An infrared optical image fusion method and system based on manifold adaptive filtering

Abstract

The application discloses an infrared optical image fusion method and system based on manifold adaptive decomposition, and belongs to the technical field of image processing and multi-modal information fusion, mainly including the steps of image structure scale feature estimation, adaptive multi-scale decomposition construction, structure perception fusion mechanism design and gradient domain reconstruction. Experimental results show that the method has significant advantages in maintaining structure integrity, enhancing details and textures, and improving fusion contrast, and can obtain high-quality fusion results with clear structure and natural vision in a complex environment.

Description

Technical Field

[0001] This invention belongs to the field of image processing and multimodal information fusion technology, specifically relating to an infrared and visible light image fusion method, and more particularly to a high-fidelity infrared and visible light image fusion method based on manifold adaptive multi-scale decomposition, Riemannian manifold structural feature modeling, and structure-aware gradient fusion mechanism. Background Technology

[0002] Existing infrared and visible light image fusion methods can be broadly categorized into two types: traditional fusion methods and deep learning-based fusion methods. Traditional fusion methods dominated early research, typically relying on manually designed feature representations and fusion rules, and offer advantages such as clear model structure and strong interpretability. However, these methods have limited adaptability to different datasets and complex application scenarios, and still have limitations in preserving the structure of the fused image, enhancing resolution, and representing details. In contrast, deep learning-based methods can automatically learn deep feature representations of multimodal images through large-scale data training, possessing end-to-end modeling capabilities and good scene generalization performance. However, they still face challenges in preserving salient targets, maintaining detail consistency, and ensuring the synergy between the fusion result and downstream visual tasks.

[0003] Traditional methods can be broadly categorized into multi-scale transformation-based, sparse representation-based, subspace-based, and saliency-based approaches. Multi-scale transformation-based methods typically employ wavelet transform, Laplacian pyramids, or non-subsampled contour waves to decompose the source image at multiple scales, then fuse and reconstruct it according to predefined rules across each scale sub-band. Their advantage lies in explicitly separating structural and detail information; however, the decomposition and reconstruction process can easily introduce edge blurring or artifacts, especially in infrared images where structural edges are weakened. Sparse representation-based methods utilize sparse representation theory to represent the image as a linear combination of a small number of dictionary atoms, fusing them in the sparse coefficient domain. These methods offer advantages in detail enhancement, but their computational complexity is high, and they are sensitive to dictionary learning quality and parameter settings, easily leading to detail smoothing or loss of structural information. Subspace-based methods, from the perspective of overall structural modeling, decompose multi-source images into common and differential subspace components through low-rank decomposition or correlation matrix decomposition to achieve information fusion. However, their subspace assumptions are often difficult to strictly satisfy in complex scenes, potentially causing structural loss or unnatural texture representation. Saliency-based methods highlight target regions and key information by constructing saliency maps or weight maps, and use these to guide the fusion process. Although they can enhance target saliency and overall contrast, saliency modeling usually relies on manually designed features and is sensitive to changes in parameters and scenes. It is also prone to fusion imbalance or artifacts when the saliency regions from multiple sources are inconsistent.

[0004] With the development of deep learning technology, deep learning-based infrared and visible light image fusion methods have gradually become a research hotspot, exhibiting diverse network structures and modeling paradigms. Autoencoder-based methods achieve multimodal information fusion by learning latent feature representations. Among them, DenseFuse networks extract features using an encoder and then employ feature addition strategies and selection strategies based on the l-norm combined with softmax to achieve relatively stable fusion results. End-to-end residual fusion nested networks (RFN-Nest) introduce residual network structures to enhance the representation of deep features, focusing on preserving complex textures and details through refined network design. However, these methods typically rely on fixed feature fusion rules or simple feature overlay methods, lacking explicit modeling of the geometric relationships of image structures. They are prone to structural blurring or ghosting in areas with significant differences in multimodal structures, and their detail recovery capabilities in areas with weak texture or low contrast remain limited.

[0005] Methods based on Convolutional Neural Networks (CNNs) typically guide the fusion process through customized loss functions. For example, the Progressive Illumination-Aware Infrared and Visible Image Fusion Network (PIAFusion) proposes an illumination-aware loss function and introduces a sub-network to evaluate scene illumination, thereby improving fusion performance under different lighting conditions. The Salience Detection-Based Fusion Network (STDFusionNet) introduces a loss function that enhances salient object masking, prioritizing the fusion of infrared salient objects and visible light texture information, thus improving the clarity and discriminative power of the fused image. However, these methods often rely on the local receptive field of convolutional operators for feature extraction, limiting their ability to model cross-scale structural relationships and global spatial consistency. In complex scenes, they are prone to insufficient structure preservation or fragmented detail textures. Furthermore, convolutional networks often focus on learning feature response intensity during the fusion process, lacking explicit constraints on multimodal gradient consistency, which can easily generate detail ghosting or artifacts in structural edge regions, thus affecting the overall structural stability of the fused image. Methods based on Generative Adversarial Networks (GANs) constrain the distribution characteristics of the fusion result through adversarial learning mechanisms. FusionGAN uses a generator to preserve infrared intensity information and texture details, while a discriminator constrains the fusion result to approximate the texture distribution of a visible light image. Dual-Discriminator Conditional Generative Adversarial Networks (DDcGANs) further introduce a dual-discriminator structure to enhance the similarity between the fusion result and source images of different modalities. However, GAN-based methods are prone to modality collapse or training instability during training. Their fusion results are highly dependent on the distribution of training data, and their stability and controllability under different scene conditions remain insufficient. Furthermore, adversarial learning primarily focuses on overall distribution consistency, with relatively weak constraints on local structural geometric relationships. This can easily introduce texture artifacts or structural distortions in detailed regions, thus affecting the structural realism and detail continuity of the fused image. In recent years, Transformer-based methods have gradually attracted attention. The correlation-driven two-branch feature decomposition and fusion network (CDDFuse) constructs a two-branch Transformer-CNN framework, extracting cross-modal features through decomposition strategies and combining the hybrid module in the efficient image restoration Transformer (Restormer) to achieve efficient information transfer. The manifold-aware fusion network (Grformer) introduces the attention mechanism into the manifold space to model semantic similarity and designs a cross-modal fusion strategy based on covariance masking, thereby achieving effective integration of complementary information from infrared and visible light.However, Transformer-type methods typically rely on large-scale data training to learn stable attention relationships. Their model parameters are large, and their computational complexity is high, placing significant demands on computational resources in practical applications. Furthermore, these methods primarily rely on feature attention weights for information selection during the fusion process, lacking explicit constraints on the image's physical structure and gradient geometry. This can lead to issues such as inconsistent edge structures or ghosting in complex regions.

[0006] In summary, existing infrared and visible light image fusion methods still have shortcomings in terms of structure preservation, detail consistency, and fusion contrast. Especially in scenarios with significant differences in multimodal structures or complex contrasts between thermal targets and the background, problems such as structural distortion, detail ghosting, and insufficient contrast in the fusion result easily arise. Therefore, it is necessary to propose a novel infrared and visible light image fusion method to achieve stable fusion of multimodal information while maintaining structural integrity. Summary of the Invention

[0007] To address the problems of structural distortion, detail ghosting, and insufficient fusion contrast in existing infrared and visible light image fusion technologies, this invention provides a manifold-adaptive multi-scale decomposition and structure-aware infrared and visible light image fusion method, applicable to fields such as intelligent monitoring, autonomous driving environmental perception, remote sensing image analysis, and target detection in complex environments. This invention performs multi-scale decomposition of images by constructing a manifold-adaptive scale space, utilizes Riemannian manifold structural features to construct saliency measures for basal layer fusion, and designs a structure-aware gradient transparency vector field fusion mechanism to achieve detail layer fusion, thereby achieving high-quality fusion of multimodal information while maintaining structural integrity.

[0008] To achieve the above objectives, the technical solution adopted by this invention is a manifold adaptive multi-scale decomposition and structure-aware infrared and visible light image fusion method, comprising the following steps: Structural scale features are estimated from the input visible light image to obtain stable distribution parameters and frequency domain features; An adaptive multi-scale decomposition model driven by manifold features is constructed. An adaptive scale sequence is generated by establishing a scale mapping mechanism based on stable distribution parameters and frequency domain features. At each scale, side window filtering is used to decompose layer by layer to obtain multi-scale detail layers and basal layers of visible light and infrared images. A structure-aware fusion mechanism is constructed at a multi-scale detail level. By modeling the activity level of the detail level and optimizing the weight graph with guided filtering, the fusion result of the detail level is obtained. A fusion model integrating Riemannian manifold structural features and global regularization constraints is constructed in the coarsest-scale basal layer. The structural saliency weight map is obtained by jointly modeling the structure tensor logarithmic mapping, local contrast and global information entropy, and the basal layer fusion result is obtained. The basal layer fusion result is superimposed with the detail layer fusion result to obtain a preliminary fused image. A structure-aware gradient vector field fusion mechanism is constructed. The selective fusion of multimodal gradient information is achieved through the gradient advantage ratio decision function. The global consistency of the fused image is restored through masked Poisson reconstruction to obtain the final fused image.

[0009] Furthermore, robust regression is performed using empirical characteristic functions to estimate the corresponding stable distribution parameter α, calculated as follows:

[0010] In the formula, Let be the characteristic function of the phase-consistent amplitude random variable, and t be the frequency sampling point of the characteristic function; parameters It describes the sparsity of the distribution and the degree of tail heaviness, reflecting the sharpness of image edges and the intensity of structural abrupt changes. Used to characterize the strength of local structures; Furthermore, the scale characteristics and texture complexity of the image are characterized in the frequency domain, where frequency domain features include characteristic wavelengths. and fractal dimension Specifically, the radial power spectrum is averaged to obtain the radial power spectrum function. The energy centroid method was used to estimate the characteristic wavelengths of the image. Meanwhile, log-linear fitting of the power spectrum in the mid-frequency band is performed to estimate the fractal dimension. The calculation formula is:

[0011] In the formula, The slope of the power spectrum fitting. The power spectral intercept term represents the constant offset when performing log-log spatial linear regression on the power spectrum in the mid-frequency band, reflecting the overall power baseline level of image texture energy at the spatial frequency baseline scale; characteristic wavelength Spatial scale and fractal dimension representing the dominant structure in an image It reflects the roughness and spatial complexity of the texture.

[0012] Furthermore, the baseline scale in the scale mapping mechanism Characteristic wavelength With stable distribution parameters The decision was made jointly, and the specific calculation formula is as follows:

[0013] In the formula, , , , It is a constant; Scale growth factor It is the fractal dimension; from this, a scale sequence can be constructed:

[0014] In the formula, To control the scaling parameters of the subsequent side-window filtering decomposition radius, the number of scaling layers... Constrained by both image size and scale upper limit; A multi-scale decomposition is constructed using side-window filtering, and the calculation formula is as follows:

[0015] In the formula, For multi-scale decomposition, the scale index is... For the two-dimensional spatial coordinates of a pixel, The input image is at the current scale and can be either a visible light image or an infrared image. This is the window orientation index for side window filtering. For the The optimal direction index determined by the criteria. The filtering results are for d directional windows; For scale The underlying basal layer, For scale The detailed layer below.

[0016] Furthermore, construct the activity level graph of the detail layer. To characterize the importance of local structures, its calculation form is as follows:

[0017] in, Represents the detail layer image. The standard deviation is A Gaussian smoothing kernel is used to aggregate local neighborhood energy. This is a Laplacian operator used to enhance the response of edge and corner structures; This is the balance coefficient; Based on the comparison of activity levels, an initial hard decision graph is obtained. Then, guided filtering is used to optimize it, transforming the hard decision into a continuous, spatially consistent weighted graph, which is thus formalized as a weighted least squares optimization problem:

[0018] in, and These represent the spatial indices of any pixel in the image and its neighboring pixels, respectively. This represents the detail layer fusion weights after weighted least squares variational optimization. This is the initial decision value, which can be either 0 or 1. The neighborhood affinity weights are calculated based on the guidance graph. For regularization parameters; For both visible light and infrared input images, the activity level is below the threshold. In the relevant regions, a soft-switching strategy is adopted, using mean fusion instead of a weight-based selection mechanism for fusion. The final detail layer fusion result is obtained by linearly weighting the multi-scale detail layers using an optimized weight map, achieving stable and continuous integration of high-frequency information.

[0019] In the formula, This is the result of detail layer fusion at scale k. and These are visible light and infrared detail layer images at the k-scale, respectively. and These refer to the combined activity levels of the visible light and infrared detail layers, respectively. for The final optimized weight graph after linear normalization.

[0020] Furthermore, the gradient fields of the visible light and infrared images are extracted using Gaussian derivative kernels, respectively. A structure tensor is constructed, and the Riemannian manifold structure features are calculated using logarithmic mapping. The calculation formula is as follows:

[0021] , These are the eigenvalues ​​of the structure tensor; Regularization terms to prevent numerical instability; Riemannian manifold structure characteristics The value reflects the geometric distance of the local structure on the Riemannian manifold; Further, local contrast features and global information entropy bias are integrated to construct a comprehensive saliency measure, which is used to characterize the relative advantages of different source images in terms of structural strength and information content, and an adaptive temperature parameter is introduced. The significant differences are modulated, and the calculation form is as follows:

[0022] In the formula, , These represent the combined saliency fields of visible light and infrared images, respectively. For data stability; This represents the mean. Indicates standard deviation; With adaptive temperature parameters To normalize the scale, a thermodynamic Softmax mapping is applied to the comprehensive significance field to obtain the initial weight map:

[0023] In the formula, This is the initial weighted graph. This is a numerically stable term used to avoid division by zero errors.

[0024] Furthermore, a global manifold regularization framework is constructed to optimize the initial weight graph, transforming the weight solution into an energy minimization problem that includes data fidelity terms and structure-guided smoothing terms. Its final form is reduced to solving the following linear equation:

[0025] Where I is the identity matrix, For the final fusion weight map of the basal layer, This is the initial weighted graph. This is the global smoothing coefficient, which controls the smoothing intensity. The graph Laplacian matrix is ​​constructed based on the basal layer guiding graph, and its value is modulated by the logarithmic gradient. By solving the above equation, the spatially consistent and structurally preserved final basal layer fusion weights are obtained.

[0026] Furthermore, the gradient advantage ratio decision function formula is as follows:

[0027] in, Refers to the spatial coordinates of a pixel. For infrared images in gradient magnitude at that point For visible light images in gradient magnitude at that point For gradient transparency decision mask; This is the decision margin coefficient, used to ensure that the infrared gradient only participates in fusion when it is significantly better than the visible light gradient; To ensure spatial continuity, a gradient transparency decision mask is applied. Gaussian smoothing is performed to obtain soft transparent weights. Based on this, a fused gradient vector field is constructed, and the calculation formula is as follows:

[0028] In the formula, The resulting gradient vector field; The reconstruction of the fused image is achieved by solving the masked Poisson equation, and its energy model is expressed as:

[0029] In the formula, For the final merged image, The prior image after pixel-domain thermal information is injected. For the gradient field of the final fused image, To fuse gradient vector fields, Used to balance pixel consistency and gradient constraints.

[0030] Furthermore, prior images The acquisition method is as follows: thermal targets are extracted from infrared images using morphological white top hat transformation, and then injected into the preliminary fused image based on the color mixing model. Before completing the reconstruction of the fused image, gradient conflicts between modes are detected by calculating the directional consistency of multimodal gradients, and gradient conflict detection is achieved by retaining the gradient of the mode with larger gradient magnitude in the region with opposite direction.

[0031] The present invention also provides an infrared optical image fusion system with manifold adaptive scaling decomposition, including a processor and a memory. The memory is used to store program instructions, and the processor is used to call the program instructions in the memory to execute the infrared optical image fusion method with manifold adaptive scaling decomposition as described in the above technical solution.

[0032] The present invention also provides a computer-readable storage medium, including a readable storage medium on which a computer program is stored, wherein when the computer program is executed, it implements the infrared optical image fusion method based on manifold adaptive scaling decomposition as described in the above technical solution.

[0033] Compared with the prior art, the present invention has the following advantages and beneficial effects: Strong structural preservation capability: This invention introduces a manifold adaptive multi-scale decomposition method, enabling the construction of the scale space to be dynamically adjusted according to the structural complexity of the image, thereby effectively avoiding the edge blurring problem caused by traditional multi-scale decomposition methods; Significantly improved fusion contrast: By utilizing the Riemannian manifold structure tensor to construct a saliency measure and combining local contrast and global information entropy for joint modeling, the structural advantages of different modal images can be more accurately characterized, thereby improving the overall contrast of the fused image; Effective suppression of detail ghosting: This invention proposes a structure-aware gradient transparency fusion mechanism, selectively fusing multimodal information in the gradient domain, avoiding the double edges and ghosting problems caused by traditional linear gradient fusion methods; Strong adaptability to complex environments: In complex scenes such as low light at night, smoke obscuration, and high reflectivity, the method of this invention can still maintain good structural integrity and visual naturalness; Natural visual effect of the fusion result: By co-modeling in the pixel domain and gradient domain, the decoupling between thermal information enhancement and structural detail preservation is achieved, so that the final fused image has good visual consistency and detail representation capability.

[0034] In summary, the method of this invention can significantly improve image contrast while maintaining structural consistency. Furthermore, qualitative experimental results in complex scenes show that the method of this invention can effectively preserve fine line structures and reflection edge information, and avoids the structural blurring and detail loss problems common in traditional methods. It demonstrates significant advantages in terms of structure preservation capability, detail representation capability, and adaptability to complex environments. Attached Figure Description

[0035] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a comparison diagram of the fusion effect of the method of the present invention. Detailed Implementation

[0036] To facilitate understanding and implementation of the present invention by those skilled in the art, the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the embodiments described herein are for illustration and explanation only and are not intended to limit the present invention.

[0037] This invention provides an infrared optical image fusion method based on manifold adaptive scaling decomposition. First, structural scale features of the input image are estimated through joint analysis in the statistical and frequency domains. Stable distribution parameters are estimated using phase consistency features and empirical feature functions, and characteristic wavelengths and fractal dimensions are obtained through power spectrum radial analysis, thus achieving a comprehensive representation of image structural complexity and scale characteristics. Second, a manifold feature-driven adaptive multi-scale decomposition model is constructed. An adaptive scale sequence is generated by establishing a scale mapping mechanism, and side-window filtering is used for layer-by-layer decomposition at each scale to obtain an edge-preserving multi-scale representation, providing a foundation for the stable fusion of structural and detail information. Subsequently, a structure-aware fusion mechanism is constructed at the detail layer. The activity level of the detail layer is jointly modeled by local energy and edge response, and the weight map is optimized by guided filtering to achieve continuous enhancement of texture details. Simultaneously, a fusion model fusing Riemannian manifold structural features and global regularization constraints is constructed at the basal layer. Structural saliency weights are obtained through joint modeling of structural tensor logarithmic mapping, local contrast, and global information entropy, achieving stable fusion of macroscopic structural information. Finally, based on the fusion results, a structure-aware gradient vector field fusion mechanism is constructed. The selective fusion of multimodal gradient information is achieved through the gradient advantage ratio decision function, and the global consistency restoration of the fused image is achieved through gradient conflict detection and masked Poisson reconstruction.

[0038] Please see Figure 1 The flowchart illustrates an infrared optical image fusion method based on manifold adaptive scaling decomposition provided in this invention, comprising the following steps: Step 1: Estimate the structural scale features of the input visible light image to obtain the stable distribution parameter α and frequency domain features, where the frequency domain features include the characteristic wavelength. and fractal dimension Phase consistency (PC) features are calculated for the input visible light image, and the stable distribution parameters of the phase consistency amplitude are estimated based on empirical characteristic functions to obtain statistical parameters α that characterize structural sparsity and local intensity. Simultaneously, radial analysis is performed on the image power spectrum to estimate characteristic wavelengths and fractal dimension to characterize the dominant structural scale and texture complexity. Through joint modeling of statistical and frequency domain parameters, a comprehensive estimate of image structural complexity and scale features is achieved, providing a basis for subsequent scale space construction.

[0039] Step 1 employs robust regression using the Empirical Characteristic Function (ECF) to estimate its corresponding stable distribution parameter α, and then performs linear regression based on the logarithmic relationship of the characteristic function's magnitude. The calculation formula is as follows:

[0040] In equation (1), Let be the characteristic function of the PC amplitude random variable, and t be the frequency sampling point of the characteristic function. The distribution of the random variable is probed by selecting several frequency values ​​in the complex frequency domain; parameters The sparsity of the distribution and the degree of tail heaviness can reflect the sharpness of image edges and the intensity of structural abrupt changes. Used to characterize the strength of local structures. By... and Estimation can statistically reflect the sparsity and abrupt changes in the local response of an image, thus providing an indirect measure consistent with manifold complexity for subsequent scale mapping.

[0041] In addition to statistical domain modeling, this invention further characterizes the scale properties and texture complexity of images in the frequency domain. Specifically, radial averaging of the image power spectrum yields the radial power spectrum function. The energy centroid method was used to estimate the characteristic wavelengths of the image. Meanwhile, log-linear fitting of the power spectrum in the mid-frequency band is performed to estimate the fractal dimension. The calculation formula is:

[0042] In equation (2), The slope of the power spectrum fitting. The power spectral intercept term represents the constant offset when performing a log-log spatial linear regression of the power spectrum in the mid-frequency band, reflecting the overall power baseline level of image texture energy at the spatial frequency baseline scale. Characteristic wavelength. The spatial scale representing the dominant structure in an image, while fractal dimension... This reflects the roughness and spatial complexity of the texture. Through joint modeling... and It can supplement the characterization of image structural complexity from the spectral level.

[0043] Step 2: Construct an adaptive multi-scale decomposition model driven by manifold features. Based on the stable distribution parameters and frequency domain features obtained in Step S1, an adaptive scale mapping mechanism is established. The baseline scale and scale growth factor are calculated by combining the feature wavelength, stable distribution parameters, and fractal dimension to generate an adaptive scale sequence. Subsequently, side-window filtering is used at each scale for layer-by-layer decomposition, thereby obtaining multi-scale detail layers and basal layers of visible light and infrared images. The detail layer includes multiple scales, while the basal layer has only one scale. The filtering response most consistent with the center pixel is obtained through a direction selection mechanism, realizing edge-preserving multi-scale representation of the image and providing a foundation for the stable fusion of subsequent structural and detail information.

[0044] Step 2 constructs an adaptive scale mapping mechanism driven by manifold features. Baseline scale. Characteristic wavelength With stable distribution parameters A joint decision, in which Used to adjust the scale's response to edge sharpness. A smaller scale is used when the image structure is sharper. The initial scale will be constrained to prevent high-frequency information from being overly smoothed during decomposition. Simultaneously, the scale growth factor is determined by the fractal dimension. To control the scale growth, the more complex the texture, the slower the growth, ensuring that important structures are fully preserved in multi-scale representations. The specific calculation formula is as follows:

[0045] In equation (3), , , , It is a constant. This is the scaling factor. From this, a scaling sequence can be constructed:

[0046] In equation (4), To control the scaling parameters of the subsequent side-window filtering decomposition radius, the number of scaling layers... The image size and scale upper limit are used to constrain the process and avoid introducing redundant smoothing and edge effects.

[0047] After obtaining the adaptive scaling sequence, this invention employs Side Window Filtering (SWF) to construct a multi-scale decomposition. Side Window Filtering calculates the response within an asymmetric window in multiple directions and selects the direction most consistent with the center pixel as the output, effectively avoiding the cross-structure smoothing at edges caused by traditional center-symmetric filtering. The calculation formula is as follows:

[0048] In equation (5), For multi-scale decomposition, the scale index is... For the two-dimensional spatial coordinates of a pixel, The input image is at the current scale and can be either a visible light image or an infrared image. This is the window direction index for the side window filtering (ranging from 1 to 8). For the The optimal direction index determined by the criteria. This is the filtering result for eight directional windows. This is the basal layer at this scale. This represents the detail layer at this scale. In multi-scale decomposition, the intermediate basal layer at each scale... Side window filtering is performed iteratively on a scale-by-scale basis, but only on the coarsest scale. As the final base layer output, this invention enables the construction of a multi-scale representation with a clear structure, well-preserved edges, and adaptive scale, providing a stable and unified structural prior for the high-fidelity fusion of subsequent base layers and detail layers.

[0049] Step 3, Detail Layer Structure-Aware Fusion. Based on the detail layers obtained from multi-scale decomposition, a saliency-guided structure-aware fusion mechanism is constructed. The activity level of the detail layers is jointly calculated by local energy and edge response to obtain the initial fusion decision. Subsequently, a variational optimization model is established using guided filtering to constrain the spatial consistency of the decision graph, generating a continuous and smooth weight map. Combined with a soft fusion strategy for low-activity regions, the multi-scale detail layers are weighted and fused, thereby achieving stable enhancement of texture details while maintaining the continuity of edge structure.

[0050] Step 3 first constructs an activity level map of the detail layer to characterize the importance of local structures. This activity is jointly characterized by local energy and edge response, and its calculation form is as follows:

[0051] In equation (6), Represents the detail layer image. The standard deviation is Gaussian smoothing kernels are used to aggregate local neighborhood energy. This is a Laplacian operator used to enhance the response of edge and corner structures; This is the balance coefficient. This activity level measure integrates the consistency of local energy with the salience of edge structure, providing a reliable physical basis for subsequent fusion decisions.

[0052] An initial hard decision graph can be obtained by comparing activity levels. However, this decision is often spatially discontinuous and prone to introducing block artifacts. Therefore, this invention further optimizes it through guided filtering, transforming the hard decision into a continuous, spatially consistent weight graph. This process can be formalized as a weighted least squares optimization problem:

[0053] In equation (7), and These represent the spatial indices of any pixel in the image and its neighboring pixels, respectively. This represents the detail layer fusion weights after weighted least squares variational optimization. This is the initial decision value, which takes the value 0 or 1 and is used to indicate pixel-level fusion selection; The neighborhood affinity weight is calculated based on the guide graph. Its value comes from the local variance of the guide graph within the neighborhood window. The guide graph is composed of pixel-by-pixel stitching of the gray values ​​of the detail layers of the two source images at this scale after the activity level decision. is the regularization parameter. Through this optimization, the weight graph achieves a smooth transition in flat regions and maintains necessary abrupt changes at edge locations, thereby significantly improving the spatial consistency of the fusion decision.

[0054] Furthermore, for regions where the activity of both source images is low, where the two source images refer to the visible light and infrared input images, and the activity of both source images is low means that, after joint measurement of Gaussian weighted local energy and Laplacian edge saliency, the combined activity value of both detail layers is simultaneously lower than a preset threshold. (In this embodiment, a flat noise region of 0.05 is used.) This invention employs a soft-switching strategy, replacing the weight-dominated selection mechanism with mean fusion, thereby effectively suppressing noise amplification in flat regions. Finally, the detail layer fusion result is obtained by linearly weighting the multi-scale detail layers using the optimized weight map, achieving stable and continuous integration of high-frequency information.

[0055] In the formula, This represents the fusion result at this scale. and These are images of the visible light and infrared detail layers at this scale, respectively. and These refer to the combined activity levels of the visible light and infrared detail layers, respectively. for The final optimized weight graph after linear normalization.

[0056] Step 4: Preserving and fusing the basal layer structure. Based on the coarsest-scale basal layer obtained from multi-scale decomposition, a fusion mechanism is constructed that integrates Riemannian manifold structural features and global regularization constraints. Gradient fields from visible and infrared images are extracted using Gaussian derivative kernels, a structure tensor is constructed, and the Riemannian manifold energy is calculated via logarithmic mapping. Simultaneously, a structural coherence measure is constructed by combining eigenvalue ratios. Based on this, a comprehensive saliency measure is constructed by fusing local contrast and global information entropy gain, and initial soft fusion weights are directly generated through adaptive temperature normalization. Subsequently, using the entropy-weighted joint guiding graph as a manifold constraint, global consistency optimization of the weights is performed through graph Laplacian anisotropic diffusion, obtaining spatially consistent and edge-preserving fusion weights, achieving a smooth transition of the basal layer and stable fusion of overall structural information.

[0057] Step 4 first utilizes the structure tensor to characterize the geometric properties of the image in its local neighborhood. By performing eigenvalue decomposition on the smoothed structure tensor and introducing a logarithmic mapping, it is embedded from Euclidean space into Riemannian manifold space, thereby enhancing the perception of weak structures and low-contrast variations. The calculation formula is as follows:

[0058] In equation (9), , These are the eigenvalues ​​of the structure tensor; To prevent numerically unstable regularization terms, logarithmic mappings amplify the relative differences between weak structures, allowing structural saliency to be expressed more evenly in the manifold space, thus enhancing the structural characteristics of Riemannian manifolds. This value reflects the "geometric distance" of the local structure on the Riemannian manifold.

[0059] Building upon this foundation, this invention further integrates local contrast features and global information entropy bias to construct a comprehensive saliency measure, which characterizes the relative advantages of different source images in terms of structural strength and information content. To accommodate the differences between different image pairs, an adaptive temperature parameter is introduced. The significant differences are modulated, and the calculation form is as follows:

[0060] In equation (9), , These represent the combined saliency fields of visible light and infrared images, respectively. This is the data stability term. This parameter dynamically adjusts the sensitivity of the weight mapping based on the overall characteristics of the significance distribution, thereby obtaining the initial soft fusion weights. Mean term. It reflects the average deviation of the two source images at the saliency level, while the standard deviation term... This characterizes the discreteness and non-uniformity of the spatial distribution of this difference. By jointly modeling the two, a robust estimate of the significant difference can be obtained, thus providing an adaptive scaling normalization basis for subsequent weight mapping.

[0061] With adaptive temperature parameters To normalize the scale, a thermodynamic softmax mapping is applied to the combined saliency field of the two source images to obtain the initial weight map:

[0062] In the formula, This is the initial weighted graph. This is a numerically stable term used to avoid division by zero errors.

[0063] However, the initial weight graph may contain spatial irregularities caused by local noise or inconsistencies. To address this, this invention constructs a global manifold regularization framework for optimization. This model transforms the weight solution into an energy minimization problem that includes data fidelity terms and structure-guided smoothing terms, and its final form can be reduced to solving the following linear equation:

[0064] Where I is the identity matrix that guarantees the existence and uniqueness of the solution and maintains data fidelity. For the final fusion weight graph, For the initial soft fusion weights, This is the global smoothing coefficient, which controls the smoothing intensity. The graph Laplacian matrix is ​​constructed based on the basal layer guide map, and its value is modulated by the logarithmic gradient. The basal layer guide map is obtained by weighting the two basal layers pixel-by-pixel using the ratio of the information entropy of the visible light and infrared basal layers as a globally fixed weight. By solving this equation, the final fusion weights that are spatially consistent and structurally preserved can be obtained. The basal layer fusion result is generated by weighted averaging of the basal layer components of the visible light and infrared images using these weights, thereby achieving a natural transition of macroscopic structure and effective enhancement of overall contrast.

[0065] Step 5: Structure-Aware Gradient Vector Field Fusion. First, the basal layer fusion result is superimposed with the detail layer fusion result to obtain a preliminary fusion result. Based on this, a structure-aware nonlinear gradient vector field fusion mechanism is constructed. Thermal target enhancement is introduced in the pixel domain to decouple energy injection from structural modeling. Subsequently, a transparency decision function based on the gradient advantage ratio is established in the gradient domain to selectively fuse infrared and visible light gradients. Continuous weights are obtained through smoothing to construct the fused gradient vector field, thereby achieving structurally consistent integration of multimodal gradient information and effectively avoiding zero-gradient traps and structural conflicts.

[0066] Step 5, to enhance the saliency of thermal targets and avoid their interference with the gradient field, employs morphological white-hat transform on the infrared image for thermal target extraction. Then, based on a screen blending model, the extracted thermal targets are injected into the initial fused image to obtain a prior image with pixel-domain thermal information injection. This method effectively suppresses the interference of edge gradients on the subsequent gradient fusion process while enhancing the brightness of the thermal target. In the gradient domain fusion stage, to avoid the infrared low-gradient region dominating gradient fusion, this invention proposes a transparency decision function based on the gradient dominance ratio, calculated as follows:

[0067] In equation (13), Refers to the spatial coordinates of a pixel. For infrared images in gradient magnitude at that point For visible light images in gradient magnitude at that point For gradient transparency decision mask; The decision margin coefficient is used to ensure that the infrared gradient only participates in fusion when it is significantly better than the visible light gradient. This decision mechanism automatically invalidates the infrared gradient in low-gradient regions such as inside the hot target, thus mathematically avoiding the zero-gradient trap. To ensure spatial continuity, a gradient transparency decision mask is applied. Gaussian smoothing is performed to obtain soft transparent weights. Based on this, a fused gradient vector field is constructed, and the calculation formula is as follows:

[0068] In equation (14), This is the fused gradient vector field. This process is not a traditional linear weighted average, but a nonlinear gradient selection mechanism driven by structure-aware decision. Its essence is to selectively "pass through" multimodal gradient information at the vector field level.

[0069] Step 6, Gradient Conflict Suppression and Poisson Reconstruction. Based on the constructed fused gradient vector field, gradient conflicts between modes are detected by calculating the directional consistency of multimodal gradients, and the gradient of the mode with larger gradient magnitude is retained in the region with opposite directions to suppress the generation of ghosting artifacts. Subsequently, using the pixel-domain prior image as a constraint, the fused gradient field is globally reconstructed by solving the Screened Poisson Equation, thereby achieving continuous restoration of detailed textures and overall consistency of the fusion result while maintaining the saliency of thermal targets.

[0070] Step 6 detects the directional consistency between modes through gradient dot product and retains only the gradient of the mode with the larger gradient magnitude in regions of directional conflict, thus suppressing ghosting at its source. Finally, the fused image is reconstructed by solving the Screened Poisson Equation, whose energy model can be expressed as:

[0071] In equation (15), For the final merged image, The prior image after pixel-domain thermal information is injected. For the gradient field of the final fused image, To fuse gradient vector fields, Used to balance pixel consistency and gradient constraints.

[0072] By introducing gradient transparency decision and direction consistency constraint in the gradient domain, the zero gradient region is avoided from dominating the Poisson reconstruction process at the gradient field construction level, and ghosting artifacts caused by multimodal gradient conflicts are effectively suppressed. While ensuring the saliency of thermal targets, the integrity of detail texture and the naturalness of the fusion result are significantly improved.

[0073] Step 7: Quantitatively and qualitatively evaluate the image fusion effect comparison across multiple datasets.

[0074] This invention constructed and selected four representative datasets for experimental evaluation. The data came from publicly available multimodal infrared-visible fusion datasets: MSRS, LLVIP, M3FD, and RoadScene. To ensure the balance and statistical reliability of the experimental samples, this invention randomly selected 100 pairs of rigorously registered infrared and visible light image pairs from each publicly available dataset, totaling 400 pairs of images for quantitative comparison. The quantitative results are shown in Table 1. Table 1. Results of fusion of four sets of infrared and visible light images. The best result is underlined and bolded, and the second best result is also bolded.

[0075]

[0076] The five evaluation metrics are: standard deviation (SD), spatial frequency (SF), average gradient (AG), entropy (EN), and visual information fidelity (VIF). Specifically, standard deviation (SD) reflects the dispersion of the image's gray-level distribution and can be used to characterize the overall contrast of the image. Spatial frequency (SF) measures the overall activity of gray-level changes in the image, reflecting the richness of image details and edges. Average gradient (AG) measures the detail and texture information of the image. Entropy (EN) measures the amount of information contained in the image. VIF quantifies the degree to which the original content is preserved, reflecting the loss of perceptual information.

[0077] As shown in Table 1, the overall results reveal significant differences in performance across different datasets and evaluation dimensions. Some methods achieve high scores on high-frequency relevant metrics (such as SF or AG), but fluctuate in structural consistency and information fidelity metrics (such as VIF), exhibiting an imbalance among metrics. In contrast, the method of this invention demonstrates a more stable and balanced performance advantage across the eight datasets, consistently achieving leading or near-optimal results on multiple key metrics, demonstrating excellent cross-scene generalization and overall optimization capabilities. Particularly noteworthy are the significant advantages of the method in the two core metrics of SD and VIF, where it maintains a dominant position in most datasets, indicating that it effectively preserves structural consistency and information integrity while enhancing image contrast.

[0078] The fusion comparison results are as follows Figure 2 As shown, to address the brightness distortion caused by abnormal infrared response, in the first group of vehicle license plate areas, due to the strong thermal radiation response of the infrared image, the other six methods generally assigned excessive weight to the bright areas during the fusion process, resulting in local brightness significantly higher than the overall vehicle brightness, producing abrupt visual discontinuities. A similar phenomenon exists in the second group of daytime bus scenes. The windshield exhibits abnormally high temperatures in the infrared image due to heat absorption and reflection effects, thus appearing as pseudo-bright areas that do not conform to visual perception in the fusion results of most methods. In contrast, the proposed method effectively suppresses the interference of abnormally high response on the fusion results through the synergistic effect of structure-aware saliency modeling and global constraints. While preserving thermal target information, it maintains the consistency of the overall brightness distribution, thereby obtaining a more natural and harmonious visual appearance.

[0079] Regarding structure and edge preservation, in the third group of nighttime playground scenes, the differences in the artificial turf area are mainly reflected in color changes in the visible light image. However, the corresponding areas show similar responses in the infrared image, resulting in a lack of significant discriminative features at the turf seams in the infrared mode. This places higher demands on the cross-modal information integration capabilities of the fusion method. Most contrast methods exhibit significant oversmoothing in this area, further blurring or even making the original boundary lines, which rely on visible light information for differentiation, difficult to identify. In contrast, the proposed method effectively fuses color and structural information from the visible light, ensuring that the turf seams remain clearly discernible in the fused result, demonstrating stronger cross-modal structure recovery capabilities. In the fourth group, glass reflection significantly interferes with the fusion result. Other methods often weaken or obfuscate the glass edges when processing reflective areas, causing the reflective structure to blend with the background. However, the proposed method effectively distinguishes between reflective information and the real structure, maintaining edge integrity while avoiding over-enhancement. The fifth set of low-light nighttime scenes further validated the method's ability to recover weak structures. Most methods exhibited missing ground markings under these conditions; although EgeFusion preserved some details, its results were accompanied by significant noise and artifacts caused by over-sharpening. In contrast, the proposed method was still able to stably recover ground markings, clearly presenting structural information with a more natural visual effect.

[0080] In terms of fine-grained texture representation, the road signs and trash can areas in the sixth evening scene contain rich and complex local details, making it an important scene to test the fine-grained representation capabilities of the fusion method. The contrasting methods generally suffer from texture blurring and structural weakening after fusion, making it difficult to clearly identify road sign information and causing varying degrees of degradation in the outline of trash cans; while the method in this paper can better preserve detailed textures and structural boundaries, making the local areas present higher clarity and integrity.

[0081] In summary, the method of this invention achieves high-fidelity infrared and visible light image fusion in multi-interference scenarios through the synergistic effect of structure perception, multi-scale fusion and physical saliency constraints, taking into account thermal information representation, structure preservation and detail naturalness, which is significantly better than the existing mainstream methods.

[0082] The above VSMWLS method is quoted from Ma J, Zhou Z, Wang B, et al. Infrared and visibleimage fusion based on visual saliency map and weighted least squareoptimization[J]. Infrared Physics&Technology, 2017, 82: 8-17. The EMMA method described above is cited in Zhao Z, Bai H, Zhang J, et al. Equivariant multi-modality image fusion[C] / / Proceedings of the IEEE / CVF conference on computer vision and pattern recognition. 2024: 25912-25921. The above EgeFusion method is quoted from Tang H, Liu G, Qian Y, et al. EgeFusion: Towards edge gradient enhancement in infrared and visible image fusion with multi-scale transform[J]. IEEE Transactions on Computational Imaging, 2024,10: 385-398. The ITFuse method described above is cited in Tang W, He F, Liu Y. ITFuse: An interactive transformer for infrared and visible image fusion[J]. Pattern Recognition, 2024, 156: 110822. The above UMCFuse method is quoted from Li X, Li X, Tan T, et al. UMCFuse: A unifiedmultiple complex scenes infrared and visible image fusion framework[J]. IEEETransactions on Image Processing, 2025. The GrFormer method described above is cited in Kang H, Li H, Wu XJ, et al. GrFormer: A novel Transformer on Grassmann manifold for infrared and visible image fusion[J]. Information Fusion, 2026, 125: 103402. Secondly, embodiments of the present invention also provide an infrared optical image fusion system with manifold adaptive scaling decomposition, including a processor and a memory. The memory is used to store program instructions, and the processor is used to call the program instructions in the memory to execute the infrared optical image fusion method with manifold adaptive scaling decomposition as described in the above technical solution.

[0083] Thirdly, embodiments of the present invention also provide a computer-readable storage medium, including a readable storage medium on which a computer program is stored, wherein when the computer program is executed, it implements the infrared optical image fusion method based on manifold adaptive scaling decomposition as described in the above technical solution.

[0084] It should be understood that any parts not described in detail in this specification belong to the prior art.

[0085] It should be understood that the above description of the preferred embodiments is quite detailed, but it should not be considered as a limitation on the scope of protection of this invention. Those skilled in the art, under the guidance of this invention, can make substitutions or modifications without departing from the scope of protection of the claims of this invention, and all such substitutions or modifications fall within the scope of protection of this invention. The scope of protection of this invention should be determined by the appended claims.

Claims

1. A method for infrared optical image fusion based on manifold adaptive scaling decomposition, characterized in that, Includes the following steps: Structural scale features are estimated from the input visible light image to obtain stable distribution parameters and frequency domain features; An adaptive multi-scale decomposition model driven by manifold features is constructed. An adaptive scale sequence is generated by establishing a scale mapping mechanism based on stable distribution parameters and frequency domain features. At each scale, side window filtering is used to decompose layer by layer to obtain multi-scale detail layers and basal layers of visible light and infrared images. A structure-aware fusion mechanism is constructed at a multi-scale detail level. By modeling the activity level of the detail level and optimizing the weight graph with guided filtering, the fusion result of the detail level is obtained. A fusion model integrating Riemannian manifold structural features and global regularization constraints is constructed in the coarsest-scale basal layer. The structural saliency weight map is obtained by joint modeling of structural tensor logarithmic mapping, local contrast and global information entropy, and the basal layer fusion result is obtained. The basal layer fusion result is superimposed with the detail layer fusion result to obtain a preliminary fused image. A structure-aware gradient vector field fusion mechanism is constructed. The selective fusion of multimodal gradient information is achieved through the gradient advantage ratio decision function. The global consistency of the fused image is restored through masked Poisson reconstruction to obtain the final fused image.

2. The infrared optical image fusion method based on manifold adaptive scaling decomposition as described in claim 1, characterized in that: Robust regression is performed using empirical characteristic functions to estimate the corresponding stable distribution parameter α. The calculation formula is as follows: In the formula, Let be the characteristic function of the phase-consistent amplitude random variable, and t be the frequency sampling point of the characteristic function; parameters It describes the sparsity of the distribution and the degree of tail heaviness, reflecting the sharpness of image edges and the intensity of structural abrupt changes. Used to characterize the strength of local structures; Furthermore, the scale characteristics and texture complexity of the image are characterized from the frequency domain, among which frequency domain features Including characteristic wavelengths and fractal dimension Specifically, the radial power spectrum is averaged to obtain the radial power spectrum function. The energy centroid method was used to estimate the characteristic wavelengths of the image. Meanwhile, log-linear fitting of the power spectrum in the mid-frequency band is performed to estimate the fractal dimension. The calculation formula is: In the formula, The slope of the power spectrum fitting. For the power spectrum intercept term, it represents the constant offset when performing log-log spatial linear regression of the power spectrum in the mid-frequency band, reflecting the overall power reference level of image texture energy on the spatial frequency reference scale. Characteristic wavelength Spatial scale and fractal dimension representing the dominant structure in an image It reflects the roughness and spatial complexity of the texture.

3. The infrared optical image fusion method based on manifold adaptive scaling decomposition as described in claim 1, characterized in that: The reference scale in the scale mapping mechanism Characteristic wavelength With stable distribution parameters The decision was made jointly, and the specific calculation formula is as follows: In the formula, , , , It is a constant; Scale growth factor It is the fractal dimension; from this, a scale sequence can be constructed: In the formula, To control the scaling parameters of the subsequent side-window filtering decomposition radius, the number of scaling layers... Constrained by both image size and scale upper limit; A multi-scale decomposition is constructed using side-window filtering, and the calculation formula is as follows: In the formula, For multi-scale decomposition, the scale index is... For the two-dimensional spatial coordinates of a pixel, The input image is at the current scale and can be either a visible light image or an infrared image. This is the window orientation index for side window filtering. For the The optimal direction index determined by the criteria. The filtering results are for d directional windows; For scale The underlying basal layer, For scale The detailed layer below.

4. The infrared optical image fusion method based on manifold adaptive scaling decomposition as described in claim 1, characterized in that: Constructing the activity level graph of the detail layer To characterize the importance of local structures, its calculation form is as follows: in, Represents the detail layer image. The standard deviation is A Gaussian smoothing kernel is used to aggregate local neighborhood energy. This is a Laplacian operator used to enhance the response of edge and corner structures; This is the balance coefficient; Based on the comparison of activity levels, an initial hard decision graph is obtained. Then, guided filtering is used to optimize it, transforming the hard decision into a continuous, spatially consistent weighted graph, which is thus formalized as a weighted least squares optimization problem: in, and These represent the spatial indices of any pixel in the image and its neighboring pixels, respectively. This represents the detail layer fusion weights after weighted least squares variational optimization. This is the initial decision value, which can be either 0 or 1. The neighborhood affinity weights are calculated based on the guidance graph. For regularization parameters; For both visible light and infrared input images, the activity level is below the threshold. In the relevant regions, a soft-switching strategy is adopted, using mean fusion instead of a weight-based selection mechanism for fusion. The final detail layer fusion result is obtained by linearly weighting the multi-scale detail layers using an optimized weight map, achieving stable and continuous integration of high-frequency information. In the formula, This is the result of detail layer fusion at scale k. and These are visible light and infrared detail layer images at the k-scale, respectively. and These refer to the combined activity levels of the visible light and infrared detail layers, respectively. for The final optimized weight graph after linear normalization.

5. The infrared optical image fusion method based on manifold adaptive scaling decomposition as described in claim 1, characterized in that: The gradient fields of visible light and infrared images are extracted using Gaussian derivative kernels, respectively. A structure tensor is constructed, and the Riemannian manifold structure features are calculated using logarithmic mapping. The calculation formula is as follows: , These are the eigenvalues ​​of the structure tensor; Regularization terms to prevent numerical instability; Riemannian manifold structure characteristics The value reflects the geometric distance of the local structure on the Riemannian manifold; Further, local contrast features and global information entropy bias are integrated to construct a comprehensive saliency measure, which is used to characterize the relative advantages of different source images in terms of structural strength and information content, and an adaptive temperature parameter is introduced. The significant differences are modulated, and the calculation form is as follows: In the formula, , These represent the combined saliency fields of visible light and infrared images, respectively. For data stability; This represents the mean. Indicates standard deviation; With adaptive temperature parameters To normalize the scale, a thermodynamic Softmax mapping is applied to the comprehensive significance field to obtain the initial weight map: In the formula, For the initial weighted graph, This is a numerically stable term used to avoid division by zero errors.

6. The infrared optical image fusion method based on manifold adaptive scaling decomposition as described in claim 5, characterized in that: A global manifold regularization framework is constructed to optimize the initial weight graph, transforming the weight solution into an energy minimization problem that includes data fidelity terms and structure-guided smoothing terms. Its final form reduces to solving the following linear equation: Where I is the identity matrix, For the final fusion weight map of the basal layer, For the initial weighted graph, This is the global smoothing coefficient, which controls the smoothing intensity. The graph Laplacian matrix is ​​constructed based on the basal layer guiding graph, and its value is modulated by the logarithmic gradient. By solving the above equation, the spatially consistent and structurally preserved final basal layer fusion weights are obtained.

7. The infrared optical image fusion method based on manifold adaptive scaling decomposition as described in claim 1, characterized in that: The formula for the gradient advantage ratio decision function is: in, Refers to the spatial coordinates of a pixel. For infrared images in gradient magnitude at that point For visible light images in gradient magnitude at that point For gradient transparency decision mask; This is the decision margin coefficient, used to ensure that the infrared gradient only participates in fusion when it is significantly better than the visible light gradient; To ensure spatial continuity, a gradient transparency decision mask is applied. Gaussian smoothing is performed to obtain soft transparent weights. Based on this, a fused gradient vector field is constructed, and the calculation formula is as follows: In the formula, The resulting gradient vector field; The reconstruction of the fused image is achieved by solving the masked Poisson equation, and its energy model is expressed as: In the formula, For the final merged image, The prior image after pixel-domain thermal information is injected. For the gradient field of the final fused image, To fuse gradient vector fields, Used to balance pixel consistency and gradient constraints.

8. The infrared optical image fusion method based on manifold adaptive scaling decomposition as described in claim 7, characterized in that: Prior image The acquisition method is as follows: thermal targets are extracted from infrared images using morphological white top hat transformation, and then injected into the preliminary fused image based on the color mixing model. Before completing the reconstruction of the fused image, gradient conflicts between modes are detected by calculating the directional consistency of multimodal gradients, and gradient conflict detection is achieved by retaining the gradient of the mode with larger gradient magnitude in the region with opposite direction.

9. An infrared optical image fusion system with manifold adaptive scaling decomposition, characterized in that: It includes a processor and a memory, the memory being used to store program instructions, and the processor being used to call the program instructions in the memory to execute the infrared optical image fusion method of manifold adaptive scaling decomposition as described in any one of claims 1-8.

10. A computer-readable storage medium, characterized in that, It includes a readable storage medium on which a computer program is stored, which, when executed, implements the infrared optical image fusion method of manifold adaptive scaling decomposition as described in any one of claims 1-8.

Images