Infrared smoke screen image enhancement method and system based on continuous extreme value mapping and dynamic range shield

Abstract

The application discloses an infrared smoke screen image enhancement method and system based on continuous extreme value mapping and dynamic range shield, which is used for solving the problems of low target contrast and serious background artifacts under smoke screen interference. The method first performs multi-scale pre-enhancement on the infrared image to obtain a stable input structure; then, a local region extreme value matrix is constructed, and a linear interpolation is performed to reconstruct a pixel-level continuous minimum / maximum extreme value mapping, so that blocky artifacts and spatial discontinuities are eliminated; then, a dynamic range shield mechanism is introduced to constrain the lower limit of the local dynamic range by a preset threshold tau, so that the contour artifact of the low dynamic background such as the sea surface and the sky and noise amplification are suppressed; then, a relaxation parameter and controlled normalization are combined to realize adaptive recovery of the dynamic range of the target region; finally, a weighted smoothing and gradient sparse constraint are performed to output a high-fidelity enhanced image. Without relying on deep learning, the method significantly improves the target contour definition and signal-to-clutter ratio in the smoke screen, effectively suppresses the grid traces, water ripples and over-enhancement artifacts caused by traditional methods, and is suitable for battlefield reconnaissance, sea surface monitoring and infrared guidance and other application scenarios.

Description

Technical Field

[0001] This invention relates to the fields of infrared image processing, infrared target image enhancement, and image restoration under smoke screen interference conditions. Specifically, it relates to an infrared smoke screen image enhancement method and system based on continuous extremum mapping and dynamic range shield. Background Technology

[0002] Infrared thermal imaging relies on the difference in thermal radiation between a target and its background to form an image. It features independence from visible light illumination, strong day-and-night continuous operation, and good adaptability to complex lighting environments. In scenarios such as battlefield reconnaissance, target surveillance, maritime observation, and precision guidance, infrared images are frequently used as an important source of information for target detection and identification.

[0003] However, suspended media such as smoke, haze, and dust can absorb, scatter, and extinct thermal infrared radiation, weakening the radiative contrast between the target and the background in the image. This manifests as reduced overall contrast, slower edge transitions, obscured texture details, blurred target outlines, and compressed local dynamic range. Especially under strong smoke occlusion, the grayscale difference between the target area and the surrounding background is significantly weakened, making it difficult for conventional enhancement methods to simultaneously maintain target saliency, background stability, and artifact suppression.

[0004] Existing infrared image enhancement methods mainly include histogram equalization, contrast-limited adaptive histogram equalization, Retinex enhancement, guided filter layered enhancement, multi-scale detail enhancement, prior constraint recovery, and deep learning enhancement. These methods can improve brightness, local contrast, or edge details in general low-contrast infrared images. However, when directly applied to infrared images with strong smoke screens, they are prone to problems such as obvious smoke residue, insufficient target recovery, synchronous amplification of background noise, and unstable enhancement results.

[0005] Among them, layered filtering methods can decompose infrared images into a base layer and a detail layer, and perform brightness enhancement and texture enhancement respectively; dynamic region extremum prior methods can use the maximum and minimum gray values ​​of local regions to describe the dynamic range of the region and constrain the image enhancement process. These methods themselves have publicly available foundations and can serve as basic processing tools in infrared enhancement. However, in strong smoke-covered thermal infrared scenes, directly combining the two may still generate new engineering problems: on the one hand, if block-level region extrema are directly used as the basis for pixel-level enhancement, discontinuities and abrupt changes can easily occur at region boundaries, forming block artifacts, spatial discontinuities, or mesh traces; on the other hand, the local maximum and minimum values ​​of large-area low-dynamic regions such as the sea surface, sky, and distant backgrounds are relatively small. If this difference is directly used as the denominator for local normalization, weak quantization errors, background noise, or texture disturbances will be abnormally amplified, resulting in contour line artifacts, water ripple artifacts, and amplified background noise. This severely interferes with target interpretation. In addition, conventional enhancement methods lack modeling of smoke screen degradation characteristics, only unilaterally improving local contrast or global brightness, failing to effectively restore the contour details of the occluded target, and the enhancement results are sensitive to parameters and have poor stability, making it difficult to achieve a balance between target salience, background stability and artifact suppression.

[0006] Therefore, although existing infrared image enhancement techniques have a certain effect in general low-contrast scenes, they are still difficult to solve the problems of spatial discontinuity caused by block-level extremum constraints and abnormal gain amplification caused by dynamic range collapse in low dynamic background areas under strong smoke interference conditions. It is necessary to propose an infrared smoke image enhancement method that does not simply rely on existing layered filtering or dynamic region extremum priors. Instead, it is necessary to construct a continuous, stable, and controlled extremum constraint enhancement mechanism to address the problems of local dynamic range collapse, block-level extremum constraint discontinuity, and abnormal gain in low dynamic background in smoke interference images, so as to improve the target recognition in smoke-covered areas while suppressing background artifacts. Summary of the Invention

[0007] The purpose of this invention is to provide an infrared smoke screen image enhancement method and system based on continuous extremum mapping and dynamic range shielding, which can solve the problems of target detail degradation, edge blurring, local dynamic range compression, and block artifacts, contour artifacts, water ripple artifacts, and low dynamic background abnormal gain in infrared images under strong smoke screen interference conditions.

[0008] To achieve the above objectives, the present invention adopts the following technical solution.

[0009] This invention provides a method for enhancing infrared smoke screen images based on continuous extremum mapping and dynamic range shielding, comprising the following steps:

[0010] S1. Obtain the infrared smoke screen image to be processed, perform grayscale and normalization processing to obtain the normalized input image I∈[0,1].

[0011] S2. Multi-scale edge-preserving filtering is used to progressively decompose the normalized input image I. Through step-by-step filtering and subtraction of adjacent layers, the original image is decomposed into a low-frequency base layer and edge and texture detail layers of different scales. Local contrast enhancement is performed on the base layer and different gains are applied to each detail layer. Then, the images are fused to obtain the pre-enhanced image P∈[0,1].

[0012] S3. Based on the pre-enhanced image P, construct dynamic regional extreme value statistics, divide the pre-enhanced image P into multiple local regions Ωi,j, and count the minimum gray value m(i,j) and maximum gray value M(i,j) in each local region, and construct the block-level minimum extreme value matrix m and the block-level maximum extreme value matrix M.

[0013] S4. Perform spatial continuous interpolation reconstruction on the block-level minimum extremum matrix m and the block-level maximum extremum matrix M to generate pixel-level continuous minimum extremum mapping Emin(x,y) and pixel-level continuous maximum extremum mapping Emax(x,y) at the pixel position (x,y) with the same resolution as the pre-enhanced image P.

[0014] S5. Calculate the local dynamic range R(x,y)=Emax(x,y)-Emin(x,y) based on Emin(x,y) and Emax(x,y), and construct the controlled normalization factor: D(x,y)=max(R(x,y), τ), where τ is the lower limit threshold of the local dynamic range, τ>0, which is used to limit the normalization gain in the low dynamic region;

[0015] S6. First, relax and correct Emin(x,y) and Emax(x,y) respectively to obtain the target minimum extreme value mapping E'min(x,y) and the target maximum extreme value mapping E'max(x,y): E'min(x,y)=max(Emin(x,y)-λmin, 0), E'max(x,y)=min(Emax(x,y)+λmax, 1), where λmin≥0 is the minimum extreme value relaxation parameter, λmax≥0 is the maximum extreme value relaxation parameter, and satisfies E'max(x,y)-E'min(x,y)≥δ, where δ is a local positive number or δ=τ;

[0016] Then, based on E'min(x,y), E'max(x,y), and D(x,y), the prior guided enhancement image Q(x,y) is calculated:

[0017] Q(x,y)=clip(((P(x,y)-E'min(x,y)) / D(x,y))·(E'max(x,y)-E'min(x,y))+E'min(x,y), 0, 1).

[0018] S7. The pre-enhanced image P and the prior guided enhancement image Q are smoothly updated to obtain an intermediate updated image. Gradient sparsity constraints are first applied to the intermediate updated image to suppress weak texture noise and local artifact accumulation. Then, frequency domain reconstruction or equivalent stabilization regularization processing is combined to obtain the enhanced infrared smoke image.

[0019] Further, the progressive decomposition in step 2 specifically involves: using the normalized input image I as the 0th layer image G0, and sequentially applying different scale radii r1, r2, ..., r... n The edge-preserving filter smooths the base image of the previous layer, resulting in G. k =F(G k -1, r k , ε k ), and let the k-th scale detail layer D k =G k -1-G k The final base layer is Gn; where F is a guided filter, self-guided filter, bilateral filter, rolling guided filter, or an equivalent filtering method with edge preservation capability, and ε k For smoothing regularization parameters.

[0020] Furthermore, the local region Ωi,j described in step 3 is a square window with a side length of w, where w is an odd number. Adjacent windows are divided using a non-overlapping regular grid or by sliding overlap with a step size s, where s≤w. Boundary regions are filled using mirror filling, edge copying, or truncating window methods.

[0021] Furthermore, the local dynamic range lower limit threshold τ mentioned in step 5 ranges from 0.05 to 0.20, and is adaptively determined based on the image sensor noise level, satisfying... ,in The standard deviation of background or dark field noise. ∈[3,6], This represents the low quantile statistics of the local dynamic range R in a flat background region or the entire image.

[0022] Furthermore, in step S4, the spatial continuity reconstruction employs bilinear interpolation, bicubic interpolation, spline interpolation, edge-preserving interpolation, or other continuous interpolation methods to extend the block-level minimum and maximum value matrices into pixel-level continuous extremum mappings. During interpolation, the center point of each local region is used as the block-level extremum sampling point. For virtual sampling points outside the image edge, edge copying, mirror extension, or nearest neighbor extrapolation are used to determine the boundary extrema.

[0023] Furthermore, the relaxation parameters λmin and λmax described in step 6 range from 0.02 to 0.10, and are adaptively adjusted according to the local smoke density or the gray difference between the target and the background: λmin and λmax are increased in areas with dense smoke or weak target contrast, and λmin and λmax are decreased in areas with flat background.

[0024] Further, the smooth update in step S7 is expressed as: U(x,y)=η·P(x,y)+(1-η)·Q(x,y), where U(x,y) is the intermediate updated image, η is the smooth update weight, and 0<η<1.

[0025] Furthermore, the gradient sparsity constraint in step S7 is used to preserve significant edge gradients and suppress weak texture noise and artifact accumulation; the frequency domain reconstruction is used to achieve stable restoration of the enhanced image while maintaining the main edge structure; the frequency domain reconstruction is based on the gradient consistency estimation of the blur kernel between the current image estimate and the reference image, or it is implemented by using an equivalent frequency domain regularization method that does not explicitly estimate the blur kernel. The reference image is preferably the pre-enhanced image P, or the normalized input image I or the smoothed intermediate image U can be selected according to the application requirements.

[0026] Furthermore, the stabilization regularization process described in step S7 includes one or more of edge-preserving smoothing, iterative amplitude-limiting update, gradient threshold shrinkage, nonlocal mean smoothing, or total variational regularization, used to suppress weak texture noise, contour line perturbation, and local artifact accumulation while preserving the main edge structure.

[0027] In another aspect, this application also proposes an infrared smoke screen image enhancement system based on continuous extremum mapping and dynamic range shield, which includes the following functional modules:

[0028] The image acquisition module is used to acquire the infrared smoke screen image to be processed, perform grayscale and normalization processing, and obtain the normalized input image I∈[0,1].

[0029] The pre-enhancement module is used to progressively decompose the normalized input image I using multi-scale edge-preserving filtering. By filtering step by step and subtracting adjacent layers, the original image is decomposed into a low-frequency base layer and edge and texture detail layers of different scales. Local contrast enhancement is performed on the base layer, and different gains are applied to each detail layer. Then, the images are fused to obtain the pre-enhancement image P∈[0,1].

[0030] The extreme value statistics module is used to construct dynamic regional extreme value statistics based on the pre-enhanced image P, divide the pre-enhanced image P into multiple local regions Ωi,j, and count the minimum gray value m(i,j) and maximum gray value M(i,j) in each local region to construct the block-level minimum extreme value matrix m and the block-level maximum extreme value matrix M.

[0031] The continuous mapping reconstruction module is used to perform spatial continuous interpolation reconstruction on the block-level minimum extremum matrix m and the block-level maximum extremum matrix M to generate pixel-level continuous minimum extremum mapping Emin(x,y) and pixel-level continuous maximum extremum mapping Emax(x,y) at pixel position (x,y) with the same resolution as the pre-enhanced image P.

[0032] The dynamic range shield calculation module calculates the local dynamic range R(x,y)=Emax(x,y)-Emin(x,y) based on Emin(x,y) and Emax(x,y), and constructs a controlled normalization factor: D(x,y)=max(R(x,y), τ), where τ is the lower limit threshold of the local dynamic range, τ>0, which is used to limit the normalization gain in the low dynamic region.

[0033] The extreme value relaxation and enhancement calculation module is used to perform relaxation corrections on Emin(x,y) and Emax(x,y) respectively, to obtain the target minimum extreme value mapping E'min(x,y) and the target maximum extreme value mapping E'max(x,y): E'min(x,y)=max E'max(x,y)=min , where λmin≥0 is the minimum extreme relaxation parameter, λmax≥0 is the maximum extreme relaxation parameter, and satisfies E'max(x,y)-E'min(x,y)≥δ, where δ is a minimum positive number or δ=τ.

[0034] Then, based on E'min(x,y), E'max(x,y), and D(x,y), the prior guided enhancement image Q(x,y) is calculated:

[0035] Q(x,y)=clip .

[0036] The smoothing and stabilization module is used to smoothly update the pre-enhanced image P and the prior guided enhancement image Q to obtain an intermediate updated image; gradient sparsity constraints are first applied to the intermediate updated image, and then combined with frequency domain reconstruction or equivalent stabilization regularization processing to obtain the enhanced infrared smoke screen image.

[0037] The above technical solution of this application first performs grayscale and normalization processing on the infrared smoke image to obtain a stable and controllable input foundation; then, it achieves layered decomposition of the image through multi-scale edge-preserving filtering, performs local contrast enhancement on the low-frequency base layer, and applies differential gain to the high-frequency detail layer, effectively preserving the smoke edge structure while improving the overall visual discernibility, resulting in a pre-enhanced image; subsequently, it constructs local regional extreme value statistics based on the pre-enhanced image to form block-level minimum and maximum extreme value matrices, and generates pixel-level continuous minimum and maximum extreme value mappings with the same resolution as the original image through spatial continuous interpolation reconstruction, thereby eliminating boundary abruptness and spatial discontinuity caused by traditional block-level extreme value division, fundamentally suppressing the generation of block artifacts; based on this, it calculates the pixel-level local dynamic range, Furthermore, a controlled normalization factor is introduced to limit excessive gain in low dynamic range regions with a preset threshold, avoiding contour artifacts, water ripple effects, and noise amplification caused by excessively small dynamic range. Further, by implementing relaxation correction on the extreme value mapping and ensuring the minimum interval between the minimum and maximum values ​​is maintained, a physically reasonable target extreme value boundary is constructed. Based on this boundary and the controlled normalization factor, a priori guided enhancement image is generated, achieving adaptive stretching and constraint of the local grayscale distribution. Finally, the pre-enhanced image and the priori guided enhancement image are smoothly fused, and combined with gradient sparsity constraints, frequency domain reconstruction, or stabilization regularization processing, unnatural abrupt changes and high-frequency oscillations during the enhancement process are further suppressed, achieving clear enhancement of weak target structures in infrared smoke images and synergistic suppression of background noise. Therefore, compared with the prior art, this invention has the following beneficial effects:

[0038] (1) This invention does not simply perform conventional layered filtering enhancement on infrared images, but further constructs continuous extreme value mapping and dynamic range shield on the basis of pre-enhancement results, so that the local dynamic range recovery process of the smoke-covered area is subject to continuous, stable and controlled extreme value constraints, which can improve the recognizability of target contours and details.

[0039] (2) The present invention reconstructs the block-level regional extrema into a pixel-level continuous extrema mapping, which reduces the extrema abrupt change between adjacent regions and can effectively suppress the block artifacts, spatial faults and grid-like traces caused by the traditional block-level extrema direct constraint.

[0040] (3) The present invention constructs a controlled normalization factor by constraining the lower limit of the local dynamic range, and limits the maximum enhancement gain of the low dynamic flat background from the normalization denominator side, which can suppress contour artifacts, water ripple artifacts and noise amplification in background areas such as sea surface and sky.

[0041] (4) This invention combines pre-enhancement, continuous extremum mapping, dynamic range shield, smooth update, gradient sparsity constraint and frequency domain reconstruction, which can achieve a good balance between smoke suppression, target detail preservation and background stability, and is suitable for high-fidelity image enhancement in infrared smoke interference scenarios.

[0042] In summary, this invention can effectively solve the problems of block artifacts and spatial discontinuities caused by block-level extreme value discontinuity in the prior art under strong smoke interference, as well as contour artifacts, water ripples and noise amplification caused by the small dynamic range in low dynamic regions. It achieves the technical effect of significantly improving the structural clarity, contrast continuity and visual stability of infrared images under complex smoke backgrounds, and enhancing the reliability of target detection and recognition.

[0043] This invention is applicable to scenarios where thermal infrared images are degraded due to battlefield smoke, sea surface smoke, haze, dust, or other suspended media. It can be used for target observation, battlefield reconnaissance, sea surface surveillance, precision guidance front-end image preprocessing, and infrared image-assisted interpretation. Attached Figure Description

[0044] Figure 1 This is a schematic diagram of the overall process of the infrared smoke screen image enhancement method of the present invention;

[0045] Figure 2 This is a schematic diagram of the layered pre-enhancement process of the present invention;

[0046] Figure 3 This is a schematic diagram of the continuous extremum mapping reconstruction process of the present invention;

[0047] Figure 4 This is a schematic diagram of the dynamic range shield mechanism of the present invention;

[0048] Figure 5 A comparison of the enhancement effects of different methods on scene 1 of infrared smoke image;

[0049] Figure 6 Comparison of enhancement effects of different methods on infrared smoke screen image scene 2;

[0050] Figure 7 A diagram illustrating the limiting effect of dynamic range shields on local normalized gain.

[0051] Figure 8 Comparison of grayscale response of region of interest (ROI) with and without dynamic range shield on flat background;

[0052] Figure 9 Comparison chart of reconstruction effects for continuous extremum mapping. Detailed Implementation

[0053] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0054] It should be noted that the designations "S1", "S2", etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having", and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0055] Existing infrared image enhancement methods, while capable of improving image contrast in scenes with strong smoke interference through histogram equalization, layered filtering, or local extremum constraints, have failed to effectively address block artifacts and spatial discontinuities caused by block-level extremum statistics, as well as contour artifacts, water ripple artifacts, and noise amplification issues arising from the limited local dynamic range of low-dynamic background regions. While enhancing target details, these methods often simultaneously amplify background noise and non-physical textures, leading to image distortion, reduced target discernibility, and an inability to achieve a balance between target restoration and background stabilization.

[0056] The term "block artifact" refers to rectangular area steps, grid marks, or local abrupt changes in brightness caused by block-level processing or block boundary discontinuities; "contour artifact" refers to ring-shaped, band-shaped, or layered gray-level faults formed in a low-dynamic flat background due to gray-level quantization or overstretching; "water ripple artifact" refers to periodic or quasi-periodic low-frequency ripples, oscillating textures, or strip-shaped disturbances appearing in the background of sea, sky, or smoke; and "spatial fault" refers to visual hierarchical breaks caused by discontinuities in gray-level or extreme value constraint fields between adjacent local areas.

[0057] Existing technologies for modeling local dynamic range suffer from two structural flaws. First, by directly extracting extrema from fixed-size discrete blocks (such as 15×15 or 25×25 windows) and applying them to pixel-level enhancement, the extremum constraint field exhibits discontinuous, blocky jumps in space. Abrupt changes in extrema between adjacent regions directly map to visual blocky artifacts and spatial discontinuities. Second, in smoke-covered and flat background areas, the true local dynamic range (Emax−Emin) approaches zero, yet existing methods still directly use this difference in extrema as the normalization denominator. This amplifies weak noise, quantization errors, or low-frequency disturbances, resulting in contour line and ripple artifacts. These two factors together constitute the fundamental contradiction of spatial discontinuity and amplified dynamic range collapse. That is, while relying on local extremum constraints to improve target contrast, the discreteness of the constraint method and the lack of protection for the denominator introduce more severe artifacts, leading to the paradox of increasing distortion with enhancement.

[0058] Based on this, this invention proposes a scheme to transform discrete block-level extremum constraints into continuous pixel-level mappings and implement a dynamic range lower limit protection scheme for the normalized denominator: instead of directly using block-level extrema as the basis for pixel enhancement, it generates continuously changing pixel-level minimum extremum mappings Emin(x,y) and maximum extremum mappings Emax(x,y) by spatially continuous interpolation reconstruction of the block-level minimum and maximum extremum matrices, thereby eliminating step abrupt changes at block boundaries and realizing the spatial continuity of extremum constraints; at the same time, it no longer allows the normalized denominator to approach zero unconstrainedly with the extremum difference, but forcibly introduces a controlled lower limit threshold τ, constructing a controlled normalization factor D(x,y)=max(R(x,y), τ), so that the normalized gain in the low dynamic region is limited to an acceptable range, fundamentally suppressing the abnormal amplification of noise and artifacts. This scheme is the first to combine spatial continuity reconstruction and denominator dynamic range shield as a pair of complementary mechanisms for infrared smoke image enhancement. It not only retains the ability of local extremum priors to guide the recovery of target contrast, but also breaks the vicious cycle of enhancement artifacts in traditional methods through the dual constraints of continuity and protection.

[0059] See Figure 1 The following describes in detail the infrared smoke screen image enhancement method based on continuous extremum mapping and dynamic range shield of the present invention through a series of embodiments.

[0060] S1. Obtain the infrared smoke screen image to be processed, and perform grayscale and normalization processing on the infrared smoke screen image to obtain a normalized input image. If the input image is a single-channel thermal infrared grayscale image, normalization processing is performed directly.

[0061] In one embodiment, the input image is a single thermal infrared grayscale image acquired under smoke interference conditions. First, the input image is converted to a grayscale image and normalized to the [0,1] interval to obtain a normalized input image I. If the input image is already a single-channel infrared grayscale image, normalization is performed directly.

[0062] In this embodiment, acquiring the infrared smoke image to be processed means receiving the raw image data output from the infrared imaging device. This data reflects the thermal radiation distribution and is affected by smoke interference, resulting in reduced contrast and degraded details. Subsequently, grayscale processing is performed, which converts the raw image into a single-channel grayscale representation to eliminate interference from multispectral or color components and ensure that subsequent processing is based on a unified radiation intensity dimension. Normalization processing then uniformly scales the grayscale values ​​to the [0,1] range, giving the image pixel values ​​a standardized numerical range. This provides a consistent numerical basis for subsequent extreme value statistics and enhancement calculations, avoiding calculation deviations or overflows caused by differences in the range of the raw data.

[0063] S2, see also Figure 2 The normalized input image I is pre-enhanced to obtain a pre-enhanced image P containing background structure and target details. This pre-enhancing improves the initial visibility of the smoke image and provides a structurally stable input for subsequent extreme value constraint enhancement. After pre-enhancing, the fusion result is subjected to amplitude limiting or linear renormalization to keep the grayscale values ​​of the pre-enhanced image P within the range of [0,1]. If local enhancement causes a small number of pixels to exceed the limit, amplitude limiting is applied first, followed by subsequent extreme value statistics and normalization calculations to avoid extreme value mapping distortion or normalization calculation failure due to grayscale exceeding the limit.

[0064] The progressive decomposition specifically involves: using the normalized input image I as the 0th layer image G0, and sequentially applying edge-preserving filters with different scale radii r1, r2, ..., rn to smooth the previous layer base image, resulting in Gk = F(Gk-1, rk, εk). The k-th scale detail layer Dk is then set to Gk-1 - Gk, ultimately forming the base layer Gn. Here, F represents guided filtering, self-guided filtering, bilateral filtering, rolling guided filtering, or an equivalent filtering method with edge-preserving capabilities, and εk is the smoothing regularization parameter. Through the above-mentioned progressive filtering and subtraction of adjacent layers, the original image can be decomposed into a low-frequency base layer and edge and texture detail layers of different scales.

[0065] In one embodiment, pre-enhancing the normalized input image I includes: progressively decomposing the normalized input image using multi-scale edge-preserving filtering to obtain a base layer and detail layers at multiple scales; performing local contrast enhancement on the base layer; applying different gains to detail layers at different scales; and fusing the enhanced base layer and the enhanced detail layers to obtain the pre-enhanced image.

[0066] Specifically, a multi-scale self-guided filtering method is used to progressively decompose the image, resulting in a base layer B and multiple detail layers D1, D2, and D3. The base layer B characterizes the low-frequency background structure and overall brightness distribution, while the detail layers D1, D2, and D3 characterize target edges, textures, and local grayscale variations. Optionally, the radii of the multi-scale filtering windows are set to r1, r2, and r3, respectively, with r1... <r2<r3。

[0067] Next, contrast-limited adaptive histogram equalization is performed on the base layer B to enhance local contrast and suppress excessive noise amplification in flat areas. Different linear gains are applied to detail layers at multiple scales, with larger gains used for smaller scale detail layers to enhance target edges and textures, and smaller gains used for larger scale detail layers to avoid over-enhancing background undulations. The enhanced base layer and the enhanced detail layers are then reconstructed according to fusion weights to obtain the pre-enhanced image P.

[0068] The fusion weights can be fixed or adaptive. In the fixed case, the base layer weight is preferably 0.35 to 0.60, and the total weight of the detail layer is preferably 0.40 to 0.65. In the adaptive case, the weights can be determined based on local gradient, local variance, or noise estimation results. The detail layer weights are increased at target edges and textured regions, while the base layer weights are increased in flat backgrounds or areas with strong noise. After fusion, P is amplitude-limited or renormalized to ensure P∈[0,1].

[0069] In this embodiment, multi-scale edge-preserving filtering is first applied to the normalized input image I. Through a step-by-step filtering approach, the image is decomposed into multiple levels of structural components. Each filtering step uses edge-preserving filters of different scales for smoothing, gradually stripping away high-frequency details while preserving low-frequency background structures. The differences between adjacent layers are calculated and extracted as detail layers, forming a layered representation composed of a low-frequency base layer and multiple scale detail layers. Subsequently, local contrast enhancement is applied to the base layer to improve its overall brightness distribution discrimination ability; simultaneously, different gain values ​​are applied to each scale detail layer to differentiate and enhance edge and texture information at different scales. Finally, the enhanced base layer and the enhanced detail layers are fused to generate a pre-enhanced image P, ensuring that its grayscale values ​​are constrained within the range of [0,1] to maintain numerical consistency in subsequent processing.

[0070] S3. Construct local region extreme value statistics based on the pre-enhanced image P: Divide the pre-enhanced image P into multiple local regions Ωi,j according to a preset window size, calculate the minimum gray value m(i,j) and maximum gray value M(i,j) in each local region, and form the block-level minimum extreme value matrix m and the block-level maximum extreme value matrix M, respectively.

[0071] The local region Ωi,j can be divided using a regular grid, or it can be adaptively divided based on edge strength, target region, local variance, or local complexity.

[0072] In one embodiment, the local region Ωi,j is preferably a square statistical window with side length w, where w is an odd-numbered pixel side length, such as 15, 21, 25, 31, or an adaptively selected window side length. Adjacent windows can be divided using a non-overlapping regular grid or by sliding overlap with a step size s, where s ≤ w. For regions at the image boundary that are less than a complete window, mirror filling, edge duplication, or window truncation are used to supplement the statistical range, ensuring that all pixels are included in at least one local region.

[0073] In this embodiment, dynamic regional extremum statistics are constructed based on the pre-enhanced image P. The image needs to be divided into multiple local regions Ωi,j, each serving as an independent statistical unit for extracting its minimum and maximum grayscale values. By traversing and comparing the grayscale values ​​of all pixels within each local region Ωi,j, the lowest and highest grayscale responses are determined, forming a set of discrete extremum pairs. These extremum pairs are organized according to their region location, forming two matrices that correspond one-to-one with the local region grid structure: a block-level minimum extremum matrix m and a block-level maximum extremum matrix M. These matrices record the minimum and maximum grayscale constraints corresponding to each local region, respectively. This process completes the discrete mapping from the image space to the block-level extremum matrix, directly providing the original extremum distribution data based on the local regions for subsequent extremum constraint enhancement.

[0074] S4, see also Figure 3 The block-level minimum and maximum extremum matrices are spatially continuous to reconstruct the image, generating pixel-level continuous minimum and maximum extremum mappings consistent with the input image resolution. This spatial continuous reconstruction employs bilinear interpolation, bicubic interpolation, spline interpolation, edge-preserving interpolation, or other continuous interpolation methods to expand the block-level minimum and maximum matrices into pixel-level continuous extremum mappings. During interpolation, the center point of each local region is used as the block-level extremum sampling point. For virtual sampling points outside the image edges, edge replication, mirror extension, or nearest-neighbor extrapolation are used to determine boundary extrema, avoiding non-physical grayscale stretching or collapse in edge regions due to the lack of adjacent blocks. Through this continuous extremum mapping, the local region extremum statistics are spatially transformed from discrete block constraints to continuous constraints, thereby reducing block effects, spatial discontinuities, and grid-like artifacts caused by abrupt changes in block boundaries.

[0075] In one embodiment, to avoid abrupt block boundary changes when block-level extrema directly affect pixels, bilinear interpolation is performed on the block-level minimum extremum matrix m and the block-level maximum extremum matrix M to obtain continuous minimum extremum mappings Emin(x,y) and continuous maximum extremum mappings Emax(x,y) with the same resolution as the pre-enhanced image P, respectively. This processing transforms the regional extremum prior from discrete block-level constraints to pixel-level continuous constraints, thereby improving the spatial continuity of the constraint field.

[0076] After obtaining the continuous extremum mapping, the local dynamic range at pixel position (x,y) is calculated as R(x,y) = Emax(x,y) - Emin(x,y). When R(x,y) is too small, directly using R(x,y) as the normalization denominator will lead to excessively high enhancement gain in low dynamic background regions. Therefore, a lower limit threshold τ for the local dynamic range is introduced, and a controlled normalization factor D(x,y) = max(R(x,y),τ) is constructed. Here, τ can be determined based on the image noise level, background flatness, sensor depth, or smoke density. For normalized grayscale images, τ is preferably between 0.05 and 0.20; alternatively, it can be calculated as follows: Adaptive determination, where The standard deviation of background or dark field noise. The preferred values ​​are 3 to 6. This represents the low quantile statistics of the local dynamic range R over a flat background region or the entire image, such as the 5th to 20th percentiles. For b-bit infrared data, where... ,in, This is the quantization step size multiple, used to prevent quantization errors from being amplified.

[0077] In this embodiment, the minimum and maximum gray values, originally statistically analyzed on a discrete region basis, are extended to every pixel location with the same resolution as the pre-enhanced image through interpolation, thereby forming continuously changing minimum and maximum extremum mappings. This process, while preserving the block-level extremum statistical trend and sampling point constraints, generates a smoothly transitioning pixel-level extremum distribution between adjacent block-level extremum points through interpolation operations. This transforms the extremum constraints, originally defined only at the center point of a local region, into a continuous minimum and maximum extremum corresponding to each pixel in the image. This operation directly realizes the mapping transformation from discrete block-level extrema to continuous pixel-level extrema, reducing abrupt changes or steps in the spatial extremum constraints upon which subsequent enhancement processing is based.

[0078] S5, see also Figure 4The local dynamic range R(x,y) = Emax(x,y) - Emin(x,y) is calculated based on the pixel-level continuous maximum extremum mapping Emax(x,y) and the pixel-level continuous minimum extremum mapping Emin(x,y). A lower limit constraint is set on the local dynamic range, and a controlled normalization factor D(x,y) = max(R(x,y), τ) in the form of a dynamic range shield is constructed. Here, Emax(x,y) is the pixel-level continuous maximum extremum mapping at pixel position (x,y), Emin(x,y) is the pixel-level continuous minimum extremum mapping at pixel position (x,y), τ is the lower limit threshold of the local dynamic range, and D(x,y) is the controlled normalization factor in the form of a dynamic range shield. By using the aforementioned lower threshold, when Emax(x,y)-Emin(x,y) is less than τ, τ is used to replace the original local dynamic range as the normalization denominator, thereby limiting the maximum stretching gain of the low dynamic range; when Emax(x,y)-Emin(x,y) is not less than τ, the original local dynamic range is still retained, so that the target area covered by the smoke can obtain the necessary contrast recovery.

[0079] The so-called dynamic range shield refers to a protection mechanism that replaces the original local dynamic range denominator with a controlled denominator of no less than a preset threshold τ at pixel locations where the local dynamic range is too small. Its function is to limit the maximum normalized gain of low dynamic flat background areas without changing the trend of continuous extreme value mapping space, and to prevent noise, quantization error and weak texture from being excessively amplified.

[0080] In this embodiment, the difference between the values ​​of the continuous minimum and maximum extrema mappings at each pixel location is calculated to obtain the local dynamic range. This difference reflects the grayscale variation capability of the local region where each pixel is located in the image. To avoid abnormally high enhancement gain in regions with excessively small local dynamic ranges due to the normalization denominator approaching zero, a lower limit threshold greater than zero is introduced. The local dynamic range is compared with this threshold, and the larger of the two values ​​is taken as the final normalization factor. This ensures that the normalization denominator at any location is not lower than the preset threshold, thereby directly limiting the overstretching effect that may occur in low dynamic regions during enhancement.

[0081] S6. Based on the pixel-level continuous extremum mapping and the controlled normalization factor, the pre-enhanced image is dynamically enhanced by regional extremum constraints to obtain a priori guided enhanced image.

[0082] In one embodiment, relaxation parameters are set for the continuous minimum and maximum extremum mappings to obtain the target minimum and maximum extremum mappings, and then the prior guided enhancement image is calculated based on the target extremum mappings and the controlled normalization factor.

[0083] The relaxation parameters λmin and λmax represent the allowable expansion of the lower and upper bounds of the local grayscale, used to compensate for the dynamic range compression caused by smoke occlusion. They can be set as global constants or adaptively adjusted according to smoke concentration, local variance, target brightness, or local dynamic range. Within the normalized grayscale range [0,1], λmin and λmax are preferably taken as 0 to 0.20, more preferably as 0.02 to 0.10; they can be appropriately increased when the smoke is denser and the difference between the target and the background is weaker, and should be decreased when the background noise is strong or there are many flat areas.

[0084] The target minimum and target maximum mappings satisfy:

[0085] E'min(x,y)=max(Emin(x,y)-λmin, 0)

[0086] E'max(x,y)=min(Emax(x,y)+λmax, 1)

[0087] Where λmin is the minimum extremum relaxation parameter, λmax is the maximum extremum relaxation parameter, E'min(x,y) is the target minimum extremum mapping, and E'max(x,y) is the target maximum extremum mapping. Both λmin and λmax are non-negative, and max(·,0) and min(·,1) ensure that the target extremum mapping remains within the range [0,1]. To prevent abnormal interpolation or parameter settings from causing the target extremum order to become invalid, E'max(x,y)-E'min(x,y)≥δ is further constrained before calculation, where δ is a very small positive number or δ=τ; if a pixel does not meet this condition, the original pre-enhanced image P(x,y) is retained, or the target extremum difference is reset to δ before normalization.

[0088] In one embodiment, the priori guidance enhancement image can be calculated in the following form:

[0089] Q(x,y)=clip(((P(x,y)-E'min(x,y)) / D(x,y))·(E'max(x,y)-E'min(x,y))+E'min(x,y), 0, 1)

[0090] Where P(x,y) is the gray value of the pre-enhanced image at pixel position (x,y), Q(x,y) is the gray value of the prior-guided enhanced image at pixel position (x,y), and clip(·,0,1) indicates that the gray value is truncated to the range [0,1]. Since P(x,y), E'min(x,y), E'max(x,y), and D(x,y) have all been constrained in their value ranges, Q(x,y) is usually kept near [0,1]. The clip operation is mainly used to suppress gray overshoot of a small number of abnormal pixels. To reduce the information loss caused by truncation, percentile linear compression, soft clipping, or prior-guided enhancement intensity reduction strategies can be used before clipping to ensure that most pixels fall within the effective gray range.

[0091] In this embodiment, relaxation corrections are first applied to the continuous minimum extremum mapping Emin(x,y) and the continuous maximum extremum mapping Emax(x,y). The target minimum extremum mapping E'min(x,y) is obtained by subtracting the non-negative minimum extremum relaxation parameter λmin from Emin(x,y) while ensuring the result is not less than 0. Simultaneously, the target maximum extremum mapping E'max(x,y) is obtained by adding the non-negative maximum extremum relaxation parameter λmax to Emax(x,y) while ensuring the result is not greater than 1. This actively expands the allowable range of the local grayscale upper and lower bounds without altering the extremum mapping spatial trend. Subsequently, the difference between the target maximum extremum mapping and the target minimum extremum mapping is constrained. The threshold value is not less than a minimum positive number δ or equal to the lower limit threshold τ of the dynamic range shield, to ensure that there is an effective dynamic range at any pixel position for subsequent normalization processing. Based on this corrected target extremum mapping and the controlled normalization factor D(x,y), a linear stretching transformation is performed on the pre-enhanced image P(x,y), and the gray values ​​of P(x,y) are proportionally mapped according to the interval from E'min(x,y) to E'max(x,y). D(x,y) is used as the normalization denominator, and the final result is restricted to the range [0,1] by a truncation operation. This achieves prior-guided enhancement of the pre-enhanced image under the constraint of the target extremum, ensuring that the enhancement process is controlled by the local extremum boundary after relaxation correction and range constraint.

[0092] This processing method can restore the dynamic range to a certain extent in areas obscured by dense smoke, while the maximum stretching gain of low-dynamic background areas such as the sea and sky is limited.

[0093] S7. The prior guided enhancement image is smoothly updated, and combined with gradient sparsity constraints, frequency domain reconstruction, or stabilization regularization processing, an enhanced infrared smoke screen image is obtained. The stabilization regularization processing includes one or more of edge-preserving smoothing, iterative amplitude-limiting update, gradient threshold shrinkage, nonlocal mean smoothing, or total variational regularization, which are used to suppress weak texture noise, contour line perturbation, and local artifact accumulation while preserving the main edge structure.

[0094] In one implementation, the smooth update is performed using a weighted approach to avoid image oscillation caused by the one-time application of prior constraints. Specifically, the enhancement result Q is weighted and updated with the pre-enhanced image P to obtain an intermediate image U. The smooth update can be expressed as: U(x,y)=η·P(x,y)+(1-η)·Q(x,y), where U(x,y) is the intermediate updated image, η is the smooth update weight, and 0<η<1. η can be a fixed value or an adaptive value: when fixed, it is preferably 0.4 to 0.8; when adaptive, it can be determined based on the local gradient magnitude, local variance, or smoke confidence. In flat backgrounds and areas with strong noise, η is increased to retain more of the pre-enhanced image P, while η is decreased in areas with target edges, obvious textures, or smoke-covered areas to enhance the effect of the prior guiding result Q. η is used in conjunction with subsequent gradient sparsity constraints; the former controls the update amplitude, while the latter suppresses weak gradient noise and artifact accumulation.

[0095] In one implementation, after smoothing the update, an approximate sparse constraint is applied to U in the gradient domain to preserve significant edge gradients and suppress weak texture noise, contour perturbations, and local artifact accumulation.

[0096] In one embodiment, the final enhancement result is obtained by combining frequency domain reconstruction, which is used to achieve stable restoration of the enhanced image while preserving the main edge structures. Specifically, frequency domain reconstruction can estimate the blur kernel based on the gradient consistency between the current image estimate and the reference image, and then reconstruct and update the image in the frequency domain; alternatively, stable restoration can be achieved by using an equivalent frequency domain regularization method without explicitly estimating the blur kernel. After reconstruction, the image grayscale is truncated to the [0,1] interval, and the enhanced infrared smoke image is output.

[0097] In one embodiment, the reference image is preferably a pre-enhanced image P, but a normalized input image I or a smoothed intermediate image U can also be selected according to application needs. When P is used as the reference image, the final result can be constrained not to deviate from the original thermal radiation distribution while preserving the pre-enhanced structural information.

[0098] In the above implementation methods, the multi-scale pre-enhancement can employ three scale windows, and the detail layer gain can be determined based on the image noise level and target texture intensity. The size of the local region statistical window refers to the side length w of the square window used for extreme value statistics, in pixels. w can be set to fixed odd values ​​such as 15, 21, 25, and 31, or it can be adaptively adjusted according to the image size and local complexity. If a sliding window statistics method is used, the step size s can be w, w / 2, or other integer values ​​not greater than w. The lower limit threshold τ of the local dynamic range can be set to a value in the range of 0.05 to 0.20, or it can be adaptively determined based on the background region grayscale variance, sensor noise level, sensor bit depth, or local dynamic range distribution. The above parameters are not limitations of the present invention, and can be adjusted according to different infrared imaging devices, smoke density, target type, and application scenarios in practical applications.

[0099] Through the above implementation methods, the present invention can further perform controlled recovery of the smoke-covered area on the basis of ordinary pre-enhancement, and suppress abnormal gain in low dynamic background areas through continuous extreme value mapping and dynamic range shielding, thereby obtaining an infrared smoke-enhanced image with high target clarity, relatively stable background, and fewer block artifacts and contour artifacts.

[0100] The following provides a further embodiment illustrating an infrared smoke screen image enhancement system based on continuous extremum mapping and dynamic range shield. The system includes an image acquisition module, a pre-enhancement module, an extremum statistics module, a continuous mapping reconstruction module, a dynamic range shield calculation module, an extremum relaxation and enhancement calculation module, and a smoothing and stabilization module. The modules work together to implement the method described in the foregoing embodiment.

[0101] The image acquisition module is used to acquire the infrared smoke screen image to be processed, perform grayscale and normalization processing, and obtain the normalized input image I∈[0,1].

[0102] The pre-enhancement module is used to progressively decompose the normalized input image I using multi-scale edge-preserving filtering. By filtering step by step and subtracting adjacent layers, the original image is decomposed into a low-frequency base layer and edge and texture detail layers of different scales. Local contrast enhancement is performed on the base layer, and different gains are applied to each detail layer. Then, the images are fused to obtain the pre-enhancement image P∈[0,1].

[0103] The extreme value statistics module is used to construct dynamic regional extreme value statistics based on the pre-enhanced image P, divide the pre-enhanced image P into multiple local regions Ωi,j, and count the minimum gray value m(i,j) and maximum gray value M(i,j) in each local region to construct the block-level minimum extreme value matrix m and the block-level maximum extreme value matrix M.

[0104] The continuous mapping reconstruction module is used to perform spatial continuous interpolation reconstruction on the block-level minimum extremum matrix m and the block-level maximum extremum matrix M to generate pixel-level continuous minimum extremum mapping Emin(x,y) and pixel-level continuous maximum extremum mapping Emax(x,y) at pixel position (x,y) with the same resolution as the pre-enhanced image P.

[0105] The dynamic range shield calculation module calculates the local dynamic range R(x,y)=Emax(x,y)-Emin(x,y) based on Emin(x,y) and Emax(x,y), and constructs a controlled normalization factor: D(x,y)=max(R(x,y), τ), where τ is the lower limit threshold of the local dynamic range, τ>0, which is used to limit the normalization gain in the low dynamic region.

[0106] The extreme value relaxation and enhancement calculation module is used to perform relaxation corrections on Emin(x,y) and Emax(x,y) respectively, to obtain the target minimum extreme value mapping E'min(x,y) and the target maximum extreme value mapping E'max(x,y): E'min(x,y)=max E'max(x,y)=min , where λmin≥0 is the minimum extreme relaxation parameter, λmax≥0 is the maximum extreme relaxation parameter, and satisfies E'max(x,y)-E'min(x,y)≥δ, where δ is a local positive number or δ=τ;

[0107] Then, based on E'min(x,y), E'max(x,y), and D(x,y), the prior guided enhancement image Q(x,y) is calculated:

[0108] Q(x,y)=clip .

[0109] The smoothing and stabilization module is used to smoothly update the pre-enhanced image P and the prior-guided enhanced image Q to obtain an intermediate updated image. Gradient sparsity constraints are first applied to the intermediate updated image, and then combined with frequency domain reconstruction or equivalent stabilization regularization processing to obtain the enhanced infrared smoke screen image.

[0110] To further illustrate the technical effectiveness of the continuous extremum mapping and dynamic range shield mechanism described in this invention in infrared smoke image enhancement, typical thermal infrared smoke interference images are selected for experimental verification. The experimental scenarios include smoke interference scenarios involving ship targets on the sea surface and smoke interference scenarios involving plate-shaped targets on mountains. The images simultaneously contain smoke-covered targets, dark backgrounds, localized smoke areas, and low-dynamic-range flat background areas, effectively reflecting typical problems such as target edge blurring, local dynamic range compression, and background artifact amplification under strong smoke conditions.

[0111] The experimental comparison methods included the original image, the CLAHE method, the traditional guided filtering enhancement method, the method without dynamic range shield, the method without continuous extremum mapping, and the method of this invention. All methods were run under the same input image and computational environment. In the experiments, multi-scale pre-enhancement used three scale windows with radii of 8, 16, and 32, respectively, and a smoothing regularization parameter of 0.01; the corresponding detail layer gain coefficients were set to 3.8, 2.5, and 1.0, respectively; the fusion weights of the base layer and detail layer were set to 0.45 and 0.55, respectively; the local region statistical window size was set to 25; and the lower limit threshold τ of the dynamic range shield was set to 0.15.

[0112] I. Analysis of Subjective Enhancement Effect

[0113] Figure 5 and Figure 6 The processing results of different enhancement methods in two types of infrared smoke interference scenarios are presented. Figure 5 and Figure 6 In the image, (a) is the original infrared image, (b) is the CLAHE method, (c) is the traditional guided filtering method, (d) is the method without dynamic range shield, (e) is the method without continuous extremum mapping, and (f) is the method of this invention. Figure 5 As can be seen, the CLAHE method can improve overall brightness and global contrast, but smoke residue is still quite noticeable, and there is some low-frequency haloing in the background area. Traditional guided filtering enhancement methods can enhance some ship edges and local textures, but they are insufficient in suppressing the main smoke. The non-dynamic range shield method can reduce smoke occlusion to some extent, but blocky or grid-like artifacts, contour artifacts, and water ripple noise are prone to appear in flat backgrounds. The non-continuous extremum mapping method, due to the direct use of block-level extremum constraints, is prone to spatial discontinuities and blocky boundary traces after enhancement. In contrast, the method f of this invention further reduces smoke occlusion while maintaining the target outline and local details, and no obvious contour artifacts or blocky discontinuities appear in the background area, resulting in a more natural and stable overall visual effect.

[0114] Depend on Figure 6 As can be seen, in the scenario of smoke interference from a plate-shaped target on a mountain, the CLAHE method enhances background texture and smoke undulations while improving local contrast. In traditional guided filtering enhancement methods, methods without dynamic range shields, and methods without continuous extremum mapping, the target and smoke background still exhibit varying degrees of aliasing, and the target boundary is not clear enough. The method of this invention can enhance the grayscale difference between the plate-shaped target and the surrounding background while maintaining relative stability in the background area, indicating that the method of this invention has good adaptability to different types of infrared smoke degradation scenarios.

[0115] II. Dynamic Range Shield Effect Analysis

[0116] To verify the suppressive effect of the dynamic range shield mechanism on anomalous gain in low dynamic backgrounds, a comparative analysis was conducted on the normalized gain and the grayscale response of the true background ROI with and without the dynamic range shield. Figure 7 As shown, when the local dynamic range approaches zero, if the difference between the continuous maximum and minimum extreme value mappings is directly used as the local normalization denominator, the normalization gain will increase rapidly, easily leading to the abnormal amplification of weak disturbances in a flat background. After introducing the dynamic range shield, the controlled normalization factor satisfies D(x,y)=max(Emax(x,y)-Emin(x,y),τ). When the local dynamic range is less than the threshold τ, τ is used to replace the original local dynamic range as the normalization denominator, thereby limiting the maximum stretching gain in the low dynamic background region to a finite range.

[0117] like Figure 8 As shown, in the real background ROI region, the grayscale response curve of the method without dynamic range shield has obvious peaks and local violent fluctuations, indicating that weak textures, quantization errors or noise in the flat background are abnormally amplified; after adopting the dynamic range shield of the present invention, the background grayscale response curve is more stable overall, and the local peaks are significantly reduced, indicating that the present invention can effectively suppress abnormal gain, contour artifacts and water ripple artifacts in low dynamic background regions.

[0118] III. Analysis of the Effect of Continuous Extremum Mapping

[0119] To verify the suppressive effect of continuous extremum mapping on block effects and spatial faults, a comparative analysis of the spatial responses of block-level extremum mapping and continuous extremum mapping was conducted. For example... Figure 9 As shown, if the block-level minimum and maximum values ​​obtained from local region statistics are directly used as pixel-level enhancement constraints, discontinuities and abrupt changes are likely to occur at the region boundaries, leading to block artifacts, grid-like traces, and spatial tomography. This invention expands the block-level minimum and maximum extrema matrices into continuous minimum and maximum extrema mappings consistent with the input image resolution through bilinear interpolation and other continuous reconstruction methods. This transforms the regional extrema prior from discrete block constraints into pixel-level continuous constraints. This processing can reduce abrupt extrema between adjacent regions, providing a smoother and more stable extrema constraint field for subsequent controlled enhancement.

[0120] IV. Objective Evaluation Data

[0121] To further quantitatively evaluate the processing effects of different enhancement methods, information entropy (EN), Tenengrad sharpness (TG), local contrast enhancement index (EME), and signal-to-noise ratio (SCR) were selected as evaluation metrics. EN is used to evaluate the amount of image information, TG to evaluate edge sharpness, EME to evaluate the degree of local contrast enhancement, and SCR to evaluate the discriminability of the target relative to the background.

[0122] Table 1. Comparison of Objective Indicators of Different Enhancement Methods in Scenario 1

[0123]

[0124] As shown in Table 1, the EN, TG, EME, and SCR values ​​of the enhanced images are generally improved compared to the original images after processing by each enhancement method, indicating that the enhancement processing can improve image information content, edge sharpness, local contrast, and target saliency. Among them, the method without dynamic range shielding has higher values ​​for EN and EME, but combined with subjective results, it is clear that it has obvious over-enhancement and background artifact amplification problems, with an SCR of only 1.6468, which is lower than the traditional guided filtering and the method without continuous extremum mapping. The method of this invention achieves a TG of 0.0026 and an SCR of 2.3403, both of which are superior, indicating that the present invention has a good effect on improving target edge sharpness and target background distinguishability.

[0125] Table 2 Comparison of Objective Indicators of Different Enhancement Methods in Scenario 2

[0126]

[0127] As shown in Table 2, although the CLAHE method has high EN and TG, its SCR is only 1.4192, indicating that while enhancing background texture, it weakens the distinguishability between the target and the background. Traditional guided filtering, methods without dynamic range shielding, and methods without continuous extremum mapping all have high EME, but their SCRs are all lower than the original image, indicating that local contrast enhancement did not effectively improve target saliency. The SCR of the method in this invention reaches 2.7943, which is higher than CLAHE, traditional guided filtering, methods without dynamic range shielding, and methods without continuous extremum mapping, indicating that the method in this invention can better maintain the difference between the target and the background while enhancing image details.

[0128] V. Experimental Conclusions

[0129] Based on a comprehensive evaluation of subjective enhancement effects, dynamic range shield verification, continuous extremum mapping verification, and objective indicators, this invention reduces the block effect and spatial discontinuity caused by block-level extremum constraints through continuous extremum mapping, limits the abnormal normalized gain in low-dynamic background regions through dynamic range shielding, and achieves a good balance between smoke suppression, target detail preservation, and background stability by combining pre-enhancement, smoothing update, and constraint reconstruction. Experimental results show that this invention can effectively improve the target discernibility of infrared smoke images and suppress contour artifacts, water ripple artifacts, block artifacts, and background noise amplification.

Claims

1. A method for enhancing infrared smoke screen images based on continuous extremum mapping and dynamic range shielding, characterized in that, Includes the following steps: S1. Obtain the infrared smoke screen image to be processed, perform grayscale and normalization processing to obtain the normalized input image I∈[0,1]; S2. Multi-scale edge-preserving filtering is used to progressively decompose the normalized input image I. Through step-by-step filtering and subtraction of adjacent layers, the original image is decomposed into a low-frequency base layer and edge and texture detail layers of different scales. Local contrast enhancement is performed on the base layer, and different gains are applied to each detail layer. Then, the images are fused to obtain the pre-enhanced image P∈[0,1]. S3. Based on the pre-enhanced image P, construct dynamic regional extreme value statistics, divide the pre-enhanced image P into multiple local regions Ωi,j, and count the minimum gray value m(i,j) and maximum gray value M(i,j) in each local region to construct a block-level minimum extreme value matrix m and a block-level maximum extreme value matrix M. S4. Perform spatial continuous interpolation reconstruction on the block-level minimum extremum matrix m and the block-level maximum extremum matrix M to generate pixel-level continuous minimum extremum mapping Emin(x,y) and pixel-level continuous maximum extremum mapping Emax(x,y) at the pixel position (x,y) with the same resolution as the pre-enhanced image P. S5. Calculate the local dynamic range R(x,y)=Emax(x,y)-Emin(x,y) based on Emin(x,y) and Emax(x,y), and construct the controlled normalization factor: D(x,y)=max(R(x,y), τ), where τ is the lower limit threshold of the local dynamic range, τ>0, which is used to limit the normalization gain in the low dynamic region; S6. First, relax and correct Emin(x,y) and Emax(x,y) respectively to obtain the target minimum extreme value mapping E'min(x,y) and the target maximum extreme value mapping E'max(x,y): E'min(x,y)=max(Emin(x,y)-λmin, 0), E'max(x,y)=min(Emax(x,y)+λmax, 1), where λmin≥0 is the minimum extreme value relaxation parameter, λmax≥0 is the maximum extreme value relaxation parameter, and satisfies E'max(x,y)-E'min(x,y)≥δ, where δ is a local positive number or δ=τ; Then, based on E'min(x,y), E'max(x,y), and D(x,y), the prior guided enhancement image Q(x,y) is calculated: Q(x,y)=clip(((P(x,y)-E'min(x,y)) / D(x,y))·(E'max(x,y)-E'min(x,y))+E'min(x,y), 0, 1) S7. The pre-enhanced image P and the prior guided enhancement image Q are smoothly updated to obtain an intermediate updated image; the intermediate updated image is combined with gradient sparsity constraints, frequency domain reconstruction and stabilization regularization processing to obtain the enhanced infrared smoke screen image.

2. The method according to claim 1, characterized in that, The progressive decomposition in step 2 specifically involves: using the normalized input image I as the 0th layer image G0, and sequentially applying different scale radii r1, r2, ..., r... n The edge-preserving filter smooths the base image of the previous layer, resulting in G. k =F(G k -1, r k , ε k ), and let the k-th scale detail layer D k =G k -1-G k The final base layer is Gn; where F is a guided filter, self-guided filter, bilateral filter, rolling guided filter, or an equivalent filter with edge preservation capability, and ε k For smoothing regularization parameters.

3. The method according to claim 1, characterized in that, The local region Ωi,j described in step 3 is a square window with a side length of w, where w is an odd number. Adjacent windows are divided using a non-overlapping regular grid or by sliding overlap with a step size s, where s≤w. Boundary regions are filled using mirror filling, edge copying, or truncating window methods.

4. The method according to claim 1, characterized in that, The local dynamic range lower limit threshold τ mentioned in step 5 ranges from 0.05 to 0.20, and is adaptively determined based on the image sensor noise level, satisfying the following conditions: ,in The standard deviation of background or dark field noise. ∈[3,6], This represents the low quantile statistics of the local dynamic range R in a flat background region or the entire image.

5. The method according to claim 1, characterized in that, In step S4, the spatial continuity reconstruction uses bilinear interpolation, bicubic interpolation, spline interpolation, edge-preserving interpolation, or other continuous interpolation methods to expand the block-level minimum and maximum value matrices into pixel-level continuous extremum mappings. During interpolation, the center point of each local region is used as the block-level extremum sampling point. For virtual sampling points outside the image edge, edge copying, mirror extension, or nearest neighbor extrapolation is used to determine the boundary extremum.

6. The method according to claim 1, characterized in that, The relaxation parameters λmin and λmax described in step 6 range from 0.02 to 0.10, and are adaptively adjusted according to the local smoke density or the gray difference between the target and the background: λmin and λmax are increased in areas with dense smoke or weak target contrast, and λmin and λmax are decreased in areas with flat background.

7. The method according to claim 1, characterized in that, The smooth update in step S7 is represented as: U(x,y)=η·P(x,y)+(1-η)·Q(x,y), where U(x,y) is the intermediate updated image, η is the smooth update weight, and 0<η<1.

8. The method according to claim 1, characterized in that, The gradient sparsity constraint in step S7 is used to preserve significant edge gradients and suppress weak texture noise and artifact accumulation; the frequency domain reconstruction is used to achieve stable restoration of the enhanced image while maintaining the main edge structure; the frequency domain reconstruction is based on the gradient consistency estimation of the blur kernel between the current image estimate and the reference image, or it is implemented by using an equivalent frequency domain regularization method that does not explicitly estimate the blur kernel; the reference image is preferably the pre-enhanced image P, or the normalized input image I or the smoothed intermediate image U can be selected according to the application needs.

9. The method according to claim 1, characterized in that, The stabilization regularization process described in step S7 includes one or more of the following: edge preservation smoothing, iterative amplitude limiting update, gradient threshold shrinkage, nonlocal mean smoothing, or total variational regularization, which are used to suppress weak texture noise, contour line perturbation, and local artifact accumulation while preserving the main edge structure.

10. An infrared smoke screen image enhancement system based on continuous extremum mapping and dynamic range shield, characterized in that: The image acquisition module is used to acquire the infrared smoke screen image to be processed, perform grayscale and normalization processing, and obtain the normalized input image I∈[0,1]; The pre-enhancement module is used to progressively decompose the normalized input image I using multi-scale edge-preserving filtering. By filtering step by step and subtracting adjacent layers, the original image is decomposed into a low-frequency base layer and edge and texture detail layers of different scales. Local contrast enhancement is performed on the base layer, and different gains are applied to each detail layer. Then, the images are fused to obtain the pre-enhancement image P∈[0,1]. The extreme value statistics module is used to construct dynamic regional extreme value statistics based on the pre-enhanced image P, divide the pre-enhanced image P into multiple local regions Ωi,j, and count the minimum gray value m(i,j) and maximum gray value M(i,j) in each local region to construct a block-level minimum extreme value matrix m and a block-level maximum extreme value matrix M. The continuous mapping reconstruction module is used to perform spatial continuous interpolation reconstruction on the block-level minimum extremum matrix m and the block-level maximum extremum matrix M to generate pixel-level continuous minimum extremum mapping Emin(x,y) and pixel-level continuous maximum extremum mapping Emax(x,y) at the pixel position (x,y) with the same resolution as the pre-enhanced image P. The dynamic range shield calculation module calculates the local dynamic range R(x,y)=Emax(x,y)-Emin(x,y) based on Emin(x,y) and Emax(x,y), and constructs a controlled normalization factor: D(x,y)=max(R(x,y),τ), where τ is the lower limit threshold of the local dynamic range, τ>0, which is used to limit the normalization gain in the low dynamic region; The extreme value relaxation and enhancement calculation module is used to perform relaxation corrections on Emin(x,y) and Emax(x,y) respectively, to obtain the target minimum extreme value mapping E'min(x,y) and the target maximum extreme value mapping E'max(x,y): E'min(x,y)=max E'max(x,y)=min , where λmin≥0 is the minimum extreme relaxation parameter, λmax≥0 is the maximum extreme relaxation parameter, and satisfies E'max(x,y)-E'min(x,y)≥δ, where δ is a local positive number or δ=τ; Then, based on E'min(x,y), E'max(x,y), and D(x,y), the prior guided enhancement image Q(x,y) is calculated: Q(x,y)=clip The smoothing and stabilization module is used to smoothly update the pre-enhanced image P and the prior guided enhancement image Q to obtain an intermediate updated image; gradient sparsity constraints are first applied to the intermediate updated image, and then combined with frequency domain reconstruction or equivalent stabilization regularization processing to obtain the enhanced infrared smoke screen image.

Images