each connected region of the dense smoke region, a preset proportion of pixels with the highest dark channel value are selected to calculate the local atmospheric light value. In the sparse smoke region, a preset proportion of pixels with the highest dark channel value are selected to calculate the atmospheric light value. For the transition region, the atmospheric light values ​​of adjacent regions are fused using an inverse distance weighting method. For the flame region, the atmospheric light value of the nearest non-flame pixel is taken. Finally, the atmospheric light distribution map of the entire image is Gaussian smoothed to obtain the final atmospheric light distribution map.

[0008] The initial transmittance is estimated based on the dark channel prior and the atmospheric light distribution map. The dark channel is calculated and the initial transmittance is obtained by dividing each channel of the normalized image by the atmospheric light value at the corresponding position.

[0009] The gradient magnitude of the dark channel map is calculated and normalized to obtain the smoke density gradient map. Based on the normalized gradient map, the kernel radius and regularization parameters of the guided filter are adaptively calculated. The initial transmittance is applied to multiple discrete scales using the denoised luminance channel as the guide map. The normalization weight is calculated based on the distance between the adaptive parameters and each discrete scale, and the results of each scale are weighted and fused. The lower bound of the fused transmittance is truncated to obtain the refined transmittance.

[0010] The scene irradiance is recovered by inverse calculation of the atmospheric scattering model using the atmospheric light distribution map and refined transmittance. The restored result is then cropped for pixel values, adaptively balanced for color, and enhanced for contrast before the final sharpened image is output.

[0011] As a preferred embodiment of the present invention, the determination rule for flame pixels in flame region detection is as follows: a flame pixel is determined when its hue value is within the range of 0° to 40° and its saturation is greater than a first saturation threshold and its brightness after noise reduction is greater than a first brightness threshold; or a flame pixel is determined when its hue value is within the range of 0° to 40° and its brightness after noise reduction is greater than a second brightness threshold. The first saturation threshold is 0.15, the first brightness threshold is 0.70, and the second brightness threshold is 0.85. This dual-condition determination rule can simultaneously capture typical bright red-orange-yellow flames and the core region of bright but low-saturation white flames, improving the completeness of flame region detection.

[0012] As a preferred embodiment of the present invention, morphological refining includes first performing a kernel size of... The closing operation is used to fill the voids in the flame mask, and then the kernel size is executed. The opening operation is used to remove isolated noise, thereby obtaining a continuous and accurate flame area mask.

[0013] As a preferred embodiment of the present invention, the adaptive threshold for smoke density grading is determined as follows: A histogram of dark channel values ​​in non-flame areas is plotted; the dark channel value corresponding to a cumulative distribution function value of 0.7 is taken as the boundary threshold between the dense smoke zone and the transition zone; and the dark channel value corresponding to a cumulative distribution function value of 0.3 is taken as the boundary threshold between the transition zone and the sparse smoke zone. The introduction of adaptive thresholds allows the grading of zones to adapt to fire scenarios with different smoke concentration distributions.

[0014] As a preferred embodiment of the present invention, the pixels with the highest dark channel values ​​in the dense smoke area and the sparse smoke area are selected as the top 0.1% of pixels, and the weight of the distance-inverse weighted fusion in the transition area is calculated according to the following formula:

[0015]

[0016] in For pixels To the The Euclidean distance of the centroid of the dense smoke region This represents the Euclidean distance to the centroid of the nearest sparsely smoked region. The inverse square weighted distance strategy ensures that spatially closer regions contribute more to atmospheric light in the transition zone, achieving a smooth transition of atmospheric light values ​​at the regional boundaries.

[0017] As a preferred embodiment of the present invention, the initial transmittance is calculated as follows:

[0018]

[0019] in The fog retention coefficient is 0.95. For size Retaining a small amount of smoke information in a local window helps enhance the sense of scene depth in the restored image.

[0020] As a preferred embodiment of the present invention, the mapping method between the kernel radius and the regularization parameter of the adaptive calculation guided filter is as follows:

[0021]

[0022]

[0023] in , , , , It is a non-linear mapping exponent. This is the normalized gradient value. This nonlinear mapping allows for the use of a small kernel and low regularization in dense smoke-sparse smoke boundary regions with sharp gradients to maintain edge sharpness, and a large kernel and high regularization in uniform regions with gentle gradients to ensure regional smoothness.

[0024] As a preferred embodiment of the present invention, the multiple discrete scales are divided into four scale levels, with kernel radii of 5, 10, 20, and 30, respectively, and corresponding regularization parameters of . The normalized weights for each scale in the weighted fusion process are calculated using the following formula: 0.005, 0.02, and 0.08.

[0025]

[0026] in The multi-scale discretization approximation strategy achieves a pixel-by-pixel variation of kernel parameters by using Gaussian kernel weighting, thus striking a balance between computational efficiency and refinement quality.

[0027] As a preferred technical solution of the present invention, the adaptive color balance processing method is as follows: the mean values ​​of the R, G, and B channels of the restored image are calculated respectively. , , Calculate the global average brightness Multiply each channel by a correction factor After eliminating the color cast introduced by smoke, perform a block size adjustment. A contrast-adaptive histogram equalization with a contrast limit threshold of 2.0 is used to further enhance local contrast.

[0028] Compared with the prior art, the present invention has the following beneficial effects:

[0029] This invention employs a zoned adaptive atmospheric light estimation method based on collaborative sensing of the flame and smoke regions. It divides the image into flame, dense smoke, transition, and sparse smoke regions. While eliminating interference from flame specular highlights, it independently estimates local atmospheric light values ​​for each region and generates a spatially continuously varying atmospheric light distribution map through distance-inverse weighted fusion. This fundamentally eliminates the interference of the flame region on atmospheric light estimation, avoiding the imbalance of overexposure in the flame region and insufficient smoke removal in the restored image. Furthermore, this invention utilizes a multi-scale adaptive transmittance refinement method driven by smoke density gradients. Based on the dark channel gradient, it adaptively adjusts the kernel size and regularization parameters of the guided filter, maintaining edge sharpness in the dense smoke-sparse smoke boundary region and ensuring smoothness and consistency in uniform regions. This effectively suppresses halo artifacts and blocky effects in the transition region, improving the edge clarity and structural integrity of the restored image. These two improvements work synergistically in the solution chain of the atmospheric scattering model. Accurate zoned atmospheric light provides a reliable benchmark for transmittance calculation, and adaptive transmittance refinement further ensures the restoration quality across the entire image. Attached Figure Description

[0030] Figure 1 This is a flowchart of a preferred embodiment of the present invention. Detailed Implementation

[0031] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings. It should be noted that the following embodiments are only used to further illustrate the present invention and should not be construed as limiting the scope of protection of the present invention. Any non-essential improvements and adjustments made by those skilled in the art based on the content of the present invention are within the scope of protection of the present invention.

[0032] like Figure 1 As shown, this invention provides a method for sharpening fire smoke images based on image enhancement. The overall processing flow includes six stages: fire smoke image preprocessing, collaborative sensing and partitioning of flame-smoke regions, partitioned adaptive atmospheric light estimation, initial transmittance estimation, smoke density gradient-driven multi-scale adaptive transmittance refinement, and image restoration and post-processing. This method is based on an atmospheric scattering model. The image restoration framework unfolds, in which For the acquired smoke degradation image, For the smoke-free, clear image to be recovered, For scene transmittance, This invention relates to atmospheric light. and transmittance The estimation of the two key parameters introduced improved mechanisms for fire smoke scenarios, forming a collaboratively enhanced recovery link.

[0033] In the preprocessing stage, the input raw fire smoke image is processed. Color space conversion, noise reduction, and normalization are performed sequentially. The image is converted from the BGR color space to the HSV color space, and the tone channels are separated. Saturation channel and brightness channel The hue and saturation channels provide color feature information for subsequent flame area detection. The brightness channel... Apply bilateral filtering and set the spatial standard deviation. , range and standard deviation The luminance channel after noise reduction is obtained. The spatial Gaussian kernel of the bilateral filter controls the range of spatial smoothing, while the value-domain Gaussian kernel assigns different weights based on pixel value differences, thus suppressing random noise interference in smoke while maintaining the edge structure from being blurred. Subsequently, the RGB three channels are normalized by dividing by 255 and mapped to... The interval is denoted as the normalized image. All subsequent calculations based on the atmospheric scattering model are performed within this normalized space.

[0034] The collaborative perception and partitioning of flame-smoke regions is the foundation of the first core improvement of this invention. In the HSV space, preliminary flame region detection is performed using the joint features of flame hue and brightness. The determination of flame pixels follows two parallel conditions; a pixel is marked as a flame pixel if either condition is met: the other is the hue value. Falling within the range of 0° to 40°, and saturation Greater than 0.15, and brightness after noise reduction The first condition, a hue value greater than 0.70, corresponds to a typical bright flame of red, orange, and yellow. The second condition, with a hue value also within the range of 0° to 40° and a brightness greater than 0.85 after noise reduction, is used to capture the core of a white flame with extremely high brightness but low saturation. The selected hue range covers the red to yellow band, consistent with the spectral radiation characteristics of flames. This dual-condition setting allows for the effective detection of flames produced at different combustion stages and with different fuel types. The resulting binary flame mask... Refinement by applying morphological operations: First with The structuring element performs a closing operation to fill small voids within the flame region caused by localized smoke obstruction or uneven sampling, ensuring the flame region remains spatially connected and intact; then... The structuring elements are opened to eliminate isolated false points caused by bright reflections or sensor noise, resulting in a refined flame mask. .

[0035] For pixels in non-flame areas, further smoke density grading is performed. First, the minimum value in the RGB three channels for each pixel is calculated, i.e., the pixel-level dark channel. Then in a size of Local window Local dark channels are obtained by taking the minimum value inside. The dark channel prior states that in a clear, fog-free image, at least one color channel has low-value pixels close to zero in a local region; therefore, the denser the smoke, the higher the dark channel value. Based on this principle, a histogram of dark channel values ​​in non-flame regions is plotted, and its cumulative distribution function is calculated. The dark channel value corresponding to a cumulative distribution function value reaching 0.7 is taken as the threshold. The dark channel value corresponding to a value of 0.3 is used as the threshold. Using these two thresholds as the dividing line, the dark channel value is greater than... Non-flame pixels are zoned into the smoke area, with dark channel values ​​between [value missing]. and The transition zone between them is defined as a dark channel value less than 1. Areas are classified as sparsely smoked zones. An adaptive threshold determination method based on the cumulative distribution function allows the classification criteria to dynamically adjust according to changes in smoke concentration distribution in different fire scenarios, eliminating the need for manually setting fixed thresholds. Combining flame masking and smoke classification results, four types of zone label maps are generated. .

[0036] The first core inventive point of this invention is the partitioned adaptive atmospheric light estimation. Traditional methods estimate a single global atmospheric light scalar value using the brightest global pixel or a few pixels with the highest dark channel values. This strategy is severely ineffective in fire scenarios due to the presence of flame highlights. This invention extends atmospheric light from a global scalar value to a spatially continuously varying distribution map. The specific estimation process is as follows: Pixels in regions labeled "Fire" are completely excluded from atmospheric light estimation, which is the primary measure to eliminate flame highlight interference. In dense smoke regions labeled "Dense," each spatially connected sub-region is processed separately, and the set of pixels with the highest dark channel values ​​in each connected region is selected. Calculate the mean of the RGB values ​​of these pixels in the normalized original image to obtain the local atmospheric light value of the connected region. Simultaneously, the centroid coordinates of the connected region are recorded. The rationale for independent estimation of each connected region is that dense smoke at different spatial locations in a fire scene may originate from different fire sources or different combustible materials, and its scattering characteristics and color tendencies may differ. Uniform estimation would mask these local differences. In the sparse smoke region labeled "Light," the same selection strategy is used to calculate the atmospheric light value. Because the smoke concentration is low in sparse smoke areas, the reliability of the dark channel prior is higher in this area, and the estimated atmospheric light value is closer to the actual atmospheric light level of the scene.

[0037] For pixels in the transition region labeled "Trans", their atmospheric light values ​​are determined using a distance-inverse weighted fusion mechanism. For each transition region pixel... Calculate the Euclidean distance from the centroid of each connected region of dense smoke. and the Euclidean distance to the centroid of the nearest sparsely smoked area. The weights are assigned according to the reciprocal of the square of the distance, i.e. , Similar calculations are performed. The atmospheric light value of pixels in the transition zone is taken as a weighted sum of the atmospheric light from each dense smoke region and the atmospheric light from each sparse smoke region. This fusion strategy ensures that the gradual change in atmospheric light values ​​from dense smoke areas to sparse smoke areas is spatially continuous, without abrupt changes at region boundaries. For flame regions labeled "Fire," although their pixels do not participate in the atmospheric light estimation calculation, atmospheric light reference values ​​still need to be assigned to them in subsequent image restoration. This invention uses the atmospheric light value from its nearest neighboring non-flame pixel. This serves as the atmospheric light reference for that location, ensuring a reasonable atmospheric light baseline for the flame area during reconstruction. After integrating the atmospheric light values ​​from all areas, a standard deviation is applied to each of the three channels. Gaussian smoothing is applied to eliminate potential discontinuities at region label boundaries, resulting in the final atmospheric light distribution map. .

[0038] In the initial transmittance estimation stage, the initial transmittance of each pixel is calculated based on the dark channel prior and using the spatially varying atmospheric light distribution map obtained above. The transmittance of each channel of the normalized image is then divided pixel-by-pixel by the corresponding atmospheric light value, and the minimum value among the RGB channels is taken. The normalized dark channel is obtained by taking the minimum value within the local window. Initial transmittance according to Calculation, where This is the fog retention factor. Setting this factor slightly less than 1 means intentionally retaining about 5% of the smoke information in the restored result. This visually helps maintain the depth and realism of the scene, avoiding an unnatural "too clean" effect after completely removing the smoke. Unlike traditional methods that use global scalar atmospheric light values, this method uses a spatially varying atmospheric light distribution map for normalized dark channel calculations. This ensures that the transmittance estimates in different regions are based on the actual atmospheric light of their respective regions, avoiding systematic errors introduced by atmospheric light deviations.

[0039] The second core inventive point of this invention is the multi-scale adaptive transmittance refinement driven by smoke density gradient. Due to the inherent block effect of the dark channel prior and the limitations of local windowing operations, the initial transmittance map exhibits an insufficiently smooth transition in the boundary region between dense and sparse smoke, and edge details are also somewhat blurred. The conventional approach is to apply a guided filter with fixed parameters to the initial transmittance map for uniform refinement across the entire image. However, the drastic spatial variations in fire smoke density require the filter parameters to be adaptable to local conditions. This invention first constructs a smoke density gradient map as the guiding signal for adaptive control. For the dark channel map... Gradient components are obtained by applying the Sobel operator along the horizontal and vertical directions respectively. and Calculate the gradient magnitude and normalize it to Interval The spatial gradient of the dark channel value directly reflects the local rate of change of smoke density. Locations with large gradient magnitudes correspond to the boundary between dense and sparse smoke or regions where smoke density changes abruptly, while locations with small gradient magnitudes correspond to regions where smoke density is relatively uniform within a local area.

[0040] Based on the normalized gradient map, the kernel radius of the guided filter is adaptively calculated for each pixel position. and regularization parameters The mapping function is , ,in , , , , This mapping has a clear physical meaning: at the dense smoke-dense smoke boundary where the gradient magnitude is large, As the value approaches 1, the mapping result approaches... and This involves performing filtering with a small kernel window and a low regularization coefficient. The small kernel window limits the spatial average range, preventing edges from being blurred by cross-region smoothing. The low regularization coefficient allows the filter to more strictly perform smoothing along the edge structure of the guiding image. In uniform regions with gentle gradients, As the value approaches 0, the mapping result approaches... and This means that filtering is performed with a large kernel window and a high regularization coefficient to ensure the spatial consistency and smoothness of transmittance values ​​within the region. The quadratic nonlinear mapping concentrates parameter changes in the interval with higher gradient values, making it more sensitive to edge regions.

[0041] Since calculating a continuously varying kernel radius pixel-by-pixel would be computationally too demanding in engineering implementation, this invention employs a multi-scale discretization approximation strategy to address this issue. The kernel radius is quantized into four discrete levels. The corresponding regularization parameter is With the noise-reduced luminance channel To guide the image, the initial transmittance was adjusted under these four sets of parameters respectively. Standard guided filtering was performed to obtain four refined transmittance images at different scales. ( The luminance channel was chosen as the guide map because it fully preserves the scene's light and dark structure and edge information. After noise reduction, it has less noise interference and can effectively guide the filter to perform edge-preserving smoothing along the object's edges. The refined results at the four scales each have their own characteristics: the small-scale results maintain a sharp transition at the edges, while the large-scale results provide better smoothing in uniform areas.

[0042] For each pixel, based on its adaptive kernel radius With four discrete levels The distance between them is used to calculate the normalized weights using the Gaussian kernel function. ,in The rate at which the weights decay with distance is controlled. The final refined transmittance is a weighted fusion of results from four scales. The advantage of Gaussian kernel weight allocation is that when the adaptive parameter is exactly equal to a certain discrete level, the weight of that level approaches 1 while the others approach 0, approximating the filtering result of that level directly; when the adaptive parameter falls between two discrete levels, the weights are smoothly distributed between adjacent levels, avoiding jumps at quantization boundaries. To prevent excessively small transmittance values ​​from causing the denominator to approach zero during image restoration and resulting in drastic noise amplification, a lower bound truncation is applied to the refined transmittance. That is, the transmittance is not less than 0.1.

[0043] In the image restoration and post-processing stage, a partitioned adaptive atmospheric light distribution map is used. and refined transmittance Recover scene irradiance using inverse operation of atmospheric scattering model The restored result was first cropped to... The effective pixel value range. Smoke scattering often introduces color cast in images, so adaptive color balancing is performed on the restored image: the mean values ​​of the R, G, and B channels are calculated separately. , , Calculate the global average brightness Multiply each channel by a correction factor ( This process aims to make the three channel means more consistent, eliminating color cast introduced by smoke. Then, contrast-limited adaptive histogram equalization is performed on the restored image, with the block size set to... The contrast limit threshold was set to 2.0 to further enhance the contrast and detail visibility in local areas without producing excessive enhancement artifacts. The final result was multiplied by 255 and rounded to the nearest integer. Sharpened image within range .

[0044] In the technical solution of this invention, the zonal adaptive atmospheric light estimation and the smoke density gradient-driven multi-scale adaptive transmittance refinement work collaboratively within the unified solution framework of the atmospheric scattering model. The former provides an accurate atmospheric light reference for the latter, ensuring that transmittance calculation is not affected by flame highlights; the latter adaptively adjusts the refinement strategy based on the local variation characteristics of smoke density, so that the final restoration result achieves a balanced clarity across the entire image, including the flame region, dense smoke region, transition region, and sparse smoke region. Under typical experimental conditions, this method can improve the peak signal-to-noise ratio by 3 to 5 dB and the structural similarity index by 0.05 to 0.12 compared to traditional dark channel prior methods.

[0045] In practical applications, this invention can be deployed on image processing terminals at fire scenes to perform real-time or near-real-time enhancement processing on video surveillance footage or images captured by drones. This assists fire commanders in quickly grasping the fire situation, identifying the location of trapped personnel and escape routes, and improving the efficiency and safety of firefighting and rescue operations. For fire scenarios of different scales, the regionally graded adaptive threshold mechanism and multi-scale transmittance refinement strategy can automatically adapt without requiring manual parameter adjustments for specific scenarios. Furthermore, each step in the processing flow of this method is based on traditional image processing operators, without relying on the training and inference of deep learning models, and has good deployability even on embedded platforms or mobile terminals with limited computing power.

[0046] The above description is merely a preferred embodiment of the present invention and does not limit the patent scope of the present invention. Any equivalent structural or procedural transformations made based on the content of the present invention's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of the present invention.

Claims

1. A method for clarifying fire smoke images based on image enhancement, characterized in that, Includes the following steps: Preprocessing is performed on the input raw fire smoke image to convert the image from BGR color space to HSV color space to obtain the hue channel, saturation channel and luminance channel. Bilateral filtering is performed on the luminance channel to obtain the denoised luminance channel. Then, the three RGB channels are normalized to obtain the normalized image. Flame region detection is performed using color-luminance joint features in HSV space. The detection results are morphologically refined to obtain a flame mask. Dark channel values ​​are calculated for non-flame regions, and the grading threshold is adaptively determined based on the cumulative distribution function of the dark channel values. The image is divided into four types of regions: flame region, dense smoke region, transition region, and sparse smoke region, and a region label map is generated. When performing atmospheric light estimation, pixels in the flame region are excluded. In each connected region of the dense smoke region, a preset proportion of pixels with the highest dark channel value are selected to calculate the local atmospheric light value. In the sparse smoke region, a preset proportion of pixels with the highest dark channel value are selected to calculate the atmospheric light value. For the transition region, the atmospheric light values ​​of adjacent regions are fused using an inverse distance weighting method. For the flame region, the atmospheric light value of the nearest non-flame pixel is taken. Finally, the atmospheric light distribution map of the entire image is Gaussian smoothed to obtain the final atmospheric light distribution map. Based on the dark channel prior and the atmospheric light distribution map, the initial transmittance is estimated. The dark channel is then calculated by dividing each channel of the normalized image by the atmospheric light value at the corresponding position, and the initial transmittance is then calculated. The gradient magnitude of the dark channel map is calculated and normalized to obtain the smoke density gradient map. Based on the smoke density gradient map, the kernel radius and regularization parameters of the guided filter are adaptively calculated. The initial transmittance is applied to multiple discrete scales using the denoised luminance channel as the guided map. The normalization weight is calculated based on the distance between the adaptive parameters and each discrete scale, and the results of each scale are weighted and fused. The lower bound of the fused transmittance is truncated to obtain the refined transmittance. The scene irradiance is recovered by inverse calculation of the atmospheric scattering model using the atmospheric light distribution map and the refined transmittance. The restored result is then cropped for pixel values, subjected to adaptive color balancing and contrast enhancement, and the final sharpened image is output.

2. The method for clarifying fire smoke images based on image enhancement according to claim 1, characterized in that: The rules for determining flame pixels in the flame region detection are as follows: when the hue value is in the range of 0° to 40° and the saturation is greater than the first saturation threshold and the brightness after noise reduction is greater than the first brightness threshold, it is determined to be a flame pixel; or when the hue value is in the range of 0° to 40° and the brightness after noise reduction is greater than the second brightness threshold, it is determined to be a flame pixel; wherein the first saturation threshold is 0.15, the first brightness threshold is 0.70, and the second brightness threshold is 0.

85.

3. The method for clarifying fire smoke images based on image enhancement according to claim 1, characterized in that: The morphological refining includes first performing a kernel size of... The closing operation is used to fill the voids in the flame mask, and then the kernel size is executed. The opening operation is used to remove isolated noise.

4. The method for clarifying fire smoke images based on image enhancement according to claim 1, characterized in that: The adaptive determination method of the grading threshold is as follows: statistically analyze the dark channel value histogram of the non-flame area, take the dark channel value corresponding to the cumulative distribution function value of 0.7 as the boundary threshold between the dense smoke area and the transition area, and take the dark channel value corresponding to the cumulative distribution function value of 0.3 as the boundary threshold between the transition area and the sparse smoke area.

5. The method for clarifying fire smoke images based on image enhancement according to claim 1, characterized in that: The preset ratio is the first 0.1%, and the weights of the distance-inverse weighted fusion are calculated using the following formula: ; in For pixels To the The Euclidean distance of the centroid of the dense smoke region The Euclidean distance to the centroid of the nearest sparsely smoked region.

6. The method for clarifying fire smoke images based on image enhancement according to claim 1, characterized in that: The initial transmittance is calculated as follows: ; in The fog retention coefficient is 0.

95. For size A local window.

7. The method for clarifying fire smoke images based on image enhancement according to claim 1, characterized in that: The mapping method for the kernel radius and regularization parameter of the adaptive computation guided filter is as follows: ; in , , , , It is a non-linear mapping exponent. This is the normalized gradient value.

8. The method for clarifying fire smoke images based on image enhancement according to claim 1, characterized in that: The multiple discrete scales are divided into four scale levels, with kernel radii of 5, 10, 20, and 30, respectively, and corresponding regularization parameters of , ... 0.005, 0.02, and 0.08; the normalized weights are calculated using the following formula: .

9. The method for clarifying fire smoke images based on image enhancement according to claim 1, characterized in that: The adaptive color balance process involves calculating the average values ​​of the R, G, and B channels of the restored image. , , Calculate the global average brightness Multiply each channel by a correction factor Eliminate color distortion caused by smoke.

10. The method for clarifying fire smoke images based on image enhancement according to claim 1, characterized in that: The contrast enhancement employs a contrast-limited adaptive histogram equalization method, with a block size of [missing value]. The contrast limit threshold is 2.0.

Images