A method for suppressing and enhancing dead pixels and thermal noise in infrared images

Abstract

This invention provides a method for suppressing and enhancing infrared image defects and thermal noise, relating to the field of image enhancement. The method aims to construct thermal target protection weights and data fidelity weights by normalizing, performing local statistics, and estimating gradient and fringe principal directions on the original infrared image. The observed image is decomposed into structural components, point-like defective pixel components, structured thermal noise components, and low-frequency thermal bias components. Through a thermal consistency-guided defective pixel weight generation mechanism, direction-aware structured thermal noise separation, background-gated low-frequency thermal bias separation, anisotropic edge-preserving enhancement for thermal target protection, and multi-scale local thermal contrast preservation, a joint optimization solution is achieved, and the enhanced image is output after local thermal contrast remapping. This method effectively suppresses defective pixels, fringe noise, and background thermal bias, reduces the risk of accidental deletion of small thermal targets, maintains the target outline and the thermal difference between the target and the background, and improves the clarity, recognizability, and engineering applicability of infrared images.

Description

Technical Field

[0001] This invention belongs to the field of image enhancement, and specifically relates to a method for suppressing and enhancing bad pixels and thermal noise in infrared images. Background Technology

[0002] With the rapid development of infrared thermal imaging and night vision imaging technologies, infrared images have been widely used in scenarios such as nighttime surveillance, target search, equipment inspection, security early warning, intelligent driving, industrial inspection, and unmanned platform perception. Compared with visible light imaging, infrared images can directly reflect the difference in thermal radiation between the target and the background, and have strong applicability in low-light, backlight, smoke, and complex background environments. Therefore, the quality of infrared images has a significant impact on subsequent target recognition, target tracking, anomaly detection, and scene understanding.

[0003] However, in actual infrared imaging, due to factors such as non-uniformity of infrared detector arrays, bad pixels, thermal drift, fixed pattern noise, electronic readout noise, and low-frequency shadow bias introduced by optical links, the original thermal images often contain isolated bright spots, isolated dark spots, random thermal scintillation spots, salt-and-pepper anomalous pixels, row and column stripe noise, low-frequency background undulations, and insufficient local thermal contrast. The above degradation directly leads to blurred target edges, distorted background textures, and reduced thermal difference between thermal targets and backgrounds, further reducing image resolvability and the accuracy of subsequent tasks.

[0004] Existing infrared image enhancement methods often employ median filtering, nonlocal filtering, total variational constraints, nonuniformity correction, or local contrast enhancement to process images. However, these methods generally suffer from the following shortcomings: First, they struggle to simultaneously address point defects, structured stripe noise, and low-frequency thermal bias within the same model. Second, traditional sparse anomaly modeling easily misclassifies small, real thermal targets as defects and removes them along with the image. Third, while ordinary smoothing constraints can suppress random noise, they can lead to excessive smoothing of thermal target edges, weakening the local thermal contrast between the target and the background. Fourth, methods relying solely on display enhancement can improve subjective visual effects but lack explicit decomposition capabilities targeting the degradation mechanism of infrared imaging, making it difficult to balance the stability, interpretability, and applicability of the enhancement. Therefore, an infrared image enhancement method is needed that can simultaneously separate point defects, structured thermal noise, and low-frequency thermal bias, while also protecting thermal targets and maintaining local thermal contrast. Summary of the Invention

[0005] This invention provides a method for suppressing and enhancing dead pixels and thermal noise in infrared images. It aims to jointly decompose and enhance point anomalies, structured noise, and low-frequency non-uniform background in infrared images by constructing a thermal consistency-guided sparse separation term for point dead pixels, a direction-aware structured thermal noise separation term, a background-gated low-frequency thermal bias separation term, an anisotropic edge-preserving enhancement term for thermal target protection, and a multi-scale local thermal contrast preservation term. This method suppresses dead pixels and thermal noise while maintaining target contours and local thermal differences.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for suppressing and enhancing infrared image defects and thermal noise, comprising the following steps.

[0007] S1. Acquire the raw infrared image. For the original infrared image Normalization, local statistics estimation, directional gradient estimation, thermal target salience estimation, and fringe principal direction estimation are performed to obtain the observed image. Local mean plot, local standard deviation plot, data fidelity weighted plot Thermal target protection weight map and stripe main direction pattern .

[0008] S2. Establish an infrared image decomposition model, and decompose the observed image. Decomposed into structural components , Point-like defect components Structured thermal noise components and low-frequency thermal bias components .

[0009] S3. Generate bad pixel weights based on pixel local isolation degree and neighborhood thermal continuity. Construct a sparse separation term for point-like bad points.

[0010] S4. According to the main direction pattern of the stripes Construct directional difference operators for structured thermal noise components Apply directional constraints and base them on the thermal target protection weight map. Constructing low-frequency thermal bias components Background gating constraints.

[0011] S5, Structural components Construct an anisotropic edge protection enhancement term for thermal target protection.

[0012] S6. Construct multi-scale local thermal contrast preservation terms for structural components. With the component for removing pinpoint defects Structured thermal noise components and low-frequency thermal bias components Consistency constraints are applied to the subsequent thermal contrast distribution.

[0013] S7. Construct the joint energy function and update the structural components using an alternating iterative method. , Point-like defect components Structured thermal noise components and low-frequency thermal bias components The optimized structural components are obtained. And based on the structural components Perform local thermal contrast remapping to output an enhanced infrared image.

[0014] Preferably, in step S1, the specific steps are as follows: S11, Transfer the original infrared image Normalized to observed images Its mathematical model is: In the formula, Define the domain of the image At any pixel position within, Define the domain for the image. and These are the height and width of the image, respectively. and Original infrared images The minimum and maximum gray values, To prevent small constants with a denominator of zero; S12, in pixels Centered on, with radius local neighborhood Calculate the local mean. and local standard deviation Their mathematical models are as follows: , In the formula, Representing the neighborhood The number of pixels in; S13, The scale parameter is: Gaussian kernel For the observed images Perform smoothing and calculate the horizontal gradient. Vertical gradient and gradient response value Their mathematical models are as follows: , , In the formula, and Let represent the horizontal first-order difference operator and the vertical first-order difference operator, respectively. This represents the convolution operation; S14. Construct a thermal target protection weight map based on local brightness temperature deviation and gradient significance. Its mathematical model is: , , In the formula, Slope factor The gradient fusion coefficient is... The threshold for identifying thermal targets. To prevent small constants with a denominator of zero; S15. Construct a data fidelity weighting graph based on local noise levels. Its mathematical model is: In the formula, and These are the minimum and maximum values ​​of the data fidelity weight, respectively. As the attenuation factor, To prevent small constants with a denominator of zero; S16. Construct the main fringe pattern using the structure tensor. Structure tensor The components are defined as follows: , , , in, This represents the local energy of the horizontal gradient component. This represents the local energy of the vertical gradient component. This indicates the local correlation between the horizontal and vertical gradients; The main direction pattern of the stripes The mathematical model is as follows: In the formula, The scale parameter is Gaussian smoothing kernel, To prevent small constants with a denominator of zero.

[0015] Preferably, in step S1, by introducing local mean, local standard deviation, gradient significance, and structural tensor direction estimation, the thermal target salience information, local noise level information, and fringe direction information can be obtained simultaneously at the enhancement front end, enabling the subsequent component decomposition process to have explicit prior input; by constructing a thermal target protection weight map... It can improve the structural protection level in the thermal target area and reduce the risk of the real target being falsely suppressed; by constructing a data fidelity weight map This can reduce the strong fidelity constraint in high-noise regions, thus providing a more stable optimization basis for bad pixel separation, stripe suppression, and background bias estimation.

[0016] Preferably, in step S2, the constructed joint energy function is: , in, , , , , , , In the formula, , , and These represent the structural component, the point defect component, the structured thermal noise component, and the low-frequency thermal bias component, respectively. , , , and These represent the weighting coefficients of the point-like defect sparse separation term, the structured thermal noise separation term, the low-frequency thermal bias separation term, the anisotropic edge preservation enhancement term, and the local thermal contrast preservation term, respectively, and all are positive real numbers; For the transdirectional smoothing coefficient of the structured thermal noise component; The target suppression coefficient for the low-frequency thermal bias component; As the background gating weight, satisfying: , For a Laplace operator, satisfying: ,in, and These represent the second-order difference operator in the horizontal direction and the second-order difference operator in the vertical direction, respectively. Let be a non-convex edge-preserving penalty function, satisfying: In the formula, To prevent small constants from causing instability when the gradient is zero; and These represent the main direction patterns of the stripes. Construct directional difference operators along the principal fringe direction and perpendicular to the principal fringe direction; and These represent the weighted graphs based on gradient response and thermal target protection, respectively. Construct anisotropic edge-preserving weights in the horizontal and vertical directions; This represents a multi-scale local thermal comparison operator.

[0017] Preferably, in step S2, by observing the image Explicit decomposition into structural components , Point-like defect components Structured thermal noise components and low-frequency thermal bias components It can incorporate different degradation types in infrared images into a unified optimization framework, avoiding insufficient processing or false suppression caused by mixing point defects, stripe noise, and low-frequency thermal bias into a single noise term; among them, structural components The main task is to reconstruct the thermal target contour and background structure, including point-like bad pixel components. Its main task is to absorb isolated anomalous pixels and structure thermal noise components. Its main tasks include handling row and column fringes and fixed-mode perturbations, as well as low-frequency thermal bias components. The main task is to separate lens shadows and temperature drift background, thereby improving the interpretability and engineering applicability of the model.

[0018] Preferably, in step S3, the point-like defect sparse separation term is determined by the local isolation degree of the pixel. Thermal continuity with neighboring regions Adaptive generation of bad pixel weights Among them, local isolation The mathematical model is as follows: In the formula, To prevent small constants with denominators of zero; neighborhood thermal continuity. The mathematical model is as follows: In the formula, Used to measure pixels The degree of continuity of thermal response with its neighborhood; the higher the thermal continuity, the more likely the pixel is to belong to a real thermal target or a continuous structure. Bad Pixel Weight The mathematical model is as follows: In the formula, and Let be the minimum and maximum values ​​of the bad pixel weights, respectively, and satisfy . ; This is the slope factor for the weight mapping; This is the suppression coefficient for local isolation. The threshold for bad pixel weight mapping; when the local isolation of a pixel... Larger, neighborhood thermal continuity When the value is small, the weight of bad pixels Adaptive approach To reduce the number of point-like bad pixels entering the corresponding pixel The penalty intensity; when the local isolation of a pixel. Smaller or neighboring thermal continuity When the value is large, the weight of bad pixels Adaptive approach To improve the accuracy of separating real thermal targets, continuous edges, and small-scale thermal structures from point-like bad spot components. The cost.

[0019] Preferably, in step S3, by simultaneously introducing two types of discrimination information, local isolation and thermal continuity, isolated abrupt defective points can be distinguished from continuous thermal response real targets. This allows the present invention to avoid simply treating all bright small areas as abnormal when processing isolated bright spots, dark spots, and random flickering points in infrared thermal images. This method is particularly suitable for infrared imaging scenarios where the thermal target is small in size, has a high brightness temperature, but still has a certain degree of thermal continuity in space, which helps to reduce the false deletion rate of small thermal targets.

[0020] Preferably, in step S4, the direction-aware structured thermal noise separation term and the background-gated low-frequency thermal bias separation term are specifically: a directional difference operator along the principal fringe direction. Directional difference operator perpendicular to the principal direction of the fringes They are defined as follows: , , In the formula, The defined fringe principal direction pattern; the mathematical model for the structured thermal noise separation term is: In the formula, the first term Used to constrain structured thermal noise components The stripes exhibit sparse or segmented smooth variations along the main direction of the stripes, the second term Used to suppress irregular fluctuations in the direction perpendicular to the main direction of the stripes; The mathematical model for the low-frequency thermal bias separation term is: In the formula, Used for low-frequency thermal bias components in the background region Apply a smoothing constraint to characterize lens shading, thermal drift, and slow background offset; Used for low-frequency thermal bias components in thermal target regions Apply suppression to avoid misclassifying the low-frequency energy of the true thermal target into the low-frequency thermal bias component. ; is the target inhibition coefficient.

[0021] Preferably, in step S4, the structured thermal noise component is... With low-frequency thermal bias components Separate modeling avoids residual artifacts and target distortion caused by the traditional single smoothing model that mixes stripes, background shadows, and thermal drift; adaptive orientation constraints in the main direction of stripes enhance the targeting of stripe / FPN removal; and a low-frequency bias separation strategy with background gating enables stable separation of background undulations and lens shadows while preserving the overall outline of the thermal target.

[0022] Preferably, in steps S5 and S6, the anisotropic edge preservation enhancement term and the multi-scale local thermal contrast preservation term for thermal target protection specifically refer to: anisotropic edge preservation weights in the horizontal and vertical directions. and The mathematical models are as follows: , In the formula, and These are the defined horizontal and vertical gradients, respectively. , and These are the horizontal gradient suppression coefficient, the vertical gradient suppression coefficient, and the thermal target protection coefficient, respectively; the mathematical model for the anisotropic edge preservation enhancement term is: Among them, when a pixel is located in a strong edge region or a hot target region, due to , or Larger, weight and Adaptive reduction, thereby reducing over-smoothing of corresponding edges and thermal targets; Multiscale local thermal contrast operator The mathematical model is as follows: In the formula, For thermal contrast scale number, For the first Neighborhood radius of scale, satisfying ; For the first The weight coefficients of each scale satisfy... and ; The mathematical model for the multi-scale local thermal contrast preservation term is as follows: In the formula, This represents the local thermal contrast response of the enhanced structural components across multiple scales. This represents the local thermal contrast response of the original thermal information at multiple scales after removing the dotted bad pixel component, structured thermal noise component, and low-frequency thermal bias component. By constraining the consistency of the two, the enhanced image still maintains the thermal difference relationship between the target and the background after bad pixel suppression and thermal noise suppression.

[0023] Preferably, in steps S5 and S6, when the pixel is located in a strong edge region or a hot target region, because... , or The weight is relatively large. and Adaptive reduction reduces over-smoothing of corresponding edges and thermal targets; at the same time, by constraining the consistency of multi-scale local thermal contrast, the enhanced image maintains the original thermal difference hierarchy between the target and the background after bad spot suppression and stripe removal, thus avoiding the problem of "the image is smoother but the target is not prominent".

[0024] Preferably, in step S7, the joint energy function is solved using an alternating minimization method, specifically including: S71. Initialization: Let... ,in, For outer iteration counting; S72, Regarding the first The outermost iteration, at a fixed , and Under the condition of updating structural components Solve the following subproblems: , in, , , , , Set the initial value of the inner iteration to The structural component quantum problem is solved using gradient descent for the inner layer. The innermost update satisfies: , in, Update the step size and gradient term for the structural components. for: , in, , , , , In the formula, This represents the Hadamard element-wise product. , and They represent , and The adjoint operator; S73, in a fixed position , and Under the condition of updating the point-like bad pixel component Solve the following subproblems: , in: The closed-form solution to the point-like defect quantum problem is: , , In the formula, To prevent small constants with a denominator of zero, Represents a symbolic function; S74, in a fixed position , and Under the condition of updating the structured thermal noise components Solve the following subproblems: , in, , , , Set the initial value of the inner iteration to The structured thermal noise component quantum problem is solved using gradient descent for the inner layer. The innermost update satisfies: , in, To update the step size and gradient term for the structured thermal noise components. for: In the formula, and They represent and The adjoint operator; S75, in a fixed position , and Under these conditions, update the low-frequency thermal bias component. Solve the following subproblems: , in, , , Set the initial value of the inner iteration to The low-frequency thermal bias quantum problem is solved using gradient descent for the inner layer. The innermost update satisfies: ,in, To update the step size and gradient term for the low-frequency thermal bias component. for: ; S76. Determine the stopping condition. If it is satisfied: Or the outer iteration count has reached the maximum iteration count. If so, then stop the iteration and let , , , Otherwise And return to step S72; where, This indicates taking the maximum value among all terms. express Norm, To prevent small constants with a denominator of zero, This is the stopping threshold for the outer iteration. This represents the maximum number of outermost iterations.

[0025] S77, Based on optimized structural components Perform local thermal contrast remapping to obtain a normalized enhanced image. Its mathematical model is: , In the formula, The local thermal contrast enhancement coefficient is used to normalize the enhanced image. Inverse mapping back to the original dynamic range yields the final enhanced infrared image. Its mathematical model is: .

[0026] Preferably, in step S7, the structural component, point-like defect component, structured thermal noise component, and low-frequency thermal bias component are solved by alternating minimization, which decomposes the originally coupled multi-component problem into several easily solvable sub-problems. Among them, the point-like defect component adopts a soft-threshold closed-loop update method, which can improve the efficiency of isolated abnormal pixels separation. The structural component, structured thermal noise component, and low-frequency thermal bias component are solved by gradient iteration, which can balance the stability of the solution and the flexibility of the model. Finally, the enhanced result is output through local thermal contrast remapping, so that the image can further improve thermal resolvability after noise reduction and defect suppression.

[0027] Compared with existing technologies, the beneficial effects of the present invention are as follows: The technical solution provided by the present invention achieves joint suppression of isolated bad pixels, stripe noise, and background thermal bias by explicitly decomposing infrared images into structural components, dot-like bad pixel components, structured thermal noise components, and low-frequency thermal bias components; by introducing a bad pixel weight generation mechanism guided by thermal consistency, the risk of real small thermal targets being misjudged as bad pixels can be reduced; through the direction-aware structured thermal noise separation term, the targeted removal effect of row and column stripes and fixed pattern noise can be enhanced; through the background-gated low-frequency thermal bias separation term, the impact of lens shadows and slow thermal drift on image quality can be reduced; through the anisotropic edge preservation enhancement term and multi-scale local thermal contrast preservation term for thermal target protection, the target contour and target-background thermal difference can be maintained while suppressing thermal noise, thereby significantly improving the clarity, recognizability, and engineering applicability of infrared images. Attached Figure Description

[0028] Figure 1 This is a flowchart of an infrared image defect and thermal noise suppression and enhancement method provided by the present invention.

[0029] Figure 2 This is a diagram of the infrared image front-end preprocessing and weight estimation structure provided by the present invention.

[0030] Figure 3 This is a schematic diagram of the joint decomposition model of structural components, point defect components, structured thermal noise components and low-frequency thermal bias components provided by the present invention.

[0031] Figure 4 This is a schematic diagram of the thermal consistency-guided sparse separation of point defects and direction-aware structured thermal noise separation module provided by the present invention.

[0032] Figure 5 This is a schematic diagram of the anisotropic edge protection enhancement and multi-scale local thermal contrast preservation module for thermal target protection provided by the present invention.

[0033] Figure 6 This is a flowchart of the alternating iterative optimization solution provided by the present invention.

[0034] Figure 7 This is a comparison diagram of the original infrared image and the enhanced image provided by the present invention.

[0035] Figure 8 This is a schematic diagram of the decomposition results of the structural component, point defect component, structured thermal noise component, and low-frequency thermal bias component provided by the present invention. Detailed Implementation

[0036] 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. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0037] Please see Figures 1 to 8 This invention provides a method for suppressing and enhancing infrared images by eliminating bad pixels and thermal noise. It aims to address the coexistence of bad pixels, stripe noise, thermal drift, and low contrast in infrared images by using thermal consistency-guided bad pixel separation, direction-aware structured thermal noise suppression, low-frequency thermal bias separation, edge-preserving enhancement for thermal target protection, and multi-scale local thermal contrast preservation to achieve joint enhancement of infrared images.

[0038] Please see Figure 1 As shown in the embodiment of this application, there is a method for suppressing and enhancing infrared image defects and thermal noise.

[0039] S1. Acquire the raw infrared image. For the original infrared image Normalization, local statistics estimation, directional gradient estimation, thermal target salience estimation, and fringe principal direction estimation are performed to obtain the observed image. Local mean plot, local standard deviation plot, data fidelity weighted plot Thermal target protection weight map and stripe main direction pattern .

[0040] Furthermore, such as Figure 2 As shown, the infrared image front-end preprocessing and weight estimation process in this embodiment includes processing the original infrared image... Normalization, local statistics estimation, directional gradient estimation, thermal target salience estimation, and fringe principal direction estimation are performed to obtain the observed image. Local mean plot, local standard deviation plot, data fidelity weighted plot Thermal target protection weight map and stripe main direction pattern .

[0041] Furthermore, an uncooled long-wave infrared thermal imager is selected to acquire the raw thermal image, with the preferred image resolution being... or The pixel dynamic range is preferably 14-bit or 16-bit.

[0042] S11, Transfer the original infrared image Normalized to observed images Its mathematical model is: In the formula, Define the domain of the image At any pixel position within, Define the domain for the image. and These are the height and width of the image, respectively. and Original infrared images The minimum and maximum gray values, To prevent small constants with a denominator of zero, The preferred value range is... arrive In this embodiment, take .

[0043] S12, in pixels Centered on, with radius local neighborhood Calculate the local mean. and local standard deviation Their mathematical models are as follows: , In the formula, Representing the neighborhood Number of pixels in the neighborhood radius The preferred value range is 3 to 9, and in this embodiment, it is taken as... .

[0044] S13, The scale parameter is: Gaussian kernel For the observed images Smooth the surface. The preferred value range is 0.8 to 1.5, and in this embodiment, it is taken as... Calculate the horizontal gradient Vertical gradient and gradient response value Their mathematical models are as follows: , , In the formula, and Let represent the horizontal first-order difference operator and the vertical first-order difference operator, respectively. This represents the convolution operation.

[0045] S14. Construct a thermal target protection weight map based on local brightness temperature deviation and gradient significance. Its mathematical model is: , , In the formula, Slope factor The preferred value range is 4 to 10, and in this embodiment, it is taken as... ; The gradient fusion coefficient is... The preferred value range is 0.2 to 1.0. In this embodiment, the value is taken as... ; The threshold for identifying thermal targets. The preferred value range is 0.6 to 1.5, and in this embodiment, it is taken as... ; To prevent small constants with a denominator of zero, The preferred value range is... arrive In this embodiment, take .

[0046] S15. Construct a data fidelity weighting graph based on local noise levels. Its mathematical model is: , and These are the minimum and maximum values ​​of the data fidelity weight, respectively. The preferred value range is 0.3 to 1.0. In this embodiment, the value is taken as... ; The preferred value range is 1.0 to 3.0. In this embodiment, the value is taken as... ; As the attenuation factor, The preferred value range is 0.2 to 2.0. In this embodiment, the value is taken as... ; To prevent small constants with a denominator of zero, The preferred value range is... arrive In this embodiment, take .

[0047] S16. Construct the main fringe pattern using the structure tensor. Structure tensor The components are defined as follows: , , ,in, This represents the local energy of the horizontal gradient component. This represents the local energy of the vertical gradient component. This represents the local correlation between the horizontal and vertical gradients; the main fringe pattern. The mathematical model is as follows: In the formula, The scale parameter is Gaussian smoothing kernel, The preferred value range is 1.0 to 2.0. In this embodiment, the value is taken as... ; To prevent small constants with a denominator of zero, The preferred value range is... arrive In this embodiment, take .

[0048] S2. Establish an infrared image decomposition model to decompose the observed image. Decomposed into structural components , Point-like defect components Structured thermal noise components and low-frequency thermal bias components .

[0049] Furthermore, such as Figure 3 As shown, this embodiment constructs structural components. , Point-like defect components Structured thermal noise components and low-frequency thermal bias components The joint decomposition model uses a joint energy function to uniformly model and collaboratively optimize each component, thereby achieving joint separation and enhancement of point defects, structured thermal noise, and low-frequency thermal bias.

[0050] Furthermore, construct the joint energy function: , in, , , , , , , In the formula, , , and These represent the structural component, the point defect component, the structured thermal noise component, and the low-frequency thermal bias component, respectively. , , , and These represent the weighting coefficients for the point-like defect sparsity separation term, the structured thermal noise separation term, the low-frequency thermal bias separation term, the anisotropic edge preservation enhancement term, and the local thermal contrast preservation term, respectively. The preferred value range is 0.01 to 0.20. In this embodiment, the value is taken as... ;parameter The preferred value range is 0.02 to 0.30. In this embodiment, the value is taken as... ;parameter The preferred value range is 0.01 to 0.15. In this embodiment, the value is taken as... ;parameter The preferred value range is 0.02 to 0.50. In this embodiment, the value is taken as... ;parameter The preferred value range is 0.01 to 0.20. In this embodiment, the value is taken as... ; For the transdirectional smoothing coefficient of the structured thermal noise component; The target suppression coefficient for the low-frequency thermal bias component; As the background gating weight, satisfying: , For a Laplace operator, satisfying: ,in, and These represent the second-order difference operator in the horizontal direction and the second-order difference operator in the vertical direction, respectively. Let be a non-convex edge-preserving penalty function, satisfying: In the formula, To prevent small constants from causing instability when the gradient is zero, The preferred value range is... arrive In this embodiment, take ;parameter The preferred value range is 0.5 to 1.0. In this embodiment, the value is taken as... ; and These represent the main direction patterns of the stripes. Construct directional difference operators along the principal fringe direction and perpendicular to the principal fringe direction; and These represent the weighted graphs based on gradient response and thermal target protection, respectively. Construct anisotropic edge-preserving weights in the horizontal and vertical directions; This represents a multi-scale local thermal comparison operator.

[0051] S3. Generate bad pixel weights based on pixel local isolation degree and neighborhood thermal continuity. Construct a sparse separation term for point-like bad points.

[0052] Furthermore, such as Figure 4 As shown, this embodiment uses a thermally consistent guided sparse separation of point-like defects and a direction-aware structured thermal noise separation module to separate point-like defect components. and structured thermal noise components Targeted constraints are applied; among them, bad pixel weights are generated based on the local isolation degree of the pixel and the thermal continuity of the neighborhood. And according to the main fringe pattern A directional difference operator is constructed to enhance the effects of bad pixel separation and structured thermal noise suppression.

[0053] Furthermore, the sparse separation term for point-like defects is determined by the local isolation degree of pixels. Thermal continuity with neighboring regions Adaptive generation of bad pixel weights Among them, local isolation The mathematical model is as follows: In the formula To prevent small constants with a denominator of zero, The preferred value range is... arrive In this embodiment, take ; Neighborhood thermal continuity The mathematical model is as follows: In the formula, Used to measure pixels The degree of continuity in thermal response between the pixel and its neighborhood; higher thermal continuity indicates that the pixel is more likely to belong to a real thermal target or continuous structure; bad pixel weight. The mathematical model is as follows: In the formula, and Let be the minimum and maximum values ​​of the bad pixel weights, respectively, and satisfy . ,parameter The preferred value range is 0.1 to 0.5. In this embodiment, the value is taken as... ;parameter The preferred value range is 0.8 to 1.5, and in this embodiment, it is taken as... ; For the weight mapping slope factor, parameters The preferred value range is 2 to 8, and in this embodiment, it is taken as... ; The suppression coefficient for local isolation, parameter The preferred value range is 0.2 to 1.5. In this embodiment, the value is taken as... ; For bad pixel weight mapping threshold, parameters The preferred value range is 0.05 to 0.5. In this embodiment, the value is taken as... When the local isolation of a pixel Larger, neighborhood thermal continuity When the value is small, the weight of bad pixels Adaptive approach To reduce the number of point-like bad pixels entering the corresponding pixel The penalty intensity; when the local isolation of a pixel. Smaller or neighboring thermal continuity When the value is large, the weight of bad pixels Adaptive approach To improve the accuracy of separating real thermal targets, continuous edges, and small-scale thermal structures from point-like bad spot components. The cost.

[0054] S4. According to the main direction pattern of the stripes Construct directional difference operators for structured thermal noise components Apply directional constraints and base them on the thermal target protection weight map. Constructing low-frequency thermal bias components Background gating constraints.

[0055] Furthermore, the direction-aware structured thermal noise separation term and the background-gated low-frequency thermal bias separation term are specifically: a directional difference operator along the principal fringe direction. Directional difference operator perpendicular to the principal direction of the fringes They are defined as follows: , In the formula, The defined main fringe pattern; The mathematical model for the structured thermal noise separation term is: In the formula, the first term Used to constrain structured thermal noise components The stripes exhibit sparse or segmented smooth variations along the main direction of the stripes, the second term Used to suppress irregular fluctuations in the direction perpendicular to the principal direction of the stripes; the mathematical model for the low-frequency thermal bias separation term is: In the formula, Used for low-frequency thermal bias components in the background region Apply a smoothing constraint to characterize lens shading, thermal drift, and slow background offset; Used for low-frequency thermal bias components in thermal target regions Apply suppression to avoid misclassifying the low-frequency energy of the true thermal target into the low-frequency thermal bias component. ;parameter The preferred value range is 0.1 to 2.0. In this embodiment, the value is taken as... ; The target suppression coefficient, parameter The preferred value range is 0.5 to 5.0. In this embodiment, the value is taken as... .

[0056] S5, Structural components Construct an anisotropic edge protection enhancement term for thermal target protection.

[0057] Furthermore, such as Figure 5 As shown, this embodiment utilizes an anisotropic edge protection enhancement and multi-scale local thermal contrast preservation module for thermal target protection to protect structural components. Edge preservation enhancement is performed, and the component for removing point-like defects is also addressed. Structured thermal noise components and low-frequency thermal bias components The thermal contrast distribution is then constrained to maintain the target contour and the local thermal difference between the target and the background while suppressing bad pixels and thermal noise.

[0058] Furthermore, the anisotropic edge-preserving weights in the horizontal and vertical directions and The mathematical models are as follows: In the formula, and These are the defined horizontal and vertical gradients, respectively. , and These are the horizontal gradient suppression coefficient, the vertical gradient suppression coefficient, and the thermal target protection coefficient, respectively. and The preferred value range is 0.5 to 2.0, and in this embodiment, it is taken as 1.2; The preferred value range is 0.2 to 1.5. In this embodiment, the value is taken as... The mathematical model for the anisotropic edge-preserving enhancement term is as follows: Among them, when a pixel is located in a strong edge region or a hot target region, due to , or Larger, weight and Adaptive reduction reduces over-smoothing of corresponding edges and thermal targets.

[0059] S6. Construct multi-scale local thermal contrast preservation terms for structural components. With the component for removing pinpoint defects Structured thermal noise components and low-frequency thermal bias components Consistency constraints are applied to the subsequent thermal contrast distribution.

[0060] Furthermore, multi-scale local thermal contrast operator The mathematical model is as follows: In the formula, For thermal contrast scale number, The preferred value range is 2 to 5, and in this embodiment, it is taken as... ;No. Neighborhood radius of each scale An incremental setting is adopted, with odd-numbered radii being preferred. In this embodiment, the radius is selected as follows: , , ;No. Weight coefficients for each scale satisfy and In this embodiment, take , , The mathematical model for the multi-scale local thermal contrast preservation term is as follows: , In the formula, This represents the local thermal contrast response of the enhanced structural components across multiple scales. This represents the local thermal contrast response of the original thermal information at multiple scales after removing the point-like defect component, structured thermal noise component, and low-frequency thermal bias component. By applying consistency constraints to both, the enhanced image maintains the thermal difference relationship between the target and the background even after defect suppression and thermal noise suppression. Here, the number of scales... The preferred value range is 2 to 4, and in this embodiment, it is taken as... The neighborhood radius is preferably set to , , The optimal weighting for each scale is set as follows: , , .

[0061] S7. Construct the joint energy function and update the structural components using an alternating iterative method. , Point-like defect components Structured thermal noise components and low-frequency thermal bias components The optimized structural components are obtained. And based on the structural components Perform local thermal contrast remapping to output an enhanced infrared image.

[0062] Furthermore, such as Figure 6 As shown, this embodiment employs an alternating iterative optimization solution process for the structural components in the joint energy function. , Point-like defect components Structured thermal noise components and low-frequency thermal bias components Alternating updates are performed to obtain optimized structural components. And based on the structural components Perform local thermal contrast remapping to output an enhanced infrared image.

[0063] Furthermore, the joint energy function is solved using an alternating minimization method, specifically including...

[0064] S71. Initialization: Let... ,in, Set the update step size for structural components to count the outer iterations. Structured thermal noise component update step size Low-frequency thermal bias component update step size The preferred value ranges are 0.01 to 0.10, 0.01 to 0.08, and 0.01 to 0.05, respectively. In this embodiment, the values ​​are respectively taken as follows: , , Maximum number of iterations The preferred value range is 20 to 100. In this embodiment, the value is taken as... Stop threshold The preferred value range is... arrive In this embodiment, take .

[0065] S72, Regarding the first The outermost iteration, at a fixed , and Under the condition of updating structural components Solve the following subproblems: , in, , , , , Set the initial value of the inner iteration to The structural component quantum problem is solved using gradient descent for the inner layer. The innermost update satisfies: , in, Update the step size and gradient term for the structural components. for: , in, , , , In the formula, This represents the Hadamard element-wise product. , and They represent , and The adjoint operator.

[0066] S73, in a fixed position , and Under the condition of updating the point-like bad pixel component Solve the following subproblems: , in: The closed-form solution to the point-like defect quantum problem is: , , In the formula, To prevent small constants with a denominator of zero, The preferred value range is... arrive In this embodiment, take , Represents a symbolic function.

[0067] S74, in a fixed position , and Under the condition of updating the structured thermal noise components Solve the following subproblems: , in, , , , Set the initial value of the inner iteration to The structured thermal noise component quantum problem is solved using gradient descent for the inner layer. The innermost update satisfies: , in, To update the step size and gradient term for the structured thermal noise components. for: In the formula, and They represent and The adjoint operator.

[0068] S75, in a fixed position , and Under these conditions, update the low-frequency thermal bias component. Solve the following subproblems: , in, , , Set the initial value of the inner iteration to The low-frequency thermal bias quantum problem is solved using gradient descent for the inner layer. The innermost update satisfies: ,in, To update the step size and gradient term for the low-frequency thermal bias component. for: ;

[0069] S76. Determine the stopping condition. If it is satisfied: Or the outer iteration count has reached the maximum iteration count. If so, then stop the iteration and let , , , Otherwise And return to step S72; where, This indicates taking the maximum value among all terms. express Norm, To prevent small constants with a denominator of zero, The preferred value range is... arrive In this embodiment, take , This is the stopping threshold for the outer iteration. This represents the maximum number of outermost iterations.

[0070] S77, Based on optimized structural components Perform local thermal contrast remapping to obtain a normalized enhanced image. Its mathematical model is: , In the formula, This represents the local thermal contrast enhancement coefficient. The preferred value range is 0.05 to 0.5. In this embodiment, the value is taken as... Normalization enhances the image. Inverse mapping back to the original dynamic range yields the final enhanced infrared image. Its mathematical model is: .

[0071] Furthermore, the method described in this embodiment can be implemented using Python, C++, or Matlab, preferably using Python combined with NumPy, OpenCV, and CUDA acceleration libraries; for a single frame The method of this invention can achieve bad pixel separation, stripe noise suppression, low-frequency offset correction, and thermal contrast enhancement on a general image processing platform or an embedded edge computing platform.

[0072] Furthermore, such as Figure 7 As shown, Figure 7 The image on the left is the original infrared image. Figure 7 The image on the right is the enhanced image processed by the method of this invention; isolated bright spots and isolated dark spots in the enhanced image are effectively suppressed, stripe artifacts and background shadows are significantly reduced, the edges of hot targets are clearer, and the local thermal difference between the target and the background is more prominent.

[0073] Furthermore, such as Figure 8 As shown, the method of the present invention can obtain the decomposition results of structural components, point defect components, structured thermal noise components, and low-frequency thermal bias components. Among them, the point defect components mainly inherit isolated anomalous pixels, the structured thermal noise components mainly inherit stripes and fixed pattern perturbations, the low-frequency thermal bias components mainly inherit background slow undulations and lens shadows, and the structural components retain the thermal target contour and background structure information, thereby verifying the interpretability of the joint decomposition model of the present invention.

[0074] The above are merely preferred embodiments of the present invention. It should be noted that those skilled in the art can make various modifications and improvements without departing from the inventive concept of the present invention, and these modifications and improvements all fall within the protection scope of the present invention.

Claims

1. A method for suppressing and enhancing dead pixels and thermal noise in infrared images, characterized in that, Includes the following steps: S1. Acquire the raw infrared image. For the original infrared image Normalization, local statistics estimation, directional gradient estimation, thermal target salience estimation, and fringe principal direction estimation are performed to obtain the observed image. Local mean plot, local standard deviation plot, data fidelity weighted plot Thermal target protection weight map and stripe main direction pattern ; S2. Establish an infrared image decomposition model, and decompose the observed image. Decomposed into structural components , Point-like defect components Structured thermal noise components and low-frequency thermal bias components ; S3. Generate bad pixel weights based on pixel local isolation degree and neighborhood thermal continuity. Construct a sparse separation term for point-like defects; S4. According to the main direction pattern of the stripes Construct directional difference operators for structured thermal noise components Apply directional constraints and base them on the thermal target protection weight map. Constructing low-frequency thermal bias components Background gating constraints; S5, Structural components Construct an anisotropic edge protection enhancement term for thermal target protection; S6. Construct multi-scale local thermal contrast preservation terms for structural components. With the component for removing pinpoint defects Structured thermal noise components and low-frequency thermal bias components Consistency constraints are applied to the subsequent thermal contrast distribution. S7. Construct the joint energy function and update the structural components using an alternating iterative method. , Point-like defect components Structured thermal noise components and low-frequency thermal bias components The optimized structural components are obtained. And based on the structural components Perform local thermal contrast remapping to output an enhanced infrared image.

2. The method for suppressing and enhancing infrared image defects and thermal noise according to claim 1, characterized in that, The specific steps of S1 are as follows: S11, Transfer the original infrared image Normalized to observed images Its mathematical model is: ， In the formula, Define the domain of the image At any pixel position within, Define the domain for the image. and These are the height and width of the image, respectively. and Original infrared images The minimum and maximum gray values, To prevent small constants with a denominator of zero; S12, in pixels Centered on, with radius local neighborhood Calculate the local mean. and local standard deviation Their mathematical models are as follows: ， ， In the formula, Representing the neighborhood The number of pixels in; S13, The scale parameter is: Gaussian kernel For the observed images Perform smoothing and calculate the horizontal gradient. Vertical gradient and gradient response value Their mathematical models are as follows: ， ， ， In the formula, and Let represent the horizontal first-order difference operator and the vertical first-order difference operator, respectively. This represents the convolution operation; S14. Construct a thermal target protection weight map based on local brightness temperature deviation and gradient significance. Its mathematical model is: ， ， In the formula, Slope factor The gradient fusion coefficient is... The threshold for identifying thermal targets. To prevent small constants with a denominator of zero; S15. Construct a data fidelity weighting graph based on local noise levels. Its mathematical model is: ， In the formula, and These are the minimum and maximum values ​​of the data fidelity weight, respectively. As the attenuation factor, To prevent small constants with a denominator of zero; S16. Construct the main fringe pattern using the structure tensor. Structure tensor The components are defined as follows: ， ， ， in, This represents the local energy of the horizontal gradient component. This represents the local energy of the vertical gradient component. This indicates the local correlation between the horizontal and vertical gradients; The main direction pattern of the stripes The mathematical model is as follows: , In the formula, The scale parameter is Gaussian smoothing kernel, To prevent small constants with a denominator of zero.

3. The method for suppressing and enhancing infrared image defects and thermal noise according to claim 2, characterized in that, The joint energy function constructed in step S2 is: ， in, , ， ， ， ， ， In the formula, , , and These represent the structural component, the point defect component, the structured thermal noise component, and the low-frequency thermal bias component, respectively. , , , and These represent the weighting coefficients of the point-like defect sparse separation term, the structured thermal noise separation term, the low-frequency thermal bias separation term, the anisotropic edge preservation enhancement term, and the local thermal contrast preservation term, respectively, and all are positive real numbers; For the transdirectional smoothing coefficient of the structured thermal noise component; The target suppression coefficient for the low-frequency thermal bias component; As the background gating weight, satisfying: , For a Laplace operator, satisfying: ,in, and These represent the second-order difference operator in the horizontal direction and the second-order difference operator in the vertical direction, respectively. Let be a non-convex edge-preserving penalty function, satisfying: In the formula, To prevent small constants from causing instability when the gradient is zero; and These represent the main direction patterns of the stripes. Construct directional difference operators along the principal fringe direction and perpendicular to the principal fringe direction; and These represent the weighted graphs based on gradient response and thermal target protection, respectively. Construct anisotropic edge-preserving weights in the horizontal and vertical directions; This represents a multi-scale local thermal comparison operator.

4. The method for suppressing and enhancing infrared image defects and thermal noise according to claim 3, characterized in that, In step S3, the sparse separation term for point-like bad pixels is determined by the local isolation degree of pixels. Thermal continuity with neighboring regions Adaptive generation of bad pixel weights Among them, local isolation The mathematical model is as follows: In the formula, To prevent small constants with a denominator of zero; Neighborhood thermal continuity The mathematical model is as follows: ; Bad Pixel Weight The mathematical model is as follows: ， In the formula, and Let be the minimum and maximum values ​​of the bad pixel weights, respectively, and satisfy . ; This is the slope factor for the weight mapping; This is the suppression coefficient for local isolation. This is the threshold for mapping bad pixel weights.

5. The method for suppressing and enhancing infrared image defects and thermal noise according to claim 4, characterized in that, In step S4, the direction-aware structured thermal noise separation term and the background-gated low-frequency thermal bias separation term are specifically as follows: Directional difference operator along the principal direction of the stripes Directional difference operator perpendicular to the principal direction of the fringes They are defined as follows: ， ， In the formula, The defined main fringe pattern; The mathematical model for the structured thermal noise separation term is: , In the formula, the first term Used to constrain structured thermal noise components The stripes exhibit sparse or segmented smooth variations along the main direction of the stripes, the second term Used to suppress irregular fluctuations in the direction perpendicular to the main direction of the stripes; The mathematical model for the low-frequency thermal bias separation term is: , In the formula, Used for low-frequency thermal bias components in the background region Apply a smoothing constraint to characterize lens shading, thermal drift, and slow background offset; Used for low-frequency thermal bias components in thermal target regions Apply suppression to avoid misclassifying the low-frequency energy of the true thermal target into the low-frequency thermal bias component. ; is the target inhibition coefficient.

6. The method for suppressing and enhancing infrared image defects and thermal noise according to claim 5, characterized in that, In steps S5 and S6, the anisotropic edge preservation enhancement term and the multi-scale local thermal contrast preservation term for thermal target protection are specifically as follows: Anisotropic edge-preserving weights in horizontal and vertical directions and The mathematical models are as follows: ， ， In the formula, and These are the defined horizontal and vertical gradients, respectively. , and These are the horizontal gradient suppression coefficient, the vertical gradient suppression coefficient, and the thermal target protection coefficient, respectively. The mathematical model for the anisotropic edge-preserving enhancement term is as follows: , Among them, when the pixel is located in a strong edge region or a hot target region, due to , or Larger, weight and Adaptive reduction, thereby reducing over-smoothing of corresponding edges and thermal targets; Multiscale local thermal contrast operator The mathematical model is as follows: ， In the formula, For thermal contrast scale number, For the first Neighborhood radius of scale, satisfying ; For the first The weight coefficients of each scale satisfy... and ; The mathematical model for the multi-scale local thermal contrast preservation term is as follows: ， In the formula, This represents the local thermal contrast response of the enhanced structural components across multiple scales. This represents the local thermal contrast response of the original thermal information at multiple scales after removing the dotted bad pixel component, structured thermal noise component, and low-frequency thermal bias component. By constraining the consistency of the two, the enhanced image still maintains the thermal difference relationship between the target and the background after bad pixel suppression and thermal noise suppression.

7. The method for suppressing and enhancing infrared image defects and thermal noise according to claim 6, characterized in that, In step S7, the joint energy function is solved using an alternating minimization method, specifically including: S71. Initialization: Let... ,in, For outer iteration counting; S72, Regarding the first The outermost iteration, at a fixed , and Under the condition of updating structural components Solve the following subproblems: ， in, , ， ， ， ， Set the initial value of the inner iteration to The structural component quantum problem is solved using gradient descent for the inner layer. The innermost update satisfies: ， in, Update the step size and gradient term for the structural components. for: ， in, , ， ， ， In the formula, This represents the Hadamard element-wise product. , and They represent , and The adjoint operator; S73, in a fixed position , and Under the condition of updating the point-like bad pixel component Solve the following subproblems: ， in: The closed-form solution to the point-like defect quantum problem is: ， ， In the formula, To prevent small constants with a denominator of zero, Represents a symbolic function; S74, in a fixed position , and Under the condition of updating the structured thermal noise components Solve the following subproblems: ， in, , ， ， Set the initial value of the inner iteration to The structured thermal noise component quantum problem is solved using gradient descent for the inner layer. The innermost update satisfies: ， in, To update the step size and gradient term for the structured thermal noise components. for: ， In the formula, and They represent and The adjoint operator; S75, in a fixed position , and Under these conditions, update the low-frequency thermal bias component. Solve the following subproblems: ， in, , ， Set the initial value of the inner iteration to The low-frequency thermal bias quantum problem is solved using gradient descent for the inner layer. The innermost update satisfies: ,in, To update the step size and gradient term for the low-frequency thermal bias component. for: ； S76. Determine the stopping condition. If it is satisfied: Or the outer iteration count has reached the maximum iteration count. If so, then stop the iteration and let , , , Otherwise And return to step S72; In the formula, This indicates taking the maximum value among all terms. express Norm, To prevent small constants with a denominator of zero, This is the stopping threshold for the outer iteration. This represents the maximum number of outermost iterations. S77, Based on optimized structural components Perform local thermal contrast remapping to obtain a normalized enhanced image. Its mathematical model is: ， ， In the formula, The local thermal contrast enhancement coefficient is used to normalize the enhanced image. Inverse mapping back to the original dynamic range yields the final enhanced infrared image. Its mathematical model is: 。

Images