Improved infrared and low-light image fusion method based on rolling guide filtering

Abstract

The invention belongs to the technical field of image fusion, and particularly relates to an infrared and low-light-level image fusion improvement method based on rolling guide filtering, which comprises the following steps of S1, acquiring a low-light-level image and an infrared image, and preprocessing the infrared image for denoising; s2, converting to an IHS space to extract an intensity component; s3, decomposing the dual-source I component into a small-scale layer, a large-scale layer and a base layer by using rolling guide filtering; s4, performing small-scale layer fusion, and combining an absolute value maximization strategy with non-local similarity weighted denoising; s5, performing large-scale layer fusion, and performing joint sparse low-rank representation by adopting principal component analysis; s6, fusion of the base layer adopts a visual saliency mapping method; and S7, after IHS inverse transformation reconstruction, realizing multi-level fusion in combination with bilateral filtering detail enhancement and a guide filtering saliency map. According to the invention, based on IHS space and rolling guide filtering multi-scale decomposition, the effect of cooperative utilization of multi-modal information is improved, and the method effectively improves the problems of insufficient target detection precision and poor scene reconstruction robustness in a dark environment.

Description

Technical Field

[0001] The present invention relates to the technical field of image fusion, and in particular to an improved method for fusion of infrared and low-light-level images based on rolling guided filtering. Background Art

[0002] Fusion of infrared and low-light-level imagery is a key approach to enhancing nighttime environmental perception. Infrared imaging captures the target's thermal radiation characteristics, enabling clear visualization of target outlines in extreme conditions such as complete darkness and smoke. However, it is susceptible to interference from ambient heat sources, resulting in artifacts, and atmospheric absorption of mid- and far-infrared wavelengths significantly reduces image contrast. Low-light-level imaging relies on ambient light enhancement from the visible to near-infrared range, preserving rich texture and color details. However, it suffers from low signal-to-noise ratios and motion blur in extremely low illumination conditions, and atmospheric absorption of shortwave radiation also diminishes the sense of scene depth. Using either sensor alone is insufficient to meet the demands of modern military reconnaissance, security surveillance, and other scenarios. Fusion technology, by integrating infrared thermal radiation characteristics with low-light-level detail information, can generate high-contrast and high-resolution images, effectively improving nighttime target recognition accuracy and scene reconstruction integrity. Current mainstream approaches can be categorized into three main categories: physics-based models, deep learning and generative adversarial networks, and lightweight methods.

[0003] The physical model-based method separates the reflection / transmission components of the infrared and low-light bands by establishing an image formation model, retaining the target thermal radiation characteristics and low-light texture details, but the physical model relies on prior assumptions and is easily restricted to certain scenarios; the deep learning method uses feature pyramids or cross-modal attention mechanisms, but requires a large amount of labeled data and has high computational overhead; the generative adversarial network and lightweight method use implicit feature mapping or model compression technology, which are suitable for real-time scenarios, but require a balance between model performance and efficiency. Summary of the Invention

[0004] (1) Technical problems solved

[0005] In response to the shortcomings of the existing technology, the present invention provides an improved method for the fusion of infrared and low-light-level images based on rolling guided filtering. By adopting IHS color space analysis combined with rolling guided filtering multi-scale decomposition, absolute value maximization feature selection, PCA dimensionality reduction and sparse representation fusion and visual saliency map enhancement technology, the effect of the collaborative utilization of multimodal information is effectively improved, thereby improving the target detection accuracy and scene reconstruction robustness in dark environments, and solving the problems raised in the above-mentioned background technology.

[0006] (2) Technical solution

[0007] In order to achieve the above-mentioned purpose, the present invention specifically adopts the following technical solutions: an improved method for fusion of infrared and low-light-level images based on rolling guided filtering, comprising the following steps:

[0008] S1. Acquire a low-light image and an infrared image, and perform denoising on the infrared image to obtain a denoised infrared image;

[0009] S2. Convert the low-light image and the denoised infrared image to the IHS color space to obtain their respective I, H, and S components;

[0010] S3. Perform rolling guided filtering on the I components of the infrared image and the low-light image, respectively, to decompose them into a small-scale layer, a large-scale layer, and a base layer;

[0011] S4. A feature selection strategy of maximizing absolute values ​​is used to fuse the infrared and low-light-level images in their small-scale layers, combined with a non-local similarity adaptive weighted denoising algorithm, to obtain a small-scale fused image.

[0012] S5. Use principal component analysis (PCA) to reduce the dimensionality of the large-scale layers of the infrared and low-light-level images, and fuse them with sparse and low-rank representations to obtain a large-scale fused image.

[0013] S6. Fusing the base layers of the infrared image and the low-light image using a visual saliency mapping method, controlling the contrast distribution of the fused image and reducing the overall contrast loss, thereby obtaining a base layer fused image;

[0014] After inverse IHS transformation and reconstruction, bilateral filtering detail enhancement and guided filtering saliency map are combined to achieve multi-level fusion and obtain the final infrared and low-light fusion image.

[0015] Furthermore, in said S1, an infrared image and a low-light image to be processed are input, and a bilateral filtering algorithm is used to denoise the infrared image;

[0016] Bilateral filtering algorithm: Its function expression is:

[0017]

[0018] Among them, W p : Normalization factor, σ d : used to represent the amount of filtering of image I, Range function, which can reduce and I p The influence of pixels p with large gray value difference, A spatial function that reduces the influence of distant pixels.

[0019] Furthermore, the IHS and RGB conversion process in S2 needs to be implemented using two intermediate variables F1 and F2, and its function expression is:

[0020]

[0021] Among them, R, G, and B are the red, green, and blue bands of the original multispectral image, respectively. LL , F1, and F2 are the I, H, and S components obtained by IHS transformation respectively.

[0022] Furthermore, in S3, the image is decomposed into multiple scales to obtain a base layer, a small-scale layer, and a large-scale layer;

[0023] First, the I component is initially filtered by a Gaussian filter to remove high-frequency noise and small texture structures, and obtain the base layer containing low-frequency information. Its function expression is:

[0024]

[0025] Where, I(q), G(p): input and output pixel values, p, q: corresponding image pixel coordinates, σ s is: the parameter that controls the scale of the Gaussian structure, N(p): the filter window centered at pixel p;

[0026] Secondly, the filtered image is subjected to multi-scale iterative processing based on the bilateral filter, gradually restoring the edge information of different scales and generating a large-scale layer representing the structural features. Its function expression is:

[0027]

[0028] Among them, J t+1 : The result of the tth iteration, J t : guided filtering, σ r : Control distance weight, t: number of iterations;

[0029] Finally, by adjusting σ z , σ r and the number of iterations t, and finally output the decomposed three-level structure.

[0030] Further, in said S4;

[0031] First, the small-scale layers are fused using the absolute value maximization strategy. The coefficient with the largest absolute value represents the prominent feature of the image. The dominant attributes can be extracted from the small-scale layers of the two images, and the important fine-scale features can be injected into the fused image to retain more texture information in the original image. It can be expressed as:

[0032]

[0033] Among them, GF(W j ,σ0) is Gaussian smoothing of the weighted coefficients to reduce noise interference, σ0 is usually set to 2, and

[0034]

[0035] Secondly, in order to further suppress the noise in the small-scale layer and avoid excessive enhancement of the noise during the fusion process, a denoising method based on non-local means (NLM) is introduced;

[0036] For each pixel i in the small-scale layer, the similarity weight w(i, j) between it and the surrounding pixels j is calculated. The similarity weight is determined by comparing the grayscale value difference of the local windows Ni and Nj centered on pixels i and j. The formula is:

[0037]

[0038] Where, h: parameter that controls weight decay; Represents the Euclidean distance between two local windows;

[0039] For each pixel i, the similarity weight w(i, j) is used to perform weighted averaging on the surrounding pixels j to obtain the denoised pixel value:

[0040]

[0041] Adjust the search window size and attenuation parameter h according to the image characteristics to balance the denoising effect and detail preservation;

[0042] Finally, the denoised small-scale layer and Substitute into the absolute value maximization fusion formula to obtain the final fusion result:

[0043]

[0044] By introducing a denoising method based on non-local means (NLM), the noise in small-scale layers is effectively suppressed while retaining the detail information of the image.

[0045] Furthermore, in said S5,

[0046] First, the large-scale layer images of infrared and low-light images are used as column vectors of the matrix. Then, each row is used as a reference and each column is used as a variable γ. The covariance matrix C of γ is obtained as follows:

[0047]

[0048] In the formula is the mean vector of γ;

[0049] Calculate the eigenvalues ​​and corresponding eigenvectors of C and record the results as λ1,λ2,…,λ n ; Find the largest eigenvalue from the eigenvalues, i.e. λ max; and use its corresponding eigenvector as the principal component; normalize the principal component as the fusion weight to obtain the final large-scale information fusion image;

[0050] Secondly, to further enhance the feature extraction capability of large-scale layers, the PCA fusion method is improved by combining sparse representation and low-rank representation. Sparse representation can effectively capture local details in the image, while low-rank representation can extract global structural information in the image. By combining these two representation methods, key visual features in large-scale layers can be more comprehensively mined.

[0051] The large-scale layer image is represented as a sparse linear combination, and the sparse coefficients are obtained by solving the following optimization problem:

[0052]

[0053] Where D is the dictionary, X is the large-scale layer image, α is the sparse coefficient, and λ is the regularization parameter;

[0054] Use sparse coefficients to reconstruct large-scale layer images and extract local detail information;

[0055] The large-scale layer image is decomposed into a low-rank matrix and a sparse matrix, and the low-rank representation is obtained by solving the following optimization problem:

[0056]

[0057] Among them, L is a low-rank matrix, S is a sparse matrix, ||·|| * represents the nuclear norm, and ||·||1 represents L 0031 norm;

[0058] Using low-rank matrices to extract global structural information;

[0059] Finally, the results of sparse representation and low-rank representation are weighted fused with the principal components extracted by PCA to obtain the final large-scale information fusion image; the fusion formula is:

[0060] F large =w1·PCA+w2·Sparse+w3·Low-Rank

[0061] Where w1, w2, w3 are weight coefficients satisfying w1+w2+w3=1;

[0062] By combining sparse representation and low-rank representation, this improved large-scale layer fusion method can more comprehensively extract local details and global structural information in the image, thereby improving the quality and visual effect of the fused image.

[0063] Furthermore, the S7 performs guided filtering on the infrared image to generate a saliency map;

[0064] In order to extract salient targets and suppress the influence of background on targets, the infrared image is used as a segmentation template. ir ) is multiplied by the saliency map mask (M1) to obtain the infrared sub-image (S), the formula is:

[0065] S=I ir ×M1

[0066] To unify the dimensions, improve the fusion effect, and facilitate subsequent processing, the infrared sub-image (S) and the low-light image detail layer (D) are normalized so that their grayscale values ​​fall within the range [0, 1]. This ensures that the features of each part are balanced during the fusion process, resulting in better visual quality and effects. The normalization formula is as follows:

[0067]

[0068] S norm is the normalized result of infrared sub-image, D norm Normalization result of detail layer of low-light image;

[0069] The weighted average method is used to calculate S norm , D norm Perform fusion to obtain the target enhanced image F, the formula is as follows:

[0070] F(i,j)=ω1S(i,j)+ω2D(i,j)

[0071] Wherein, i, j: represent the rows and columns of image pixels; ω1, ω2: represent weighting coefficients, and ω1+ω2=1. After experimental analysis, ω1=0.6, ω2=0.4.

[0072] The comprehensive enhanced image and the target enhanced image are weightedly fused to obtain the final fused image, which improves the visual effect and clarity of the image.

[0073] (3) Beneficial effects

[0074] Compared with the existing technology, the present invention provides an improved method for fusion of infrared and low-light-level images based on rolling guided filtering, which has the following beneficial effects:

[0075] The present invention separates the low-light-level image and the infrared image into the intensity component (I), hue component (H) and saturation component (S) by adopting the IHS color space, realizing independent processing of intensity and color and avoiding color distortion during the fusion process. Secondly, the I component is decomposed into multiple scales by rolling guided filtering to accurately separate the detail layer, structure layer and base layer, and specifically retain texture, edge and low-frequency information. The small-scale layers are fused by combining non-local mean denoising with the absolute value maximization strategy to suppress noise interference. The large-scale layers are optimized based on PCA and sparse-low rank joint representation to improve the matching degree between global structure and local details. Finally, a high-quality infrared-low-light-level fused image is generated by inverse IHS transform and detail enhancement fusion.

[0076] Compared to saliency map fusion methods, this method achieves an 8.6% improvement in mean, a 4.7% improvement in standard deviation, and a 5.5% increase in entropy. The fused image retains enhanced detail and optimizes color fidelity, effectively improving target recognition in complex scenes and addressing issues such as noise amplification, blurred edges, and contrast imbalance found in traditional methods. BRIEF DESCRIPTION OF THE DRAWINGS

[0077] Figure 1 This is a flowchart of the steps of the image fusion method of the present invention;

[0078] Figure 2 A detailed flowchart of the image fusion method of the present invention is provided;

[0079] Figure 3 This is a comparison diagram of the fusion effects of different methods of the present invention.

[0080] In the figure: (1) low-light image; (2) infrared image; (3) pixel value maximization fusion; (4) saliency map-based fusion; (5) the fusion method in this paper. DETAILED DESCRIPTION

[0081] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0082] Example

[0083] like Figure 1-3As shown, an embodiment of the present invention proposes an improved method for fusion of infrared and low-light-level images based on rolling guided filtering. By decomposing infrared and low-light-level images into multiple scales and adopting an adaptive feature fusion strategy for different scale layers, the advantages of multimodal features are complementary and fused. The rolling guided filtering algorithm is used to improve the image detail retention capability and suppress noise interference, and the multi-level feature collaborative optimization is achieved by combining visual saliency and non-local similarity.

[0084] An improved method for infrared and low-light-level image fusion based on rolling guided filtering includes the following steps:

[0085] S1. Obtain a low-light image and an infrared image from the same viewing angle. Apply bilateral filtering to denoise the infrared image, suppressing thermal noise while retaining edge features. The function expression is:

[0086]

[0087] Among them, W p : Normalization factor, σ d : used to represent the amount of filtering of image I, Range function, which can reduce and I p The influence of pixels p with large gray value difference, A spatial function that reduces the influence of distant pixels.

[0088] Bilateral filtering achieves noise suppression and edge protection through a dual-weight kernel function.

[0089] S2. Convert the low-light image and the denoised infrared image to the IHS color space and extract the intensity component (I component), hue component (H component), and saturation component (S component). The IHS to RGB conversion process requires the use of two intermediate variables, F1 and F2, and its function expression is:

[0090]

[0091] Among them, R, G, and B are the red, green, and blue bands of the original multispectral image, respectively. LL , F1, and F2 are the I, H, and S components obtained by IHS transformation respectively.

[0092] S3. Perform Rolling Guidance Filtering (RGF) on the I component of the infrared image and the low-light image, respectively, decomposing them into three levels: a small-scale layer containing high-frequency details, a large-scale layer representing structural features, and a base layer reflecting illumination distribution.

[0093] First, the I component is initially filtered by a Gaussian filter to remove high-frequency noise and small texture structures, and obtain the base layer containing low-frequency information. Its function expression is:

[0094]

[0095] Where, I(q), G(p): input and output pixel values, p, q: corresponding image pixel coordinates, σ s is: the parameter that controls the scale of the Gaussian structure, N(p): the filter window centered at pixel p;

[0096] Secondly, the filtered image is subjected to multi-scale iterative processing based on the bilateral filter, gradually restoring the edge information of different scales and generating a large-scale layer representing the structural features. Its function expression is:

[0097]

[0098] Among them, J t+1 : The result of the tth iteration, J t : guided filtering, σ r : Control distance weight, t: number of iterations;

[0099] Finally, by adjusting σ z , σ r and the number of iterations t, and finally output the decomposed three-level structure.

[0100] S4. A feature selection strategy based on maximizing absolute values ​​is adopted for the small-scale layers of infrared and low-light-level images, which is then fused with a non-local similarity adaptive weighted denoising algorithm to suppress noise accumulation while retaining significant edge features.

[0101] First, the small-scale layers are fused using the absolute value maximization strategy. The coefficient with the largest absolute value represents the prominent feature of the image. The dominant attributes can be extracted from the small-scale layers of the two images, and the important fine-scale features can be injected into the fused image to retain more texture information in the original image. It can be expressed as:

[0102]

[0103] Among them, GF(W j ,σ0) is Gaussian smoothing of the weighted coefficients to reduce noise interference, σ0 is usually set to 2, and

[0104]

[0105] Secondly, in order to further suppress the noise in the small-scale layer and avoid excessive enhancement of the noise during the fusion process, a denoising method based on non-local means (NLM) is introduced.

[0106] For each pixel i in the small-scale layer, calculate its similarity weight w(i, j) with the surrounding pixels j. The similarity weight is determined by comparing the grayscale value difference of the local windows Ni and Nj centered on pixels i and j, and the formula is:

[0107]

[0108] Where, h: parameter that controls weight decay; Represents the Euclidean distance between two local windows.

[0109] For each pixel i, the similarity weight w(i, j) is used to perform weighted averaging on the surrounding pixels j to obtain the denoised pixel value:

[0110]

[0111] The search window size and attenuation parameter h are adjusted according to the image characteristics to balance the denoising effect and detail preservation.

[0112] Finally, the denoised small-scale layer and Substitute into the absolute value maximization fusion formula to obtain the final fusion result:

[0113]

[0114] By introducing a denoising method based on non-local means (NLM), noise in small-scale layers is effectively suppressed while preserving image details. Experimental results show that this method significantly reduces the interference of noise on the fusion result while improving the quality of the fused image.

[0115] S5. Principal component analysis (PCA) technology is used to reduce the dimensionality of the large-scale layers of infrared images and low-light images. The main energy information is extracted through feature space transformation. A joint optimization model is constructed by combining sparse representation and low-rank representation for fusion to enhance structural consistency.

[0116] First, the large-scale layer images of infrared and low-light images are used as column vectors of the matrix. Then, each row is used as a reference and each column is used as a variable γ. The covariance matrix C of γ is obtained as follows:

[0117]

[0118] In the formula is the mean vector of γ.

[0119] Calculate the eigenvalues ​​and corresponding eigenvectors of C and record the results as λ1,λ2,…,λ n . Find the largest eigenvalue from the eigenvalues, i.e. λ maxThe corresponding eigenvector is used as the principal component. The principal component is normalized and used as the fusion weight to obtain the final large-scale information fusion image.

[0120] Secondly, to further enhance the feature extraction capabilities of large-scale layers, the PCA fusion method is improved by combining sparse representation and low-rank representation. Sparse representation effectively captures local details in an image, while low-rank representation extracts global structural information. By combining these two representation methods, key visual features in large-scale layers can be more comprehensively mined.

[0121] The large-scale layer image is represented as a sparse linear combination, and the sparse coefficients are obtained by solving the following optimization problem:

[0122]

[0123] Where D is the dictionary, X is the large-scale layer image, α is the sparse coefficient, and λ is the regularization parameter.

[0124] The sparse coefficients are used to reconstruct large-scale layer images and extract local detail information.

[0125] The large-scale layer image is decomposed into a low-rank matrix and a sparse matrix, and the low-rank representation is obtained by solving the following optimization problem:

[0126]

[0127] Among them, L is a low-rank matrix, S is a sparse matrix, ||·|| * represents the nuclear norm, and ||·||1 represents L 0031 norm.

[0128] Utilize low-rank matrices to extract global structural information.

[0129] Finally, the results of sparse representation and low-rank representation are weighted fused with the principal components extracted by PCA to obtain the final large-scale information fusion image. The fusion formula is:

[0130] F large =w1·PCA+w2·Sparse+w3·Low-Rank

[0131] Among them, w1, w2, w3 are weight coefficients satisfying w1+w2+w3=1.

[0132] By combining sparse representation and low-rank representation, this improved large-scale layer fusion method can more comprehensively extract local details and global structural information in the image, thereby improving the quality and visual effect of the fused image.

[0133] S6. The base layers of infrared images and low-light-level images are fused using a visual saliency mapping method. By performing regional saliency detection and adaptive weight allocation, the infrared thermal radiation characteristics and the low-light-level scene illumination distribution are retained.

[0134] S7. Linearly reorganize the fused small-scale layer, large-scale layer, and base layer and perform an inverse IHS transform with the H and S components of the low-light image to reconstruct a preliminary fused image. Subtract the original low-light image from the bilaterally filtered low-light image pixel by pixel. High-frequency residual calculation is used to extract the detail layer of the low-light image, preserving texture detail information. Guided filtering is performed on the denoised infrared image, combined with visual saliency detection to extract the target saliency map. Normalization is performed to obtain an infrared sub-image with enhanced target features.

[0135] In order to extract salient targets and suppress the influence of background on targets, the infrared image is used as a segmentation template. ir ) is multiplied by the saliency map mask (M1) to obtain the infrared sub-image (S), the formula is:

[0136] S=I ir ×M1

[0137] To unify dimensions, improve fusion effects, and facilitate subsequent processing, the infrared sub-image (S) and the low-light image detail layer (D) are normalized so that their grayscale values ​​fall within the range [0, 1]. This ensures balanced features across the fusion process, resulting in better visual quality and results. The normalization formula is as follows:

[0138]

[0139] S norm is the normalized result of infrared sub-image, D norm Normalization result of detail layer of low-light image.

[0140] The weighted average method is used to calculate S norm , D norm Perform fusion to obtain the target enhanced image F, the formula is as follows:

[0141] F(i,j)=ω1S(i,j)+ω2D(i,j)

[0142] Wherein, i, j: represent the rows and columns of image pixels; ω1, ω2: represent weighting coefficients, and ω1+ω2=1. After experimental analysis, ω1=0.6, ω2=0.4.

[0143] The infrared sub-image and the detail layer of the low-light-level image are fused at multi-resolution to generate a detail enhancement map. Finally, the preliminary fused image and the detail enhancement map are fused by weighted superposition to obtain the final infrared and low-light-level fused image with high dynamic range, clear details and significant target features.

[0144] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art will be able to modify the technical solutions described in the aforementioned embodiments or substitute equivalents for some of the technical features. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention shall be included within the scope of protection of the present invention.

Claims

1. An improved method for infrared and low-light-level image fusion based on rolling guided filtering, characterized in that: The following steps are included: S1. Acquire a low-light image and an infrared image, and perform denoising on the infrared image to obtain a denoised infrared image; S2. Convert the low-light image and the denoised infrared image to the IHS color space to obtain their respective I, H, and S components; S3. Perform rolling guided filtering on the I components of the infrared image and the low-light image, respectively, to decompose them into a small-scale layer, a large-scale layer, and a base layer; S4. A feature selection strategy of maximizing absolute values ​​is used to fuse the infrared and low-light-level images in their small-scale layers, combined with a non-local similarity adaptive weighted denoising algorithm, to obtain a small-scale fused image. S5. Use principal component analysis (PCA) to reduce the dimensionality of the large-scale layers of the infrared and low-light-level images, and fuse them with sparse and low-rank representations to obtain a large-scale fused image. S6. Fusing the base layers of the infrared image and the low-light image using a visual saliency mapping method, controlling the contrast distribution of the fused image and reducing the overall contrast loss, thereby obtaining a base layer fused image; After inverse IHS transformation and reconstruction, bilateral filtering detail enhancement and guided filtering saliency map are combined to achieve multi-level fusion and obtain the final infrared and low-light fusion image.

2. The improved method for infrared and low-light-level image fusion based on rolling guided filtering according to claim 1, characterized in that: In S1, an infrared image and a low-light image to be processed are input, and the infrared image is denoised using a bilateral filtering algorithm; The function expression of the bilateral filtering algorithm is: Among them, W p : Normalization factor, σ d : used to represent the amount of filtering of image I, Range function, which can reduce and I p The influence of pixels p with large gray value difference, A spatial function that reduces the influence of distant pixels.

3. The improved method for infrared and low-light-level image fusion based on rolling guided filtering according to claim 1, characterized in that: The IHS and RGB conversion process in S2 needs to be implemented using two intermediate variables F1 and F2, and its function expression is: Among them, R, G, and B are the red, green, and blue bands of the original multispectral image, respectively. LL , F1, and F2 are the I, H, and S components obtained by IHS transformation respectively.

4. The improved method for infrared and low-light-level image fusion based on rolling guided filtering according to claim 1, characterized in that: In S3, the image is decomposed into multiple scales to obtain a base layer, a small-scale layer, and a large-scale layer; First, the I component is initially filtered by a Gaussian filter to remove high-frequency noise and small texture structures, and obtain the base layer containing low-frequency information. Its function expression is: Where, I(q), G(p): input and output pixel values, p, q: corresponding image pixel coordinates, σ s is: the parameter that controls the scale of the Gaussian structure, N(p): the filter window centered at pixel p; Secondly, the filtered image is subjected to multi-scale iterative processing based on the bilateral filter, gradually restoring the edge information of different scales and generating a large-scale layer representing the structural features. Its function expression is: Among them, J t+1 : The result of the tth iteration, J t : guided filtering, σ r : Control distance weight, t: number of iterations; Finally, by adjusting σ z , σ r and the number of iterations t, and finally output the decomposed three-level structure.

5. The improved method for infrared and low-light-level image fusion based on rolling guided filtering according to claim 1, characterized in that: In S4, the small-scale layers are first fused using the absolute value maximization strategy. The coefficient with the largest absolute value represents the prominent feature of the image. The dominant attributes can be extracted from the small-scale layers of the two images, and the important fine-scale features can be injected into the fused image to retain more texture information in the original image. It can be expressed as: Among them, GF(W j ,σ0) is Gaussian smoothing of the weighted coefficients to reduce noise interference, σ0 is usually set to 2, and Secondly, in order to further suppress the noise in the small-scale layer and avoid excessive enhancement of the noise during the fusion process, a denoising method based on non-local means (NLM) is introduced; For each pixel i in the small-scale layer, the similarity weight w(i, j) between it and the surrounding pixels j is calculated. The similarity weight is determined by comparing the grayscale value difference of the local windows Ni and Nj centered on pixels i and j. The formula is: Where, h: parameter that controls weight decay; Represents the Euclidean distance between two local windows; For each pixel i, the similarity weight w(i, j) is used to perform weighted averaging on the surrounding pixels j to obtain the denoised pixel value: Adjust the search window size and attenuation parameter h according to the image characteristics to balance the denoising effect and detail preservation; Finally, the denoised small-scale layer and Substitute into the absolute value maximization fusion formula to obtain the final fusion result: By introducing a denoising method based on non-local means (NLM), the noise in small-scale layers is effectively suppressed while retaining the detail information of the image.

6. The improved method for infrared and low-light-level image fusion based on rolling guided filtering according to claim 1, characterized in that: In S5, first, the large-scale layer images of the infrared and low-light images are used as column vectors of the matrix, and then each row is used as a reference and each column is used as a variable γ. The covariance matrix C of γ is obtained as follows: In the formula is the mean vector of γ; Calculate the eigenvalues ​​and corresponding eigenvectors of C and record the results as λ1,λ2,…,λ n ; Find the largest eigenvalue from the eigenvalues, i.e. λ max ; and use its corresponding eigenvector as the principal component; normalize the principal component as the fusion weight to obtain the final large-scale information fusion image; Secondly, to further enhance the feature extraction capability of large-scale layers, the PCA fusion method is improved by combining sparse representation and low-rank representation. Sparse representation can effectively capture local details in the image, while low-rank representation can extract global structural information in the image. By combining these two representation methods, key visual features in large-scale layers can be more comprehensively mined. The large-scale layer image is represented as a sparse linear combination, and the sparse coefficients are obtained by solving the following optimization problem: Where D is the dictionary, X is the large-scale layer image, α is the sparse coefficient, and λ is the regularization parameter; Use sparse coefficients to reconstruct large-scale layer images and extract local detail information; The large-scale layer image is decomposed into a low-rank matrix and a sparse matrix, and the low-rank representation is obtained by solving the following optimization problem: Among them, L is a low-rank matrix, S is a sparse matrix, ||·|| * represents the nuclear norm, and ||·||1 represents L 0031 norm; Use low-rank matrix to extract global structural information; Finally, the results of sparse representation and low-rank representation are weighted fused with the principal components extracted by PCA to obtain the final large-scale information fusion image; the fusion formula is: F large =w1·PCA+w2·Sparse+w3·Low-Rank Where w1, w2, w3 are weight coefficients satisfying w1+w2+w3=1; By combining sparse representation and low-rank representation, this improved large-scale layer fusion method can more comprehensively extract local details and global structural information in the image, thereby improving the quality and visual effect of the fused image.

7. The improved method for infrared and low-light-level image fusion based on rolling guided filtering according to claim 1, characterized in that: In S6, guided filtering is performed on the infrared image to generate a saliency map; In order to extract salient targets and suppress the influence of background on targets, the infrared image is used as a segmentation template. ir ) is multiplied by the saliency map mask (M1) to obtain the infrared sub-image (S), the formula is: S=I ir ×M1 To unify the dimensions, improve the fusion effect, and facilitate subsequent processing, the infrared sub-image (S) and the low-light image detail layer (D) are normalized so that their grayscale values ​​fall within the range [0, 1]. This ensures that the features of each part are balanced during the fusion process, resulting in better visual quality and effects. The normalization formula is as follows: S norm is the normalized result of infrared sub-image, D norm Normalization result of detail layer of low-light image; The weighted average method is used to calculate S norm , D norm Perform fusion to obtain the target enhanced image F, the formula is as follows: F(i,j)=ω1S(i,j)+ω2D(i,j) Wherein, i, j: represent the rows and columns of image pixels; ω1, ω2: represent weighting coefficients, and ω1+ω2=1. After experimental analysis, ω1=0.6, ω2=0.

4.

8. The improved method for infrared and low-light-level image fusion based on rolling guided filtering according to claim 1, characterized in that: In the step S7 , the comprehensive enhanced image and the target enhanced image are weightedly fused to obtain a final fused image, thereby improving the visual effect and clarity of the image.

Images